Datasets:

timaeus
/

pile-arxiv

Dataset card Data Studio Files Files and versions Community

Split (1)

train · 100k rows

text stringlengths 4 2.78M	meta dict
--- abstract: 'The main purpose of this paper is to study the structure of the well-known non-iterative MUltiple SIgnal Classification (MUSIC) algorithm for identifying the shape of extended electromagnetic inclusions of small thickness located in a two-dimensional homogeneous space. We construct a relationship between the MUSIC-type imaging functional for thin inclusions and the Bessel function of integer order of the first kind. Our construction is based on the structure of the left singular vectors of the collected multistatic response matrix whose elements are the measured far-field pattern and the asymptotic expansion formula in the presence of thin inclusions. Some numerical examples are shown to support the constructed MUSIC structure.' author: - 'Won-Kwang Park[^1]' title: 'Asymptotic properties of MUSIC-type imaging in two-dimensional inverse scattering from thin electromagnetic inclusions' --- Introduction ============ One of the goals of the inverse scattering problem is to identify unknown properties (e.g., the shapes, material properties, locations, and constitutions) of electromagnetic targets from measured scattered field data. This problem is generally solved by Newton-type iteration schemes or level-set method involving minimization of the difference between the measured scattered data and the computed data by generating an admissible cost functional. Related works can be found in [@AGJKLY; @BHL; @B; @CR; @DL; @GH2; @K; @M2; @PL4; @VXB; @SZ] and references therein. To execute these schemes, the Fr[é]{}chet derivative must be evaluated at each iteration step, so the computational costs are large. Unfortunately, non-convergence or the appearance of several minima arises in the iteration procedure owing to the non-convex nature of the cost functional. Furthermore, a priori information on unknown targets is essential to guarantee a successful reconstruction, and even when the above conditions are fulfilled, reconstruction will fail if the iteration process is begun with a bad initial guess. Hence, generating a good initial guess close to the expected conditions is a priority. For this purpose, alternative non-iterative imaging algorithms such as the linear sampling method, single- and multi-frequency based Kirchhoff and subspace migrations, and the topological derivative strategy have been developed. In pioneering research [@D], the MUltiple SIgnal Classification (MUSIC) algorithm has been investigated to find the locations of point-like scatterers. It was recently applied to various problems, for example, detection of antipersonnel mines buried in the ground [@AIL], searching for the locations of small inclusions [@AILP; @CZ; @GH1; @HM; @SCC; @ZC], identifying internal corrosion in a pipeline [@AKKLV], and shape reconstruction of arbitrarily shaped thin inclusions, cracks, and extended targets [@AGKPS; @AKLP; @HSZ; @JP; @PL1; @PL3]. On the basis of these results, the locations of small inclusions can be accurately identified using MUSIC, but owing to the intrinsic resolution limit, the complete shape of extended targets cannot be imaged. Hence, the obtained results were based on good initial guesses, and iteration-based algorithms such as the level-set method were successfully executed (see [@AGJKLY; @BHL; @PL4]). Although the MUSIC algorithm offers good results for small and extended targets, a detailed structural analysis must be attempted because some phenomena cannot be explained using the traditional approach, for example, the unexpected appearance of artifacts or of two curves along the boundary of targets instead of the true shape (see [@HSZ Figure 9(b)]), or an image with poor resolution (see [@PL3 Section 4.4]). Numerical results in existing works motivate us to explore some properties of the MUSIC-type algorithm for imaging the arbitrarily shaped thin penetrable electromagnetic inclusions and perfectly conducting cracks considered in [@PL1; @PL3]. Our exploration is based on the rigorously derived asymptotic expansion formula in the presence of a thin inclusion [@BF] and physical factorization of the so-called multistatic response (MSR) matrix [@HSZ]. With this, we will establish a relationship between the MUSIC-type imaging functional and the Bessel function of integer order of the first kind, and identify its properties. This paper is organized as follows. In section \[sec:2\], we introduce two-dimensional direct scattering problems, the asymptotic expansion formula in the presence of thin inclusions, and the MUSIC-type algorithm for imaging thin electromagnetic inclusions. In section \[sec:3\], we identify the structure of a MUSIC-type functional focused on imaging of thin penetrable electromagnetic inclusions by constructing a relationship with the Bessel function of integer order of the first kind, and discuss its properties. In section \[sec:4\], numerical simulation results are presented to support the identified structure. This paper ends with a short conclusion in section \[sec:5\]. Direct scattering problems and MUSIC algorithm {#sec:2} ============================================== Two-dimensional direct scattering problems and asymptotic expansion formula --------------------------------------------------------------------------- Suppose that an extended electromagnetic inclusion $\Gamma$ with a small (with respect to the given wavelength) thickness $2h$ is located in the two-dimensional homogeneous space $\mathbb{R}^2$. We assume that the shape of $\Gamma$ is characterized by the supporting smooth curve $\gamma$ such that (see FIG. \[ThinInclusion\]) $$\Gamma=\{{\mathbf{x}}+\eta{\mathbf{n}}({\mathbf{x}}):{\mathbf{x}}\in\gamma,~\eta\in(-h,h)\}.$$ ![\[ThinInclusion\]Sketch of two-dimensional thin electromagnetic inclusion $\Gamma$.](ThinInclusion.eps){width="35.00000%"} Throughout this paper, we assume that $\Gamma$ and $\mathbb{R}^2$ are classified by their dielectric permittivity and magnetic permeability at a given frequency $\omega=2\pi/\lambda$, where $\lambda$ denotes the wavelength. Let $0<{\varepsilon}_0<+\infty$ and $0<\mu_0<+\infty$ denote the permittivity and permeability of $\mathbb{R}^2$, respectively; analogously, $0<{\varepsilon}<+\infty$ and $0<\mu<+\infty$ represent those of $\Gamma$. Then, we can define the piecewise constant dielectric permittivity $\varepsilon({\mathbf{x}})$ and magnetic permeability $\mu({\mathbf{x}})$, $$\label{EPSMU} \varepsilon({\mathbf{x}})=\left\{\begin{array}{ccl} \varepsilon_0&\mbox{for}&{\mathbf{x}}\in\mathbb{R}^2\backslash\overline{\Gamma}\\ \varepsilon&\mbox{for}&{\mathbf{x}}\in\Gamma \end{array}\right. \quad\mbox{and}\quad \mu({\mathbf{x}})=\left\{\begin{array}{ccl} \mu_0&\mbox{for}&{\mathbf{x}}\in\mathbb{R}^2\backslash\overline{\Gamma}\\ \mu&\mbox{for}&{\mathbf{x}}\in\Gamma, \end{array}\right.$$ respectively. For convenience, we set ${\varepsilon}_0=\mu_0=1$, ${\varepsilon}>{\varepsilon}_0$, and $\mu>\mu_0$. For a given fixed frequency $\omega$ (we assume that the wave number $k=\omega\sqrt{{\varepsilon}_0\mu_0}=\omega$), we let $$\label{IncidentField} u_0({\mathbf{x}},{\boldsymbol{\theta}};\omega):=e^{i\omega{\boldsymbol{\theta}}\cdot{\mathbf{x}}},\quad{\mathbf{x}}\in\mathbb{R}^2$$ be a plane-wave incident field with the incident direction ${\boldsymbol{\theta}}\in\mathbb{S}^1$, where $\mathbb{S}^1$ denotes the unit circle. Let $u({\mathbf{x}},{\boldsymbol{\theta}};\omega)=u_0({\mathbf{x}},{\boldsymbol{\theta}};\omega)+u_s({\mathbf{x}},{\boldsymbol{\theta}};\omega)$ denote the time-harmonic total field that satisfies the following Helmholtz equation: $$\label{TotalField} \nabla\cdot\bigg(\frac{1}{\mu({\mathbf{x}})}\nabla u({\mathbf{x}},{\boldsymbol{\theta}};\omega)\bigg) +\omega^2{\varepsilon}({\mathbf{x}})u({\mathbf{x}},{\boldsymbol{\theta}};\omega)=0,$$ with transmission conditions on the boundary of $\Gamma$. Here, $u_s({\mathbf{x}},{\boldsymbol{\theta}};\omega)$ denotes the unknown scattered field that satisfies the Sommerfeld radiation condition $$\lim_{r\to0}\sqrt{r}\bigg(\frac{{\partial}u_s({\mathbf{x}},{\boldsymbol{\theta}};\omega)}{{\partial}r}-i\omega u_s({\mathbf{x}},{\boldsymbol{\theta}};\omega)\bigg)=0,\quad{\boldsymbol{\vartheta}}=\frac{{\mathbf{x}}}{\|{\mathbf{x}}\|}\in\mathbb{S}^1$$ uniformly in all directions ${\boldsymbol{\vartheta}}$. Notice that the above radiation condition implies the asymptotic behavior $$u_s({\mathbf{x}},{\boldsymbol{\theta}};\omega)=\frac{e^{i\omega\|{\mathbf{x}}\|}}{\sqrt{\|{\mathbf{x}}\|}}u_{\infty}({\boldsymbol{\vartheta}},{\boldsymbol{\theta}};\omega)+o\left(\frac{1}{\sqrt{\|{\mathbf{x}}\|}}\right)\quad\mbox{for all}\quad\|{\mathbf{x}}\|\longrightarrow+\infty.$$ Then, according to [@BF], the far-field pattern $u_{\infty}({\boldsymbol{\vartheta}},{\boldsymbol{\theta}};\omega)$ can be written using the following asymptotic expansion formula: $$u_{\infty}({\boldsymbol{\vartheta}},{\boldsymbol{\theta}};\omega)=h\frac{\omega^2(1+i)}{4\sqrt{\omega\pi}} \int_\gamma\bigg((-{\boldsymbol{\vartheta}})\cdot\mathbb{M}({\mathbf{x}})\cdot{\boldsymbol{\theta}}+({\varepsilon}-1)\bigg)e^{i\omega({\boldsymbol{\theta}}-{\boldsymbol{\vartheta}})\cdot{\mathbf{x}}}d\gamma({\mathbf{x}})+o(h).$$ Here, the $2\times2$ symmetric matrix $\mathbb{M}({\mathbf{x}})$ is defined as follows: for ${\mathbf{x}}\in\gamma$, let ${\mathbf{t}}({\mathbf{x}})$ and ${\mathbf{n}}({\mathbf{x}})$ denote the unit tangent and normal vectors to $\gamma$ at ${\mathbf{x}}$, respectively. Then - $\mathbb{M}({\mathbf{x}})$ has eigenvectors ${\mathbf{t}}({\mathbf{x}})$ and ${\mathbf{n}}({\mathbf{x}})$. - The eigenvalue corresponding to ${\mathbf{t}}({\mathbf{x}})$ is $2\left(\frac{1}{\mu}-1\right)$. - The eigenvalue corresponding to ${\mathbf{n}}({\mathbf{x}})$ is $2\left(1-\mu\right)$. MUSIC-type imaging algorithm ---------------------------- Next, we introduce the MUSIC algorithm for imaging $\Gamma$. For simplicity, suppose that we have $N$ incident and observation directions ${\boldsymbol{\theta}}_l$ and ${\boldsymbol{\vartheta}}_j$, respectively, for $j,l=1,2,\cdots,N$, and the incident and observation directions are the same, i.e., ${\boldsymbol{\vartheta}}_j=-{\boldsymbol{\theta}}_j$. In this paper, we consider the full-view inverse problem. Hence, we assume that ${\left\{{\boldsymbol{\theta}}_n:n=1,2,\cdots,N\right\}}$ spans unit circle $\mathbb{S}^1$. Moreover, we assume that the supporting curve $\gamma$ is divided into $M$ different segments with sizes on the order of half the wavelength, $\lambda/2$. Then, keeping in mind the Rayleigh resolution limit from the far-field data, any detail less than one-half of the wavelength in size cannot be seen, and only one point at each segment is expected to contribute to the image space of the response matrix $\mathbb{K}$ (see [@AKLP; @HSZ; @PL1; @PL3], for instance). Each of these points, say ${\mathbf{x}}_j$ for $j=1,2,\cdots,M$, can be imaged by MUSIC-type imaging. With this in mind, we consider the collected MSR matrix such that $$\label{MSR} \mathbb{K}=\left[K_{jl}\right]_{j,l=1}^{N}=\left[\begin{array}{cccc} u_{\infty}({\boldsymbol{\vartheta}}_1,{\boldsymbol{\theta}}_1;\omega) & u_{\infty}({\boldsymbol{\vartheta}}_1,{\boldsymbol{\theta}}_2;\omega) & \cdots & u_{\infty}({\boldsymbol{\vartheta}}_1,{\boldsymbol{\theta}}_N;\omega)\\ u_{\infty}({\boldsymbol{\vartheta}}_2,{\boldsymbol{\theta}}_1;\omega) & u_{\infty}({\boldsymbol{\vartheta}}_2,{\boldsymbol{\theta}}_2;\omega) & \cdots & u_{\infty}({\boldsymbol{\vartheta}}_2,{\boldsymbol{\theta}}_N;\omega)\\ \vdots&\vdots&\ddots&\vdots\\ u_{\infty}({\boldsymbol{\vartheta}}_N,{\boldsymbol{\theta}}_1;\omega) & u_{\infty}({\boldsymbol{\vartheta}}_N,{\boldsymbol{\theta}}_2;\omega) & \cdots & u_{\infty}({\boldsymbol{\vartheta}}_N,{\boldsymbol{\theta}}_N;\omega)\\ \end{array}\right].$$ Because the incident and observation directions are the same, $K_{jl}$ becomes $$\begin{aligned} \begin{aligned}\label{StructureofMSRmatrix} K_{jl}=&u_{\infty}(-{\boldsymbol{\theta}}_j,{\boldsymbol{\theta}}_l;\omega)\\ &\approx h\frac{\omega^2(1+i)}{4\sqrt{\omega\pi}} \int_\gamma\bigg({\boldsymbol{\theta}}_j\cdot\mathbb{M}({\mathbf{x}})\cdot{\boldsymbol{\theta}}_l+({\varepsilon}-1)\bigg)e^{i\omega({\boldsymbol{\theta}}_j+{\boldsymbol{\theta}}_l)\cdot{\mathbf{x}}}d\gamma({\mathbf{x}})\\ =&h\frac{\omega^2(1+i)}{4\sqrt{\omega\pi}}\frac{\mbox{length}(\gamma)}{M} \sum_{m=1}^{M}\bigg[2\left(\frac{1}{\mu}-1\right){\boldsymbol{\theta}}_j\cdot{\mathbf{t}}({\mathbf{x}}_m){\boldsymbol{\theta}}_l\cdot{\mathbf{t}}({\mathbf{x}}_m)\\ &+2(1-\mu){\boldsymbol{\theta}}_j\cdot{\mathbf{n}}({\mathbf{x}}_m){\boldsymbol{\theta}}_l\cdot{\mathbf{n}}({\mathbf{x}}_m)+({\varepsilon}-1)\bigg]e^{i\omega({\boldsymbol{\theta}}_j+{\boldsymbol{\theta}}_l)\cdot{\mathbf{x}}}d\gamma({\mathbf{x}}), \end{aligned}\end{aligned}$$ where $\mbox{length}(\gamma)$ denotes the length of $\gamma$ (refer to [@PL3]). With this background, the MUSIC algorithm can be introduced as follows. Let us perform the singular value decomposition of $\mathbb{K}$: $$\mathbb{K}=\mathbb{USV}^\approx\sum_{m=1}^M\sigma_m{\mathbf{U}}_m{\mathbf{V}}_m^,$$ where ${\mathbf{U}}_m$ and ${\mathbf{V}}_m$ are the left and right singular vectors of $\mathbb{K}$, respectively, and $\sigma_m$ denotes the non-zero singular values. Then, on the basis of the structure of the MSR matrix (\[StructureofMSRmatrix\]), we define a vector ${\mathbf{f}}({\mathbf{z}})\in\mathbb{C}^{N\times1}$ as $$\label{Vecf} {\mathbf{f}}({\mathbf{z}})=\bigg[\mathbf{c}_1\cdot[1,{\boldsymbol{\theta}}_1]^Te^{i\omega{\boldsymbol{\theta}}_1\cdot{\mathbf{z}}}, \mathbf{c}_2\cdot[1,{\boldsymbol{\theta}}_2]^Te^{i\omega{\boldsymbol{\theta}}_2\cdot{\mathbf{z}}},\cdots, \mathbf{c}_N\cdot[1,{\boldsymbol{\theta}}_N]^Te^{i\omega{\boldsymbol{\theta}}_N\cdot{\mathbf{z}}}\bigg]^T,$$ where the selection of $\mathbf{c}_n\in\mathbb{R}^3\backslash{\left\{\mathbf{0}\right\}}$, $n=1,2,\cdots,N$, depends on the shape of the supporting curve $\gamma({\mathbf{x}})$. In fact, this is a linear combination of ${\mathbf{t}}({\mathbf{x}}_m)$ and ${\mathbf{n}}({\mathbf{x}}_m)$. Let us define a projection operator onto the noise subspace: $$P_{\mathrm{noise}}({\mathbf{f}}({\mathbf{z}})):=\left(\mathbb{I}_{N}-\sum_{m=1}^{M}{\mathbf{U}}_m{\mathbf{U}}_m^\right){\mathbf{f}}({\mathbf{z}}).$$ Then, the MUSIC-type imaging functional can be introduced: $$\label{MUSICfunction} \mathbb{E}({\mathbf{z}}):=\frac{1}{\|P_{\mathrm{noise}}({\mathbf{f}}({\mathbf{z}}))\|}.$$ Note that $\mathbb{K}$ is symmetric, but it is not self-adjoint. Therefore, we form a self-adjoint matrix $$\mathbb{A}=\mathbb{K^K}=\mathbb{\overline{K}K},$$ where $$ denotes the adjoint, and the bar denotes the complex conjugate. Then, with a careful choice of $\mathbf{c}_n$, the range of $\mathbb{A}$ is spanned by the $M$ vectors ${\left\{{\mathbf{f}}({\mathbf{x}}_1),{\mathbf{f}}({\mathbf{x}}_2),\cdots,{\mathbf{f}}({\mathbf{x}}_M)\right\}}$ (see [@AK; @C] for instance). Therefore, $${\mathbf{f}}({\mathbf{z}})\in\mbox{Range}(\mathbb{\overline{K}K})\quad\mbox{if and only if}\quad{\mathbf{z}}\in{\left\{{\mathbf{x}}_1,{\mathbf{x}}_2,\cdots,{\mathbf{x}}_M\right\}};$$ i.e., equivalently $\|P_{\mathrm{noise}}({\mathbf{f}}({\mathbf{z}}))\|=0$. Thus, the map of (\[MUSICfunction\]) will show a large magnitude (theoretically, $+\infty$) at ${\mathbf{x}}_m\in\gamma$. The results of the numerical simulations in [@P1; @PL1; @PL3] indicate that the selection of $\mathbf{c}_n$ is a strong prerequisite. The selection depends on the shape of the supporting curve $\gamma$. For purely dielectric contrast, $\mathbf{c}_n=[1,0,0]^T$ is a good choice. However, for purely magnetic contrast, $\mathbf{c}_n$ must be in the form $\mathbf{c}_n=[0,\mathbf{b}]^T$, where $\mathbf{b}$ is a linear combination of ${\mathbf{t}}({\mathbf{x}}_m)$ and ${\mathbf{n}}({\mathbf{x}}_m)$ for $m=1,2,\cdots,M$. Unfortunately, we have no a priori* information on the shape of $\gamma$. Therefore, in [@HSZ; @PL1], a large number of directions are applied to find an optimal vector $\mathbf{b}$. Applying this $\mathbf{b}$ yields a good result, but this process incurs large computational costs. Hence, motivated by recent work [@HSZ], we assume that $\mathbf{c}_n$ satisfies $\mathbf{c}_n\cdot[1,{\boldsymbol{\theta}}_n]^T=1$ for all $n$, i.e., $$\label{VecF} {\mathbf{f}}({\mathbf{z}})=\bigg[e^{i\omega{\boldsymbol{\theta}}_1\cdot{\mathbf{z}}},e^{i\omega{\boldsymbol{\theta}}_2\cdot{\mathbf{z}}},\cdots,e^{i\omega{\boldsymbol{\theta}}_N\cdot{\mathbf{z}}}\bigg]^T,$$ and explore some properties of the MUSIC-type imaging algorithm. Structure of certain properties of MUSIC-type imaging function {#sec:3} ============================================================== In this section, we identify the structure of the MUSIC-type imaging function. Before starting, we recall a useful result derived in [@G]. This plays a key role in our identification of the structure. \[TheoremBessel\] Assume that ${\left\{{\boldsymbol{\theta}}_n:n=1,2,\cdots,N\right\}}$ spans unit circle $\mathbb{S}^1$. Then, the following identities hold for sufficiently large $N$ and ${\boldsymbol{\xi}},{\mathbf{x}}\in\mathbb{R}^2$. $$\begin{aligned} &\frac{1}{N}\sum_{n=1}^{N}e^{i\omega{\boldsymbol{\theta}}_n\cdot{\mathbf{x}}}=\frac{1}{2\pi}\int_{\mathbb{S}^1}e^{i\omega{\boldsymbol{\theta}}\cdot{\mathbf{x}}}dS({\boldsymbol{\theta}})=J_0(\omega\|{\mathbf{x}}\|),\\ &\frac{1}{N}\sum_{n=1}^{N}({\boldsymbol{\theta}}_n\cdot{\boldsymbol{\xi}})e^{i\omega{\boldsymbol{\theta}}_n\cdot{\mathbf{x}}}=\frac{1}{2\pi}\int_{\mathbb{S}^1}({\boldsymbol{\theta}}\cdot{\boldsymbol{\xi}})e^{i\omega{\boldsymbol{\theta}}\cdot{\mathbf{x}}}dS({\boldsymbol{\theta}})=i\left(\frac{{\mathbf{x}}}{\|{\mathbf{x}}\|}\cdot{\boldsymbol{\xi}}\right)J_1(\omega\|{\mathbf{x}}\|),\end{aligned}$$ where $J_p$ denotes the Bessel function of integer order $p$ of the first kind. Pure dielectric permittivity contrast case: ${\varepsilon}\ne{\varepsilon}_0$ and $\mu=\mu_0$ --------------------------------------------------------------------------------------------- First, we consider the dielectric permittivity contrast case; i.e., we assume that ${\varepsilon}\ne{\varepsilon}_0$ and $\mu=\mu_0$. The proof is similar to the result in [@JP Theorem 3.3]. \[Theorem1\] For sufficiently large $N$ ($>M$) and $\omega$, (\[MUSICfunction\]) can be written as follows: $$\label{MUSIC1} \mathbb{E}_{{\varepsilon}}({\mathbf{z}}):=\frac{1}{\|P_{\mathrm{ noise}}({\mathbf{f}}({\mathbf{z}}))\|}\approx\frac{1}{\sqrt{N}}\left(1-\sum_{m=1}^{M}J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2+o(h)\right)^{-1/2}.$$ In this case, if we let $\mu=\mu_0$ in (\[StructureofMSRmatrix\]), the left singular vectors are of the form $${\mathbf{U}}_m\approx\frac{1}{\sqrt{N}}\bigg[e^{i\omega{\boldsymbol{\theta}}_1\cdot {\mathbf{x}}_m},e^{i\omega{\boldsymbol{\theta}}_2\cdot {\mathbf{x}}_m},\cdots,e^{i\omega{\boldsymbol{\theta}}_N\cdot {\mathbf{x}}_m}\bigg]^T+{\mathbf{O}}(h),$$ where ${\mathbf{O}}(h)$ denotes a $N\times1$ vector whose elements are $o(h)$. Then, for sufficiently large $\omega$, we can observe that $$\label{orthogonal1} {\mathbf{U}}_m\cdot\overline{{\mathbf{U}}}_{m'}\approx\frac{1}{N}\sum_{n=1}^{N}e^{i\omega{\boldsymbol{\theta}}_n\cdot({\mathbf{x}}_m-{\mathbf{x}}_{m'})}\approx J_0(\omega\|{\mathbf{x}}_m-{\mathbf{x}}_{m'}\|)+o(h)\approx\left\{\begin{array}{ccl} 1&\mbox{if}&m=m'\\ \noalign{\medskip}0&\mbox{if}&m\ne m'. \end{array}\right.$$ Following elementary calculus, $P_{\mathrm{noise}}$ can be written as $$\begin{aligned} P_{\mathrm{noise}}({\mathbf{f}}({\mathbf{z}}))&=\left(\mathbb{I}_{N}-\sum_{m=1}^{M}{\mathbf{U}}_m\overline{{\mathbf{U}}}_m^T\right){\mathbf{f}}({\mathbf{z}})\\ &=\left[\begin{array}{c} e^{i\omega{\boldsymbol{\theta}}_1\cdot{\mathbf{z}}} \\ e^{i\omega{\boldsymbol{\theta}}_2\cdot{\mathbf{z}}} \\ \vdots \\ e^{i\omega{\boldsymbol{\theta}}_N\cdot{\mathbf{z}}} \\ \end{array}\right] -\frac{1}{N}\sum_{m=1}^{M}\left[\begin{array}{c} \displaystyle e^{i\omega{\boldsymbol{\theta}}_1\cdot{\mathbf{z}}}+\sum_{n\in\mathbb{N}_1}e^{i\omega{\boldsymbol{\theta}}_1\cdot {\mathbf{x}}_m}e^{i\omega{\boldsymbol{\theta}}_n\cdot({\mathbf{z}}-{\mathbf{x}}_m)}+o(h)\\ \displaystyle e^{i\omega{\boldsymbol{\theta}}_2\cdot{\mathbf{z}}}+\sum_{n\in\mathbb{N}_2}e^{i\omega{\boldsymbol{\theta}}_1\cdot {\mathbf{x}}_m}e^{i\omega{\boldsymbol{\theta}}_n\cdot({\mathbf{z}}-{\mathbf{x}}_m)}+o(h)\\ \vdots \\ \displaystyle e^{i\omega{\boldsymbol{\theta}}_N\cdot{\mathbf{z}}}+\sum_{n\in\mathbb{N}_N}e^{i\omega{\boldsymbol{\theta}}_1\cdot {\mathbf{x}}_m}e^{i\omega{\boldsymbol{\theta}}_n\cdot({\mathbf{z}}-{\mathbf{x}}_m)}+o(h) \end{array}\right],\end{aligned}$$ where $\mathbb{N}_n=\{1,2,\cdots,N\}\backslash\{n\}$. Because $$e^{i\omega{\boldsymbol{\theta}}_n\cdot{\mathbf{z}}}=e^{i\omega{\boldsymbol{\theta}}_n\cdot {\mathbf{x}}_m}e^{i\omega{\boldsymbol{\theta}}_n\cdot ({\mathbf{z}}-{\mathbf{x}}_m)},$$ $P_{\mathrm{noise}}$ can be expressed as $$P_{\mathrm{noise}}({\mathbf{f}}({\mathbf{z}}))=\left[ \begin{array}{c} \displaystyle e^{i\omega{\boldsymbol{\theta}}_1\cdot{\mathbf{z}}}-\sum_{m=1}^{M}e^{i\omega{\boldsymbol{\theta}}_1\cdot {\mathbf{x}}_m}J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)+o(h) \\ \displaystyle e^{i\omega{\boldsymbol{\theta}}_2\cdot{\mathbf{z}}}-\sum_{m=1}^{M}e^{i\omega{\boldsymbol{\theta}}_2\cdot {\mathbf{x}}_m}J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)+o(h) \\ \vdots \\ \displaystyle e^{i\omega{\boldsymbol{\theta}}_N\cdot{\mathbf{z}}}-\sum_{m=1}^{M}e^{i\omega{\boldsymbol{\theta}}_N\cdot {\mathbf{x}}_m}J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)+o(h) \\ \end{array} \right].$$ Therefore, we can obtain $$\|P_{\mathrm{noise}}({\mathbf{f}}({\mathbf{z}}))\|=\bigg(P_{\mathrm{noise}}({\mathbf{f}}({\mathbf{z}}))\cdot\overline{P_{\mathrm{noise}}({\mathbf{f}}({\mathbf{z}}))}\bigg)^{1/2} =\left(\sum_{n=1}^{N}\bigg(1-\Phi_1+\Phi_2+o(h)\bigg)\right)^{1/2},$$ where $$\begin{aligned} \Phi_1&=\sum_{m=1}^{M}\bigg(e^{i\omega{\boldsymbol{\theta}}_n\cdot({\mathbf{z}}-{\mathbf{x}}_m)}+e^{-i\omega{\boldsymbol{\theta}}_n\cdot({\mathbf{z}}-{\mathbf{x}}_m)}\bigg)J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)\\ \Phi_2&=\left(\sum_{m=1}^{M}e^{i\omega{\boldsymbol{\theta}}_n\cdot{\mathbf{x}}_m}J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)\right) \left(\sum_{m=1}^{M}e^{-i\omega{\boldsymbol{\theta}}_n\cdot{\mathbf{x}}_m}J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)\right).\end{aligned}$$ Because $$\begin{aligned} \begin{aligned}\label{term1} \sum_{n=1}^{N}\Phi_1&=\sum_{n=1}^{N}\sum_{m=1}^{M}\bigg(e^{i\omega{\boldsymbol{\theta}}_n\cdot({\mathbf{z}}-{\mathbf{x}}_m)}+e^{-i\omega{\boldsymbol{\theta}}_n\cdot({\mathbf{z}}-{\mathbf{x}}_m)}\bigg)J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)\\ &=2N\sum_{m=1}^{M}J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2, \end{aligned}\end{aligned}$$ and on the basis of the orthogonal property (\[orthogonal1\]), we can evaluate $$\begin{aligned} \begin{aligned}\label{term2} \sum_{n=1}^{N}\Phi_2&=\left(\sum_{m=1}^{M}e^{i\omega{\boldsymbol{\theta}}_n\cdot{\mathbf{x}}_m}J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)\right) \left(\sum_{m=1}^{M}e^{-i\omega{\boldsymbol{\theta}}_n\cdot{\mathbf{x}}_m}J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)\right)\\ &=\sum_{n=1}^{N}\sum_{m=1}^{M}\sum_{m'=1}^{M}\bigg(e^{-i\omega{\boldsymbol{\theta}}_n\cdot{\mathbf{x}}_m}J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)\bigg) \bigg(e^{i\omega{\boldsymbol{\theta}}_n\cdot{\mathbf{x}}_{m'}}J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_{m'}\|)\bigg)\\ &=\sum_{m=1}^{M}\sum_{m'=1}^{M}\sum_{n=1}^{N}\bigg(e^{i\omega{\boldsymbol{\theta}}_n\cdot({\mathbf{x}}_{m'}-{\mathbf{x}}_m)}J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_{m'}\|)\bigg)\\ &=N\sum_{m=1}^{M}\sum_{m'=1}^{M}J_0(\omega\|{\mathbf{x}}_{m'}-{\mathbf{x}}_m\|)J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_{m'}\|)\\ &=N\sum_{m=1}^{M}J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2. \end{aligned}\end{aligned}$$ Therefore, using (\[term1\]) and (\[term2\]), we can obtain the following result: $$\|P_{\mathrm{noise}}({\mathbf{f}}({\mathbf{z}}))\|=\sqrt{N}\left(1-\sum_{m=1}^{M}J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2+o(h)\right)^{1/2}.$$ Note that $J_0(x)$ has its maximum value $1$ at $x=0$. This means that plots of $\mathbb{E}_{{\varepsilon}}({\mathbf{z}})$ will show peaks of large ($+\infty$ in theory) and small magnitude at ${\mathbf{x}}_m\in\gamma$ and at ${\mathbf{x}}\notin\gamma$, respectively (see FIG. \[PlotMUSIC1\]). This is why the MUSIC algorithm offers a good result for the pure dielectric contrast case of the full-view inverse scattering problem. We refer to FIG. \[MapMUSIC1\] and various results in [@AKLP; @HSZ; @PL1; @PL3]. Pure magnetic permeability contrast case: ${\varepsilon}={\varepsilon}_0$ and $\mu\ne\mu_0$ ------------------------------------------------------------------------------------------- Next, we consider the magnetic permeability contrast case; i.e., we assume that ${\varepsilon}={\varepsilon}_0$ and $\mu\ne\mu_0$. The result is as follows. \[Theorem2\] For sufficiently large $N$ ($>2M$) and $\omega$, (\[MUSICfunction\]) can be written as follows: $$\label{MUSIC2} \mathbb{E}_{\mu}({\mathbf{z}})=\frac{1}{\sqrt{N}}\left(1-\sum_{m=1}^{M}\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot({\mathbf{t}}({\mathbf{x}}_m)+{\mathbf{n}}({\mathbf{x}}_m))\right)^2J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2+o(h)\right)^{-1/2}.$$ In this case, the left singular vectors are of the form (see [@AGKPS; @PL3]) $${\mathbf{U}}_{2(m-1)+s}\approx\frac{1}{\sqrt{N}}\bigg[{\boldsymbol{\theta}}_1\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m)e^{i\omega{\boldsymbol{\theta}}_1\cdot{\mathbf{x}}_m}, \cdots,{\boldsymbol{\theta}}_N\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m)e^{i\omega{\boldsymbol{\theta}}_N\cdot{\mathbf{x}}_m}\bigg]^T+{\mathbf{O}}(h),$$ where $${\boldsymbol{\xi}}_s({\mathbf{x}}_m):=\left\{\begin{array}{rcl} {\mathbf{t}}({\mathbf{x}}_m) & \mbox{if} & s=1 \\ \noalign{\medskip}{\mathbf{n}}({\mathbf{x}}_m) & \mbox{if} & s=2. \end{array} \right.$$ Then, based on the orthonormal property of singular vectors, we can observe the following: because $\omega$ is sufficiently large, if $m\ne m'$ or $s\ne s''$, then $$\begin{gathered} \label{orthogonal2} {\mathbf{U}}_{2(m-1)+s}\cdot\overline{{\mathbf{U}}}_{2(m'-1)+s''}\approx\frac{1}{N} \sum_{n=1}^{N}\bigg({\boldsymbol{\theta}}_n\cdot({\boldsymbol{\xi}}_s({\mathbf{x}}_m)+{\boldsymbol{\xi}}_{s''}({\mathbf{x}}_{m'})\bigg) e^{i\omega{\boldsymbol{\theta}}_n\cdot({\mathbf{x}}_m-{\mathbf{x}}_{m'})}\\ \approx i\frac{{\mathbf{x}}_m-{\mathbf{x}}_{m'}}{\|{\mathbf{x}}_m-{\mathbf{x}}_{m'}\|}\cdot\bigg({\boldsymbol{\xi}}_s({\mathbf{x}}_m)+{\boldsymbol{\xi}}_{s''}({\mathbf{x}}_{m'})\bigg)J_1(\omega\|{\mathbf{x}}_m-{\mathbf{x}}_{m'}\|)+o(h)\approx0.\end{gathered}$$ Because we selected ${\mathbf{f}}({\mathbf{z}})$ as (\[VecF\]), $P_{\mathrm{noise}}$ can be written as $$\begin{aligned} &P_{\mathrm{noise}}({\mathbf{f}}({\mathbf{z}})) =\left(\mathbb{I}_{N}-\sum_{m=1}^{M}\sum_{s=1}^{2}{\mathbf{U}}_{2(m-1)+s}\overline{{\mathbf{U}}}_{2(m-1)+s}^T\right){\mathbf{f}}({\mathbf{z}})\\ &\approx\left[\begin{array}{c} e^{i\omega{\boldsymbol{\theta}}_1\cdot{\mathbf{z}}} \\ e^{i\omega{\boldsymbol{\theta}}_2\cdot{\mathbf{z}}} \\ \vdots \\ e^{i\omega{\boldsymbol{\theta}}_N\cdot{\mathbf{z}}} \\ \end{array}\right]- \frac{1}{N}\sum_{m=1}^{M}\sum_{s=1}^{2}\left[\begin{array}{c} \displaystyle({\boldsymbol{\theta}}_1\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))e^{i\omega{\boldsymbol{\theta}}_1\cdot{\mathbf{x}}_m}\sum_{n=1}^{N}({\boldsymbol{\theta}}_n\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))e^{i\omega{\boldsymbol{\theta}}_n\cdot({\mathbf{z}}-{\mathbf{x}}_m)}+o(h)\\ \displaystyle({\boldsymbol{\theta}}_2\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))e^{i\omega{\boldsymbol{\theta}}_2\cdot{\mathbf{x}}_m}\sum_{n=1}^{N}({\boldsymbol{\theta}}_n\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))e^{i\omega{\boldsymbol{\theta}}_n\cdot({\mathbf{z}}-{\mathbf{x}}_m)}+o(h)\\ \vdots\\ \displaystyle({\boldsymbol{\theta}}_N\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))e^{i\omega{\boldsymbol{\theta}}_N\cdot{\mathbf{x}}_m}\sum_{n=1}^{N}({\boldsymbol{\theta}}_n\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))e^{i\omega{\boldsymbol{\theta}}_n\cdot({\mathbf{z}}-{\mathbf{x}}_m)}+o(h) \end{array}\right]\\ &=\left[\begin{array}{c} \displaystyle e^{i\omega{\boldsymbol{\theta}}_1\cdot{\mathbf{z}}}-i\sum_{m=1}^{M}\sum_{s=1}^{2}({\boldsymbol{\theta}}_1\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m)\right)e^{i\omega{\boldsymbol{\theta}}_1\cdot {\mathbf{x}}_m}J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)+o(h)\\ \displaystyle e^{i\omega{\boldsymbol{\theta}}_2\cdot{\mathbf{z}}}-i\sum_{m=1}^{M}\sum_{s=1}^{2}({\boldsymbol{\theta}}_2\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m)\right)e^{i\omega{\boldsymbol{\theta}}_2\cdot {\mathbf{x}}_m}J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)+o(h)\\ \vdots\\ \displaystyle e^{i\omega{\boldsymbol{\theta}}_N\cdot{\mathbf{z}}}-i\sum_{m=1}^{M}\sum_{s=1}^{2}({\boldsymbol{\theta}}_N\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m)\right)e^{i\omega{\boldsymbol{\theta}}_N\cdot {\mathbf{x}}_m}J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)+o(h) \end{array}\right].\end{aligned}$$ Hence, $$\|P_{\mathrm{noise}}({\mathbf{f}}({\mathbf{z}}))\|=\left(\sum_{n=1}^{N}\bigg(1+\Psi_1-\overline{\Psi}_1+\Psi_2\overline{\Psi}_2+o(h)\bigg)\right)^{1/2},$$ where $$\begin{aligned} \Psi_1&=i\sum_{m=1}^{M}\sum_{s=1}^{2}({\boldsymbol{\theta}}_n\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m)\right) e^{i\omega{\boldsymbol{\theta}}_n\cdot({\mathbf{z}}-{\mathbf{x}}_m)}J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)\\ \Psi_2&=\sum_{m=1}^{M}\sum_{s=1}^{2}({\boldsymbol{\theta}}_n\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m)\right)e^{i\omega{\boldsymbol{\theta}}_n\cdot {\mathbf{x}}_m}J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|). $$ Because $$\begin{aligned} &\sum_{n=1}^{N}(\Psi_1-\overline{\Psi}_1)\\ =&i\sum_{n=1}^{N}\sum_{m=1}^{M}\sum_{s=1}^{2}({\boldsymbol{\theta}}_n\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m)\right) e^{i\omega{\boldsymbol{\theta}}_n\cdot({\mathbf{z}}-{\mathbf{x}}_m)}J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)\\ &+i\sum_{n=1}^{N}\sum_{m=1}^{M}\sum_{s=1}^{2}({\boldsymbol{\theta}}_n\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m)\right) e^{-i\omega{\boldsymbol{\theta}}_n\cdot({\mathbf{z}}-{\mathbf{x}}_m)}J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)\\ =&\sum_{m=1}^{M}\sum_{s=1}^{2}\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m)\right)\sum_{n=1}^{N}\bigg(i({\boldsymbol{\theta}}_n\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))e^{i\omega{\boldsymbol{\theta}}_n\cdot({\mathbf{z}}-{\mathbf{x}}_m)}\bigg)J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)\\ &+\sum_{m=1}^{M}\sum_{s=1}^{2}\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m)\right)\sum_{n=1}^{N}\bigg(i({\boldsymbol{\theta}}_n\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))e^{-i\omega{\boldsymbol{\theta}}_n\cdot({\mathbf{z}}-{\mathbf{x}}_m)}\bigg)J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|), $$ applying Theorem \[TheoremBessel\], we can obtain $$\begin{aligned} \begin{aligned}\label{term3} \sum_{n=1}^{N}(\Psi_1-\overline{\Psi}_1)&=-2N\sum_{m=1}^{M}\sum_{s=1}^{2}\bigg(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m)\bigg)^2J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2\\ &=-2N\sum_{m=1}^{M}\bigg(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot\bigg({\mathbf{t}}({\mathbf{x}}_m)+{\mathbf{n}}({\mathbf{x}}_m)\bigg)\bigg)^2J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2. \end{aligned}\end{aligned}$$ Next, on the basis of the orthogonal property (\[orthogonal2\]), we can evaluate $$\begin{aligned} \sum_{n=1}^{N}\Psi_2\overline{\Psi}_2=&\sum_{n=1}^{N}\left(\sum_{m=1}^{M}\sum_{s=1}^{2}({\boldsymbol{\theta}}_n\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m)\right)e^{i\omega{\boldsymbol{\theta}}_n\cdot {\mathbf{x}}_m}J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)\right)\\ &\times\left(\sum_{m'=1}^{M}\sum_{s''=1}^{2}({\boldsymbol{\theta}}_n\cdot{\boldsymbol{\xi}}_{s''}({\mathbf{x}}_{m'}))\left(\frac{{\mathbf{z}}-{\mathbf{x}}_{m'}}{\|{\mathbf{z}}-{\mathbf{x}}_{m'}\|}\cdot{\boldsymbol{\xi}}_{s''}({\mathbf{x}}_{m'})\right)e^{-i\omega{\boldsymbol{\theta}}_n\cdot{\mathbf{x}}_{m'}}J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_{m'}\|)\right)\\ =&\sum_{m=1}^{M}\sum_{n=1}^{N}\left(\sum_{s=1}^{2}({\boldsymbol{\theta}}_n\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m)\right)e^{i\omega{\boldsymbol{\theta}}_n\cdot {\mathbf{x}}_m}J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)\right)\\ &\times\left(\sum_{{s''}=1}^{2}({\boldsymbol{\theta}}_n\cdot{\boldsymbol{\xi}}_{s''}({\mathbf{x}}_{m'}))\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot{\boldsymbol{\xi}}_{s''}({\mathbf{x}}_m)\right)e^{-i\omega{\boldsymbol{\theta}}_n\cdot {\mathbf{x}}_{m}}J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_{m}\|)\right)\\ =&2\sum_{m=1}^{M}\sum_{n=1}^{N}\left(({\boldsymbol{\theta}}_n\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))^2\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m)\right)^2J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2\right)\\ =&2N\sum_{m=1}^{M}\sum_{s=1}^{2}\left(\frac{1}{N}\sum_{n=1}^{N}({\boldsymbol{\theta}}_n\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))^2\right)\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot{\boldsymbol{\xi}}_s({\mathbf{x}})\right)^2J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2\\ =&\frac{N}{\pi}\sum_{m=1}^{M}\sum_{s=1}^{2}\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot{\boldsymbol{\xi}}_s({\mathbf{x}})\right)^2\int_{\mathbb{S}^1}({\boldsymbol{\varphi}}\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))^2dS({\boldsymbol{\varphi}})J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2.\end{aligned}$$ Now, we consider polar coordinates; because ${\boldsymbol{\varphi}},{\boldsymbol{\xi}}_s\in\mathbb{S}^1$, let ${\boldsymbol{\varphi}}=[\cos\phi,\sin\phi]^T$ and ${\boldsymbol{\xi}}_s=[\cos\psi,\sin\psi]^T$; then elementary calculus yields $$\int_{\mathbb{S}^1}({\boldsymbol{\varphi}}\cdot{\boldsymbol{\xi}}_s({\mathbf{x}}_m))^2dS({\boldsymbol{\varphi}})=\int_0^{2\pi}\cos^2(\phi-\psi)d\phi=\pi.$$ Hence, we can obtain $$\label{term4} \sum_{n=1}^{N}\Psi_2\overline{\Psi}_2=N\sum_{m=1}^{M}\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot\bigg({\mathbf{t}}({\mathbf{x}}_m)+{\mathbf{n}}({\mathbf{x}}_m)\bigg)\right)^2J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2.$$ Therefore, from (\[term3\]) and (\[term4\]), we can obtain $$\|P_{\mathrm{ noise}}({\mathbf{f}}({\mathbf{z}}))\|=\sqrt{N}\left(1-\sum_{m=1}^{M}\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot({\mathbf{t}}({\mathbf{x}}_m)+{\mathbf{n}}({\mathbf{x}}_m))\right)^2J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2+o(h)\right)^{1/2}.$$ Unlike the permittivity contrast case, the map of (\[MUSIC2\]) shows two curves in the neighborhood of $\gamma$ because $$\lim_{{\mathbf{z}}\to{\mathbf{x}}_m}\frac{J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}=0,$$ and $J_1(x)^2$ is maximum at two points, $x_1$ and $x_2$, which are symmetric with respect to $x=0$. This is why two ghost replicas with large magnitude and many artifacts with small magnitude appear instead of the true shape of the supporting curve $\gamma$ (see FIG. \[PlotMUSIC2\]). Some numerical simulation results can be found in FIG. \[MapMUSIC2\] and in [@HSZ Section 5]. Note that $J_1(x)^2\ne1$ for all $x\in\mathbb{R}$. Hence, in contrast to the permittivity contrast case, (\[MUSIC2\]) does not blow up. Finally, by combining (\[MUSIC1\]) and (\[MUSIC2\]), we can immediately obtain the following result. \[Theorem3\] Let ${\varepsilon}\ne{\varepsilon}_0$ and $\mu\ne\mu_0$. Then, for sufficiently large $N$ ($>3M$) and $\omega$, (\[MUSICfunction\]) can be written as follows: $$\begin{gathered} \label{MUSIC3} \mathbb{E}_{{\varepsilon},\mu}({\mathbf{z}}):=\frac{1}{\sqrt{N}}\left(1-\sum_{m=1}^{M}J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2\right.\\ \left.+\sum_{m=1}^{M}\bigg(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot({\mathbf{t}}({\mathbf{x}}_m)+{\mathbf{n}}({\mathbf{x}}_m))\bigg)^2J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2+o(h)\bigg)\right)^{-1/2}. \end{gathered}$$ This result shows that plots of (\[MUSIC3\]) show a large magnitude at ${\mathbf{z}}$ if ${\mathbf{z}}\ne{\mathbf{x}}_m$ and $$\sum_{m=1}^{M}\left(J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2+\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot({\mathbf{t}}({\mathbf{x}}_m)+{\mathbf{n}}({\mathbf{x}}_m))\right)^2J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2+o(h)\right)=1.$$ Thus, a result with poor resolution will appear; refer to the examples of numerical simulation in [@PL3 Section 4.4]. On the basis of recent work [@P1], the structure of so-called subspace migration is as follows. 1. Permittivity contrast case: $$\mathbb{E}_{\mathrm{SM}}({\mathbf{z}})=\sum_{m=1}^{M}J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2.$$ 2. Permeability contrast case: $$\mathbb{E}_{\mathrm{SM}}({\mathbf{z}})=\sum_{m=1}^{M}\left\{\bigg(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\bigg)\cdot\bigg({\mathbf{t}}({\mathbf{x}}_m)+{\mathbf{n}}({\mathbf{x}}_m)\bigg)J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)\right\}^2.$$ Hence, we can observe the following relationship between MUSIC and subspace migration, which is derived in [@AGKLS Formula (7.4)]: Let us select the unit vector ${\mathbf{f}}({\mathbf{z}})$ of (\[VecF\]) as $${\mathbf{f}}({\mathbf{z}})=\frac{1}{\sqrt{N}}\bigg[e^{i\omega{\boldsymbol{\theta}}_1\cdot{\mathbf{z}}},e^{i\omega{\boldsymbol{\theta}}_2\cdot{\mathbf{z}}},\cdots,e^{i\omega{\boldsymbol{\theta}}_N\cdot{\mathbf{z}}}\bigg]^T.$$ Then, $$\mathbb{E}({\mathbf{z}})=\bigg(1-\mathbb{E}_{\mathrm{SM}}({\mathbf{z}})\bigg)^{-1/2}.$$ Imaging of perfectly conducting cracks -------------------------------------- Here, let $\Gamma$ be a smooth curve that describes the crack: for an injective piecewise smooth function $\boldsymbol{\phi}:[-1,1]\longrightarrow\mathbb{R}^2$, $$\label{PCrack} \Gamma=\{\boldsymbol{\phi}(x):-1\leq x\leq1\}.$$ Let $u({\mathbf{x}},{\boldsymbol{\theta}};\omega)$ be the single-component electric field that satisfies the Helmholtz equation: $$\Delta u({\mathbf{x}},{\boldsymbol{\theta}};\omega)+\omega^2 u({\mathbf{x}},{\boldsymbol{\theta}};\omega)=0\quad\mbox{in}\quad\mathbb{R}^2\backslash\Gamma.$$ For the sound-soft arc \[transverse magnetic (TM) polarization\], $u({\mathbf{x}},{\boldsymbol{\theta}};\omega)$ satisfies the Dirichlet boundary condition on $\Gamma$ (see [@K]): $$u({\mathbf{x}},{\boldsymbol{\theta}};\omega)=0\quad\mbox{on}\quad\Gamma$$ and for the sound-hard arc \[transverse electric (TE) polarization\], $u({\mathbf{x}},{\boldsymbol{\theta}};\omega)$ satisfies the Neumann boundary condition on $\Gamma$ (see [@M2]): $$\frac{{\partial}u({\mathbf{x}},{\boldsymbol{\theta}};\omega)}{{\partial}{\mathbf{n}}({\mathbf{x}})}=0\quad\mbox{on}\quad\Gamma\backslash\{\boldsymbol{\phi}(-1),\boldsymbol{\phi}(1)\},$$ where ${\mathbf{n}}({\mathbf{x}})$ is the unit normal vector to $\Gamma$ at ${\mathbf{x}}$. Then the far-field pattern $u_{\infty}({\boldsymbol{\vartheta}},{\boldsymbol{\theta}};\omega)$ for the scattering of an incident field $u_0({\mathbf{x}},{\boldsymbol{\theta}};\omega)=e^{i\omega{\boldsymbol{\theta}}\cdot{\mathbf{x}}}$ from $\Gamma$ is given by $$u_{\infty}({\boldsymbol{\vartheta}},{\boldsymbol{\theta}};\omega)=\left\{ \begin{array}{ll} \displaystyle\medskip\frac{1+i}{4\sqrt{\pi\omega}}\int_\Gamma e^{-i\omega{\boldsymbol{\vartheta}}\cdot{\mathbf{x}}}\varphi({\mathbf{x}},{\boldsymbol{\theta}};\omega)d{\mathbf{x}}& \mbox{: sound-soft arc}\\ \displaystyle\medskip\frac{(1-i)\sqrt{\omega}}{4\sqrt{\pi}}\int_\Gamma {\boldsymbol{\vartheta}}\cdot{\mathbf{n}}({\mathbf{x}})e^{-i\omega{\boldsymbol{\vartheta}}\cdot{\mathbf{x}}}\psi({\mathbf{x}},{\boldsymbol{\theta}};\omega)d{\mathbf{x}}& \mbox{: sound-hard arc}. \end{array} \right.$$ According to the physical factorization in [@HSZ; @PL1], if the incident and observation directions are the same, the left singular vector of the MSR matrix is of the form $${\mathbf{U}}_m=\left\{ \begin{array}{ll} \medskip\bigg[e^{i\omega{\boldsymbol{\theta}}_1\cdot{\mathbf{x}}_m},e^{i\omega{\boldsymbol{\theta}}_2\cdot{\mathbf{x}}_m},\cdots,e^{i\omega{\boldsymbol{\theta}}_N\cdot{\mathbf{x}}_m}\bigg]^T & \mbox{: sound-soft arc} \\ \medskip\bigg[({\boldsymbol{\theta}}_1\cdot{\mathbf{n}}({\mathbf{x}}_m))e^{i\omega{\boldsymbol{\theta}}_1\cdot{\mathbf{x}}_m},\cdots,({\boldsymbol{\theta}}_N\cdot{\mathbf{n}}({\mathbf{x}}_m))e^{i\omega{\boldsymbol{\theta}}_N\cdot{\mathbf{x}}_m}\bigg]^T & \mbox{: sound-hard arc}. \end{array} \right.$$ Thus, the structure of the left singular vectors for the sound-soft and sound-hard arcs is almost the same as in the permittivity contrast and permeability contrast cases (except for the absence of a unit tangential vector), respectively. Hence, we can obtain the following result. \[Theorem4\] Let $N$ and $\omega$ be sufficiently large. Then (\[MUSICfunction\]) can be written as follows. 1. Sound-soft arc (or TM) case: $$\mathbb{E}_{\mathrm{TM}}({\mathbf{z}})\approx\frac{1}{\sqrt{N}}\left(1-\sum_{m=1}^{M}J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2\right)^{-1/2}.$$ 2. Sound-hard arc (or TE) case: $$\mathbb{E}_{\mathrm{TE}}({\mathbf{z}})\approx\frac{1}{\sqrt{N}}\left(1-\sum_{m=1}^{M}\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot{\mathbf{n}}({\mathbf{x}}_m)\right)^2J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2\right)^{-1/2}.$$ Imaging of small electromagnetic inclusions ------------------------------------------- We briefly consider the use of MUSIC for imaging electromagnetic inclusions $\Sigma_m$, $m=1,2,\cdots,M$, with small diameter $r$: $$\Sigma_m:={\mathbf{x}}_m+r\mathbf{B}_m,$$ where $\mathbf{B}_m$ is a simply connected smooth domain containing the origin. We assume that $\Sigma_m$ are sufficiently separate from each other, and denote the collection of such inclusions as $\Sigma$. As in section \[sec:2\], we let $u({\mathbf{x}},{\boldsymbol{\theta}};\omega)$ satisfy (\[TotalField\]) in the presence of $\Sigma$, and $u_0({\mathbf{x}},{\boldsymbol{\theta}};\omega)$ is given by (\[IncidentField\]). Then, the far-field pattern can be written as the following asymptotic expansion formula (see [@AK]): $$\label{smallexpansion} u_{\infty}({\boldsymbol{\vartheta}},{\boldsymbol{\theta}};\omega)\approx r^2\frac{\omega^2(1+i)}{4\sqrt{\omega\pi}} \sum_{m=1}^{M}\|\mathbf{B}_m\|\bigg((-{\boldsymbol{\vartheta}})\cdot\mathbb{A}({\mathbf{x}}_m)\cdot{\boldsymbol{\theta}}+({\varepsilon}-{\varepsilon}_0)\bigg)e^{i\omega({\boldsymbol{\theta}}-{\boldsymbol{\vartheta}})\cdot{\mathbf{x}}_m}+o(r^2).$$ Here, $\mathbb{A}({\mathbf{x}}_m)$ denotes the polarization tensor corresponding to $\Sigma_m$. Then, we can obtain the following results in a similar manner: \[Theorem5\] Let $N$ and $\omega$ be sufficiently large. Then (\[MUSICfunction\]) can be written as follows. 1. Dielectric permittivity contrast case: $$\mathbb{E}_{{\varepsilon}}({\mathbf{z}})\approx\frac{1}{\sqrt{N}}\left(1-\sum_{m=1}^{M}J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2+o(r^2)\right)^{-1/2}.$$ 2. Magnetic permeability contrast case: $$\mathbb{E}_{\mu}({\mathbf{z}})\approx\frac{1}{\sqrt{N}}\left(1-\sum_{m=1}^{M}\left\{\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot({\mathbf{e}}_1+{\mathbf{e}}_2)\right)J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)\right\}^2+o(r^2)\right)^{-1/2},$$ where $\left\{{\mathbf{e}}_1=[1,0]^T,{\mathbf{e}}_2=[0,1]^T\right\}$ denotes an orthonormal basis in $\mathbb{R}^2$. 3. Both permittivity and permeability contrast case: $$\begin{gathered} \mathbb{E}_{{\varepsilon},\mu}({\mathbf{z}})\approx\frac{1}{\sqrt{N}}\left(1-\sum_{m=1}^{M}J_0(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2\right.\\ \left.+\sum_{m=1}^{M}\left(\frac{{\mathbf{z}}-{\mathbf{x}}_m}{\|{\mathbf{z}}-{\mathbf{x}}_m\|}\cdot({\mathbf{e}}_1+{\mathbf{e}}_2)\right)^2J_1(\omega\|{\mathbf{z}}-{\mathbf{x}}_m\|)^2+o(r^2)\right)^{-1/2}. \end{gathered}$$ If $\Sigma_m$ denotes a perfectly conducting inclusion with a small diameter, the asymptotic expansion formulas for the TM and TE cases are similar to (\[smallexpansion\]); refer to [@G Theorem 3.1]. Hence, the structure of (\[MUSICfunction\]) will be the same as in Theorem \[Theorem5\]. Numerical examples {#sec:4} ================== In this section, we present some numerical simulation results. For this, we choose a thin inclusion $\Gamma_1=\{{\mathbf{x}}+\eta{\mathbf{n}}(x):{\mathbf{x}}\in\gamma_1,~\eta\in(-h,h)\}$ with a smooth supporting curve: $$\gamma_1=\left\{\left[x+0.2,x^3+x^2-0.3\right]^T:-0.5\leq x\leq0.5\right\}.$$ The thickness $h$ of the thin inclusion $\Gamma_1$ is set to $0.02$, and the following parameters are chosen: ${\varepsilon}_0=\mu_0=1$, and ${\varepsilon}=\mu=5$. For the illumination and observation directions ${\boldsymbol{\theta}}_n$, they are chosen as $${\boldsymbol{\theta}}_n=\left[\cos\frac{2n\pi}{N},\sin\frac{2n\pi}{N}\right]^T$$ for $n=1,2,\cdots,N$. The total number of directions is $N=24$, and the applied frequency is $\omega=2\pi/\lambda$ with a wavelength of $\lambda=0.4$. The data set for the MSR matrix $\mathbb{K}$ in (\[MSR\]) is collected by solving the forward problem introduced in [@NK]. FIG. \[MapMUSIC\] shows maps of $\mathbb{E}_{{\varepsilon}}({\mathbf{z}})$ and $\mathbb{E}_{\mu}({\mathbf{z}})$ in the presence of $\Gamma_1$. This result demonstrates that the MUSIC algorithm offers a very accurate result for the permittivity contrast case. For the permeability contrast case, as we saw in Theorem \[Theorem2\], we cannot recognize the true shape of $\Gamma_1$. However, using Theorem \[Theorem2\], we can obtain an approximate shape of $\Gamma_1$ from the two identified curves. For a numerical example of a perfectly conducting crack, we selected the following smooth curve in the form of (\[PCrack\]): $$\Gamma_2=\left\{\left[x,\frac{1}{2}\cos\frac{x\pi}{2}+\frac{1}{5}\sin\frac{x\pi}{2}-\frac{1}{10}\cos\frac{3x\pi}{2}\right]^T:-1\leq x\leq1\right\}.$$ The data set for the MSR matrix $\mathbb{K}$ in (\[MSR\]) is collected by solving the forward problems introduced in [@N Chapter 3] and [@N Chapter 4] for the sound-soft and sound-hard arcs, respectively. FIG. \[MapMUSICPC\] shows maps of $\mathbb{E}({\mathbf{z}})$ for $N=40$ directions and $\lambda=0.4$. By comparing the results in FIG. \[MapMUSIC\], we can observe that Theorems \[Theorem1\] and \[Theorem2\] hold for the sound-soft and sound-hard arcs, respectively. Additional numerical results can be found in recent works [@PL1; @PL3]. Concluding remarks {#sec:5} ================== On the basis of the structure of the left singular vectors of the MSR matrix, we investigated the structure of the MUSIC-type imaging function by establishing a relationship between it and the Bessel function of integer order of the first kind. Using this relationship, we examined certain properties of the MUSIC algorithm. It is worth emphasizing that the MUSIC algorithm can be applied in limited-view inverse scattering problems. However, its structure has been identified for the sound-soft arc of small length [@JKP]. Hence, exploring the structure of MUSIC for the extended, sound-hard arc will be a future work. Finally, we have been considering the imaging of two-dimensional thin electromagnetic inclusions or perfectly conducting cracks. The analysis could be extended to a three-dimensional problem; refer to [@AILP] for related work. [99]{} , [A new optimal control approach for the reconstruction of extended inclusions]{}, SIAM J. Control. Optim., 51 (2013), pp. 1372–1394. , [Multistatic imaging of extended targets]{}, SIAM J. Imaging Sci., 5 (2012), pp. 564–600. , [Imaging schemes for perfectly conducting cracks]{}, SIAM J. Appl. Math, 71 (2011), pp. 68–91. , [A MUSIC algorithm for locating small inclusions buried in a half-space from the scattering amplitude at a fixed frequency]{}, Multiscale Model. Simul., 3 (2005), pp. 597–628. , [MUSIC type electromagnetic imaging of a collection of small three-dimensional inclusions]{}, SIAM J. Sci. Comput. 29 (2007), pp. 674–709. , [Reconstruction of Small Inhomogeneities from Boundary Measurements]{}, Lecture Notes in Mathematics, 1846, Springer-Verlag, Berlin, 2004. , [A direct algorithm for ultrasound imaging of internal corrosion]{}, SIAM J. Numer. Anal., 49 (2011), pp. 1177–1193. , [Asymptotic imaging of perfectly conducting cracks]{}, SIAM J. Sci. Comput., 32 (2010) pp. 894–922. , [Inverse scattering by a continuation method with initial guesses from a direct imaging algorithm]{}, J. Comput. Phys., 227 (2007), pp. 755–762. , [Asymptotic formulas for perturbations of the electromagnetic fields in the presence of thin imperfections]{}, Contemp. Math., 333 (2003), pp. 49–63. , [A level set method for inverse problems]{}, Inverse Problems, 17 (2001), pp. 1327–1355. , [Solving inhomogeneous inverse problems by topological derivative methods]{}, Inverse Problems, 24 (2008), 045014. , [MUSIC electromagnetic imaging with enhanced resolution for small inclusions]{}, Inverse Problems, 25 (2009), 015008. , [The linear sampling method and the MUSIC algorithm]{}, Inverse Problems 17 (2001), pp. 591–595. , [Super-resolution processing of multi-static data using time-reversal and MUSIC]{}, available at <http://www.ece.neu.edu/faculty/devaney/preprints/paper02n_00.pdf> , [Level set methods for inverse scattering]{}, Inverse Problems, 22 (2006), pp. R67–R131. , [A shape reconstruction method for electromagnetic tomography using adjoint fields and level sets]{}, 16 (2000), pp. 1119–1156. , [Multi-frequency orthogonality sampling for inverse obstacle scattering problems]{}, Inverse Problems, 27 (2011), 085005. , [MUSIC-characterization of small scatterers for normal measurement data]{}, Inverse Problems, 25 (2009), 075012. , [A regularized Newton method for locating thin tubular conductivity inhomogeneities]{}, Inverse Problems, 27 (2011), 115008. , [Identification of small inclusions from multistatic data using the reciprocity gap concept]{}, Inverse Problems, 28 (2012), 045011. , [A direct imaging algorithm for extended targets]{}, Inverse Problems, 22 (2006), pp. 1151–1178. , [MUSIC-type imaging of perfectly conducting cracks in limited-view inverse scattering problems]{}, Appl. Math. Comput., in revision. , [Structural behavior of the MUSIC-type algorithm for imaging perfectly conducting cracks]{}, Prog. Electromagn. Res., 138 (2013), pp. 211–226. , [Inverse scattering from an open arc]{}, Math. Methods Appl. Sci., 18 (2003), pp. 267–293. , [On the inverse acoustic scattering problem by an open arc: the sound-hard case]{}, Inverse Problems, 13 (1997), pp. 1379–1392. , [Singular Integral equations in Diffraction Theory]{}, Karpenko Physicomechanical Institute, Ukrainian Academy of Sciences, 210, Lviv, 1994. , [Mathematical modelling of electromagnetic scattering from a thin penetrable target]{}, Prog. Electromagn. Res., 55 (2005), pp. 95–116. , [Analysis of a multi-frequency electromagnetic imaging functional for thin, crack-like electromagnetic inclusions]{}, Appl. Numer. Math, 77 (2014), pp. 31–42. , [Electromagnetic MUSIC-type imaging of perfectly conducting, arc-like cracks at single frequency]{}, J. Comput. Phys., 228 (2009), pp. 8093–8111. , [MUSIC-type imaging of a thin penetrable inclusion from its far-field multi-static response matrix]{}, Inverse Problems, 25 (2009), 075002. , [Reconstruction of thin electromagnetic inclusions by a level set method, Inverse Problems]{}, 25 (2009), 085010. , [On the topological derivative in shape optimization]{}, SIAM J. Control Optim., 37 (1999), pp. 1251–1272. , [Imaging three-dimensional anisotropic scatterers in multi-layered medium by MUSIC method with enhanced resolution]{}, J. Opt. Soc. Am. A, 29 (2012), pp. 1900–1905. , [A vector level set method and new discontinuity approximations for crack growth by EFG]{}, Int. J. Numer. Meth. Engng, 54 (2002), pp. 923–944. , [MUSIC imaging and electromagnetic inverse scattering of multiple-scattering small anisotropic spheres]{}, IEEE Trans. Antennas Propag., 55 (2007), pp. 3542–3549. [^1]: e-mail: [email protected]	{ "pile_set_name": "ArXiv" }
--- abstract: 'The performance of neural network classifiers is determined by a number of hyperparameters, including learning rate, batch size, and depth. A number of attempts have been made to explore these parameters in the literature, and at times, to develop methods for optimizing them. However, exploration of parameter spaces has often been limited. In this note, I report the results of large scale experiments exploring these different parameters and their interactions.' author: - \| Thomas M. Breuel\ Google, Inc.\ [[email protected]]{} bibliography: - 'mlpbench.bib' title: The Effects of Hyperparameters on SGD Training of Neural Networks --- Datasets and Libraries ====================== All experiments reported here were carried out using the Torch library [@collobert2002torch] and CUDA (some of the experiments have been reproduced on a smaller scale with other libraries). The dataset for all the experiments is MNIST [@lecun1998mnist; @keysers2007comparison]. Characters were deskewed prior to all experiments. Deskewing significantly reduces error rates in nearest neighbor classifiers. Skew corresponds to a simple one-parameter family of linear transformations in feature space and causes decision regions to become highly anisotropic. Without deskewing, differences in performance between different architectures might primarily reduce to their ability to “learn deskewing”. With deskewing, MNIST character classification become more of an instance of a typical classification problem. Prior results on classifying deskewed MNIST data both with neural networks and with other methods are shown in the table below. Method Test Error Preprocessing Reference ---------------------------------- ------------ --------------------------------------------------- ---------------------------- Reduced Set SVM deg 5 polynomial 1 deskewing LeCun et al. 1998 SVM deg 4 polynomial 1.1 deskewing LeCun et al. 1998 K-nearest-neighbors, L3 1.22 deskewing, noise removal, blurring, 2 pixel shift Kenneth Wilder, U. Chicago K-nearest-neighbors, L3 1.33 deskewing, noise removal, blurring, 1 pixel shift Kenneth Wilder, U. Chicago 2-layer NN, 300 HU 1.6 deskewing LeCun et al. 1998 : \[tab-mnist-results\] Other previously reported results on the MNIST database. Logistic vs Softmax Outputs =========================== Multi-Layer Perceptrons (MLPs) used for classification usually attempt to approximate posterior probabilities and use those as their discriminant function. Two common approaches to this are the use of least square regression with logistic output units trained with a least square error measure (“logistic outputs”) and a softmax output layer (“softmax outputs”). In the limit of infinite amounts of training data, both approaches converge to true posterior probability estimates. Softmax output layers have the property that they are guaranteed to produce a normalized posterior probability distribution across all classes, while least square regression with logistic output units generates independent probability estimates for each class membership without any guarantees that these probabilities sum up to one. Softmax is often preferred, although there is no obvious theoretical reason why it should yield better discriminant functions or lower classification error for finite training sets. In OCR and speech recognition, some practitioners have observed that logistic outputs yield better posterior probability estimates and better results when combined with probabilistic language models. In addition, when the sum of the posterior probability estimates derived from logistic outputs differs significantly from unity, that is a strong indication that the input lies outside the training set and should be rejected. ![\[mlpsoftmax\] Training and test error for MLPs with logistic outputs (blue) and softmax output (red). Note that softmax outputs achieve down to zero percent training error (assigned to an error of 1e-4) but logistic outputs give overall better performance on new training samples. ](mlpbench-img/mlpbench-img1.png){height="3in"} Figure \[mlpsoftmax\] shows a scatterplot of test vs training error for a large number of MLPs with one hidden layer at different learning rates, different number of hidden units, and different batch sizes. Such scatterplots show what error rates are achievable by the different architectures, hyperparameter choices, initializations, and order of sample presentations. The lowest points in the vertical direction indicate the lowest test set error achievable by the architecture in this set of experiments. The scatterplot shows that logistic outputs achieve test set error rates of about 1.0% vs 1.1% for softmax outputs. At the same time, logistic outputs never achieve zero percent training set error, while softmax outputs frequently do. ![\[lrbatch\] Learning rate and batch size effects depending on output layer type. This is a scatterplot for all networks that yield test set error rates of less than 1.5%, with color indicating the test set error. Softmax outputs yield the best results for a learning rate that is about an order of magnitude smaller than logistic outputs. ](mlpbench-img/mlpbench-img2.png){width="\textwidth"} In order to ascertain that the difference in test set error between the two architectures is due to the architectures themselves, it is important to ensure that the space of hyperparameters (learning rates, batch sizes, number of hidden units) has been explored sufficiently. Note that, as Figure \[lrbatch\] shows, the in order to yield low error rates, softmax outputs require learning rates that are about an order of magnitude lower than logistic outputs. ![\[parange\] The complete range of parameters explored in the MLP experiments reported in this section. In this plot, color indicates error rate; circle size indicates the number of hidden units; a circle with a border represents softmax outputs. ](mlpbench-img/mlpbench-img3.png){height="3in"} Figure \[parange\] demonstrates that the range of parameters (learning rates, batch sizes) has been explored fully; above learning rates of 1e1, both softmax and logistic output models diverge, and at small learning rates, both fail to learn in a reasonable amount of time. The fact that logistic outputs yield 10% lower relative error rates on such a simple and widely studied problem and architecture compared to softmax outputs does not prove that “logistic outputs are better than softmax outputs”, but it suggests that it is worth testing both logistic and softmax outputs on any particular problem to see which one yields lower test set error. Batch size Effects ================== Optimizing the weights of a neural network can be carried out by stochastic gradient descent (updating after each sample) or by full gradient descent (computing a gradient on the parameters from the entire training set). In between those two extremes is batched gradient descent. In batched gradient descent, we update the parameters of the MLP to reduce the error for a small sample (“batch”) of training samples, typically consisting of between 10 and 1000 samples. Computationally, using batches instead of individual training samples allows for greater parallelism; each layer of an MLP computes effectively a function like: $$y = \sigma(M\cdot x)$$ For simple SGD, $x$ is some d-dimensional vector, but for batched gradient descent, $x$ is a $d\times b$ dimensional matrix, where $b$ is the batch size. The matrix multiplication $M\cdot x$ can be evaluated much more efficiently and in parallel than evaluating $b$ individual matrix-vector products in single-sample updates. In fact, if we use $b$ processors, we can compute updates for $b$ samples in roughly the same time as we would otherwise use for a single sample in SGD. When using batch training, a common convention is to rescale the learning rate $\lambda \rightarrow \frac{\lambda}{b}$ This means that as we increase the batch size $b$, we need to scale up the learning rate proportionately. This is the convention we use in these experiments. ![\[bslr\] Batch size and learning rate vs. error rate (logistic outputs). See the text for an explanation. ](mlpbench-img/mlpbench-img15.png){height="3in"} Figure \[bslr\] shows a scatterplot of trained networks with good test set error vs. batch size and learning rate. There are three apparent limits on performance: 1. At the lower end, we have a soft transition from well performing networks to poorly performing networks. The explanation of this is that at low learning rates, the network learns too slowly to reach low test set errors within the limited number of training steps. The reason this soft transition slopes upwards is due to the use of normalized learning rates. Without learning rate normalization, this line would remain horizontal and independent of batch size. 2. At a batch size of 1, there is a maximum learning rate; beyond that learning rate, the stochastic gradient descent optimization diverges. Without batch size normalization of learning rate, this upper limit would exist independent of batch size; due to batch size normalization, this line of divergence slopes upwards, parallel to the soft lower bound on learning rates. 3. There is a third, unexpected, constant limit on the batch-normalized learning rate. The original single sample learning rate determines the speed of convergence of the stochastic gradient descent algorithm; within its region of convergence, halving the learning rate approximately doubles the time needed to reach a given test set error, since each gradient update represents simply a step towards the minimum along some path. This reasoning also applies for batch updates. The constant upper limit on the batch-normalized learning rate corresponds to a single sample learning rate that decreases proportionally to batch size. The consequence is that, as long as the maximum usable batch normalized learning rate increases proportionally with batch size, we benefit from parallelization in terms of overall speedup of learning. Once we enter the regime where the upper limit of batch normalized learning rates is independent of batch size, further parallelization does not speed up training. Without further experimentation, we can only guess at the source of the upper limit on the batch-normalized learning rate. As a simplified model, assume that the limit on the learning rate for single sample updates is due to some subset of the input vectors (e.g., vectors that generate particularly large gradients). For batch sizes that contain, on average, only one of those input vectors, we can continue to use the original learning rate, but once we use a batch size that contains, on average, two of those vectors, we have to cut the learning rate in half in order to keep the magnitude of the update within the range that allows convergence. (In practice, we are not necessarily looking at individual samples but subspaces of the input.) This analysis suggests possible strategies for improving training performance with large batch sizes that will be explored elsewhere. Regardless of the speed of optimization, we can also ask the question of how the test set error of the final network depends on batch size. This is shown in Figure \[bste\]. We see that increasing batch size generally results in worse test set errors for both logistic outputs and softmax outputs. The dependence is somewhat stronger for logistic outputs. In addition, logistic outputs appear to yield networks with a higher variability. Note that the differences in error rates in this plot are much smaller than the differences in error rates found in the previous learning rate plots; this plot makes small but significant differences among the very best models visible. ![\[bste\] Test error by batch size for logistic and softmax outputs. Note that logistic output units achieve the lowest error. Also note that for larger batch sizes, the advantage of logistic output units over softmax output units disappears. ](mlpbench-img/mlpbench-img16.png){width="\textwidth"} The observations on the relationship between batch sizes and learning rates above also have implications for hyperparameter optimization. In particular, the hyperparameter search at large batch sizes becomes harder because the range of learning rates that yield good networks is considerably smaller than it is at small batch sizes (Figure \[hypnar\]). Batch sizes that are too large therefore not only waste computational resources through parallelism that does not result in a speedup for learning, they may actually make the hyperparameter search harder. ![\[hypnar\] At large batch sizes, the range of hyperparameter resulting in good test set errors is considerably smaller than at small batch sizes. ](mlpbench-img/mlpbench-img17.png){height="3in"} Convolutional Layers ==================== The above results were all obtained for non-convolutional networks with a single hidden layer. How do they generalize to convolutional networks? There are, of course, many different kinds of convolutional architectures we could investigate. The simplest architecture places a single convolutional layer at the input of the network. ![\[convscatter\] Training error vs test set error for non-convolutional (blue) and convolutional (red) networks. ](mlpbench-img/mlpbench-img18.png){height="3in"} Not surprisingly, adding a convolutional layer results in significantly lower test set error (0.69% test set error) compared to non-convolutional networks (1% test set error), as seen in Figure \[convscatter\]. ![\[bsconv\] Test set error by batch size for convolutional networks. Output units are softmax, and the plot shows the best networks across all learning rates and number of convolutional units. Note the significant increase in best achievable error rates with increasing batch size. ](mlpbench-img/mlpbench-img19.png){height="3in"} For convolutional networks, we also observe that small batch sizes yield the best test set errors. In fact, large batch sizes never reach comparably low error rates (Figure \[bsconv\]). ReLU Units (Not Convolutional) ============================== Another popular architectural choice is to replace the sigmoidal nonlinearity in hidden layers with rectifying linear units (ReLU). To explore the effects of this on the results, networks were trained with all four combinations of sigmoid/ReLU hidden units and softmax/sigmoid output units. The scatterplot of results is shown in Figure 9. We see in Figure \[sigrel\] that softmax outputs achieve zero percent test set error for both kinds of hidden layer nonlinearities. Furthermore, ReLU hidden units outperform sigmoidal hidden units for either kind of output layer. The overall best performing combination was logistic output units with ReLU hidden units, resulting in a test set error of 0.92%. ![\[tsbatch\] Dependence of test set error on batch size. We see that for sigmoidal hidden units (blue, cyan), large batch sizes perform considerably worse than small batch sizes. For ReLU hidden units (red, magenta), the dependence is considerably weaker, although smaller batch sizes still seem to have a slight advantage. ](mlpbench-img/mlpbench-img22.png){height="3in"} ![\[tshidden\] Dependence of test set error on the number of hidden units. For sigmoidal hidden units (blue, cyan), there is little improvement of test set error with increasing numbers of hidden units. For ReLU hidden units, there is a strong improvement in test set error with the number of hidden units. The maximum number of hidden units tested was 2000, although it looks like larger numbers of hidden units might result in even better performance. ](mlpbench-img/mlpbench-img23.png){height="3in"} ![\[tsall\] Error rates (indicated by color) vs batch size, learning rate, and number of hidden units (indicated by circle size). These scatterplots suggest that the parameter ranges for all four conditions were explored fairly completely. ](mlpbench-img/mlpbench-img24.png){height="3in"} ReLU Units (Convolutional) ========================== The previous results for ReLU units are interesting, given the low test set error rate and low dependence on batch size. It’s interesting to see whether we can reproduce those results for convolutional networks. However, in the case of convolutional networks, we find a significant batch size dependence, and that the difference between sigmoidal and ReLU hidden units is considerably smaller than for non-convolutional networks. Furthermore, some softmax networks also come close in performance (Figure 13). ![\[bsconvrelu\] Test error vs. batch size for convolutional networks with ReLU units. ](mlpbench-img/mlpbench-img25.png){width="3in"} ![\[convparm\] The batch size and learning rate parameter space explored for the comparison of convolutional ReLU and sigmoidal networks. ](mlpbench-img/mlpbench-img26.png){width="4in"} Deep ReLU Networks ================== ReLU networks also appear to be highly successful for training deep networks. We explore this architectural variant in these experiments by comparing deep networks with ReLU hidden units and sigmoidal hidden units. We find that the performance of deep networks with sigmoidal units degrades with depth, while deep networks with ReLU networks yield good performance even at large (eight layer) depths (Figure 15). However, increasing depth for ReLU networks does not result in better test set performance. The deterioration of test set error with increasing depth for sigmoidal hidden units is probably due to effects like vanishing gradients, something that ReLU networks do not seem to suffer from to the same degree. ![\[deepneterr\] Test set error by network depth for networks with sigmoidal hidden units (blue) and ReLU hidden units (red). ](mlpbench-img/mlpbench-img27.png){width="\textwidth"} For both types of deep networks, we can ask again how test set error depends on batch size. Figure 16 shows that for both types of deep networks and across a range of hidden layers, the best achievable test set error generally increases with increasing batch size. ![\[deepnetsize\] Test set error vs batch size for deep networks with ReLU (red) and sigmoidal (blue) hidden layers. Circle size indicates number of hidden layers. Notice that there is a significant decrease in test set performance with increasing batch size. ](mlpbench-img/mlpbench-img28.png){height="3in"} ![\[deepnetbs\] Test set error (indicated by color with red being high, blue being low) vs \# hidden layers and batch size. Note that deep networks are much more sensitive to large batch sizes during training than shallow networks. ](mlpbench-img/mlpbench-img29.png){width="\textwidth"} ![\[deepnetlr\] Parameters spaces explored by the deep ReLU and sigmoidal networks. The parameter spaces show that the learning rate parameters were explored sufficiently well for both network types. ](mlpbench-img/mlpbench-img30.png){width="\textwidth"} Conclusions and Discussion ========================== Machine learning algorithms cannot be totally ordered by performance and there is no single best learning algorithm. Nevertheless, machine learning benchmarks in general, and benchmarks on MNIST in particular, tell us something about how machine learning algorithms compare on typical classification tasks, and what kind of architectural features influence performance significantly. That is, the conclusions we can draw from such benchmarks are not so much about which algorithm is “better”, but rather which algorithmic choices may affect the outcome positively or negatively. Perhaps the most important result from these benchmarks is how complex the interaction between different architectural features and conditions is; performance improvements that can be demonstrated in simple networks do not add up when combined together into the same architecture. Furthermore, many benchmarks that have been carried out in the literature may have been hampered by limited sets of training conditions. For example, benchmarking logistic vs. softmax outputs at larger batch sizes suggests that there is little difference between the two methods; however, at small batch sizes, logistic outputs significantly outperform softmax outputs on MNIST data (Figure \[bste\]). Unless both methods are tested at small batch sizes, the significantly better performance of logistic outputs is not revealed. As a second example, for some architectures, ReLU networks show no batch size dependencies, while other network architectures do show such dependencies. It is important to remember that the MNIST dataset is not necessarily representative of other classification problems: it has a small number of classes, the prior probability is uniform, the number of training samples is small compared to many other problems, all geometric variability (translations, rotations, skew) has been removed, and the input vectors are binary. Therefore, more important than the results about what works are the result of what unexpectedly doesn’t work well even in such a simple case. Based on the experiments reported here, we observe: - For many problems, increasing batch sizes in a parallel implementation results in no speedup in training because the per-sample learning rate needs to be scaled down proportionately to batch size. Furthermore, large batch sizes may intrinsically limit the performance of networks. Finally, hyperparameter optimization may get harder for larger batch sizes, as the range of feasible learning rates (and other parameters) gets narrower. Therefore, it is a good idea to carry out experiments with single sample updates and small batch sizes. - Softmax outputs may yield lower training errors than logistic outputs, but often also yield higher test set errors. Therefore, it is a good idea to try both kinds of outputs when training neural networks on different tasks. In doing so, it is important to try a wide range of learning rates, since the optimal learning rates for the two kinds of outputs are different. - For non-convolutional networks, ReLU hidden layer units perform significantly better than sigmoidal hidden layer units in these experiments; they also show lower batch size dependencies and scale to much larger numbers of hidden units. For convolutional networks, however, these effects were not observed, suggesting that both sigmoidal and ReLU non-linearities should be tried. - It is much easier to train deep networks using ReLU hidden layers than hidden layers with sigmoidal non-linearities. However, additional depth does not improve test set error for either sigmoidal or ReLU units. For deep networks, we also observed batch size dependencies. In addition, the range of good learning rates shifts and becomes smaller for deeper networks. Generally, these results suggest the following strategy for training new networks: - Start with single sample updates, both during initial exploration and hyperparameter search. - When exploring new problems, compare softmax and logistic outputs, as well as ReLU and sigmoidal hidden units. - Although deep ReLU networks can be trained, be sure to try shallow ReLU networks as well, and test a wide range of learning rates. We note that there has been an extensive literature on various improved optimization methods for neural network learning, methods for learning hyperparameters, and benchmarks of MLP performance. It is impossible to do this literature justice in this technical report. However, a few general observations should suffice: - The methods described in this paper all relied on simple SGD training, yet yield excellent performance compared to other reported results. In particular, optimization or hyperparameter selection methods that yield significantly worse results than those reported here are of questionable utility. - Generally speaking, hyperparameter optimization for these these kinds of problems does not seem to be particularly critical; networks yield similar performance over a broad range of hyperparameters. - Hyperparameter optimization should not optimize for the best expected test set error of the resulting networks, but for the minimal error over a collection of multiple trained models.	{ "pile_set_name": "ArXiv" }
--- abstract: 'High momentum hadron suppression is considered to be an excellent probe of jet-medium interactions in QCD matter created in ultra-relativistic heavy ion collisions. We previously showed that our dynamical energy loss formalism can accurately explain suppression measurements at 200 GeV Au+Au collisions at RHIC and 2.76 TeV Pb+Pb collisions at the LHC. With the upcoming LHC measurements at notably higher collision energies, there is a question of what differences, with respect to the current (2.76 TeV) measurements, can be expected. In this paper we concentrate on heavy flavor suppression at the upcoming 5.1 TeV Pb+Pb collisions energy at the LHC. Naively, one would expect a notably ($\sim 30\%$) larger suppression at 5.1 TeV collision energy, due to estimated (significant) energy loss increase when transitioning from 2.76 to 5.1 TeV. Surprisingly, more detailed calculations predict nearly the same suppression results at these two energies. We show that this unexpected result is due to an interplay of the following two effects, which essentially cancel each other: [i]{}) flattening of the initial distributions with increasing collision energies, and [ii]{}) significantly slower than naively expected increase in the energy loss. Therefore, the obtained nearly the same suppression provides a clear (qualitative and quantitative) test of our energy loss formalism.' author: - Magdalena Djordjevic - Marko Djordjevic title: 'Heavy flavor suppression predictions at 5.1 TeV Pb+Pb collisions at LHC' --- Introduction ============ High energy heavy flavor suppression [@Bjorken] is considered to be an excellent probe of QCD matter created in ultra-relativistic heavy ion collisions at RHIC and LHC. One of the major goals of these experiments is mapping the QGP properties, which requires comparing available suppression data with the theoretical predictions [@STE; @STE1; @STE2]. Such comparison tests different theoretical models and provides an insight into the underlying QGP physics. Having this in mind the upcoming 5.1 TeV Pb+Pb measurements at LHC (expected at the end of 2015) - and their comparison with theoretical predictions - will provide an additional important insight in the jet-medium interactions in QGP created in such collisions. With this motivation in mind, the goal of this paper is providing the heavy flavor suppression predictions, and physical interpretation behind the obtained results, for the upcoming high-luminosity experimental data at 5.1 TeV Pb+Pb collisions at LHC. In particular, we aim assessing the differences in the predicted suppression with respect to the already available 2.76 TeV measurements at LHC and compare the results of state-of-the-art calculations with simple expectations/estimates. To generate the theoretical predictions we will use our recently developed dynamical energy loss formalism, which includes: i) dynamical scattering centers, ii) QCD medium of a finite size [@DynEL; @DynEL1], iii) both radiative [@DynEL; @DynEL1] and collisional [@MD_Coll] energy losses, iv) finite magnetic mass effects [@MagM] and v) running coupling [@RunnC]. This energy loss formalism is based on the pQCD calculations in finite size optically thin dynamical QCD medium, and has been incorporated into a numerical procedure [@RunnC] that allows generating state-of-the art suppression predictions. The model has shown to be successful in explaining a wide range of angular averaged observables [@HFLHC; @RunnC; @CRHIC; @NCLHC] at both RHIC and LHC. Since the angular averaged $R_{AA}$s are largely insensitive to the medium evolution, angular averaged $R_{AA}$ can be considered an excellent probe for jet-medium interactions [@Thorsten; @Molnar; @Footnote1]; consequently, the suppression predictions at 5.1 TeV Pb+Pb collisions at LHC, and their comparison with the measurements, will allow further testing of our energy loss formalism. Overview of the computational framework ======================================= For generating the suppression predictions, we use the computational procedure from [@RunnC]. The main features are briefly summarized below, while the full account of the procedure is provided in [@RunnC]. The quenched spectra of heavy flavor observables are calculated according to the generic pQCD convolution: $$\begin{aligned} \frac{E_f d^3\sigma}{dp_f^3} = \frac{E_i d^3\sigma(Q)}{dp^3_i} \otimes {P(E_i \rightarrow E_f )} \otimes D(Q \to H_Q) \otimes f(H_Q \to e, J/\psi). \; \label{schem} \end{aligned}$$ In the equation above subscripts “i” and “f” correspond, respectively, to “initial” and “final”, and $Q$ denotes heavy quarks. $E_i d^3\sigma(Q)/dp_i^3$ denotes the initial heavy quark spectrum, which is computed at next to leading order according to [@Cacciari:2012; @FONLL]. $P(E_i \rightarrow E_f )$ is the energy loss probability; this probability includes both radiative and collisional energy loss in a finite size dynamical QCD medium, multi-gluon [@GLV_suppress] and path-length fluctuations [@WHDG] and running coupling [@RunnC]. $D(Q \to H_Q)$ is the fragmentation function of heavy quark $Q$ to hadron $H_Q$, where for D and B mesons we use, BCFY [@BCFY] and KLP [@KLP] fragmentation functions, respectively. Finally, decay of B mesons to experimentally measured non-prompt $J/\psi$ is represented by $f(H_Q \to J/\psi)$ and obtained according to [@Cacciari:2012]. The expression for the radiative energy loss in a finite size dynamical QCD medium is extracted from Eq. (10) in [@MagM], while the collisional energy loss is extracted from Eq. (14) in [@MD_Coll]. Path length distributions are taken from [@Dainese]. The angular averaged $R_{AA}$ is a clear jet-medium interaction probe, i.e. it is not sensitive on the details of the medium evolution [@Molnar; @Thorsten], so we model the medium by assuming constant average temperature of QGP. To determine the average temperatures at 0-10% most central collisions, we start from $T{\,=\,}304$MeV (the effective temperature extracted by ALICE [@LHC_T] for 0-40% centrality), and use the procedure outlined in [@NCLHC] (based on gluon rapidity density) to determine the temperatures at central collisions at 2.76 and 5.1 TeV Pb+Pb collisions; for 2.76 TeV 0-10% centrality, this leads to the average temperature of 313 MeV. To determine the temperature at 5.1 TeV, note that it is expected that the gluon rapidity density will be 25% higher at 5.1 TeV than at 2.76 TeV in Pb+Pb collisions at LHC [@LHC_Mult]. Since the temperature is proportional to the gluon rapidity density, i.e. $T \sim (dN_g/dy)^{1/3}$, this leads to $\sim 7\%$ higher temperature at 5.1 TeV compared to 2.76 TeV at the LHC, i.e. 335 MeV for 0-10% central 5.1 TeV Pb+Pb collisions at LHC. Note that, in our energy loss calculations, this is the only parameter that differs between the two systems; i.e. all the other parameters that enter in the calculations (stated in the next paragraph) are the same for the two systems, and correspond to the standard literature values (i.e. no parameters are determined through fitting the data). The following parameters are used in the numerical calculations: QGP with effective light quark flavors $n_f{\,=\,}3$ and perturbative QCD scale of $\Lambda_{QCD}=0.2$ GeV. The Debye mass is taken to be $\mu_E \approx 0.9$ ($\mu_E \approx 0.97$) GeV for 2.76 (5.1) TeV collision energy, and is obtained by self-consistently solving Eq. (7) in [@Peshier]. The value for magnetic to electric mass ratio $\mu_M/\mu_E$ is extracted from non-perturbative calculations [@xb1; @xb2; @xb3; @xb4] $0.4 < \mu_M/\mu_E < 0.6$; the gluon mass is $m_g=\mu_E/\sqrt{2}$ [@DG_TM], while the charm and the bottom mass are, respectively, $M{\,=\,}1.2$GeV and $M{\,=\,}4.75$GeV. Path-length distribution, parton production, fragmentation functions and decays, which are used in the numerical calculations, are specified above. Results and discussion ====================== To get an insight of what results we expect at 5.1 TeV collisions at the LHC, we will first provide a simple analytic estimate for heavy flavor suppression at this collision energy. For that purpose, note that radiative energy loss is widely considered to be a dominant energy loss mechanism in QGP, so we will use only the radiative contribution for the estimate. Since it is also widely assumed that radiative energy loss is proportional to $T^3$ (see e.g. [@Betz]), one can estimate that the energy loss at 5.1 TeV should be $ \sim 25\%$ higher than at 2.76 TeV. Based on this, and if we assume that initial distributions can be approximated by power low distributions, i.e. $d\sigma/dp_\perp^2 \sim 1/p_\perp^n$, we can make an estimate on how much larger/smaller suppression one would expect at 5.1 TeV compared to the already observed results at 2.76 TeV. It was previously shown that, for radiative energy loss and power low initial distributions, suppression can be roughly estimated by using the following simple formula [@GLV_suppress; @Footnote2]: $$\begin{aligned} (1-\frac{1}{2} \frac{\Delta E}{E})^{(n-2)} &\approx& (1-\frac{n-2}{2} \frac{\Delta E}{E}), \label{suppSimpolified}\end{aligned}$$ where $\Delta E/E$ is the fractional energy loss. If we assume that, at 2.76 TeV, typical fractional energy loss for charm is $\sim 30\%$ and for bottom $\sim 15\%$, and that charm and bottom distributions do not notably change between these two collision energies, with $n \sim 6.5$ ($n \sim 6$) for charm (bottom), the above estimate will straightforwardly lead to the expectation of $\sim 30 \%$ ($\sim 10 \%$) larger suppression for charm (bottom) at 5.1 TeV compared to 2.76 TeV Pb+Pb collisions at the LHC. Contrary to these expectations, Figure \[SuppFig1\] shows that our suppression calculations - obtained from the energy loss formalism outlined in the previous section - provide substantially different predictions. From this figure, we actually do not observe any suppression increase between 2.76 to 5.1 TeV collisions at the LHC. That is, we obtain the same suppression patterns for both charm and bottom probes (D mesons and non-prompt $J/\psi$) at these two collision energies. This then leads to the question of, why the increase in the collision energy by almost a factor of 2, leads to the same predicted suppression patterns between the two collisional energies, despite the estimated significant (i.e. $\sim 30\%$ for charm, see above) increase in the suppression? To address this question, in Fig. \[ptdist\] we first compare charm and bottom initial distributions between these two collision energies. From this figure, we see that the distributions at 5.1 TeV are slightly flatter than at 2.76 TeV, for both charm and bottom, which will have the tendency to somewhat lower the suppression at 5.1 TeV compared to 2.76 TeV. Note that only the shape of the distributions contributes to the suppression predictions, and from Fig. \[ptdist\], one can observe that the differences in the shape of the distributions are not large. Still, this difference in the distributions has a notable (though again not large, i.e. $\sim 5 \%$) effect on the suppression predictions, as can be seen in the left panel of Fig. \[DecreaseIncrease\]; therefore, it should be taken into account in the suppression calculations. However, what we further see from the right panel of Fig. \[DecreaseIncrease\] is that the effect on the suppression coming from the energy loss increase between 2.76 and 5.1 TeV (due to the increase in average temperature) is also notable but not large, i.e. it corresponds to $5$ and $10 \%$. That is, the energy loss effect on the suppression has about the same magnitude, but an opposite direction, compared to the effect of different distributions between these two collision energies. The first question that we want to address is why the effect of the energy loss increase on jet suppression is not larger between these two collision energies, at least not for charm quark. That is, based on the common $T^3$ assumption, we have estimated that the energy loss increase should be on the order of $25\%$, which should, therefore, have a more prominent (estimated 30%) effect on the suppression. Regarding the $T^3$ estimate for the radiative energy loss, note that, while widely used, this estimate does not have to be justified. That is, from Eq. (10) in [@MagM], which shows the radiative energy loss expression in a finite size dynamical QCD medium, it can be straightforwardly observed that the expression nontrivially depends on T. That is, while one can recover a part with explicit dependence on $T^3$ in this expression, the rest of the expression also depends on T, where this extra term considerably modifies the temperature dependence. Additionally, the collisional energy loss effect, while smaller compared to the radiative, is still important, and this effect also has to be taken into account in the suppression calculations. Note that, for the collisional energy loss, it is commonly assumed that it has a quadratic ($T^2$) dependence on the temperature. However, similarly to the above discussion for the radiative energy loss, Eq. (14) from [@MD_Coll] shows a nontrivial temperature dependence, so we will below also test whether this simple ($T^2$) assumption is justified. With these goals, in Fig. \[ELossRatioDiffTemp\], we plot the relative heavy flavor energy loss increase between 2.76 and 5.1 collision energies at the LHC. Figures also contain dashed horizontal lines, which represent what would be the energy loss increase, if it would indeed have $T^2$ or $T^3$ dependence. For radiative energy loss, we see that, contrary to the common expectations, energy loss increase is far from $T^3$ dependence; i.e. it is between linear (for low jet energy regions) and quadratic (which can be reached for asymptotically high jet energies). Consequently, for the high momentum heavy flavor hadrons that will be studied at these two collision energies at the LHC, the expected energy loss increase is notably smaller than quadratic, i.e. it is in the region between $5-10\%$ (note that the average temperature increase between these two collision energies is $\sim 7\%$). For the collisional energy loss, we also see that energy loss increase is far from quadratic, i.e. the increase of $\sim 8.5 \%$ is constant with momentum and it has slightly larger than linear dependence on temperature. Consequently, contrary to the common expectation, the total energy loss has also a modest temperature dependence, which is close to linear, i.e. between 6 and 10% depending on the jet momentum. This modest energy loss increase between these two collisional energies consequently leads to a modest increase in the suppression which we observe in the right panel of Fig. \[DecreaseIncrease\]. Finally, in Fig. \[Suppression\], we study the combined effect of the differences in the distributions and the energy loss on jet suppression. On the two left panels, we see the effect of the difference in the distributions on the jet suppression, while the energy loss is kept fixed. On the two central panels, we keep the same distribution, but change the energy loss, while in the two right panels both the distributions and the energy loss are changed between the two collision energies. From the panels, we see that, while the change in the distribution has the tendency to reduce the suppression, the energy loss increase increases the suppression for about the same amount, so that the resultant suppression at 5.1 TeV collision energy is almost the same as at 2.76 TeV. The above obtained numerical result can also be directly estimated from Eq. \[suppSimpolified\]. For this purpose, we will take that the energy loss between 2.76 and 5.1 TeV collision energy increases by factor $\eta$, where from Fig. \[ELossRatioDiffTemp\], we see that $\eta \approx 10\%$ for both charm and bottom. Additionally, we will take that the power factor in the initial parton distributions decrease by $\delta$; by fitting the power low to the ratio of the momentum distributions in Fig. \[ptdist\], we obtain $\delta \approx 0.4$. By applying these factors into Eq. \[suppSimpolified\], one can straightforwardly obtain $$\begin{aligned} R_{AA} (5.1 \, {\rm TeV}) \approx R_{AA} (2.76\, {\rm TeV})+\frac{1}{2}\frac{\Delta E}{E}(\delta-\eta (n-2)), \label{suppComparison}\end{aligned}$$ where for $\delta$ and $\eta$ estimated above the second additive in the above becomes close to zero. Consequently, this estimate also recovers the conclusion of the same heavy flavor suppression at 2.76 and 5.1 TeV Pb+Pb collision energies at the LHC. Conclusion ========== In this paper, we provided heavy flavor suppression predictions for the upcoming 5.1 TeV Pb+Pb collisions at the LHC. Based on our energy loss formalism, we predict the same heavy flavor suppression patterns for 2.76 and 5.1 TeV collision energies. This result is surprising since, based on the commonly used assumption, a notable increase of the suppression is expected at the higher collision energy. We showed that the same suppression is a consequence of the interplay between the following two effects: [i]{}) a decrease in the suppression due to flattening of the initial momentum distributions, and [ii]{}) an increase in the suppression - though more moderate than expected - due to higher energy loss. Consequently, this unexpected, but simple, suppression prediction provides a direct (both quantitative and qualitative) test of our understanding of the medium interactions in QCD medium created in these collisions. [Acknowledgments:]{} This work is supported by Marie Curie International Reintegration Grant within the $7^{th}$ European Community Framework Programme PIRG08-GA-2010-276913 and by the Ministry of Science and Technological Development of the Republic of Serbia, under projects No. ON171004 and ON173052. J.D. Bjorken, FERMILAB-PUB-82-059-THY (1982) 287-292. N. Brambilla et al., Eur. Phys. J. C [74]{} (2014) 2981. M. Gyulassy, Lect. Notes Phys. [583]{} (2002) 37. D. d’Enterria and B. Betz, Lect. Notes Phys. [785]{} (2010) 285. M. Djordjevic, Phys. Rev. C [80]{} (2009) 064909. M. Djordjevic and U. Heinz, Phys. Rev. Lett. [101]{} (2008) 022302. M. Djordjevic, Phys. Rev. C [74]{} (2006) 064907. M. Djordjevic and M. Djordjevic, Phys. Lett. B [709]{} (2012) 229. M. Djordjevic and M. Djordjevic, Phys. Lett. B [734]{} (2014) 286. M. Djordjevic and M. Djordjevic, Phys. Rev. C [90]{} (2014) 034910. M. Djordjevic, Phys. Rev. Lett. [112]{} (2014) 4, 042302. M. Djordjevic, M. Djordjevic and B. Blagojevic, Phys. Lett. B [737]{} (2014) 298. D. Molnar and D. Sun, Nucl. Phys. A [932]{} (2014) 140; [910-911]{} (2013) 486. T. Renk, Phys. Rev. C [85]{} (2012) 044903. This is in distinction to high-momentum elliptic flow, which is sensitive to the medium evolution, and therefore is a main probe for the properties of the medium created in ultra-relativistic heavy ion collisions. M. Cacciari, S. Frixione, N. Houdeau, M. L. Mangano, P. Nason and G. Ridolfi, JHEP [1210]{}, 137 (2012) M. Cacciari, M. Greco and P. Nason, JHEP [9805]{}, 007 (1998) M. Cacciari, S. Frixione and P. Nason, JHEP [0103]{}, 006 (2001) M. Gyulassy, P. Levai and I. Vitev, Phys. Lett. B [538]{}, 282 (2002). S. Wicks, W. Horowitz, M. Djordjevic and M. Gyulassy, Nucl. Phys. A [784]{}, 426 (2007). M. Cacciari, P. Nason, JHEP [0309]{}, 006 (2003), E. Braaten, K.-M. Cheung, S. Fleming and T. C. Yuan, Phys. Rev. D [51]{}, 4819 (1995) V.G. Kartvelishvili, A.K. Likhoded, V.A. Petrov, Phys. Lett. B [78]{}, 615 (1978). A. Dainese, Eur. Phys. J. C [33]{}, 495 (2004). M. Wilde \[for the ALICE Collaboration\], Nucl. Phys. A [904-905]{} (2013) 573c. K. Aamodt [et al.]{} \[ALICE Collaboration\], Phys. Rev. Lett. [105]{}, 252301 (2010). This formula should not be used for reliable predictions - we here use it only for the purpose of analytical estimate. A. Peshier, arXiv:hep-ph/0601119 (2006). A. Nakamura, T. Saito and S. Sakai, Phys. Rev. D [69]{} (2004) 014506. A. Hart, M. Laine and O. Philipsen, Nucl. Phys. B [586]{} (2000) 443. D. Bak, A. Karch and L.G. Yaffe, JHEP [0708]{} (2007) 049. Yu. Maezawa [et al.]{} \[WHOT-QCD Collaboration\], PoS Lattice (2008) 194. M. Djordjevic and M. Gyulassy, Phys. Rev. C [68]{}, 034914 (2003). B. Betz and M. Gyulassy, JHEP [1408]{}, 090 (2014) \[Erratum-ibid. 1410, 043 (2014)\].	{ "pile_set_name": "ArXiv" }
--- abstract: 'We show that the topology of a simply-connected smooth projective surface is determined by its algebraic structure modulo p.' author: - Dosang Joe and Minhyong Kim title: On the topology of algebraic surfaces and reduction modulo p --- Observation =========== The goal of this note is to make the following simple [Observation:]{} Let $X$ and $Y$ be simply-connected smooth projective surfaces. Suppose they are isomorphic modulo $p$ for some prime $p \neq 2$ of good reduction. Then $X$ and $Y$ are homeomorphic by a map preserving the complex orientations. Here, isomorphic modulo $p$ means the following: We can find a common finitely-generated ring of definition $R$ for $X$ and $Y$ (that is, such that $X$ and $Y$ can both be defined by equations with coefficients in $R$), and a prime ideal with residue field $k$ of characteristic $p$ such that $$X\otimes_R \bar{k}\simeq Y\otimes_R \bar{k}$$ That is, $X$ and $Y$ are isomorphic when the coefficients ae regarded as lying in the algebraic closure of $k$. ‘Good reduction’ means the two varieties appearing in this isomorphism are smooth over $\bar{k}$. For example, they might both be defined over $\Z$, in which case the meaning of everything is clear. Here is the proof: By Freedman’s theorem ([@Fr], theorem 1.5) the oriented homeomorphism-type of a simply-connected oriented smooth 4-manifold $X$ is determined by the unimodular intersection pairing $$B: H^2(X,\Z) \otimes H^2(X,\Z) \ra H^4(X,\Z) \simeq \Z$$ on the free abelian group $H^2(X,\Z) $. We examine this for algebraic surfaces. By the smooth and proper base-change theorems ([@Gr], chapters 12-14), $$B_l: [H^2(X,\Z)\otimes \Z_l] \otimes [H^2(X,\Z)\otimes \Z_l] \ra \Z_l,$$ the $\Z_l$-linear extension of $B$, is determined by $X$ mod $p$ for any $l\neq p$. In particular, the rank and type (even or odd, by taking $l=2$) of $B$ is determined by reduction mod $p$. On the other hand, the signature of $X$ is determined by Hirzebruch’s signature theorem ([@Hi], theorem 8.2.2) to be $$(c_1^2(X)-2c_2(X))/3$$ The numerical invariants in this formula are preserved under specialization. Therefore, we conclude that the rank, type, and signature of $B$ are determined by reduction mod $p$. This concludes the proof when $B$ is indefinite ([@Se], chapter 5, theorem 6). But if $B$ is definite, Donaldson’s theorem ([@Do],[@DK] theorem 1.3.1) says that $$B \cong \pm (x_1^2+x_2^2+\cdots x_r^2)$$ where $r$ is the rank. But whether or not $B$ is definite (as well as the sign) is also determined by the rank and signature, and hence, by the reduction modulo $p$. So we are done. Comments ======== Determination of topological invariants of varieties by modulo $p$ arithmetic is of course a well-known side-effect of modern arithmetic geometry. For a smooth and proper variety, the Betti numbers, for example, are determined by the reduction of the variety modulo a (good) prime $p$. More intricate invariants can also be brought in (where we always assume that the variety is smooth and proper and the prime is good): -The Hodge numbers, for example, are determined by reduction modulo an [ordinary]{} prime $p$ ([@BK] Theorem (0.7)). -The integral cohomology groups are determined by reduction modulo [two]{} primes: this is because two primes are sufficient to determine all $H^i(X,\Z_p)$, and then, $H^i(X,\Z)$ by the universal coefficient theorem. -For simply-connected varieties, the rational homotopy groups are determined by reduction modulo $p$ ([@AM], [@DGMS], [@KH]). -For simply-connected varieties, the integral higher homotopy groups are determined by reduction modulo two primes [@AM]. But the case of simply-connected surfaces is the only one we know of where something as refined as the homeomorphism-type is actually determined by reduction modulo $p$. Our proof is of course a consequence of the very powerful classification theorems for four-manifolds. As such, it appears essentially to be an accident. On the other hand, it would be interesting to seek out other non-trivial examples of such theorems, if only to probe their accidental nature. We point out also that there are parallel results where topological invariants are preserved by [conjugation]{}, that is, hitting the coefficients of some defining equations with an automorphism of the complex numbers. As is obvious from our proof, this would be true, for example, for the homeomorphism type of simply-connected smooth projective surfaces. One final remark: As is well-known, as a consequence of the Weil conjectures [@De], the Betti numbers of a smooth proper variety are determined just by the zeta-function of its reduction mod $p$. So in this very rigid case of simply-connected surfaces, it is natural to ask if the homeomorphism type is determined by the zeta function. However, the simplest possible case of $\P^1 \times \P^1$ and $\P^2$ blown up at a point provides a counter-example. In fact, this shows that even the rational homotopy type cannot be determined by the zeta function. [Acknowlegments:]{} We are grateful to Igor Dolgachev for interesting conversations that led to the observation of this paper. We thank Jong-il Park for suggesting the last counter-example. M.K. was supported in part by the National Science Foundation. [10]{} Artin, M.; Mazur, B. Etale homotopy. Lecture Notes in Mathematics, No. 100 Springer-Verlag, Berlin-New York 1969 iii+169 pp. Bloch, Spencer; Kato, Kazuya $p$-adic étale cohomology. Inst. Hautes Etudes Sci. Publ. Math. No. 63 (1986), 107-152. Deligne, Pierre La conjecture de Weil. I. Inst. Hautes Etudes Sci. Publ. Math. No. 43, (1974), 273–307. Deligne, Pierre; Griffiths, Phillip; Morgan, John; Sullivan, Dennis Real homotopy theory of Kähler manifolds. Invent. Math. 29 (1975), no. 3, 245–274. Donaldson, S. K. Connections, cohomology and the intersection forms of $4$-manifolds. J. Differential Geom. 24 (1986), no. 3, 275–341. Donaldson, S. K.; Kronheimer, P. B. [The Geometry of Four-Manifolds ]{}. Oxford University Press, Oxford (1990). ix+440 pp. Freedman, Michael Hartley The topology of four-dimensional manifolds. J. Differential Geom. 17 (1982), no. 3, 357–453. Grothendieck, Alexandre; Artin, Michael; Verdier, Jean-Louis [Séminaire de Géometrie Algébrique]{} 4, Lecture Notes in Mathematics 269, 270, 305, Springer-Verlag, Heidelberg (1972-73). Hirzebruch, F. Neue topologische Methoden in der algebraischen Geometrie. (German) Ergebnisse der Mathematik und ihrer Grenzgebiete (N.F.), Heft 9. Springer-Verlag, Berlin-Göttingen-Heidelberg, 1956. viii+165 pp. Kim, Minhyong; Hain, Richard A De Rham-Witt approach to crystalline rational homotopy theory. Preprint (2002) available as math.AG/0105008. Serre, Jean-Pierre [A course in arithmetic]{}. Springer-Verlag, New York (1973), viii+115 pp. M.K.: DEPARTMENT OF MATHEMATICS, UNIVERSITY OF ARIZONA, TUCSON, AZ 85721 and KOREA INSTITUTE FOR ADVANCED STUDY, 207-43 CHEONGRYANGRI-DONG DONGDAEMUN-GU, SEOUL, KOREA, 130-012. EMAIL:[email protected] D.J.: DEPARTMENT OF MATHEMATICS, POHANG UNIVERSITY OF SCIENCE AND TECHNOLOGY, POHANG, KYOUNGBUK, KOREA 790-784. EMAIL: [email protected]	{ "pile_set_name": "ArXiv" }
[Exact Relation between Einstein and Quadratic Quantum Gravity]{}[^1]\ [E. T. Tomboulis]{}[^2]\ [Department of Physics\ University of California, Los Angeles\ Los Angeles, California 90095-1547]{}\ Abstract {#abstract .unnumbered} ======== We show the exact equality of the path integral of the general renormalizable fourth order gravitational action to the path integral of the Einstein action coupled to a massive spin-0 field and a massive spin-2 ghost-like field with non-polynomial interactions. The metric in the Einstein version is a highly nonlinear function of the metric in the quadratic version. Both massive excitations are unstable. The respective cosmological constant terms in the two versions can be very different. Some implications are briefly discussed. Gravitational actions including higher than linear powers of the Riemann tensor are, for a variety of reasons, interesting both as classical and quantum field theories. They, also, arise generically in ( four-dimensional) string one-loop effective actions (see eg ref [@Ov]). The relation between the Einstein-Hilbert action and such nonlinear extensions has been addressed by many authors. One of the earliest results is apparently that of ref. [@H]. It establishes that the addition of a term quadratic in the curvature scalar is classically equivalent to the minimal coupling of a scalar field to the E-H action plus a scalar field potential. At the classical level, the most complete considerations were given in [@Mag], [@KJ], where it was shown that the fourth order action involving $R^2$ and $R^2_{\mu\nu}$ terms is equivalent to the E-H action with a different metric and coupled to a symmetric rank-2 tensor ’matter’ field. Equivalence here means that the two actions lead to equivalent equations of motion. These results were reproduced, in a somewhat different formalism, in [@HOv], where it was further demonstrated that the content of the tensor field is presisely that of a massive pure spin-2 ghost-like field, and a massive spin-0 field. In this paper we consider the quantum theory. Our approach is motivated by that of [@Mag]-[@KJ], but, in the quantum context, we will obtain a rather stronger result. We will derive the exact, albeit formal, equality of the functional integral of the most general quadratic action (eq. (\[l1\]) below) to that of the Einstein theory with a new metric, and coupled to a massive spin-2 ghost-like field and a massive spin-0 field with nonpolynomial interactions. The metric in the Einstein version turns out to be a highly nonlinear function of the metric in the quadratic theory. It is somewhat remarkable that such an exact transformation of the functional integral, where, of course, one integrates over all metrics without regard of the equations of motion, can be given in closed form. This exact equivalence has various physical implications that we briefly discuss below. The quadratic action is renormalizable, whereas the Einstein action with matter is not (by power counting). Both massive fields turn out to be unstable, the massive spin-2 ghost being actually unstable at tree level. The cosmological constant in the Einstein version can be very different from that in the quadratic action. The same method can be used to examine the relation between other gravitational theories, for example, theories involving arbitrary powers of the Ricci tensor. Here, however, we restrict ourselves to the quadratic action, the protoype for this type of transformation. Our starting point is the path integral for the general fourth order gravitational theory $$Z = \int [Dg_{\mu\nu}] \exp\left( i\int d^4 x {\cal L}(g)\right) \quad,\label{z1}$$ where[^3] $${\cal L}(g) = \sqrt{-g}\;[\: \frac{\gamma}{\kappa^2}R -aR_{\mu\nu}R^{\mu\nu} +bR^2 +\frac{\Lambda}{\kappa^4}\: ] \quad .\label{l1}$$ The inclusion of appropriate gauge fixing and associated ghost terms does not affect the derivation below and need not be indicated explicitly. We now introduce an auxiliary, non-propagating field $\chi_{\mu\nu}$ (of mass dimension 2), and rewrite (\[z1\])- (\[l1\]) in the form: $$Z = \int [Dg_{\mu\nu}][D\chi_{\mu\nu}] C^{1/2} \exp\left( i\int d^4 x {\cal L} (g,\chi)\right)\quad, \label{z2}$$ where $$\begin{aligned} {\cal L}(g,\chi)= \sqrt{-g}\;[\: \frac{\gamma}{\kappa^2}R & - & a(R_{\mu\nu}R^{\mu\nu}- \chi_{\mu\nu}g^{\mu\kappa}g^{\nu\lambda}\chi_{\kappa\lambda})\nonumber\\ & + & b(R^2- (g^{\mu\nu}\chi_{\mu\nu})^2) -\frac{\lambda}{\kappa^2}g^{\mu\nu}\chi_{\mu\nu}\: ] \quad .\label{l2}\end{aligned}$$ In (\[l2\]) $$\lambda^2= \Lambda (4b-a)\quad ,\label{nlam}$$ and $C\equiv Det\left(\sqrt{-g}[a(g^{\mu\kappa}g^{\nu\lambda} +g^{\mu\lambda}g^{\nu \kappa})/2 -bg^{\mu\nu}g^{\kappa\lambda}]\right)$ - such purely local determinants, which are actually equal to unity in dimensional regularization, can be absorbed in the definition of the uncoupled measure $[Dg_{\mu\nu}]$. Integration over $\chi_{\mu\nu}$ gives, of course, back (\[z1\])-(\[l1\]). We now introduce the quantity $${\cal R}_{\mu\nu}(g,\chi) \equiv R_{\mu\nu} + \chi_{\mu\nu}\;\;\;, \qquad {\cal R}=g^{\mu\nu}{\cal R}_{\mu\nu}\quad , \label{sR}$$ in terms of which one has $$\begin{aligned} {\cal L}(g, \chi) & = & \sqrt{-g}\left[ \frac{\gamma}{\kappa^2}{\cal R} - a{\cal R}_{\mu\nu}{\cal R}^{\mu\nu} + b{\cal R}^2 \right]\nonumber\\ & & \quad - \: \sqrt{-g}\left[ \frac{1}{\kappa^2} ( \gamma + \lambda)g^{\mu\nu} - 2a{\cal R}^{\mu\nu} + 2b{\cal R}g^{\mu\nu}\right] \chi_{\mu\nu} \nonumber \\ & \equiv & \hat{{\cal L}}(g, {\cal R}) - \frac{1}{\kappa^2} {\cal F}^{\mu\nu}(g,\chi) \chi_{\mu\nu} \quad. \label{l3} \end{aligned}$$ In (\[l3\]), we have set $$\frac{1}{\kappa^2}{\cal F}^{\mu\nu}(g,\chi) = \sqrt{-g}\left[ \frac{1}{\kappa^2} ( \gamma + \lambda)g^{\mu\nu} - 2a{\cal R}^{\mu\nu} + 2b{\cal R}g^{\mu\nu}\right] \; . \label{sF}$$ The r.h.s. of (\[sF\]) defines a composite tensor density field ${\cal F}^{\mu\nu}$, which we want to use as a new metric field. To this end we insert unity in the integrand in (\[z2\]) in the form of a $\delta$-function integration over a tensor density field $\hat{h}^{\mu\nu}$: $$\begin{aligned} Z = \int [Dg_{\mu\nu}][D\chi_{\mu\nu}] [D\hat{h}^{\mu\nu}] & C^{1/2} & \prod_{x,\mu,\nu}\delta \left[ \sqrt{-h}h^{\mu\nu} - {\cal F}^{\mu\nu} (g,\chi)\right] \nonumber\\ & & \exp \left(i\int d^4x {\cal L}(g,\chi)\right)\; . \label{z3}\end{aligned}$$ The tensor field $h^{\mu\nu}$ is, uniquely, defined through $ \hat{h}^{\mu\nu} = \sqrt{-h}h^{\mu\nu} \;,\; h\equiv det\,h_{\mu\nu} $, with $h_{\mu\nu}$ the inverse of $h^{\mu\nu}$. (We could, of course, work throughout in terms of $\hat{h}^{\mu\nu}$, but it is more convenient to express the equations below in terms of $h^{\mu\nu}$.) Now, in the integrand in (\[z3\]), the $\delta$-function allows one to write: $$\begin{aligned} {\cal L} & = & \hat{{\cal L}}(g,{\cal R}) - \frac{1}{\kappa^2}\sqrt{-h}h^{\mu\nu}\chi_{\mu\nu} \nonumber \\ & = & \frac{1}{\kappa^2}\sqrt{-h}h^{\mu\nu}R_{\mu\nu}(h) + \frac{1}{\kappa^2}\sqrt{-h}h^{\mu\nu}[R_{\mu\nu}(g) - R_{\mu\nu}(h)]\nonumber \\ & & \qquad + \:\hat{{\cal L}}(g, {\cal R}) - \frac{1}{\kappa^2} \sqrt{-h}h^{\mu\nu}{\cal R}_{\mu\nu}(g,\chi) \quad.\label{l4}\end{aligned}$$ Furthermore, one can invert (\[sF\]) to express ${\cal R}(g,\chi)$ in terms of $g_{\mu\nu}$ and ${\cal F}^{\mu\nu}(g,\chi) = \sqrt{-h}h^{\mu\nu}$. The result is $$\begin{aligned} {\cal R}_{\mu\nu} = - \frac{1}{2}\frac{1}{\kappa^2} (4b -a)^{-1}(\gamma + \lambda)g_{\mu\nu} & - & \frac{1}{2a} \frac{1}{\kappa^2}\frac{1}{\sqrt{-g}}\left[ \sqrt{-h}h^{\alpha \beta}g_{\alpha\mu}g_{\beta\nu}\right. \nonumber \\ & & \left. - b(4b-a)^{-1}\sqrt{-h}h^{\alpha\beta} g_{\alpha\beta}g_{\mu\nu} \right] \quad, \label{isR} \end{aligned}$$ and may be substituted in (\[l4\]) to express ${\cal L}$ entirely in terms of $g_{\mu\nu}$ and $h_{\mu\nu}$. One finds $$\begin{aligned} -{\cal V}(g,h) & \equiv & \hat{{\cal L}}(g, {\cal R}) - \frac{1}{\kappa^2} \sqrt{-h}h^{\mu\nu}{\cal R}_{\mu\nu}(g,\chi) \nonumber \\ & = & -\frac{1}{\kappa^4}(\gamma^2 - \lambda^2) (4b-a)^{-1}\sqrt{-g} + \frac{1}{\kappa^4}\frac{\gamma}{2} (4b-a)^{-1}\sqrt{-h}h^{\alpha\beta} g_{\alpha\beta} \nonumber \\ & & \; +\frac{1}{\kappa^4}\frac{1}{4a}\sqrt{-h}\sqrt{\left( \frac{-h}{-g}\right)}\left[ h^{\mu\nu}h^{\alpha\beta}g_{\mu\alpha} g_{\nu\beta}- b(4b-a)^{-1}\left(h^{\alpha\beta}g_{\alpha\beta} \right)^2\right]. \label{pot}\end{aligned}$$ Also, using a standard formula of Riemannian geometry for the difference between the Ricci tensors of two different metrics $g_{\mu\nu}$ and $h_{\mu\nu}$, one obtains $$\begin{aligned} {\cal L}_{kin} & \equiv & \frac{1}{\kappa^2}\sqrt{-h}h^{\mu\nu} [ R_{\mu\nu}(g) - R_{\mu\nu}(h) ] \nonumber \\ & = & \frac{1}{4}\frac{1}{\kappa^2}\sqrt{-h}h^{\mu\nu} g^{\rho\sigma}g^{\kappa\lambda} \left[ 2\nabla_\kappa g_{\sigma\nu} \nabla_\rho g_{\mu\lambda} - 2\nabla_\kappa g_{\sigma\nu}\nabla_\lambda g_{\mu\rho} \right.\nonumber \\ & & \quad + \left.\nabla_\kappa g_{\rho\sigma} \nabla_\lambda g_{\mu\nu} - 2\nabla_\kappa g_{\rho\sigma}\nabla_\nu g_{\mu\lambda} + \nabla_\mu g_{\rho\lambda}\nabla_\nu g_{\sigma\kappa}\right] \nonumber\\ & & + (\mbox{total divergence}) \quad , \label{kin}\end{aligned}$$ with covariant derivatives $\nabla$ computed with the metric $h_{\mu\nu}$. From (\[l4\]), (\[pot\]) and (\[kin\]) one sees that the only remainng dependence on $\chi_{\mu\nu}$ in (\[z3\]) is in the argument of the $\delta$-function, where, from (\[sR\]), (\[sF\]), $\chi_{\mu\nu}$ enters linearly. Assuming, as usual, that one may interchange the order of integrations in the functional integral (\[z3\]), one may now integrate over $\chi_{\mu\nu}$ to obtain $$Z = \int [Dg_{\mu\nu}][D\hat{h}^{\mu\nu}] C^{-1/2} \exp\left( i\int d^4 x {\cal L} (g,h)\right)\quad, \label{z4}$$ where $${\cal L}(g,h) = \frac{1}{\kappa^2}\sqrt{-h}h^{\mu\nu}R_{\mu\nu}(h) + {\cal L}_{kin}(g,h) - {\cal V}(g,h) \quad , \label{l5}$$ with ${\cal L}_{kin},\; {\cal V}$ given by (\[kin\]), (\[pot\]). (\[l5\]) is the Hilbert-Einstein Lagrangian for the metric $h_{\mu\nu}$ coupled to the ’matter’ field $g_{\mu\nu}$, and fully reproduces the results of [@Mag],[@KJ], but now obtained as an exact transformation of the functional integral. To extract the spin content of the rank-2 tensor field $g_{\mu\nu}$, we decompose it into its trace and traceless components with respect to $h_{\mu\nu}$: $$\phi \equiv \frac{1}{4}\frac{1}{\kappa}h^{\mu\nu}g_{\mu\nu} \quad , \qquad \frac{1}{\kappa}g_{\mu\nu} = \phi_{\mu\nu} + h_{\mu\nu}\phi \;\; . \label{dec}$$ Then we write $$g_{\mu\nu} = \kappa\phi I_\mu^{\ \beta} h_{\beta\nu} \:, \quad g^{\mu\nu}= \kappa^{-1}\phi^{-1}h^{\mu\alpha}I^{-1}\ _\alpha^{\ \nu} \; , \quad g = \kappa^4 \phi^4Ih \;\;, \label{dec1}$$ with $$I_{\mu}^{\ \nu}\equiv \left[ \delta_\mu^\nu + \phi^{-1}\phi_{\mu\alpha}h^{\alpha\nu}\right] \;, \quad I^{-1}\ _\mu^{\ \alpha} I_\alpha^{\ \nu} =\delta_\mu^\nu \;, \quad I\equiv detI_\mu^{\ \nu}\; . \label{defI}$$ Substituting the decomposition (\[dec\]),(\[dec1\]) in (\[pot\]) one obtains (working now in terms of the ’matter’ fields $\phi_{\mu\nu}$, $\phi$ all indices are raised and lowered by the metric $h_{\mu\nu}$): $$\begin{aligned} {\cal V} & = & -\frac{1}{\kappa^4}\frac{v_0}{\gamma}\Lambda\sqrt{-h} + \frac{\gamma}{\kappa^2}v_0^{-1}(4b-a)^{-1}\sqrt{-h}\left(\phi - \frac{1}{\kappa}v_0\right)^2 \nonumber \\ & & +\frac{1}{\kappa^4}(4b-a)^{-1}\sqrt{-h}(I^{-1/2} -1) + \frac{\gamma}{\kappa^2}v_0^{-1}(4b-a)^{-1}\sqrt{-h}(I^{1/2} - 1)\phi^2 \nonumber \\ & &- \frac{1}{\kappa^4}\frac{1}{4a}\sqrt{-h}I^{-1/2}\phi^{-2} \phi_{\mu\nu}\phi^{\mu\nu} \; , \label{pot1}\end{aligned}$$ with $$v_0 \equiv \left[\gamma(1 - \lambda^2/\gamma^2)\right]^{-1}\; .$$ Note that, since $$I^{1/2} = 1 - \frac{1}{4}\phi^{-2}\phi_{\mu\nu}\phi^{\mu\nu} + \: \cdots\; , \label{expI}$$ the last three terms in (\[pot1\]) contain linear and higher powers of $\phi_{\mu\nu}\phi^{\mu\nu}$. Accordingly, we shift $$\phi = \frac{1}{\kappa}v_0 (1 + \kappa\varphi) \; ,\qquad \phi_{\mu\nu} = v_0 \varphi_{\mu\nu}\; , \label{exp}$$ so as to absorb the linear term in $\phi$ in ${\cal V}$ in (\[pot1\]). ${\cal V}$ provides then the nonvanishng background value for $\phi$ which is necessary for consistency, since $g_{\mu\nu}$ must possess an inverse. We also conveniently scaled the fluctuating fields $\varphi$, $\varphi_{\mu\nu}$ so as to give standard normalization to their kinetic terms in ${\cal L}_{kin}$. Thus $$g_{\mu\nu} = v_0h_{\mu\nu} + \kappa v_0(\varphi h_{\mu\nu} + \varphi_{\mu\nu})\; , \qquad \mbox{and} \qquad I_\mu^{\ \nu}= \left[\delta_\mu^{\ \nu} + \kappa\frac{\varphi_{\mu\alpha}h^{\alpha\nu}}{(1 +\kappa\varphi)}\right] \; .\label{exp1}$$ Inserting (\[exp1\]) in (\[z4\]), (\[l5\]), (\[pot1\]), (\[kin\]), we finally obtain: $$\begin{aligned} Z & = & \int [D\varphi_{\mu\nu}][D\varphi][D\hat{h}^{\kappa\lambda}] C^{-1/2} \exp\left( i\int d^4 x {\cal L}(\{\varphi\},h)\right)\; \; ,\label{z5} \\ {\cal L}(\{\varphi\},h) & = & \frac{1}{\kappa^2}\sqrt{-h}R(h) + \frac{1}{\kappa^4}\frac{v_0}{\gamma}\Lambda\sqrt{-h} \nonumber \\ & & + \frac{3}{2}\sqrt{-h}h^{\mu\nu}\nabla_\mu\varphi\nabla_\nu\varphi - \sqrt{-h}\nabla_\mu\varphi\nabla_\nu\varphi^{\mu\nu} - \frac{1}{2}m_0^2\sqrt{-h}\varphi^2 \nonumber \\ & & + \sqrt{-h}\left[ -\frac{1}{4}h^{\mu\nu}\nabla_\mu \varphi_{\alpha\beta}\nabla_\nu\varphi^{\alpha\beta} + \frac{1}{2}\nabla_\nu\varphi^{\mu\alpha}\nabla_\mu\varphi_\alpha^{\ \nu} \right] + \frac{1}{4}m^2\sqrt{-h}\varphi_{\mu\nu}\varphi^{\mu\nu} \nonumber\\ & & + {\cal L}_{kin}^I - {\cal V}^I \; , \label{l6} \end{aligned}$$ with $$m_0^2 =2\gamma v_0(4b-a)^{-1} \frac{1}{\kappa^2}\; , \qquad m^2 = \left[ \frac{1}{a} + \frac{v_0}{\gamma}\Lambda \right]\frac{1}{\kappa^2} \; .\label{mass}$$ In (\[l6\]) we wrote out explicitly only the parts of ${\cal L}_{kin}$ and ${\cal V}$ that are bilinear in the fields $\varphi$, $\varphi_{\mu\nu}$. ${\cal L}_{kin}^I$, and ${\cal V}^I$ then denote the non-polynomial interaction terms containing the trilinear and higher couplings in $\varphi$, $\varphi_{\mu\nu}$ from the expansion of ${\cal L}_{kin}$, and ${\cal V}$, resp., in powers of $\varphi$, $\varphi_{\mu\nu}$ upon insertion of (\[exp1\]) in (\[kin\]), (\[pot1\]). (The unexpanded expressions (\[kin\]), (\[pot1\]), with the replacements (\[exp\]), (\[exp1\]), give this nonpolynomial Lagrangian in closed form.) We have then obtained the exact transfomation of the path integral of the general fourth order theory (\[z1\])-(\[l1\]) to the form (\[z5\])-(\[l6\]). Note that any convenient gauge-fixing term, e.g. $ (\Box\partial_\mu g^{\mu\nu})^2$, plus associated FP-ghost terms in (\[z1\]) is tacitly carried along in the above derivation, and is, at the very end, reexpressed through (\[exp1\]) in terms of $\varphi$, $\varphi_{\mu\nu}$ and $h_{\mu\nu}$. A convienient choice of gauge in (\[z1\])-(\[l1\]) will, of course, not translate, in general, into a convenient gauge for computations in terms of the metric $h_{\mu\nu}$ in the theory in the form (\[z5\])-(\[l6\]). According to the standard FP argument, however, the path integral $Z$ is actually independent of the gauge-fixing term choice, so, once the equivalence of (\[z1\]) to (\[z5\]) is obtained in one gauge, one may change this gauge to any other in (\[z5\])-\[l6\]), i.e. the equivalence holds independently of the gauge choice. The same is, of course, true for any gauge invariant correlation functions. Now (\[l6\]) is the Hilbert-Einstein action for the metric $h_{\mu\nu}$ with a cosmological term and coupled to a massive spin-2 field $\varphi_{\mu\nu}$, and a massive spin-0 field $\varphi$. Note that the $\varphi_{\mu\nu}$ kinetic plus mass terms in (\[l6\]) are not in (the curved-space generalization of) the canonical Pauli-Fierz form[^4]. But they are an equally good formulation of the standard Fierz description of a pure spin-2 field, i.e. traceless $\varphi_{\mu\nu}$, and $\nabla_\mu \varphi^\mu_{\ \nu} $ obeying a constraint equation, which follows from the equations of motion, and involves only [first]{} derivatives of the fields $\varphi_{\mu\nu} , \varphi, h_{\mu\nu} $[^5] [@F]. The non-polynomial interaction terms ensure full gauge invariance and continued absence of a spin-1 component beyond the (curved-space) linearized approximation. The lagrangian (\[l6\]) thus provides a realization of a complete, consistent coupling of a massive spin-2 field to a gravitational background, a noteworthy fact ([@O], [@HOv]). The spin-2 and spin-0 field kinetic terms in (\[l6\]) come with opposite signs. The overall sign is set by the (Newtonian limit of the) EH term, which makes the spin-2 have the wrong sign, ie be a ghost-like field. We thus find, in the full non-linear theory, the same field content as in the linearized approximation to (\[l1\]) [@S]. Note, however, that in the full theory this can only be achieved by introducing a new metric field which is a highly nonlinear function of the metric in (\[l1\]); the exact relation between (\[l1\]) and (\[l6\]) is non-perturbative. It is natural to view (\[l1\]) as the theory formulated in terms of field variables suitable for the UV regime (energies higher than Planck scale energy), and (\[l6\]) as the theory in terms of variables appropriate to the IR region (energies at or below Planck scale). Indeed, recall that (\[l1\]) is a renormalizable [@S], and in fact asympotically free (in the coupling $ \alpha^2 \equiv 1/a$) lagrangian [@TT], [@FT], [@A]. Its loop perturbative expansion is, therefore, applicable in the deep UV region. (\[l6\]), on the other hand, has, by power counting, the usual non-renormalizable behavior of the Einstein theory coupled to matter. It is not clear, in the absence of explicit computations, how the renormalizability of (\[l1\]) appears in (\[l6\]). The ’matter’ fields circulating in loops must serve as regulators. Still, since the metric in (\[l6\]) is a highly nonlinear composite field in terms of that in (\[l1\]), simple direct order-by-order cancellation of the nonrenormalizable divergences in the loop expansion of (\[l6\]) presumably does not occur. Rather, one expects that the divergences are cancelled on mass-shell upon nonlinear shifts of field variables, and/or resummation of appropriate subclasses of graphs. In any case, we seem to have an interesting example of the equivalence of a renormalizable to a (by power counting) non-renormalizable lagrangian. The Einstein version (\[l6\]) separates out the massless graviton, which dominates at large distance scales, from the massive fields, and hence is suitable for consideration of the IR regime. It is here that the difficult dynamical issues of the S-matrix asymptotic states and unitarity become relevant. At tree level, the massive fields have masses naturally of the order of the Planck mass. Now, the asymptotically free coupling $\alpha$ grows large in the IR, and indeed tends to diverge below the Planck scale. Since the mass of the spin-2 particle grows with it, this leads to the possibility that this particle disappears from the spectrum at large distances; ie there is “confinement” of the massive spin-2 ghost-like excitation, as has often been suggested in the literature. In any event, at the very least, the following situation should apply. With $\alpha$ large, the Compton wavelength of the massive spin-2 becomes comparable or less than its Schwarzschild radius. This suggests that, by well-known results [@Y], already at the [classical]{} level, collapse of the surrounding spacetime and formation of a trapped surface must occur. In such a case, the usual expansion about the tree level description of a bare particle propagating on some given background is clearly not meaningful. Rather, the 0-th order approximation must already include enough interaction effects to correctly describe the appearance of such a highly dressed object, essentially a mini black hole. An even more basic question is that of the stability of these massive particles, ie whether they can appear in the true asymtotic states at all. Inspection of the interaction terms in the Lagrangian (\[l6\]) shows that [both the spin-2 and the spin-0 particles are unstable]{}: there are trilinear vertices in ${\cal L}_{kin}$ that, for $b \geq 9a/4$, allow the tree-level decay of the $\varphi_{\mu\nu}$-particle into two $\varphi$-particles (plus gravitons); and both particles can decay into gravitons through radiative loop corrections. Now the S-matrix can, strictly speaking, be defined only between in- and out-states containing solely stable particles. In the standard field theoretic treatment of unstable particles [@V], this S-matrix connecting stable particles only is constructed in terms of complete, dressed propagators for the unstable particles; and can then be shown [@V] to be unitary and causal. This formalism must be applied here too. In fact, the spin-2 particle appears as the simplest example of an unstable particle, ie decaying at tree level (so it [has]{} to be treated in terms of dressed propagators), except for the fact that its bare propagator has negative residue. In this connection it might be also useful to recall that, in the quantum theory, a negative residue can be traded for negative energy flowing through the propagator. But in a gravitational field there is no invariant meaning of local energy density. So there is nothing immediately inconsistent in allowing localized negative energies, corresponding to the occurance of an unstable excitation, if they do not affect the asymptotic values of the fields. Another noteworthy feature of the equivalence of (\[l1\]) to (\[l6\]) is the relation between their respective cosmological terms. Both vanish if $\Lambda$ is fine-tuned to zero. For $\Lambda \neq 0$, however, the cosmological constant $v_0\Lambda/\gamma$ in (\[l6\]) can be very different from $\Lambda$. If, in particular, the couplings run appropriately in the IR, it can tend to zero at large scales, even for large values of $\Lambda$. It is easily seen that there is more than one senario for the renormalization group flows of the couplings $ \gamma,\: a,\: b,\: \Lambda $ that could lead to this behavior. The asymptotic freedom of $1/a$ is firmly established, but the present state of the computations of the renormalization of the other couplings, [@FT], [@A], does not yet allow one to draw any definitive conclusions. More investigation is needed to ascertain if this interesting possibility is actually realized. The method presented here can be used to examine the relation between other gravitational theories. One obvious application is to the supersymmetric version of (\[l1\]) (cp ref [@Ov]). A more intriguing case is that of more general theories involving arbitrary polynomials in the Ricci tensor. Results will be presented elsewhere. [99]{} A. Hindawi, B. Ovrut, and D. Waldram, Un. of Penn preprint UPR-680T, hep-th/9509054 (1995). P. W. Higgs, Nuovo Cimento [11]{} (1959) 816. G. Magnano, M. Ferraris, and M. Francaviglia, Gen. Rel. Grav. [19]{} (1987) 465; M. Ferraris, M. Francaviglia, and G. Magnano, Class. Quantum Grav. [5]{} (1988) L95. A. Jakubiec and J. Kijowski, Phys. Rev. D [37]{} (1988) 1406. A. Hindawi, B. Ovrut, and D. Waldram, Un. of Penn preprint UPR-660T, hep-th/9509142 (1995). M. Fierz, Helv. Phys. Acta, [12]{} (1939) 3; M. Fierz and W. Pauli, Proc. Roy. Soc. A [173]{} (1939) 211. V. I. Ogievetsky and I. V. Polubarinov, Ann. Phys. [35]{} (1965) 167; D. Boulware and S. Deser, Phys. Rev. D [6]{} (1972) 3368. K. Stelle, Phys. Rev. D [16]{} (1977) 953; Gen. Rel. Grav. [9]{} (1978) 353. E. T. Tomboulis, Phys. Lett. [97B]{} (1980) 77. E. S. Fradkin and A. A. Tseytlin, Nucl. Phys. [B201]{} (1982) 469. I. G. Avramidi and A. O. Barvinsky, Phys. Lett. [159B]{} (1985) 269; I. G. Avramidi, Sov. J. Nucl. Phys. [44]{} (1986) 160. R. Shoen and S.-T. Yau, CMP [90]{} 1983) 575. M. Veltman, Physica [29]{} (1963) 186. [^1]: Research supported in part by NSF Grant PHY89-15286 [^2]: e-mail address: [email protected] [^3]: We use metric signature (+ - - -), $R_{\mu\nu} = \partial_\nu \Gamma^\lambda_{\mu\lambda} - \ldots $, and $\kappa^2 \equiv 16\pi G$, where $G$ is Newton’s constant. [^4]: To go over to that form, express (\[l6\]) in terms of the traceful field $ \psi \equiv \varphi h_{\mu\nu} + \varphi_{\mu\nu}$. The bilinear parts of (\[l6\]) are then precisely the massive spin-2 action of ref. [@S], which, as shown there, can be brought to the Pauli-Fierz mass form by a somewhat different field decomposition of $\psi_{\mu\nu}$. The resulting scalar field is a mixture of our scalar field and $\nabla_\mu\nabla_\nu\varphi^{\mu\nu}$. Note that this affects the definition of the scalar mass term. We do not discuss such alternative formulations since they are not pertinent to our main point here. [^5]: The tracelessness of $\varphi_{\mu\nu}$ must be remembered when performing the variation.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Anatomical landmark segmentation and pathology localisation are important steps in automated analysis of medical images. They are particularly challenging when the anatomy or pathology is small, as in retinal images (e.g. vasculature branches or microaneurysm lesions) and cardiac MRI, or when the image is of low quality due to device acquisition parameters as in magnetic resonance (MR) scanners. We propose an image super-resolution method using progressive generative adversarial networks (P-GANs) that can take as input a low-resolution image and generate a high resolution image of desired scaling factor. The super resolved images can be used for more accurate detection of landmarks and pathologies. Our primary contribution is in proposing a multi stage model where the output image quality of one stage is progressively improved in the next stage by using a triplet loss function. The triplet loss enables stepwise image quality improvement by using the output of the previous stage as the baseline. This facilitates generation of super resolved images of high scaling factor while maintaining good image quality. Experimental results for image super-resolution show that our proposed multi stage P-GAN outperforms competing methods and baseline GANs. The super resolved images when used for landmark and pathology detection result in accuracy levels close to those obtained when using the original high resolution images. We also demonstrate our method’s effectiveness on magnetic resonance (MR) images, thus establishing its broader applicability' author: - \| Dwarikanath Mahapatra$^{1}$, Behzad Bozorgtabar$^{2}$\ $^{1}$IBM Research - Australia, $^{2}$EPF Laussane, Switzerland bibliography: - 'ISR\_GAN\_Ref.bib' title: Progressive Generative Adversarial Networks for Medical Image Super resolution --- Introduction {#sec:intro} ============ Retinal fundus image analysis is essential for diagnosis of retinal conditions such as glacuoma and diabetic retinopathy. An important component of automated diagnosis of retinal conditions is the detection of pathologies (haemorrhages, microaneurysms, exudates) and landmarks (vasculature, optic cup and disc, fovea). Fundus image resolution is high enough to detect and segment prominent landmarks (e.g., optic disc, fovea, main vessels) and pathologies (e.g., hard exudates). However there are many pathologies which cover a very small area in the fundus images (e.g., microaneurysms, haemorrhages) or are not clearly visible (e.g. soft exudates and certain neovascularizations). Smaller branches of the vasculature are difficult to segment in normal fundus images and hence it is desirable to have higher resolution local image patches covering the specific pathologies to facilitate detailed disease analysis. We propose a image super-resolution algorithm using progressive generative adversarial networks (P-GANs) that takes as input a low-resolution image patch and outputs a high-resolution image to facilitate more accurate diagnosis. Our method’s effectiveness is demonstrated using retinal vasculature segmentation and microaneurysm detection. We also show results on magnetic resonance (MR) images for cardiac left ventricle segmentation, thus establishing the wider applicability of our method. Related Work ------------ Interpolation of medical images (such as MRI) leads to partial volume effects that affect the final segmentation. There exist many methods using super-resolution (SR) on medical images [@Jog7; @Jog8]. A common approach to image super-resolution (ISR) are example based methods [@Jog8] which leverage the information from high-resolution (HR) images together with a low-resolution (LR) image and generate an SR version approximating the original HR image. Self similarity [@Jog7] and generative models [@Jog5] have also been used for SR. These methods are too reliant on external data which may not always be available, thus putting them at an disadvantage. Single image based SR methods downsample a given image to create a LR image and learn the mapping between the original and LR version. The learnt mapping is then applied to the original image to generate a SR image. In [@Jog9] HR and LR dictionaries are learned from MRI to generate SR images. These methods depend on learning the dictionaries on external LR-HR images and assume that the test image is a representative of the training data. Since this is not always the case the results are unsatisfactory. These approaches are computationally demanding as the candidate patches have to be searched in the training dataset to find the most suitable HR candidate. Instead, compact and generative models can be learned from the training data to define the mapping between LR and HR patches. Parametric generative models, such as coupled-dictionary learning based approaches, have been proposed to upscale MR brain [@Jog9] and cardiac [@Oktay3] images. These methods benefit from sparsity constraint to express the link between LR and HR. Similarly, random forest based non-linear regressors have been proposed to predict HR patches from LR data and have been successfully applied on diffusion tensor images [@TannoMICCAI16]. Recently, convolutional neural network (CNN) models [@Oktay5] have been put forward to replace the inference step as they have enough capacity to perform complex nonlinear regression tasks. Even by using a shallow network composed of a few layers, [@Oktay5] achieve superior results over other state-of-the art SR methods. Recent works have proposed image SR methods based on training data free approach using Fourier burst accumulation [@Jog], CNNs [@Oktay] and generative adversarial networks (GANs) [@Srgan; @Mahapatra_MICCAI17]. Our Contribution ---------------- In an earlier work [@Mahapatra_MICCAI17] we proposed a GAN based image SR method that incorporated saliency maps. The saliency map based approach had some limitations such as: 1) choice of optimal window size and weights for saliency map calculation depended on the specific image and was non-optimal; 2) consequently, the super-resolution output was unsatisfactory for certain cases. To overcome these limitations and to avoid a heuristic approach, in this paper we propose a novel image SR method based on multi-stage GANs that uses a triplet loss function. We call our method progressive GAN (P-GAN) since it leverages the triplet loss function to progressively improve the quality of super resolved images. Our current method is different from [@Mahapatra_MICCAI17] since: 1) it uses progressive GANs and triplet loss function; 2) it does not use any heuristic based parameter values; and 3) it outperforms [@Mahapatra_MICCAI17] for different scaling factors. GANs [@Srgan21] are used to learn a generative model of images that is similar to a given set of training images. They have been used in various applications such as image super resolution [@Srgan], image registration [@MahapatraGAN_ISBI18], active learning [@MahapatraAL_MICCAI18], image synthesis and image translation using conditional GANs (cGANs) [@CondGANs] and cyclic GANs (cycleGANs) [@CyclicGANs]. Multiple blocks of deep residual network (ResNet) [@Srgan27] with skip connections are used to construct the generator network. In our proposed multi-stage approach we use the output of one stage as the input to the next stage. The image generation framework in the next stage uses triplet loss to improve the quality of the image from the previous stage. This ensures that good quality images are generated for high scaling factors. Combining the generator network with a discriminator encourages solutions that preserve the information content and perceptual information of an image. This leads to HR images that do not compromise on perceptual clarity and result in better retinal vasculature segmentation and microaneurysm (MA) detection results, as well as MRI organ segmentation. Our paper makes the following contributions: 1) a novel P-GAN architecture using multiple stages of GANs is proposed that enables generation of SR images of high scale factors (upto $32$); 2) While GAN based methods use conventional mean square error (MSE) and CNN feature loss values to generate SR images, we use an additional triplet loss function to improve image quality from one stage to the next. This ensures that image quality is not compromised despite generating images of high scale factors. A combination of multi stage GANs and triplet loss helps us outperform conventional GAN based methods [@Srgan]. The rest of the paper is organized as follows: Section \[sec:met\] describes basics of GANs and our novel contribution, followed by experimental results in Section \[sec:expt\] and our conclusion in Section \[sec:concl\]. Methodology {#sec:met} =========== In this section we will first give a brief outline of GANs (Section \[met:GANs\]) and then explain our contribution in the form of progressive-GANs (Section \[met:pGANs\]) Generative Adversarial Networks {#met:GANs} ------------------------------- Super resolution estimates a high-resolution image $I^{SR}$ from a low-resolution input image $I^{LR}$. Figure \[fig:PGan1\] shows the architecture of a progressive GAN setup where the output of the first stage is used as input to the second stage, and the triplet loss is used from the second stage onwards to improve super resolution results. Each super-resolution stage consists of a generator and discriminator network which are depicted in Figure \[fig:Gan\]. For training, $I^{LR}$ is the low-resolution version of the high-resolution counterpart $I^{HR}$, obtained by applying a Gaussian filter to $I_{HR}$ followed by downsampling with factor $\textbf{r}$. The generator network is a feed-forward CNN, $G_{\theta_G}$, parametrized by $\theta_G = {W;b}$, the weights and biases of a L-layer network. The parameters are obtained by, $$\widehat{\theta}=\arg \min_{\theta_G} \frac{1}{N} \sum_{n=1}^{N} l^{SR}\left(G_{\theta_G}(I_n^{LR}),I_n^{HR}\right), \label{eq:theta1}$$ where $l^{SR}$ is the loss function and $I_n^{HR}$, $I_n^{LR}$ are the set of high-resolution and low-resolution images. In the generator (Figure \[fig:Gan\] (a)) the input image $I^{LR}$ is passed through a convolution block followed by ReLU activation. The output is passed through a residual block with skip connections. Each block has convolutional layers with $3\times3$ filters and $64$ feature maps, followed by batch normalization and ReLU activation. This output is subsequently passed through multiple residual blocks. Their output is passed through a series of upsampling stages, where each stage doubles the input image size. The output is passed through a convolution stage to get the super resolved image $I^{SR}$. Depending upon the desired scaling, the number of upsampling stages can be changed. The adversarial min-max problem is defined by, $$% \mathop{{}\mathbb{E}} \min_{\theta_G} \max_{\theta_D} \mathop{\mathbb{E}}_{I^{HR}~p_{train}(I^{HR})}[\log D_{\theta_D}(I^{HR})] + \mathop{\mathbb{E}}_{I^{LR}~p_{G}(I^{HR})}[\log (1-D_{\theta_D}(G_{\theta_G}(I^{HR}))] \label{eq:cf1}$$ This trains a generative model $G$ with the goal of fooling a differentiable discriminator $D$ that is trained to distinguish super-resolved (SR) images from real images. $G$ creates solutions that are very similar to real images and thus difficult to classify by $D$. This encourages perceptually superior solutions and is superior to solutions obtained by minimizing pixel-wise MSE. $D$ solves the maximization problem in Eqn. \[eq:cf1\]. The discriminator network (Figure \[fig:Gan\] (b)) has multiple convolutional layers with the kernels increasing by a factor of $2$ from $64$ to $512$. Leaky ReLU is used and strided convolutions reduce the image dimension when the number of features is doubled. The resulting $512$ feature maps are followed by two dense layers and a final sigmoid activation to obtain a probability map, which is used to classify the image as real or fake. ---------------------------------------------------------------------------------------------------------------------------------- ![Depiction of Progressive GAN architecture. []{data-label="fig:PGan1"}](MultiStageGAN.jpg "fig:"){height="5.2cm" width="9.4cm"} ---------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![(a) Generator Network; (b) Discriminator network. $n64s1$ denotes $64$ feature maps (n) and stride (s) $1$ for each convolutional layer. []{data-label="fig:Gan"}](GeneratorNetwork_mod.jpg "fig:"){height="4.2cm" width="9.9cm"} (a) ![(a) Generator Network; (b) Discriminator network. $n64s1$ denotes $64$ feature maps (n) and stride (s) $1$ for each convolutional layer. []{data-label="fig:Gan"}](DiscriminatorNetwork_mod.jpg "fig:"){height="4.2cm" width="9.9cm"} (b) ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Loss Function ------------- $l^{SR}$ is a combination of content loss ($l_{Cont}^{SR}$) and adversarial or generative loss ($l_{Gen}^{SR}$), balanced by a factor $\alpha=0.01$, and is given by : $$l^{SR} = l_{Cont}^{SR} + \alpha l_{Gen}^{SR} \label{eq:cf2}$$ ### Content Loss - $l_{Cont}^{SR}$ The first component of $l_{Cont}^{SR}$ is the mean square error (MSE) loss ($l_{MSE}$), $$l_{MSE}=\frac{1}{WH} \sum _{x=1}^{W} \sum _{y=1}^{H} (I^{HR}_{x,y} -G_{\theta_G} (I^{LR})_{x,y} )^2, \label{eq:wlMSE}$$ where $I^{HR}$ is the high-resolution image and $G_{\theta_G} (I^{LR})$ is the super resolved image. Next, a CNN loss [@Srgan] is calculated as the $L2$ distance between SR image and ground-truth HR image using all $512$ feature maps of Relu $4-1$ layer of a pre-trained $VGG-16$ [@VGG]. It is defined as, $$l^{SR}_{CNN} =\frac{1}{W_{i,j}H_{i,j}} \sum _{x=1}^{W_{i,j}} \sum _{y=1}^{H_{i,j}} (\phi_{i,j}(I^{HR})_{x,y} -\phi_{i,j}(G_{\theta_G} (I^{LR}))_{x,y} )^2 \label{eq:lCNN}$$ $\phi_{i,j}$ the feature map obtained by the $j-$th convolution (after activation) before the $i-$th max pooling layer and $W_{i,j}$ and $H_{i,j}$ are the dimensions of $\phi$. ### Adversarial Loss The generative (or adversarial) loss $l^{SR}_{Gen}$ over all training samples is $$l^{SR}_{Gen}=\sum_{n=1}^N -\log D_{\theta_D}(G_{\theta_G}(I^{LR}))$$ $D_{\theta_D}(G_{\theta_G}(I^{LR}))$ is probability that $G_{\theta_G}(I^{LR})$ is a natural HR image. This network favours solutions in the manifold of retinal images. Convergence is facilitated by minimizing $-\log D_{\theta_D}(G_{\theta_G}(I^{LR}))$ instead of $-\log [1-D_{\theta_D}(G_{\theta_G}(I^{LR}))]$. Progressive Generative Adversarial Networks {#met:pGANs} ------------------------------------------- In the previous sections we have described the working of a conventional GAN with a coupled generator and discriminator. The GAN based approach of [@Srgan] does not produce good quality results for scaling factors above $8$. An obvious way to generate higher scale images is to couple successive generator networks or upsampling stages to obtain images of desired scaling factor. The shortcoming of this approach is that any errors of one stage are propagated to the other step as there is no correction mechanism. We propose a multi-stage architecture with a triplet loss based in-built correction mechanism that helps to compensate any quality degradation of images generated st each stage. In this section we explain our novel P-GAN method. We denote our method as progressive GAN for image super-resolution ($P-SRGAN$). $P-SRGAN$ works by first taking a LR image as input which is processed through the first generator network $G1$ to obtain a super resolved image of scale factor $2$. Similar to a conventional GAN there is a discriminator network $D1$ that helps $G1$ generate good quality images by comparing with high-resolution (HR) images. The loss function combines MSE and CNN loss terms. We denote the SR image from $G1$ as $I^{SR}_{1}$. $I^{SR}_{1}$ is passed through another network $G2$ to generate an image further upsampled by a factor $2$, which we denote as $I^{SR}_{2}$. However, in generating $I^{SR}_{2}$ we make changes to the cost function by including an additional triplet loss component [@TripletLoss]. Note that for each stage we upsample the image by $2$x, which implies that there is only one upsampling block in the generator. The triplet loss has three variables in it - positive, negative and anchor. The cost function is such that it minimizes the distance between the anchor ($x^{a}$) and positive ($x^{p}$), while maximizing the distance between the anchor and the negative ($x^{n}$). Thus the loss being minimized is, $$\sum_i^{N} \left[ \left\\| f(x^{a}_i) - f(x^{p}_i)\right\\|^{2}_2 - \left\\| f(x^{a}_i) - f(x^{n}_i)\right\\|^{2}_2 \right]$$ In our formulation the anchor is $I_{2}^{SR}$, the negative is $I^{SR}_{1}$ while the positive is the ground truth HR image $I^{HR}$. We would like the generated image at the second (and subsequent) stage, $I^{SR}_{2}$, to be as different as possible from the first (or previous) generated image, $I^{SR}_{1}$. At the same time we enforce that $I^{SR}_{2}$ should be as similar as possible to $I^{HR}$. The rationale behind such an approach is that $I^{SR}_{2}$ should improve upon $I^{SR}_{1}$ and both of them should be as different as possible. At the same time $I^{SR}_{2}$ should be as similar to $I^{HR}$ in order to improve upon the quality of $I^{SR}_{1}$. These dual constraints ensure that $I^{SR}_{2}$ is a quantitative and qualitative improvement over $I_{1}^{SR}$ and closer to $I^{HR}$. For subsequent steps of upsampling by factor $2$ we employ the same loss function where the anchor is the generated image at stage $n$ $I_{n}^{SR}$, the negative is $I_{n-1}^{SR}$ and the positive image is always the ground truth $I^{HR}_n$ for stage $n$. In addition to the triplet loss, MSE and CNN loss are used for all stages $n\geq2$. Training Details: ----------------- We trained all networks on a NVIDIA Tesla $K40$ GPU having $12$ GB RAM. The LR images were obtained by downsampling the original HR images with a bicubic kernel of varying downsampling factors. The intensity of all images (fundus and MRI) was scaled to $[0,1]$. VGG feature maps were rescaled by a factor of $\frac{1}{12.75}$ (based on actual values of the two feature maps) to obtain VGG losses of a scale that is comparable to the MSE loss. For optimization we use Adam with $\beta_1 = 0.9$. The SRResNet networks were trained with a learning rate of $10^{-4}$ for $10^{6}$ update iterations. We employed the trained MSE-based SRResNet network as initialization for the generator when training the actual GAN to avoid undesired local optima. GAN based methods such as ours and [@Srgan] are trained with $10^{5}$ update iterations at a learning rate of $10^{-4}$ and another $10^{5}$ iterations at a lower rate of $10^{-5}$. During test time batch-normalization update is off such that the output is deterministic and depends on the input. Our entire pipeline was implemented with Python and TensorFlow was used for the neural network architectures. Experiments And Results {#sec:expt} ======================= In this section we demonstrate the effectiveness of our super resolution algorithm by reporting results on different kinds of images. We report ISR performance on retinal colour fundus images and magnetic resonance (MR) images. We also show results for segmentation of different landmarks and pathologies in fundus and MR images. Retinal Dataset Description --------------------------- To test the effectiveness of our image super resolution algorithm we train it on $5000$ retinal fundus images taken from EYEPACS [@kaggle]. The original images were first resized to $1024\times1024$ pixels which was the reference high resolution (HR) image. Subsampled images of scaling factors (r) $\frac{1}{2},\frac{1}{4},\frac{1}{8},\frac{1}{16}$,$\frac{1}{32}$. The high resolution image was then downsampled using a bicubic kernel by a factor of $2$ along the rows and columns. This gave images of size $512\times512$, which are then further downsampled by factors of $2$ to get images of size $256\times256$, $128\times128$, etc. We train our network on the entire set of $5000$ images and use a separate set of $1000$ images for testing. The average training time for a single batch of $5000$ images is $8$ hrs for $r=2$, $15$ hours for $r=4$, $21$ hours for $r=8$, $28$ hours for $r=16$ and $35$ hours for $r=32$. The time taken to generate a super resolved image is $1$ ms for $r=2$, $1.4$ ms for $r=4$, $1.9$ ms for $r=8$, $2.5$ ms for $r=16$ and , $3.3$ ms for $r=32$. The training and test was performed on a NVIDIA Tesla K$40$ GPU with $12$ GB RAM. Image Super Resolution Results ------------------------------ For comparative analysis, results for super resolution are shown for the following methods: 1. $P-SRGAN$ - our proposed progressive method using a multi stage architecture in combination with triplet loss to generate super resolved images of desired scaling factor. Our method’s difference from [@Srgan] is in the use of a multi stage architecture and triplet loss. 2. $SRGAN_{Sal}$ - the saliency map based SR method of [@Mahapatra_MICCAI17]. 3. $P-SRGAN_{MSE}$ - our proposed progressive method with the exception that instead of triplet loss we use MSE loss in every stage. This is designed to highlight the importance of triplet loss in improving quality of super resolved images. 4. $SRGAN_{Ledig}$ - the original GAN based super resolution algorithm of [@Srgan]. Based on the original work, the desired scaling factor is obtained by additional pixel upsampling blocks in the original architecture 5. $SRCNN$- the CNN based image super resolution algorithm of [@SRCNN]. 6. $SR-RF$- the random forest based image super resolution method of [@SR-RF]. 7. $SSR$- the self super resolution method of [@Jog]. The results of the different methods are compared using the following measures: 1) peak signal to noise ratio (PSNR); 2) structural similarity (SSIM) [@SSIM]; and 3) $S3$ - the sharpness metric of [@S3Metric]. These metrics are calculated between the generated SR image and ground truth HR image. Since the original retinal fundus images had different dimensions, they are first resized to $1024\times1024$ since this was the smallest image dimension in the dataset. These images act as the ground truth HR images. They are then downsampled by different scaling factors to generate LR images. These LR images are used to generate SR images of the corresponding scaling factor, and are compared with the ground truth HR images. The quantitative results for scaling factors $4-32$ are presented in Table \[tab:ISR\_res1\]. It is quite obvious that our proposed approach performs the best among all methods, even outperforming $SRGAN_{Sal}$. An interesting observation is that replacing the triplet loss by conventional MSE loss does not result in much improved results compared to $SRGAN_{Ledig}$. This clearly demonstrates that the use of triplet loss is an important contribution in improving the quality of super resolved images. While upscaling by a factor of $2$ the performance of all methods is similar and hence we do not show the corresponding results. The difference in performance between our approach and other competing methods is noticeable for scaling factor $4$ and to a larger degree for scaling factors above $8$. Figure \[fig:RetSR2\] shows super resolution results for retinal fundus images. The image has been downsampled by a factor of $16$ each along rows and columns, and the original image is obtained using different methods. $SRGAN_{Sal}$ shows better reconstruction results (Fig. \[fig:RetSR2\] (d) )than most competing methods. However the reconstruction is a bit blurry for small retinal vessels using $SRGAN_{Sal}$ as indicated by the yellow arrow. This defect is overcome using the proposed $P-SRGAN$ architecture (Fig. \[fig:RetSR2\] (c) ), thus justifying the use of progressive GANs over saliency maps for image super resolution. ------------------------------ ---------- ---------- ---------- -------------- [SSIM]{} [PSNR]{} [S3]{} [$p$]{} [(dB)]{} [$P-SRGAN$]{} [0.91]{} [46.1]{} [0.85]{} [-]{} [$SRGAN_{Sal}$]{} 0.88 [44.3]{} 0.82 $<0.001$ [$P-SRGAN_{MSE}$]{} [0.77]{} [37.0]{} [0.68]{} [-]{} [$SRGAN_{Ledig}$ [@Srgan]]{} [0.76]{} 36.7 0.66 $<0.001$ [$SRCNN$ [@SRCNN]]{} [0.74]{} 34.3 0.61 $<0.0009$ [SR-RF [@SR-RF]]{} [0.72]{} 30.3 0.59 $<0.0009$ [SSR [@Jog]]{} [0.68]{} 27.2 0.57 $<0.001$ [SSIM]{} [PSNR]{} [S3]{} [$p$]{} [(dB)]{} [$P-SRGAN$]{} [0.81]{} [40.3]{} [0.75]{} [-]{} [$SRGAN_{Sal}$]{} 0.77 36.2 0.70 $<0.001$ [$P-SRGAN_{MSE}$]{} [0.72]{} [31.9]{} [0.62]{} [$<0.001$]{} [$SRGAN_{Ledig}$ [@Srgan]]{} 0.71 31.4 0.60 $<0.001$ [$SRCNN$ [@SRCNN]]{} 0.65 28.4 0.57 $<0.001$ [SR-RF [@SR-RF]]{} 0.63 25.6 0.55 $<0.001$ [SSR [@Jog]]{} 0.60 22.3 0.51 $<0.001$ [SSIM]{} [PSNR]{} [S3]{} [$p$]{} [(dB)]{} [$P-SRGAN$]{} [0.77]{} [36.8]{} [0.68]{} [-]{} [$SRGAN_{Sal}$]{} 0.71 31.8 0.60 $<0.001$ [$P-SRGAN_{MSE}$]{} [0.66]{} [30.1]{} [0.56]{} [$<0.001$]{} [$SRGAN_{Ledig}$ [@Srgan]]{} 0.66 29.4 0.55 $<0.001$ [$SRCNN$ [@SRCNN]]{} 0.60 23.2 0.52 $<0.001$ [SR-RF [@SR-RF]]{} 0.57 21.5 0.50 $<0.001$ [SSR [@Jog]]{} 0.52 19.4 0.46 $<0.001$ [SSIM]{} [PSNR]{} [S3]{} [$p$]{} [(dB)]{} [$P-SRGAN$]{} [0.72]{} [31.9]{} [0.61]{} [-]{} [$SRGAN_{Sal}$]{} 0.67 28.6 0.57 $<0.001$ [$P-SRGAN_{MSE}$]{} [0.62]{} [24.2]{} [0.54]{} [$<0.001$]{} [$SRGAN_{Ledig}$ [@Srgan]]{} 0.61 23.7 0.53 $<0.001$ [$SRCNN$ [@SRCNN]]{} 0.55 19.4 0.49 $<0.001$ [SR-RF [@SR-RF]]{} 0.52 17.4 0.47 $<0.001$ [SSR [@Jog]]{} 0.49 15.8 0.45 $<0.001$ ------------------------------ ---------- ---------- ---------- -------------- : Comparative results of different methods for image super resolution.[]{data-label="tab:ISR_res1"} --------------------------------------------------------- --------------------------------------------------------- ------------------------------------------------------------ ----------------------------------------------------------- ![image](IRet2_HR.jpg){height="3.2cm" width="3.2cm"} ![image](IRet2_LR_4.jpg){height="3.2cm" width="3.2cm"} ![image](IRet2_SR_PGan2.jpg){height="3.2cm" width="3.2cm"} ![image](IRet2_SR_Sal2.jpg){height="3.2cm" width="3.2cm"} (a) (b) (c) (d) ![image](IRet2_SRGAN.jpg){height="3.2cm" width="3.2cm"} ![image](IRet2_SRCNN.jpg){height="3.2cm" width="3.2cm"} ![image](IRet2_SRRF.jpg){height="3.2cm" width="3.2cm"} ![image](IRet2_SSR.jpg){height="3.2cm" width="3.2cm"} (e) (f) (g) (h) --------------------------------------------------------- --------------------------------------------------------- ------------------------------------------------------------ ----------------------------------------------------------- Robustness to Noise ------------------- In this section we show results for cases when noise was added to the LR images and the high resolution image was reconstructed by using $P-SRGAN$, $SRGAN_{Sal}$ and $SRGAN_{Ledig}$. We added the following types of noise: 1) Gaussian noise with zero mean and standard deviation varying from $[0.001,.01]$ in steps of $0.001$; 2) salt and pepper noise of density varying from $[0.01,0.05]$ increasing in steps of $0.01$; 3) speckle noise of variance from $[0.01,0.05]$ increasing in steps of $0.01$ to the LR image (subsampled to factor $8$) and attempted to obtain the original high resolution image through super resolution methods. Table \[tab:ISR\_noise\] summarizes the average performance for each noise type. We observe that as noise intensity increases the performance worsens. Although the overall performance is not as good compared to the noise free images, the results for gaussian noise case are still within acceptable limits for our proposed method. ------------------------------ ------------------- ------------------ ------------------- -------------- [SSIM]{} [PSNR]{} [S3]{} [$p$]{} [(dB)]{} [$P-SRGAN$]{} [0.81$\pm$0.04]{} [39.5$\pm$2.1]{} [0.78$\pm$0.05]{} [$<0.002$]{} [$SRGAN_{Sal}$]{} 0.73$\pm$0.06 36.1$\pm$4.2 0.66$\pm$0.06 $<0.001$ [$SRGAN_{Ledig}$ [@Srgan]]{} [0.72$\pm$0.06]{} 32.4$\pm$4.9 0.61$\pm$0.06 $<0.001$ [SSIM]{} [PSNR]{} [S3]{} [$p$]{} [(dB)]{} [$P-SRGAN$]{} [0.74$\pm$0.07]{} [35.8$\pm$4.5]{} [0.68$\pm$0.07]{} [$<0.001$]{} [$SRGAN_{Sal}$]{} 0.69$\pm$0.1 31.4$\pm$5.7 0.63$\pm$0.1 $<0.001$ [$SRGAN_{Ledig}$ [@Srgan]]{} 0.60$\pm$0.11 23.4$\pm$7.6 0.56$\pm$0.12 $<0.001$ [SSIM]{} [PSNR]{} [S3]{} [$p$]{} [(dB)]{} [$P-SRGAN$]{} [0.72$\pm$0.09]{} [33.5$\pm$8.2]{} [0.63$\pm$0.09]{} [$<0.001$]{} [$SRGAN_{Sal}$]{} 0.67$\pm$0.11 28.8$\pm$9.8 0.58$\pm$0.10 $<0.001$ [$SRGAN_{Ledig}$ [@Srgan]]{} 0.62$\pm$0.13 2.4$\pm$10.2 0.51$\pm$0.11 $<0.001$ ------------------------------ ------------------- ------------------ ------------------- -------------- : Comparative results for image super resolution at subsampling factor $8$ and added noise - mean and variance values are provided.[]{data-label="tab:ISR_noise"} Retinal Blood Vessel Segmentation Results ----------------------------------------- A major application of super resolution of retinal images is better image analysis as in accurate landmark segmentation and pathology detection. We present retinal vessel segmentation results on our dataset [@kaggle]. $40$ images from the dataset had the retinal vessels manually annotated by an expert. This is comparable to other public datasets where a maximum number of $40$ images had the retinal vessels manually delineated. We take the original images (HR images) and downsample them by factors of $4-16$ to get LR images. All the different algorithms were used to generate SR images of the original scale. The super resolution models were trained on the $5000$ training images described before. The SR images are then used to train a U-Net based method for retinal vessel segmentation [@Ret_Unet]. We use the method in [@Ret_Unet] since it shows the best results on the DRIVE dataset. The idea of developing ISR algorithms is to demonstrate that they generate good quality high resolution images from a low resolution. A possible use case is in tele-ophthalmology where images acquisition is done by a low quality camera. Even if a high resolution device is used, the image still needs to be compressed from high to low resolution for transmission over a network. In this case, the quality of the image received by the clinician is often not ideal for visualisation and interpretation. In such cases we would expect the generated SR images would do a better job at tasks like segmentation and pathology detection. To evaluate our hypothesis we adopt the following evaluation procedure for the SR images: 1. As described before we use the SR images generated from downsampled images to train different UNet segmentation networks. Additionally we also generate segmentation results using the original HR images to train another UNet architecture. Segmentation results from this network give an upper bound on the best possible super resolution performance. 2. To ascertain whether the super resolved images lead to improved image analysis over LR images, we also train UNets on the downsampled LR images by using the correspondingly downsampled ground truths. The results on these images (for scales $4,8,16$) act as a baseline to compare super resolution algorithms. If ISR is successful then segmentation performance of SR images should be better than the results from these LR images, and reasonably close to segmentation results obtained from original HR images. Table \[tab:SrRes\_vasc\] shows the average segmentation performance using training images from different ISR algorithms. In Table \[tab:SrRes\_vasc\] $HR$ denotes the fact that the network was trained on the reference HR images (of dimension $1024\times1024$). $LR_{n}$ denotes the training and testing was done on subsampled images at scale $n$ (i.e., image dimension $1024/n\times1024/n$). Results of the other methods denote the segmentations on low resolution images super resolved to dimension $1024\times1024$. Retinal vessel performance of different methods was evaluated using accuracy ($Acc$) and sensitivity ($Sen$). The segmentation results for different scaling factors gives interesting insight into the different super resolution algorithms. As expected, the HR images give the best possible results since they represent the original images while other images have some level of information loss due to subsampling and super resolution. Our proposed method gives better segmentation performance than $LR_n$ which clearly indicates that our ISR algorithm generates better quality images than the low resolution versions. Interestingly methods such as $SSR$ and $SR-RF$ perform only slightly better than $LR_n$. A major factor is the super resolution algorithm. These two methods use traditional approaches based on defining a mathematical model. The other methods that perform better than $LR_n$ however use neural network models for ISR. Amongst the neural network based models the ones using generator networks perform better than $SRCNN$ because the generative models learn to generate better images based on the similarity with the higher resolution images. This clearly indicates that the generator models do a much better job in learning the underlying the image representation, while mathematical models only learn a limited aspect of the image. Comparing between the generative models our proposed method does a better job than [@Srgan] because of the progressive stages used to generate the images. From the second stage onward our use of triplet loss improves image quality by making it as similar as possible to the HR image. Hence we obtain improved image quality. Note that results of $P-SRGAN_{MSE}$ are very similar to that of $SRGAN_{Ledig}$. This clearly indicates the improvements brought about by using the triplet loss function Figure \[fig:res1\] (a) shows an example retinal image followed by its ground truth manual segmentation in Figure \[fig:res1\] (b). Figure \[fig:res1\] (c) shows segmentation result for scaling factor $8$ when using the original HR images to train the U-Net followed by the results when trained on the super resolved images generated by $P-SRGAN$, (Figure \[fig:res1\] (d)), $SRGAN_{Ledig}$ (Figure \[fig:res1\] (e)), $SRCNN$ (Figure \[fig:res1\] (f)), SR-RF (Figure \[fig:res1\] (g)), SSR (Figure \[fig:res1\] (h)) and $LR_8$ (Figure \[fig:res1\] (i)). Obviously the results from $P-SRGAN$ provide results most similar to those of HR images. This is also validated by the quantitative results in Table \[tab:SrRes\_vasc\]. The areas where different methods are unable to obtain accurate segmentation are highlighted by yellow arrows. Due to poor quality of super resolved images most of the methods do not segment the finer vasculature structures, while SSR and SR-RF are unable to segment some of the major arteries. Importantly, our method performs much better than the low resolution image ($LR_8$) which performs poorly due to low resolution. ---------------------------- ---------- ---------- --------- --------- --------- --------- [Acc]{} [Sen]{} [Acc]{} [Sen]{} [Acc]{} [Sen]{} HR 0.98 0.91 0.98 0.90 0.97 0.84 [$P-SRGAN$]{} [0.96]{} [0.89]{} 0.89 0.85 0.84 0.80 [$SRGAN_{Sal}$]{} [0.93]{} [0.86]{} 0.86 0.81 0.82 0.76 [$P-SRGAN_{MSE}$]{} [0.92]{} [0.84]{} 0.85 0.79 0.79 0.74 $SRGAN_{Ledig}$ ([@Srgan]) 0.91 0.83 0.85 0.80 0.79 0.74 $SRCNN$ ([@SRCNN]) 0.88 0.81 0.83 0.78 0.76 0.73 SR-RF ([@SR-RF]) 0.85 0.77 0.80 0.73 0.74 0.70 SSR ([@Jog]) 0.83 0.75 0.77 0.69 0.71 0.68 $LR_n$ 0.81 0.72 0.75 0.67 0.70 0.66 ---------------------------- ---------- ---------- --------- --------- --------- --------- : Comparative vasculature segmentation results of different super resolution methods at different scale factors.[]{data-label="tab:SrRes_vasc"} --------------------------------------------------------------- --------------------------------------------------------------- -------------------------------------------------------------- ------------------------------------------------------------- ----------------------------------------------------------------- ![image](VascSeg_Img.jpg){height="2.7cm" width="2.6cm"} ![image](VascSeg_GT.jpg){height="2.7cm" width="2.6cm"} ![image](VascSeg_Res_HR.jpg){height="2.7cm" width="2.6cm"} ![image](VascSeg_Res_Our.jpg){height="2.7cm" width="2.6cm"} ![image](VascSeg_Res_Sal.jpg){height="2.7cm" width="2.6cm"} (a) (b) (c) (d) (e) ![image](VascSeg_Res_SRGAN.jpg){height="2.7cm" width="2.6cm"} ![image](VascSeg_Res_SRCNN.jpg){height="2.7cm" width="2.6cm"} ![image](VascSeg_Res_SRRF.jpg){height="2.7cm" width="2.6cm"} ![image](VascSeg_Res_SSR.jpg){height="2.7cm" width="2.6cm"} ![image](VascSeg_Res_SSR_LR4.jpg){height="2.7cm" width="2.6cm"} (f) (g) (h) (i) (j) --------------------------------------------------------------- --------------------------------------------------------------- -------------------------------------------------------------- ------------------------------------------------------------- ----------------------------------------------------------------- Results for Retinal Microaneurysm Detection ------------------------------------------- We also present results for microaneurysm (MA) detection. $300$ images from [@kaggle] are annotated by outlining the boundaries of different occurences of MAs. MA’s are very difficult to detect as they cover only a very small area in a high dimensional retinal image. The intuition behind applying image super resolution to MA detection is if the resolution of the patch containing MA’s can be increased then detection accuracy by machine learning algorithms would improve. We design a pipeline where a U-Net architecture is trained to segment these MAs. For each image we extract $256\times256$ patches covering the annotated MAs. These are the original HR images from which LR images of size $128\times128$, $64\times64$ and $32\times32$ are obtained. The different super resolution algorithms are used to generate $256\times256$ patches from the LR patches leading to scaling factors of $2,4,8$. For every generated SR image of different scaling factors, we train U-Nets on images from the different super resolution algorithms. We also train a U-Net segmentation framework with the original $256\times256$ HR images which gives the lower bound of performance error. Similar to vessel segmentation we also generate segmentation results on the LR images, $LR_{n}$, that act as the baseline metric. For our experiments we employ $5-$fold cross validation. Multiple patches extracted from the entire are transformed by rotation and translation such that the final training set size was $100,000$ patches. Note that UNet is a patch based segmentation framework and hence we use patches for its training. After the segmentations are obtained for each set of test images, we calculate the following metrics: sensitivity ($Sen$), Specificity ($Spe$) and area under curve (AUC). These values are summarized in Table \[tab:MA\_res\] with respect to scaling factors $4,8$ since the results are very similar for scaling factor $2$. It is obvious from the results that super resolution is highly effective for scaling factor $4$, and to a lesser extent for factor $8$. We can infer that super resolved images by all the competing methods for scaling factor $8$ changes the information content in the image to the extent that their performance is significantly worse than the HR images. Despite these constraints our algorithm gives the best performance amongst all methods, and most importantly it performs better than $LR_n$. It is important to note that for pathologies, super resolution for scale factor higher than $8$ is not feasible since the LR image dimensions are too small. [\|c\|c\|c\|c\|c\|]{}\ & [HR]{} & [$P-SRGAN$]{} & [$SRGAN$]{} & [$SRCNN$]{}\ & & & [[@Srgan]]{} & [[@SRCNN]]{}\ & [0.83]{} & 0.81 & [0.77]{} & [0.73]{}\ & [0.80]{} & [0.77]{} & [0.72]{} & 0.69\ & [0.94]{} & [0.92]{} & [0.86]{} & 0.83\ & [-]{} & [$<0.001$]{} & $<0.0023$ & $<0.0001$\ & [$SRGAN_{Sal}$]{} & [$SR-RF$]{} & [$SSR$]{} & [$LR_4$]{}\ & [[@Mahapatra_MICCAI17]]{} & [[@SR-RF]]{} & [[@Jog]]{} &\ & 0.79 & 0.71 & [0.68]{} & [0.67]{}\ & 0.74 & 0.65 & 0.63 & [0.63]{}\ & 0.88 & 0.80 & 0.78 &[0.77]{}\ & [$<0.002$]{} & $<0.003$ & $<0.009$ & $<0.009$\ \ & [HR]{} & [$P-SRGAN$]{} & [$SRGAN$]{} & [$SRCNN$]{}\ & & & [[@Srgan]]{} & [[@SRCNN]]{}\ & & [0.80]{} & 0.76 & [0.72]{}\ & & 0.74 & 0.69 & 0.66\ & & 0.86 & 0.81 & 0.77\ & [-]{} & $<0.00065$ & $<0.0002$ & $<0.0001$\ & [$SRGAN_{Sal}$]{} & [$SR-RF$]{} & [$SSR$]{} & [$LR_4$]{}\ & [[@Mahapatra_MICCAI17]]{} & [[@SR-RF]]{} & [[@Jog]]{} &\ & [0.77]{} & [0.68]{} & 0.67 & [0.65]{}\ & [0.70]{} & 0.64 & 0.63 & 0.61\ & [0.83]{} & 0.74 & 0.70 & 0.69\ & [$<0.004$]{} & $<0.0087$ & $<0.0001$ & $<0.0001$\ Figure \[fig:MASeg1\] shows results of MA segmentation using the images generated by different super resolution methods. Figure \[fig:MASeg1\] (a) shows the original full sized image with the regions having majority of the MAs outlined by a red square. Figure \[fig:MASeg1\] (b) shows the cropped image region with yellow arrows identifying location of MAs. Figure \[fig:MASeg1\] (c) shows manually drawn contours around the MAs. Figure \[fig:MASeg1\] (d) shows the segmentation results obtained using the original HR images as part of the training and test set. The ground truth manual contours are shown in green while the segmentations obtained using the U-Net algorithm is shown in red. The performance on the HR images is an indication of the minimum error (or best possible segmentation performance). Figures \[fig:MASeg1\] (e)-(h) show, respectively, the super resolved images obtained by $P-SRGAN$, $SRGAN_{Sal}$, $SRGAN_{Ledig}$ and $SRCNN$ along with the super imposed segmentation results. It is quite obvious that the results obtained using $P-SRGAN$ are the best. It is interesting to note that the SR images obtained using $SRGAN_{Ledig}$ and $SRCNN$ lead to blurred edges of the blood vessels and MAs, although in the case of $SRGAN_{Sal}$ the SR images are not blurred. The other two methods, $SR-RF$ and $SSR$ result in such poor quality images that the MAs are not even detected for this particular example. The MA segmentation results clearly show that if we generate a high resolution image from a low resolution image using our method and then run any segmentation or detection algorithm on the output, the results will be very close to what we would get if we had acquired the image in the high resolution setting. This has immense significance in clinical applications that require detection of small pathologies in a high resolution image. The clinician can simply select the suspect area and our algorithm can generate a high resolution image that can be used for subsequent analysis. ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![Example results for microaneurysm segmentation using super resolved images obtained using different methods. (a) original image with microaneurysm region outline by red square; (b) cropped version of original image with yellow arrows denoting MA locations; (c) manual annotations of MA regions; super resolution and segmentation results obtained when trained on (d) original HR images; (e) SR images by $P-SRGAN$; (f) SR images by $SRGAN_{Sal}$ (g) SR images by $SRGAN_{Ledig}$ ([@Srgan]); (h) SR images using $SRCNN$ ([@SRCNN]). Green contours are ground truth annotations and algorithm segmentations are denoted by red contours.[]{data-label="fig:MASeg1"}](MA_full.jpg "fig:"){height="2.9cm" width="2.7cm"} ![Example results for microaneurysm segmentation using super resolved images obtained using different methods. (a) original image with microaneurysm region outline by red square; (b) cropped version of original image with yellow arrows denoting MA locations; (c) manual annotations of MA regions; super resolution and segmentation results obtained when trained on (d) original HR images; (e) SR images by $P-SRGAN$; (f) SR images by $SRGAN_{Sal}$ (g) SR images by $SRGAN_{Ledig}$ ([@Srgan]); (h) SR images using $SRCNN$ ([@SRCNN]). Green contours are ground truth annotations and algorithm segmentations are denoted by red contours.[]{data-label="fig:MASeg1"}](MA_arrow.jpg "fig:"){height="2.9cm" width="2.7cm"} ![Example results for microaneurysm segmentation using super resolved images obtained using different methods. (a) original image with microaneurysm region outline by red square; (b) cropped version of original image with yellow arrows denoting MA locations; (c) manual annotations of MA regions; super resolution and segmentation results obtained when trained on (d) original HR images; (e) SR images by $P-SRGAN$; (f) SR images by $SRGAN_{Sal}$ (g) SR images by $SRGAN_{Ledig}$ ([@Srgan]); (h) SR images using $SRCNN$ ([@SRCNN]). Green contours are ground truth annotations and algorithm segmentations are denoted by red contours.[]{data-label="fig:MASeg1"}](MA_annot.jpg "fig:"){height="2.9cm" width="2.7cm"} ![Example results for microaneurysm segmentation using super resolved images obtained using different methods. (a) original image with microaneurysm region outline by red square; (b) cropped version of original image with yellow arrows denoting MA locations; (c) manual annotations of MA regions; super resolution and segmentation results obtained when trained on (d) original HR images; (e) SR images by $P-SRGAN$; (f) SR images by $SRGAN_{Sal}$ (g) SR images by $SRGAN_{Ledig}$ ([@Srgan]); (h) SR images using $SRCNN$ ([@SRCNN]). Green contours are ground truth annotations and algorithm segmentations are denoted by red contours.[]{data-label="fig:MASeg1"}](MA_Seg_HR.jpg "fig:"){height="2.9cm" width="2.7cm"} (a) (b) (c) (d) ![Example results for microaneurysm segmentation using super resolved images obtained using different methods. (a) original image with microaneurysm region outline by red square; (b) cropped version of original image with yellow arrows denoting MA locations; (c) manual annotations of MA regions; super resolution and segmentation results obtained when trained on (d) original HR images; (e) SR images by $P-SRGAN$; (f) SR images by $SRGAN_{Sal}$ (g) SR images by $SRGAN_{Ledig}$ ([@Srgan]); (h) SR images using $SRCNN$ ([@SRCNN]). Green contours are ground truth annotations and algorithm segmentations are denoted by red contours.[]{data-label="fig:MASeg1"}](MA_Seg_Our.jpg "fig:"){height="2.9cm" width="2.7cm"} ![Example results for microaneurysm segmentation using super resolved images obtained using different methods. (a) original image with microaneurysm region outline by red square; (b) cropped version of original image with yellow arrows denoting MA locations; (c) manual annotations of MA regions; super resolution and segmentation results obtained when trained on (d) original HR images; (e) SR images by $P-SRGAN$; (f) SR images by $SRGAN_{Sal}$ (g) SR images by $SRGAN_{Ledig}$ ([@Srgan]); (h) SR images using $SRCNN$ ([@SRCNN]). Green contours are ground truth annotations and algorithm segmentations are denoted by red contours.[]{data-label="fig:MASeg1"}](MA_Seg_Sal.jpg "fig:"){height="2.9cm" width="2.7cm"} ![Example results for microaneurysm segmentation using super resolved images obtained using different methods. (a) original image with microaneurysm region outline by red square; (b) cropped version of original image with yellow arrows denoting MA locations; (c) manual annotations of MA regions; super resolution and segmentation results obtained when trained on (d) original HR images; (e) SR images by $P-SRGAN$; (f) SR images by $SRGAN_{Sal}$ (g) SR images by $SRGAN_{Ledig}$ ([@Srgan]); (h) SR images using $SRCNN$ ([@SRCNN]). Green contours are ground truth annotations and algorithm segmentations are denoted by red contours.[]{data-label="fig:MASeg1"}](MA_Seg_Ledig.jpg "fig:"){height="2.9cm" width="2.7cm"} ![Example results for microaneurysm segmentation using super resolved images obtained using different methods. (a) original image with microaneurysm region outline by red square; (b) cropped version of original image with yellow arrows denoting MA locations; (c) manual annotations of MA regions; super resolution and segmentation results obtained when trained on (d) original HR images; (e) SR images by $P-SRGAN$; (f) SR images by $SRGAN_{Sal}$ (g) SR images by $SRGAN_{Ledig}$ ([@Srgan]); (h) SR images using $SRCNN$ ([@SRCNN]). Green contours are ground truth annotations and algorithm segmentations are denoted by red contours.[]{data-label="fig:MASeg1"}](MA_Seg_SRCNN.jpg "fig:"){height="2.9cm" width="2.7cm"} (e) (f) (g) (h) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Results on MRI Dataset ---------------------- Although retinal fundus images have a high resolution during acquisition time, ISR is still relevant to detect small pathologies and landmarks. ISR’s impact can be further judged by its performance on MRI since MR images have low-resolution during acquisition time. Furthermore, the anatomies to be detected in MRI cover an even smaller area. Hence there is greater need for reliable ISR algorithms which facilitate better performance in detection or segmentation of pathologies. We demonstrate the relevance of ISR on cardiac MR images. ### Cardiac MR Data We used images from the Sunnybrook Cardiac Dataset [@Sunnybrook]. The images used in this evaluation study are cine steady state free precession (SSFP) MR short axis (SAX) images. All images were obtained on a 1.5T GE Signa MR scanner during 10 to 15s breath holds with a temporal resolution of 20 cardiac phases over the heart cycle (thickness=8-10mm, FOV= $320 \times 320$mm, matrix=$256 \times 256$). The experimental procedures were approved by the Institutional Review Board. The ground truth of the inner and outer boundaries of the left ventricle (endocardium and epicardium respectively) is manually delineated by the clinical experts from Sunnybrook Health Sciences Centre. The training set has images from $15$ patients, and the number of slices for each patient varied from $6-12$ We train our ISR network from scratch. Each volume slice is treated as a separate image and transformed by random rotation and translation. We rotate the images between $[0,180^{\circ}]$ in steps of $9^{\circ}$. For each rotation we translate the image by $[0,50]$ pixels in steps of $2$. Thus on an average we get $20*25=500$ transformation for each image. Thus an average of $10$ slices per image gives a total of $15\times10\times500=75,000$ cardiac MR images. We show results on ISR for scale factor $4,8$. Results for ISR are summarized in Tables \[tab:cardiacISR\]. Following the approach for retinal landmark and pathology segmentation, we also show results for cardiac left ventricle segmentation (Table \[tab:LVSeg\]). For each segmentation approach we employ UNets as the segmentation framework and show results for different super-resolution methods as well as the low-resolution images ($LR_n$). Dice metric values for segmentation accuracy are shown in each case. Similar to retinal pathology segmentation we extract a $256\times256$ patch covering the pathology and apply super-resolution for scale factors $4,8$. Figure \[fig:CardSRSeg\] shows results for segmenting the cardiac LV from MRI. For each case we present results on the original HR images, SR images obtained by each of the $5$ methods being compared and also when using the LR images (scale factor $4$). It is quite obvious that the LR images are very fuzzy and don’t give accurate information on the anatomical boundaries. On the other hand the SR images from our method can predict a highly accurate reconstruction of the actual image. Other ISR methods show some degree of blur in the SR images. It is remarkable that deep neural network based methods are able to reconstruct original high quality images despite limited information in LR images. This is possible because of the ability of the generator networks to learn the relation between HR and LR images. -------------------------------- ---------- ---------- ---------- ----------- [SSIM]{} [PSNR]{} [S3]{} [$p$]{} [(dB)]{} [$P-SRGAN$]{} [0.83]{} [43.7]{} [0.81]{} [-]{} [$SRGAN_{Sal}$]{} 0.78 [38.4]{} 0.75 $<0.001$ [$SRGAN_{Ledig}$ ([@Srgan])]{} [0.70]{} 34.1 0.70 $<0.001$ [$SRCNN$ ([@SRCNN])]{} [0.69]{} 32.7 0.65 $<0.0009$ [SR-RF ([@SR-RF])]{} [0.66]{} 31.2 0.63 $<0.0009$ [SSR ([@Jog])]{} [0.63]{} 28.3 0.61 $<0.001$ [SSIM]{} [PSNR]{} [S3]{} [$p$]{} [(dB)]{} [$P-SRGAN$]{} [0.79]{} [41.3]{} [0.78]{} [-]{} [$SRGAN_{Sal}$]{} 0.74 37.0 0.72 $<0.001$ [$SRGAN_{Ledig}$ ([@Srgan])]{} 0.67 31.7 0.66 $<0.001$ [$SRCNN$ ([@SRCNN])]{} 0.64 29.8 0.61 $<0.001$ [SR-RF ([@SR-RF])]{} 0.62 28.2 0.59 $<0.001$ [SSR ([@Jog])]{} 0.59 26.1 0.58 $<0.001$ -------------------------------- ---------- ---------- ---------- ----------- : Comparative results of different methods for cardiac image super-resolution.[]{data-label="tab:cardiacISR"} --------------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- [Scale (n)]{} [DM]{} [HD]{} [DM]{} [HD]{} [DM]{} [HD]{} [DM]{} [HD]{} [$4$]{} [91.3]{} [6.9]{} [90.1]{} [7.2]{} [86.4]{} [9.0]{} [88.2]{} [8.1]{} [$8$]{} [87.4]{} [8.8]{} [86.8]{} [9.2]{} [83.0]{} [10.6]{} [84.7]{} [9.8]{} [Scale (n)]{} [DM]{} [HD]{} [DM]{} [HD]{} [DM]{} [HD]{} [DM]{} [HD]{} [$4$]{} [84.8]{} [9.7]{} [82.3]{} [11.0]{} [81.1]{} [11.4]{} [83.4]{} [10.4]{} [$8$]{} [81.1]{} [11.5]{} [79.8]{} [11.9]{} [78.7]{} [12.4]{} [80.7]{} [11.6]{} --------------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- : Comparative cardiac LV segmentation results with different super-resolution methods at different scale factors.[]{data-label="tab:LVSeg"} ------------------------------------------------------------ ----------------------------------------------------------- ---------------------------------------------------------- ------------------------------------------------------------ ![image](Card_SF4_HR.jpg){height="3.2cm" width="3.2cm"} ![image](Card_SF4_Our.jpg){height="3.2cm" width="3.2cm"} ![image](Card_SF4_Sal.jpg){height="3.2cm" width="3.2cm"} ![image](Card_SF4_Ledig.jpg){height="3.2cm" width="3.2cm"} (a) (b) (c) (d) ![image](Card_SF4_SRCNN.jpg){height="3.2cm" width="3.2cm"} ![image](Card_SF4_SRRF.jpg){height="3.2cm" width="3.2cm"} ![image](Card_SF4_SSR.jpg){height="3.2cm" width="3.2cm"} ![image](Card_SF4_LR4.jpg){height="3.2cm" width="3.2cm"} (e) (f) (g) (h) ------------------------------------------------------------ ----------------------------------------------------------- ---------------------------------------------------------- ------------------------------------------------------------ Conclusion {#sec:concl} ========== We have proposed a novel method for super-resolution of different medical images based on progressive generative adversarial networks that combines triplet loss with the conventional MSE and CNN loss in a multi stage architecture. The core novelty of our method lies in incorporating the triplet loss into the cost function of PGANs. The triplet loss ensures that super resolved images produced in one stage undergo quality improvement for the next stage. Consequently our method is able to preserve the high quality of images for high scaling factors (greater than $8$). The super resolved images obtained using our method are of much better quality than the original GAN framework that uses MSE and CNN loss. The quality improvement of our super resolved images is evident from the image quality metrics. Our proposed method’s superior performance is also demonstrated in the case of retinal vessel segmentation and microaneurysm detection, as well MRI super-resolution and cardiac LV segmentation. The resulting super resolved images can be used to increase the size and resolution of low dimensional images, and different image analysis algorithms can be applied to the super resolved images. As demonstrated through our results, the SR images obtained by our method performs better than the low-resolution images from which the super resolved images are obtained. The results from using our super resolved images will be very close to that when using the original high-resolution images for segmentation or pathology detection. References ==========	{ "pile_set_name": "ArXiv" }
--- abstract: 'Octupole deformation results in a strongly enhanced collective Schiff moment in $^{229}$Th nucleus. An additional enhancement of time reversal (T) and parity (P) violating effects (such as T,P-violating electric dipole moments) appears in the ground $^1\Sigma$ state and in the metastable $^3\Delta_1$ state of diatomic molecule $^{229}$ThO. Similar enhancements exist in molecular ions $^{229}$ThOH$^+$, $^{229}$ThF$^+$ and $^{225,223}$RaOH$^+$. Corresponding experiments may be used to test CP-violation theories predicting T,P-violating nuclear forces and to search for axions.' author: - 'V.V. Flambaum $^{1,2}$' title: 'Enhanced nuclear Schiff moment and time reversal violation in $^{229}$Th - containing molecules ' --- . Measurements of T,P and CP -violating electric dipole moments (EDM) of elementary particles, nuclei and atoms provide crucial tests of unification theories and have already cornered many popular models of CP-violation including supersymmetry [@PR; @ERK]. Corresponding effects are very small, therefore, we are looking for the enhancement mechanisms - see e.g. [@Khriplovich; @KL; @GF]. According to the Schiff theorem the nuclear EDM is completely screened in neutral atoms [@Schiff]. EDM of diamagnetic atoms is produced by interaction of electrons with the nuclear Schiff moment. The Schiff moment is a vector multipole which produces electric field nside the nucleus. It appears in the third order of the multipole expansion of the nuclear electrostatic potential with added electron screening term [@Sandars; @Hinds; @SFK; @FKS1985; @FKS1986]. The distribution of the Schiff moment electric field inside the nucleus, the Hamiltonian describing interaction of the Schiff moment with relativistic atomic electrons and the finite nuclear size corrections to the formula for the Schiff moment have been considered in Ref. [@FG]. Refs. [@Sandars; @Hinds] calculated the Schiff moment due to proton EDM. Refs. [@SFK; @FKS1985; @FKS1986] calculated (and named) the nuclear Schiff moment produced by the P,T-odd nuclear forces. It was shown in [@SFK] that the contribution of the P,T-odd forces to the nuclear EDM and Schiff moment is $\sim 40$ times larger than the contribution of the nucleon EDM. Further enhancement of the nuclear Schiff moment may be due to the close nuclear levels of opposite parity with the same angular momentum which can be mixed by the T,P-odd nuclear forces [@SFK] (nuclear EDM and magnetic quadrupole can also be enhanced [@HH]). Nuclear T,P-odd moments such as magnetic quadrupoles may also be enhanced due to their collective nature in deformed nuclei [@F1994]. However, the largest enhancement ($\sim 10^2 - 10^3$ times) happens in nuclei with the octupole deformation where both the close nuclear level effect and the collective effect work together [@Auerbach; @Spevak]. [Calculation of collective Schiff moment]{}. The Schiff moment is defined by the following expression [@SFK]: $$\label{S} {\bf S}=\frac{e}{10} [<r^2 {\bf r}> - \frac{5}{3Z}<r^2><{\bf r}>],$$ where $<r^n> \equiv \int \rho({\bf r}) r^n d^3r$ are the moments of the nuclear charge density $\rho$. The second term originates from the electron screening and contains nuclear mean squared charge radius $<r^2>/Z$ and nuclear EDM $d=e<{\bf r}>$, where $Z$ is the nuclear charge. If a nucleus has an octupole deformation $\beta_3$ and a quadrupole deformation $\beta_2$, in the fixed-body (rotating) frame the Schiff moment $S_{intr}$ is proportional to the octupole moment $O_{intr}$, i.e. it has a collective nature [@Auerbach; @Spevak]: $$\label{Sintr} S_{intr} \approx \frac{3}{5 \sqrt{35}} O_{intr} \beta_2 \approx \frac{3}{20 \pi \sqrt{35}} e Z R^3 \beta_2 \beta_3 ,$$ where $R$ is the nuclear radius. However, in the laboratory frame EDM and Schiff moment are forbidden by the parity and time reversal invariance. Indeed, EDM and Schiff moment are polar $T$-even vectors which must be directed along the nuclear spin $I$ which is $T$-odd pseudovector. A nucleus with an octupole deformation and non-zero nucleon angular momentum has a doublet of close opposite parity rotational states $\|I^{\pm}>$ with the same angular momentum $I$ ($\| I^{\pm} >=\frac{1}{\sqrt{2}} (\|\Omega> \pm \|-\Omega>)$, where $\Omega$ is the projection of $I$ on to the nuclear axis). The states of of this doublet are mixed by $P,T-$violating interaction $W$. The mixing coefficient is: $$\label{alpha} \alpha=\frac{<I^-\| W\| I^+>}{E_+ - E_-} .$$ This mixing polarises nuclear axis ${\bf n}$ along the nuclear spin ${\bf I}$, $<n_z>= 2 \alpha \frac{I_z}{I+1}$, and the intrinsic Schiff moment shows up in the laboratory frame [@Auerbach; @Spevak]: $$\label{Scol} S= 2 \alpha \frac{I}{I+1} S_{intr}.$$ A nucleus with an octupole deformation also has a small intrinsic EDM $D$ due to a difference between the proton and neutron distributions which results in the laboratory frame nuclear EDM $d=2 \alpha \frac{I}{I+1} D$ [@Auerbach; @Spevak]. A similar $\Omega$- doublet mixing mechanism produces huge enhancement of electron EDM $d_e$ and T,P-odd interactions in polar molecules, such as ThO. Interaction of $d_e$ with molecular electric field produces the mixing coefficient $\alpha$ resulting in the orientation of large intrinsic molecular EDM $D \sim e a_B$ along the molecular angular momentum ${\bf J}$, and we obtain $d=2 \alpha \frac{J}{J+1} D \sim \alpha e a_B$ [@SushkovFlambaum], where $a_B$ is the Bohr radius. As a result, the $T,P$- violating molecular EDM $d$ exceeds electron EDM $d_e$ by 10 orders of magnitude. In the papers [@Auerbach; @Spevak] the numerical calculations of the Schiff moments and estimates of atomic EDM produced by electrostatic interaction between electrons and these moments have been done for $^{223}$Ra, $^{225}$Ra, $^{223}$Rn, $^{221}$Fr, $^{223}$Fr, $^{225}$Ac and $^{229}$Pa. The Schiff moment of $^{225}$Ra exceeds the Schiff moment of $^{199}$Hg (where the most accurate measurements of the Schiff moment have been performed [@HgEDM]) 200 times. Even larger enhancement of the $^{225}$Ra Schiff moment has been obtained in Ref. [@EngelRa]. For other nuclei the enhancement factors relative to Hg are between 30 and 700. Atomic calculations of EDM induced by the Schiff moment in Hg, Xe, Rn, Ra and Pu atoms have been performed in Refs. [@SFK; @FlambaumRa; @DzubaRa; @AtomicSchiff; @AtomicSchiff2] and include additional atomic enhancement mechanisms. It is useful to make an analytical estimate of the Schiff moment. According to Ref. [@Spevak] the T,P-violating matrix element is approximately equal to $$\label{W} <I^-\| W\| I^+> \approx \frac{\beta_3 \eta}{A^{1/3}} \textrm{eV}.$$ Here $\eta$ is the dimensionless strength constant of the nuclear $T,P$- violating potential $W$: $$\label{eta} W= \frac{G}{\sqrt{2}} \frac{\eta}{2m} ({\bf \sigma \nabla}) \rho ,$$ where $G$ is the Fermi constant, $m$ is the nucleon mass and $\rho$ is the nuclear number density. Eqs. (\[Sintr\],\[alpha\],\[Scol\],\[W\]) give the analytical estimate for the Schiff moment: $$\label{San} S \approx 1. \cdot 10^{-4} \frac{I}{I+1} \beta_2 \beta_3^2 Z A^{2/3} \frac{\textrm{KeV}}{E_- - E_+} e \,\eta \, \textrm{fm}^3,$$ This estimate gives $S=280 \, e \,\eta \, \textrm{fm}^3$ for $^{225}$Ra, very close to the result of the numerical calculation in Ref. [@Spevak] $S=300 \, e \,\eta \, \textrm{fm}^3$. The values of the Schiff moments for the nuclei with octupole deformation listed above vary from 45 to 1000 $10^{-8} e \eta$ fm$^3$ [@Spevak]. For spherical nuclei $^{199}$Hg, $^{129}$Xe, $^{203}$Tl and $^{205}$Tl, where the Schiff moment measurements have been performed, the calculations [@SFK; @FKS1985; @FKS1986] give the Schiff moment $S \sim 1 \times 10^{-8} $ $e \eta$ fm$^3$. The Schiff moment in Eq. (\[San\]) is proportional to the squared octupole deformation parameter $\beta_3^2$ which is about $(0.1)^2$. According to Ref. [@Engel2000], in nuclei with a soft octupole vibration mode the squared dynamical octupole deformation $<\beta_3^2 > \sim (0.1)^2$, i.e. it is the same as the static octupole deformation. This means that a similar enhancement of the Schiff moment may be due to the dynamical octupole effect [@Engel2000; @FZ; @soft2] in nuclei where $<\beta_3>=0$. Unfortunately, the nuclei with the octupole deformation and non-zero spin have a short lifetime. Several experimental groups have considered experiments with $^{225}$Ra and $^{223}$Rn. The only published EDM measurement [@RaEDM] has been done for $^{225}$Ra which has 15 days half-life. In spite of the Schiff moment enhancement the $^{225}$Ra EDM measurement has not reached yet the sensitivity to the T,P-odd interaction Eq. (\[eta\]) comparable to the Hg EDM experiment [@HgEDM]. The experiments continue, however, the instability of $^{225}$Ra and a relatively small number of atoms available may be a problem. To have a breakthrough in the sensitivity we need a more stable nucleus and a larger number of atoms. An excellent candidate is $^{229}$Th nucleus which lives 7917 years and is very well studied in numerous experiments and calculations (this nucleus is the only candidate for the nuclear clock which is expected to have a precision significantly better than atomic clocks [@Peik; @Thclock], has strongly enhanced effects of “new physics” [@FlambaumTh; @FlambaumTh1] and may be used for a nuclear laser [@Tkalya]). $^{229}$Th is produced in macroscopic quantities by the decay of $^{233}$U (see e.g. [@production]), and its principal use is for the production of the medical isotopes $^{225}$Ac and $^{213}$Bi. According to Ref. [@Minkov] the $^{229}$Th nucleus has the octupole deformation with the parameters $\beta_3$=0.115, $\beta_2$=0.240, $I=5/2$ and the interval between the opposite parity levels $E(5/2^-)- E(5/2^+)$=133.3 KeV. The analytical formula in Eq. (\[San\]) allows us to scale the value of the Schiff moment from the numerical calculations for $^{225}$Ra which has $\beta_3$=0.099, $\beta_2$=0.129, $I=1/2$ and interval between the opposite parity levels $E(1/2^-)- E(1/2^+)$=55.2 KeV [@Spevak]. Then Eq. (\[San\]) gives: $$\label{SThRa} S( ^{229}\textrm{Th})= 2 \, S( ^{225}\textrm{Ra}),$$ Using $S( ^{225}\textrm{Ra})=300 \times 10^{-8} $ $e \eta$ fm$^3$ [@Spevak] we obtain $S( ^{229}\textrm{Th})=600 \times 10^{-8} $ $e \eta$ fm$^3$. Within the meson exchange theory the $\pi$-meson exchange gives the dominating contribution to the T,P-violating nuclear forces [@SFK]. According to Ref. [@FDK] the neutron and proton constants in the T,P-odd potential (\[eta\]) may be presented as $\eta_n \approx - \eta_p \approx 5 \times 10^6 ( -0.2 g {\bar g}_0 + g {\bar g}_1 + 0.4 g {\bar g}_2$). In Refs. [@Auerbach; @Spevak] we have not separated the proton and neutron contributions. Majority of the nucleons are neutrons, so it make sense to take $\eta=\eta_n$. However, the proton interaction constant has an opposite sign and may cancel a part of the neutron contribution, so we multiply the interaction constant by $((N-Z)/N)=0.36$ and use $\eta=0.36\eta_n$. This way we can obtain a rough estimate: $S( ^{225}\textrm{Ra})= ( - 2.2 g {\bar g}_0 + 11 g {\bar g}_1 + 4 g {\bar g}_2)\, e\, \textrm{fm}^3$, $S(^{229}\textrm{Th})= ( - 4.4 g {\bar g}_0 + 22 g {\bar g}_1 + 8 g {\bar g}_2)\, e\, \textrm{fm}^3$. A more accurate job has been done in Ref. [@EngelRa] where they presented the Schiff moment as $S( ^{225}\textrm{Ra})= (a_0 g {\bar g}_0 + a_1 g {\bar g}_1 + a_2 g {\bar g}_2) e $ fm$^3$. To estimate the error the authors of Ref. [@EngelRa] have done the calculations using 4 different models of the strong interaction. They obtained the following 4 sets of the coefficients: $a_0= -1.5,\, -1.0,\, -4.7,\, -3.0$; $a_1= 6.0,\, 7.0,\, 21.5,\, 16.9$; $a_2= -4.0,\, -3.9,\, -11.0,\, -8.8$. Taking the average values of the coefficients and using Eq. (\[SThRa\]) we obtain: $$\begin{aligned} \label{SThRag} S( ^{225}\textrm{Ra})= ( - 2.6 g {\bar g}_0 + 12.9 g {\bar g}_1 -6.9 g {\bar g}_2)\, e\, \textrm{fm}^3,\\ S( ^{229}\textrm{Th})= ( - 5.1 g {\bar g}_0 + 25.7 g {\bar g}_1 -13.9 g {\bar g}_2)\, e\, \textrm{fm}^3. \end{aligned}$$ We will use these expressions as our final values for the Ra and Th Schiff moments. We can express the results in terms of the more fundamental parameters such as the QCD $\theta$-term constant ${\bar \theta}$ and the quark chromo-EDMs ${\tilde d_u}$ and ${\tilde d_d}$ using the relations $g {\bar g}_0=- 0.37{\bar \theta}$ [@Witten] and $g {\bar g}_0 = 0.8 \cdot10^{15}({\tilde d_u} +{\tilde d_d})$/cm, $g {\bar g}_1 = 4 \cdot 10^{15}({\tilde d_u} - {\tilde d_d})$/cm [@PR]: $$\begin{aligned} \label{SRatheta} S( ^{225}\textrm{Ra})= 1.0 \,{\bar \theta} \, e\, \textrm{fm}^3,\\ \label{SThtheta} S( ^{229}\textrm{Th})= 2.0 \, {\bar \theta} \, e\, \textrm{fm}^3,\\ \label{SRaD} S( ^{225}\textrm{Ra})= 10^4 ( 0.50 \,{\tilde d}_u - 0.54 \,{\tilde d}_d )\, e\, \textrm{fm}^2,\\ \label{SThD} S( ^{229}\textrm{Th})= 10^4 ( 1.0 \,{\tilde d}_u - 1.1\, {\tilde d}_d )\, e\, \textrm{fm}^2. \end{aligned}$$ Note that the contributions of ${\bar \theta}$ and ${\tilde d_{u,d}}$ should not be added to avoid double counting since ${\tilde d_{u,d}}$ may be induced by ${\bar \theta}$. [Molecular enhancement]{}. Atomic EDM $d_a$ produced by the Schiff moment $S$ very rapidly increases with the nuclear charge $Z$ [@SFK; @Khriplovich; @KL; @Spevak]: $$\label{dZ} d_a \propto Z^2 (\frac{a_B}{2ZR})^{2- 2 \gamma} S,$$ where $R$ is the nuclear radius, $a_B$ is the Bohr radius, $\gamma=\sqrt{1 - (Z \alpha)^2}$. Th and Ra have close nuclear charges, $Z=88$ and $90$, and similar electronic structure up to last filled $7s^2$ subshell. Two extra $6d^2$ electrons in Th have high angular momenta, do not penetrate the nucleus and do not interact with the Schiff moment directly (up to many-body corrections). Therefore, $d_a/S$ for Th is approximately equal to $d_a/S$ for Ra. Using calculations of Ra atom EDM from Refs. [@AtomicSchiff; @AtomicSchiff2] we have $$\label{dS} d_a(\textrm{Th}) \approx -9 \cdot 10^{-17} \frac{S}{\|e\|\, \textrm{fm}^3} \|e\| \, \textrm{cm}= -2 \cdot 10^{-16} {\bar \theta}\, \|e\| \, \textrm{cm}.$$ $ d_a$(Th) as a function of other T,P and CP–violating interaction constants $\eta,\, {\bar g}, \, {\tilde d}$ can be found by the substitution of the Th Schiff moment from the equations in the nuclear Schiff moment section above. This value of Th EDM is 3 orders of magnitude larger than Hg EDM and 4 orders of magnitude larger than Xe EDM. However, Th atom has non-zero electron angular momentum, $J=2$, and this reduces the signal coherence time and increases systematic errors. In principle, one may use Th$^{4+}$ ion which has closed shells or look for zero electron angular momentum Th ions in solid state materials. Note that the measurements of the effects produced by the $^{229}$Th Schiff moment may be used to search for axions. Indeed, the axion dark matter produces oscillating neutron EDM [@Graham] and oscillating Schiff moment [@Stadnik], the latter is enhanced in $^{229}$Th by the same octupole mechaninism. Indeed, the axion dark matter field $a(t)=a_0 cos(m_a t)$ ($m_a$ is the axion mass) generates oscillating nuclear forces which are similar to the T,P-odd nuclear forces producing the Schiff moments. To obtain the result for the oscillating Schiff moments and EDM it is sufficient to replace the constant ${\bar \theta}$ by $a(t)/f_a$, where $f_a$ is the axion decay constant [@Graham; @Stadnik]. Search for the effects produced by the oscillating axion-induced Schiff moments in solid state materials is in progress [@Casper]. A promising direction here may be to use $^{229}$ThO molecule placed in a matrix of Xe (or other) atoms. A proposal to use paramagnetic molecules in the matrix of rare-gas atoms for the electron EDM search has been described in Ref. [@Kozlov]. Promising objects for the Th Schiff moment measurement may be ThO molecule and ThOH$^+$ molecular ion. Both molecules have zero electron angular momentum in the ground state and very close opposite parity levels which enhance T,P-violating EDM. Use of polar diatomic molecules for the measurement of the nuclear Schiff moment was suggested by Sandars [@Sandars; @Hinds] because electric field inside polarised molecule exceeds external electric field $\epsilon $ by several orders of magnitude and has the same direction. The molecular polarisation is $P \sim D\epsilon/(E_- - E_+)$, where $D \sim e a_B$ is the intrinsic electric dipole moment of the polar molecule. Therefore, to have a significant polarization degree $P$ the interval between the opposite parity molecular rotational levels $(E_- - E_+)$ should be sufficiently small. Indeed, the rotational interval in molecules is 3-5 orders of magnitude smaller than a typical interval between the opposite parity levels in atoms. We may interpret the molecular enhancement in a different way [@SushkovFlambaum]: interaction between the Schiff moment and electrons mixes close opposite parity levels in the molecule, polarises the molecule along its angular momentum and creates T,P-violating EDM proportional to the large intrinsic electric dipole moment $D$ - see the discussion below Eq. (\[Scol\]). This enhanced EDM interacts with the external electric field $\epsilon $. The experiment has been performed with the TlF molecule [@TlFexperiment]. In the paper Ref. [@RaO] it was proposed to study molecule $^{225}$RaO where the effect may be 500 times larger than in TlF due to the enhanced Schiff moment and larger nuclear charge $Z$. The best sensitivity to the electron EDM has been obtained using molecules ThO [@TheEDM] and HfF$^+$ [@Cornell] in the excited metastable electronic state $^3\Delta_1$ which contains doublets of very close opposite parity levels. Finally, in the recent paper [@MOH+] it was suggested that linear molecules MOH, molecular ions MOH$^+$ (M is a heavy atom, e.g. Ra in the molecule RaOH$^+$ ) and symmetric top molecules (such as MCH$_3$ or MOCH$_3$) may be better systems than molecules MO since such polyatomic molecules have a doublet of the close opposite parity energy levels in the bending mode and may be polarised by a weak electric field. The reduction of the strength of the necessary electric field simplifies the experiment and dramatically reduces systematic effects. The T,P-violating effect in $^{229}$ThO is much larger than in TlF due to the enhanced Schiff moment and larger nuclear charge. An additional advantage may appear in ThOH$^+$ which is expected to have very close opposite parity states (similar to RaOH$^+$). Another possibility may be to use the doublet in $^3\Delta_1$ metastable state of $^{229}$ThO (used to improve the limit on electron EDM ) and the ground state doublet $^3\Delta_1$ in ThF$^+$. The interaction constant $W_S$ for the effective T,P-violating interaction in molecules $$\label{WS} W_{T,P}=W_S\frac{S}{I} {\bf I \cdot n}$$ (here ${\bf I}$ is the nuclear spin, ${\bf n}$ is the unit vector along the molecular axis) in ThO, ThOH$^+$ and RaOH$^+$ may be estimated by the comparison with the RaO molecule. Calculation of $W_S$ for RaO has been done in Ref. [@RaOTitov]: $W_S(\textrm{RaO})=45192$ atomic units (here a.u.=$e/a_B^4$). In RaOH$^+$ ion the electron density on the Ra nucleus is slightly smaller than in RaO (since a part of the electron charge density moves to hydrogen), therefore, we assume $W_S(\textrm{RaOH}^+) \approx 30000$ a.u. In ThO and in ThOH$^+$ the electron density on the Th nucleus is expected to be slightly larger than that for Ra due to the higher Th charge and two extra electrons. Therefore, we assume $W_S(\textrm{ThO}) \approx 50000$ a.u. and $W_S(\textrm{ThOH}^+) \approx 30000$ a.u. Note that the electron wave function in the bending molecular mode of RaOH$^+$ and ThOH$^+$ is the same as in their gound states, therefore, the parameters $W_S$ are practically not affected by these bending vibrations (where we have the doublet of the opposite parity levels). The parameter $W_S$ in the $^3\Delta_1$ state of the $^{229}$ThO molecule should have comparable value to their ground state values since $^3\Delta_1$ and the ground $^1\Sigma_0$ state differ by one electron orbital only. The $^3\Delta_1$ state in ThF$^+$ molecular ion is similar to the $^3\Delta_1$ state in $^{229}$ThO molecule. The estimates presented above are based on comparison with the numerical calculations of the Schiff moment contribution in RaO. Estimates based on the $Z$ dependence extrapolation Eq. (\[dZ\]) from TlF give 2 times larger results. Substitution of the Schiff moment (\[SThtheta\]) to the energy shift $W_{T,P}=W_S\frac{S}{I} {\bf I \cdot n}$ gives for the fully polarised molecule the energy difference between the $I_z=I$ and $I_z=-I$ states in $^{229}$ThO: $$\label{WTP} 2 W_S S=1. \cdot 10^7 {\bar \theta} \, \textrm{Hz} %2<W_{T,P}(^{229}\textrm{ThO})>=2 W_S S= = 1. \cdot 10^7 {\bar \theta} Hz %1. \cdot 10^5 (e/a_B)^4 {\bar \theta} \, e\, \textrm{fm}^3$$ A similar estimate is valid for molecular ions ThF$^+$ and ThOH$^+$. The measured shift in the 1991 TlF experiment [@TlFexperiment] was $-0.13 \pm 0.22$ mHz. The same sensitivity in the $^{229}$ThO, ThF$^+$ or ThOH$^+$ experiments would allow one to improve the current limit $\| {\bar \theta} \| < 10^{-10}$ and also the limits on other fundamental parameters of the CP-violation theories such as the strength of T,P-violating potential $\eta$, the $\pi NN$ interaction constants ${\bar g}$ and the quark chromo-EDMs ${\tilde d}$. [Comaparison with existing and proposed experiments]{}. We should compare suggested experiments with $^{229}$ThO, ThF$^+$ and $^{229}$ThOH$^+$ molecules with other existing and proposed experiments. The best limit on the nuclear Schiff moment has been obtained in the measurement of Hg EDM [@HgEDM]. However, there is a theoretical problem here: the most recent sophisticated calculation [@Engel] was not able to find out even the sign of the Hg Schiff moment, different interaction models give very different results. There are two reasons for this: firstly, the Schiff moment is determined by the charge distribution of the protons. However, it is directed along the nuclear spin which in $^{199}$Hg is carried by the valence neutron, i.e. the Schiff moment in $^{199}$Hg is determined by the many-body effects which are harder to calculate. The second reason is in the formula for the Schiff moment defined by Eq. (\[S\]). There are two terms of opposite sign in this formula which tend to cancel each other, the main term and the screening term (remind the reader that the screening term kills the nuclear EDM contribution to the atomic EDM). If we do not know each term sufficiently accurately, the final sign and the magnitude of the Schiff moment are unknown. Recently the interest in EDM experiments has moved towards molecules where the effects are very strongly enhanced by the close rotational levels and very strong internal “effective electric field”. For example, the limit on electron EDM in ThO and HgF$^+$ experiments have been improved by more than an order of magnitude in comparison with the atomic EDM experiments. The Tl nuclear Schiff moment has been measured in the TlF experiment [@TlFexperiment]. Similar to $^{199}$Hg, calculations of the $^{203,205}$Tl Schiff moments suffer from the problem of the cancellation between two approximately equal terms in Eq. (\[S\]) and the problem of the nuclear core polarization contribution (since there is a strong cancellation between the two terms in the valence proton contribution in Tl [@SFK; @FKS1986; @FDK]). Actually, the interpretation of the TlF experiment [@TlFexperiment] was done in terms of the proton EDM. However, here we probably have even a more serious problem (below we will follow the discussion in Ref. [@FDK]). Firstly, calculations with different choices of the strong interaction give different signs and magnitudes of the Schiff moment $S_p$ induced by the proton EDM ( since we also have here the cancellation between the main and the screening contributions). The authors of the molecular calculation [@Coveney] selected the maximal value out of 4 numbers calculated by A. Brown (this maximal number leads to the strongest limit on the proton EDM), and this value of $S_p$ was used in all other molecular calculations for TlF [@Parpia; @Quiney; @Petrov] (see also [@AtomicSchiff]). There was no such accuracy investigation for the proton EDM contribution to the Hg Schiff moment but naively we may expect that the accuracy is actually lower than in Tl since the valence nucleon in $^{199}$Hg is neutron. The second problem is that in practically any model the contribution of the T,P-violating nuclear forces to the Schiff moment is 1-2 orders magnitude larger than the proton EDM contribution (the ratio is “model independent” since the $\pi NN$ interaction constant appears in both contributions and cancels out in the ratio). Therefore, to obtain the limit on the proton EDM we neglect much larger contribution of the P,T-odd nuclear forces. Thus, in the Particle Data tables the limit on the proton EDM is presented assuming that there are no other contributions to atomic and molecular EDM. However, if we wish to test CP-violation theories such limits on the proton EDM from Hg and Tl EDM can hardly be used. These theoretical problems do not exist for the collective Schiff moments in the nuclei with the octupole deformation. The second screening term is very small in this case since it is proportional to a very small intrinsic dipole moment $D$ of the “frozen” nucleus. If the distributions of the neutrons and protons are the same, $D=0$. Thus, there is no cancellation and the intrinsic Schiff moment is proportional to the known electric octupole moment (which may actually be measured using probabilities of the octupole transitions between the rotational levels). Then the calculation is reduced to the expectation value of the T,P-odd interaction $<\Omega\|W\|\Omega>$, here $\|\Omega>$ is the ground state of the “frozen” deformed nucleus [@Auerbach; @Spevak]. Calculation of one expectation value looks more reliable than the calculation of the infinite sum $\sum_n \frac{\|n><n\|W\|0>}{E_0 -E_n}$ in nuclei where there is no single dominating contribution. Thus, calculations of the collective Schiff moments look more “clean” theoretically. More importantly, the collective Schiff moment is enhanced by 2-3 orders of magnitude. [Conclusion]{}. We propose to search for the T,P and CP violating effects in the molecule $^{229}$ThO where the effects are 2-3 orders of magnitude larger than in TlF due to the enahnced Schiff moment of the $^{229}$Th nucleus and large nuclear charge. An additional advantage may be in $^{229}$ThOH$^+$ molecular ion, in $^3\Delta_1$ state of the $^{229}$ThO molecule and in $^{229}$ThF$^+$ molecular ion which have very close opposite parity energy levels and may be polarised by a weak electric field. The $^3\Delta_1$ state of the $^{229}$ThO molecule has already been used to measure electron EDM. The enhanced effects in these molecules may also be used to search for axions. $^{229}$Th lives 7917 years, may be produced in macroscopic quantities (as it is done for the medical applications) and is very well studied in numerous experiments. This work is supported by the Australian Research Council and Gutenberg Fellowship. I am grateful to M. Kozlov, N. Hutzler, A. Palffy, Jun Ye, D. DeMiIle, H. Feldmeier, N. Minkov, A. Afanasiev, P. Ring and TACTICA collaboration for useful discussions. [99]{} M. Pospelov, A. Ritz, Ann. Phys. (Amsterdam) [318]{}, 119 (2005). J. Engel, M.J. Ramsey-Musolf, U. van Kolck, Prog.Part.Nucl.Phys. [71]{}, 21 (2013). I.B. Khriplovich, Parity Nonconservation in Atomic Phenomena (Gordon and Breach, Amsterdam, 1991). I.B. Khriplovich, S.K. Lamoreaux, CP violation without strangeness. (Springer-Verlag, Berlin, 1997). Violation of fundamental symmetries in atoms and tests of unification theories of elementary particles. J.S.M. Ginges, V.V. Flambaum. Phys. Rep. [397]{}, 63 (2004). L.I. Schiff, Phys. Rev. [132]{}, 2194 (1963). P. G. H. Sandars, Phys.Rev.Lett. [ bf 19]{}, 1396 (1967). E. A. Hinds and P. G. H. Sandars, Phys.Rev.A [21]{}, 471 (1980). On the possibility of investigation of P- and T-odd nuclear forces in atomic and molecular experiments. O.P. Sushkov, V.V. Flambaum, I.B. Khriplovich. Zh. Exp. Teor. Fiz. [87]{}, 1521 (1984) \[ Sov. JETP [60]{}, 873 (1984)\]. V.V. Flambaum, I.B. Khriplovich, O.P. Sushkov, Phys. Lett. B [162]{}, 213 (1985). On the P- and T-nonconserving nuclear moments, V.V. Flambaum, I.B. Khriplovich, O.P. Sushkov. Nucl. Phys. A [449]{}, 750, 1986. Nuclear Schiff moment and time invariance violation in atoms. V.V. Flambaum and J.S.M. Ginges, Phys. Rev. A [65]{}, 032113 (2002). W.C. Haxton, E.M. Henley, Phys. Rev. Lett. [51]{}, 1937 (1983). Spin hedgehog and collective magnetic quadrupole moments induced by parity and time invariance violating interaction. V. V. Flambaum, Phys. Lett. B [320]{}, 211 (1994). Collective $T$- and $P$-odd electromagnetic moments in nuclei with octupole deformations. N. Auerbach, V. V. Flambaum, and V. Spevak, Phys. Rev. Lett. [76]{}, 4316 (1996). Enhanced T-odd P-odd electromagnetic moments in reflection assymmetric nuclei. V. Spevak, N. Auerbach, and V.V. Flambaum, Phys. Rev. C [56]{}, 1357 (1997). Reduced Limit on the Permanent Electric Dipole Moment of $^{199}$Hg. B. Graner, Y. Chen, E.G. Lindahl, and B.R. Heckel. Phys. Rev. Lett. 116, 161601 (2016); Phys. Rev. Lett. 119, 119901 (E) (2017). Time-reversal violating Schiff moment of 225Ra. J. Engel, M. Bender, J. Dobaczewski, J. H. de Jesus, and P. Olbratowski Phys. Rev. C [68]{}, 025501 (2003). Enhancement of parity and time invariance violation in the radium atom. V.V. Flambaum. Phys. Rev. A60, R2611-2613 (1999). Calculation of parity and time invariance violation in the radium atom. V.A. Dzuba, V.V. Flambaum, J.S.M. Ginges, Phys. Rev. A61, 062509-1 - 062509-10 (2000). Electric dipole moments of Hg, Xe,Rn,Ra,Pu and TlF induced by the nuclear Schiff moment and limits on time-reversal violating interactions. V.A. Dzuba, V.V. Flambaum, J.S.M. Ginges and M.G. Kozlov. Phys. Rev. A66,012111 (2002) . Calculation of P,T-odd electric dipole moments for diamagnetic atoms $^{129}$Xe, $^{171}$Yb, $^{199}$Hg, $^{211}$Rn, and $^{225}$Ra V.A. Dzuba, V.V. Flambaum, S.G. Porsev, Phys Rev A80,032120 (2009). Nuclear octupole correlations and the enhancement of atomic time-reversal violation. J. Engel, J.L. Friar, and A.C. Hayes, Phys. Rev. C [61]{}, 035502 (2000). Enhancement of nuclear Schiff moment and time reversal violation in atoms due to soft nuclear octupole vibrations. V.V. Flambaum, V.G. Zelevinsky. Phys. Rev. C [68]{}, 035502 (2003). Nuclear Schiff moment in nuclei with soft octupole and quadrupole vibrations. N. Auerbach, V.F. Dmitriev, V.V. Flambaum, A. Lisetskiy, R.A. Sen’kov, V.G. Zelevinsky, Phys. Rev. C [74]{}, 025502 (2006). Effects of parity nonconservation in diatomic molecules. O.P. Sushkov, V.V.Flambaum. Zh. Exp. Teor. Fiz. [75]{}, 1208, (1978) \[Sov. JETP [48]{}, 608 (1978)\]. First Measurement of the Atomic Electric Dipole Moment of $^{225}$Ra. R.H. Parker, M.R. Dietrich, M.R. Kalita, N.D. Lemke, K.G. Bailey, M. Bishof, J.P. Greene, R.J. Holt, W. Korsch, Z.-T. Lu, P. Mueller, T.P. O’Connor, and J.T. Singh, Phys. Rev. Lett. 114, 233002 (2015). Nuclear laser spectroscopy of the 3.5eV transition in Th$^{229}$, E. Peik, Chr. Tamm, Europhys. Lett., 61, 181 (2003) A Single-Ion Nuclear Clock for Metrology at the 19th Decimal Place . C. J. Campbell, A. G. Radnaev, A. Kuzmich, V. A. Dzuba, V. V. Flambaum, and A. Derevianko, Phys. Rev. Lett.108, 120802 (2012). Enhanced effect of temporal variation of the fine structure constant and strong interaction in 229Th. V.V. Flambaum, Phys. Rev. Lett. 97, 092502 (2006). Enhancing the effect of Lorentz invariance and Einstein’s equivalence principle violation in nuclei and atoms, V.V. Flambaum. Phys. Rev. Lett. 117, 072501 (2016). Proposal for a Nuclear Gamma-Ray Laser of Optical Range. E. V. Tkalya. Phys. Rev. Lett. 106, 162501 (2011). Energy Splitting of the Ground-State Doublet in the Nucleus 229Th. B. R. Beck, J. A. Becker, P. Beiersdorfer, G. V. Brown, K. J. Moody, J. B. Wilhelmy, F. S. Porter, C. A. Kilbourne, and R. L. Kelley. Phys. Rev. Lett. 98, 142501 (2007). Reduced transition probabilities for gamma decay of the 7.8 eV isomer in $^{229}$Th. N. Minkov, A. Palffy, Phys. Rev. Lett, 118, 212501(2017). Fully self-consisten calculations of nuclear Schiff moments. S. Ban, J. Dobaczewski, J. Engel, and A. Shukla Phys. Rev. C [82]{}, 015501 (2010). V. V. Flambaum, D. DeMille and M.G. Kozlov Phys. Rev. Lett. [113]{}, 103003 (2014). R.J. Crewther, P. di Vecchia, G. Veneziano, E. Witten. Phys. Lett. B [91]{}, 487 (1980). Axion Dark-Matter Detection with Cold Molecules, P. W. Graham and S. Rajendran, Phys. Rev. D84, 055013 (2011). Axion-induced effects in atoms, molecules and nuclei: Parity nonconservation, anapole moments, electric dipole moments, and spin-gravity and spin-axion momentum couplings. Y.V. Stadnik, V.V. Flambaum, Phys. Rev. D[89]{}, 043522 (2014). Proposal for a Cosmic Axion Spin Precession Experiment (CASPEr). Dmitry Budker, Peter W. Graham, Micah Ledbetter, Surjeet Rajendran, and Alexander O. Sushkov, Phys. Rev. X 4, 021030 (2014) . M.G. Kozlov, A. Derevianko. Phys. Rev. Lett. [97]{}, 063001 (2006). D. Cho, K. Sangster, E.A. Hinds, Phys. Rev. A [44]{}, 2783 (1991). Electric dipole moments of actinide atoms and RaO molecule. V.V. Flambaum, Phys.Rev. A[77]{},024501 (2008). Calculation of P,T-odd interaction effect in 225RaO. A.D. Kudashov, A.N. Petrov, L.V. Skripnikov, N.S. Mosyagin, A.V. Titov, and V.V. Flambaum, Phys. Rev. A[87]{}, 020102 (R) (2013). Precision measurement of time-reversal symmetry violation with laser-cooled polyatomic molecules. I. Kozyrev, N.R. Hutzler, Phys. Rev. Lett., [119]{}, 133002 (2017). J. Baron et al (ACME collaboration), Science [343]{}, 269 (2014). William B. Cairncross, Daniel N. Gresh, Matt Grau, Kevin C. Cossel, Tanya S. Roussy, Yiqi Ni, Yan Zhou, Jun Ye, and Eric A. Cornell. Phys. Rev. Lett. [119]{}, 153001 (2017). P. V. Coveney and P. G. H. Sandars, J. Phys. B16, 3727 (1983). F. A. Parpia, J. Phys. B30, 3983 (1997). H. M. Quiney, J. K. Laerdahl, K. Faegri, Jr., and T. Saue, Phys. Rev. A57, 920 (1998). A. N. Petrov, N. S. Mosyagin, T. A. Isaev, A. V. Titov, V. F. Ezhov, E. Eliav, and U. Kaldor, Phys. Rev. Lett. 88, 073001 (2002).	{ "pile_set_name": "ArXiv" }
--- abstract: 'We consider a non-critical five dimensional string setup which could provide a dual description of QCD in the limit of large number of colors and flavors. The model corresponds to $N_c$ color D3-branes and $N_f$ D4/anti D4-brane pairs supporting flavor degrees of freedom. The matching of the string model spectrum with the dual field theory one is considered. We discuss the consequences of the possible matching of the gravity modes with the light glueballs and the interpretation of the brane spectrum in Yang-Mills and QCD.' --- [Aldo L. Cotrone]{} Departament ECM, Facultat de Física, Universitat de Barcelona and\ Institut de Fisica d’Altes Energies,\ Diagonal 647, E-08028 Barcelona, Spain.\ and\ Institute of Theoretical Physics, KU Leuven\ Celestijnenlaan 200D, B-3001 Leuven, Belgium.\ [email protected] Introduction {#secintro} ============ Planar four dimensional YM theory and QCD may have a description as string theories in at least five dimensions [@thooft; @veneziano; @polyakov]. We will follow the prejudice that such a description exists and that five dimensions are sufficient. Such a dual to YM or QCD must have a background with curvature of the geometry of order one in $\alpha'$ units. Thus, a gravitational description of the system is not a good approximation and we should really solve the associated sigma-model. Of course, at present this is an impossible task. The best we can do is to try and learn some qualitative features of these theories, using the minimal amount of assumptions we can formulate about the actual string theory dual of YM or QCD.[^1] Besides defining better the problem, this exercise could provide a more stringy-inspired framework for the phenomenological models that go under the name of AdS/QCD [@katz; @pomarol]. For example, one often encounters in these models some five dimensional lagrangians which include a scalar field dual to the quark bilinear, that have a natural derivation from the expansion of the DBI action for systems of branes/anti-branes supporting the flavor degrees of freedom, the quark bilinear being the open string tachyon.[^2] In these notes, we will formulate the setup that we believe is the most natural candidate for describing YM and QCD in five dimensions, essentially borrowing all the data from our knowledge of the ten dimensional cases [@tutti1; @tutti2; @tutti3; @tutti4]. Clearly, since we have no control on the theory, some of the material will be speculative; nevertheless, we believe it is worthwhile to begin such a discussion. In the next section we will describe the basic features of the 5d models, in order to settle the background for a qualitative comparison of (some part of the) spectra of string theory with the ones of YM and QCD. The scalar and vectorial mesonic sector of the latter has recently been described in [@poli; @ckp], so we will just briefly report the proposal at the end of section \[secsetup\]. Then, in the following sections, we will discuss the possible qualitative matching of the gravity and non-perturbative spectrum of string theory with the YM one. Interestingly, despite our ignorance about strings in 5d, it seems that, assuming that the theory is not very different from the known 10d ones, we can have a (almost) consistent picture of the spectrum of YM theory from the string. Moreover, we can deduce some interesting possible implications for both theories. For example, the glueball spectrum indicates that among the string fields there should not be any tachyon dual to a light glueball, nor any degenerate doubling of the RR spectrum typical of a Type 0 theory in 10d, as also argued recently in [@Thorn:2008ay]. Moreover, the consistency of the D-brane spectrum with the YM one indicates that the YM coupling is likely to blow up in the far IR, rather than relaxing to a constant value of order one. Unfortunately, we could not give a complete discussion of the spectrum in these notes. Among the interesting features we do not discuss there are important issues such as the topological susceptibility and the gluon condensate. We hope to be able to discuss these features in the future. As a final remark, we report that in [@gk; @gkn] some of the issues discussed here are approached in the context of an effective gravity model in five dimensions, in the spirit of AdS/QCD. Our perspective has the complementary advantage of pointing out that some results do not rely on a specific gravity model (which is not fully under control as a string approximation in five dimensions). The setup {#secsetup} ========= The building blocks of YM and QCD are expected to be as follows. The strongly curved five dimensional background dual to $SU(N_c)$ YM is generated by a stack of $N_c$ color D3-branes [@tutti1; @tutti2; @tutti3; @tutti4; @wall]. It includes at least a running dilaton and a space-time filling five-form RR field $F_5$, proportional to the number of colors $N_c$. There could also be a non-trivial axion turned on, giving a non-zero $\theta_{YM}$ angle. The metric in string frame reads [^3] $$\label{metric} ds^2 = e^{2f(r)}dx_\mu dx^\mu + dr^2\ ,$$ where the warp factor for the Minkowski part of the metric, $e^{2f(r)}$, approaches a non-vanishing constant at the bottom of the space (this is necessary for confinement [@wym; @kinar]), that we will put at $r=0$. This corresponds to the IR of the dual field theory. At large radius, that is in the UV of the field theory, the metric should approach $AdS_5$, $e^{2f(r)}\sim e^{2 k r}$ (with radius $R^2=1/k^2$ of order one in $\alpha'$ units), since we require asymptotic freedom; nevertheless, we are not going to use this assumption. Since pure YM theory has no space-time fermions, while presumably the string theory consistency requires world-sheet supersymmetry [@wall], the dual string theory could be similar to the ten dimensional Type 0. The latter is known to contain a closed string tachyon $T$ and his spectrum of RR fields is doubled: there are two sets of Dp-branes, named “electric” and “magnetic”, for every p [@kt1]. It is not clear if the doubling of the RR spectrum persists in 5d [@armoni; @ferretti2; @Thorn:2008ay]. If the doubling survives in 5d, we have to see the background above as generated by one type of D3, say the “electric” ones. Finally, the tachyon $T$ could be possibly non-trivial [@armoni; @ghor1]. On this background, the glueballs are described by the perturbative string states, while the D-brane spectrum should be matched with other field theory degrees of freedom, as we will try and do in the following sections. In order to describe QCD, we must include fundamental fermions. Following the proposal in [@km] for the six dimensional dual of SQCD, it was proposed in [@poli] that in QCD the $N_f$ fundamental and anti-fundamental quarks are described by $N_f$ pairs of space-time filling D4/anti D4-branes. If the theory is some kind of Type 0, in the spectrum of the strings connecting “electric” D3 and “magnetic” D4-branes there should be only fundamental fermions [@kt1; @kt3; @costa]. Instead, in the spectrum of strings with both end-points on the D4-branes (the mesons) there should be only bosons (that is, there should not be “fermionic mesons” as in the usual 10d models). The two groups of branes/anti-branes account naturally for the flavor symmetry $U(N_f)\times U(N_f)$ prior to chiral symmetry breaking. The latter is triggered by the quark condensate, that is a vev for the scalar field in the bifundamental of the two groups (the pion). In the string description the chiral condensate is nothing else that a non-trivial profile for the open string tachyon $\tau$ in the brane/anti-brane system [@sugimoto; @KarKatz].[^4] Its presence in fact breaks the flavor group to a diagonal subgroup, $U(N_f)\times U(N_f) \rightarrow SU(N_{f})_D \times U(1)_B \times U(1)_A$, where $U(1)_B$ is of course the baryonic $U(1)$ and the axial $U(1)_A$ is anomalous [@ckp]. The scalar and vector meson spectrum is naturally accounted for by the open string tachyon $\tau$ and the gauge fields on the D4/anti D4-brane world-volume [@poli]. The complex tachyon gives a full tower of scalar and pseudo-scalar mesons, comprising the Goldstone pions, while the world-volume gauge fields account for a full tower of vector and axial-vector mesons. This was shown in [@ckp], where it was pointed out that this picture is consistent provided that the open string tachyon $\tau$ is diverging at the bottom of the space $r=0$, and that this condition can give “linear confinement” for highly excited mesons. Finally, the quark mass is introduced by turning on the non-normalizable mode of $\tau$ at large radius. This part of the QCD spectrum is reported in Table \[qcdmesons\]. $J^{PC}$: $0^{++}$ $0^{-+}$ $1^{--}$ $1^{++}$ ----------- ------------------ --------------------- --------------------- --------------------- DBI mode: $\tau+\tau^\dag$ $i(\tau-\tau^\dag)$ $\frac12 (A_L+A_R)$ $\frac12 (A_L-A_R)$ : Matching of the scalar and vector QCD mesons with D4/anti D4 world-volume modes. \[qcdmesons\] In the table, $A_L, A_R$ are the gauge fields living on the world-volume of the $N_f$ flavor branes and anti-branes respectively (since the branes are space-time filling, there are no scalar fields on the world-volume). The pseudo-scalars $0^{-+}$ are of course the pions. We are not going to discuss the flavor sector anymore in these notes, even if we lack some basic ingredients (for example, the dual to the mixed quark-gluon condensate). YM spectrum matching ==================== Having reported the recent results on the meson spectrum in the previous section, here we analyze the pure glue $SU(N_c)$ YM theory. Besides the assumptions on the form of the YM dual formulated above, here we will add the following: first, even if gravity is not a good approximation to the theory, and so the [usual]{} gravity modes are not much lighter than the other string states, we will reasonably assume that they are still among the lighter modes in the true string spectrum; second, we will assume that if the RR spectrum is doubled, one of the two RR fields of each type is much heavier than the other. Concerning the latter assumption, we will encounter no indication for a doubling of the RR (and brane) spectrum. Moreover, even if the spectrum is doubled, the properties of the “electric” and “magnetic” objects would be different, since the background is generated by “electric” branes only. Thus, we would expect anyway a non-degenerate mass spectrum. On the other hand, we notice that at least some of the 10d gravity duals of YM-like theories, such as the ones in [@wym; @min2], have a “doubled” brane spectrum in 5d. In fact, considering also the branes wrapped on the internal cycles, they have two kinds of branes of the same dimensionality in 5d. For example, in the Type IIA model of [@wym], besides the ordinary D0-branes, one has D4-branes wrapped on the internal $S^4$, that are “D0-like” in five dimensions; analogously, besides the D2-branes, one has wrapped D6, and so on. Note that these branes of the same dimension in 5d have different physical properties, which would be the case also in the present setup because of their “electric” or “magnetic” nature. It would be interesting to understand whether this apparently frequent “doubling” of the brane spectrum in the duals of YM-like theories has some real implication for the actual YM theory. Anyway, in the following we will not consider the possibility of this “doubling”. We will begin by the matching of the gravity spectrum and then come to the matching of the non-perturbative string spectrum of branes. We will adapt ten dimensional techniques to the five dimensional case. This naive procedure provides a somewhat consistent picture. Glueballs vs. gravity spectrum ------------------------------ The perturbative string spectrum should be matched with the glueball spectrum of YM. As usual, almost every glueball is dual to a genuinely string state. If nevertheless we insist on the assumption that the usual gravity modes are among the lighter ones in the spectrum, we should be able to match their spectrum with the lighter glueballs. The latter are known from lattice simulations. Since we do not have any computable string model at hand, we do not compare masses,[^5] but just search for the qualitative matching of the gravity and lattice modes. ### Lattice data We use the data on the $SU(8)$ theory (the nearest to the large $N_c$ limit), that can be found in [@mt; @meyer]. We report the relevant data for the continuum spectrum in Table \[glueballs\]. $J^{PC}$: $0^{++}$ $2^{++}$ $0^{++}$ $0^{-+}$ $0^{++}$ $2^{-+}$ $1^{+-}$ -------------------- ---------- ---------- ----------- ---------- ------------ ---------- ---------- $m/\sqrt{\sigma}$: 3.32(15) 4.65(19) 4.71(29) 4.72(32) $\sim $ 5.67(40) 5.70(29) : Continuum $SU(8)$ glueball spectrum. \[glueballs\] In the table, the masses are normalized with respect to the string tension $\sigma$. The reported masses are the average ones. We include only the lightest modes up to the first glueball that is surely dual to a string mode, namely the $2^{-+}$: the only spin two field in the gravity spectrum is the graviton, dual to the $2^{++}$ glueball. As an exception, we also report the $1^{+-}$ glueball, since the errors are such that it could be lighter than the $2^{-+}$. In any case, since the sting theory is at strong curvature, the string modes will not be much heavier than the gravity ones, consistently with the actual glueball spectrum. Of course, it can well be that some stringy modes are lighter than some of the usual gravity modes. In fact, this is what most probably happens. We are nevertheless interested in matching only the gravity modes with the appropriate glueball states. The first glueball we do not report is the $2^{++}$ excitation, with mass around $6.5$. We instead include the $0^{++*}$, since it appears in the table 7.15 of [@meyer], even if it is not reported in the continuum spectrum. Its mass is slightly above the $0^{-+}$ one. As a general remark, the $SU(3)$ spectrum is very similar to the $SU(8)$ one, but for the masses of the $0^{++}$ and $0^{++}$, which are much heavier. In particular, the first one is heavier than the $1^{+-}$. As such, these $0^{++}$ states could be peculiar to the large $N_c$ limit, that is, they could be the first excitations of k-strings [@meyer]; their masses fit well this explanation. We will discuss this point further in the following. ### Gravity modes {#secgravity} In 5d the gravity modes have been argued in [@kt1; @ferretti2] to include: the graviton $g$ (with five physical degrees of freedom), the dilaton $\Phi$, the NSNS two-form potential $B$ which is dual to a one-form potential $A_1^{NS}$ (with three physical degrees of freedom), and possibly a real closed string tachyon $T$ from the NSNS sector; the axion $\chi$, a one-form potential $A_1^{R}$ dual to a two-form potential $C_2$, and some fields with no propagating degrees of freedom (as it is evident considering the duals) as the four-form potential $C_4$ and the five-form potential $C_5$, in the RR sector. As said, we make the assumption that if the RR sector is doubled, one of the two fields of any kind is much heavier that the other, so we will not consider it. The $P, C$ quantum numbers of the dual glueballs of these gravity fields can be inferred from the way the latter are coupled to the YM field in the DBI and WZ action for a probe brane [@ooguri; @brower]. In our case the probe brane is of course a D3, so the terms in the WZ action that contain the YM field $F$ are $$S_{WZ} \sim \int d^4x\, (\chi \ F \wedge F + \chi \ F \wedge B + C_2 \wedge F )\ .$$ Under a parity transformation the YM gauge field transforms as: $A_i\rightarrow -A_i$ for $i=1,2,3$ and $A_t\rightarrow A_t$. Thus, every $F_{\mu\nu}$ field with both indexes in $x_{1,2,3}$ is parity even, while $F_{it}$ is parity odd. Charge conjugation acts as: $A_\mu \rightarrow -A_\mu^T$ (in matrix notation). Since for more than one $F_{\mu\nu}$ this implies that the order of the fields is reversed, it follows that symmetric products of an odd number of fields will have $C=-1$ (antisymmetric products of an odd number of fields will have $C=1$, but these do not enter the game). On the other hand, products of an even number of fields will have $C=1$. From this, the parity assignment of the fields follow. Let us first consider the field $B$. From the coupling $\chi F \wedge B$, since $F_{it}$ is parity odd and $\chi$ is odd too (see below), it follows that $B$ is even (here we use the gauge where the $t$ components of the gravity fields are zero). Furthermore, from the expansion of the square root in the standard DBI action, one reads that the coupling of $B$ with $F$ is of the form $B_{\mu\nu}\cdot Sym[F_{\mu\nu}W]$, where $Sym$ stands for “symmetric product” and $W$ contains an even number of $F$ fields (it is present only in the non-abelian case). Thus, one immediately reads that $B$ is parity even (note that here it has the same indexes as $F$) and that it is charge conjugation odd, since there is an odd number of $F$s. Analogously, from the WZ non-abelian action it follows that the coupling of the two-form $C_2$ is $\epsilon^{ijkt}C_{ij}\cdot Sym[F_{kt}W]$, so that $C_2$ is parity odd and charge conjugation odd. The axion is instead coupled like $\epsilon^{ijkt}\chi\cdot Sym[F_{ij}F_{kt}W]$ so it is P-odd and C-even. The graviton, that couples to the energy-momentum tensor[^6] as $g_{ij}\cdot T_{ij}= g_{ij}\cdot Tr(F_{ik}F_{j}^k)$, is P-even and C-even, as the dilaton, that couples as $\Phi\cdot TrF^2$. The possible closed string tachyon $T$ would presumably mix with the dilaton and again would be P and C-even.[^7] These results are summarized in Table \[gravity modes\]. Field: $\Phi$ $g$ $T$ $\chi$ $B$ $C_2$ ------------- ---------- ---------- ---------- ---------- ---------- ---------- $J^{PC}$: $0^{++}$ $2^{++}$ $0^{++}$ $0^{-+}$ $1^{+-}$ $1^{--}$ Components: 1 5 1 1 3 3 : Gravity spectrum in 5d. \[gravity modes\] ### (Partial) Matching By comparison of Table \[glueballs\] and Table \[gravity modes\], some identifications of glueballs with gravity modes are straightforward. As usual, the graviton is obviously the $2^{++}$, five-component object dual to a massive spin two glueball with the same quantum numbers. Analogously, the identification of the axion with the $0^{-+}$ is straightforward. Moreover, using the duals of the two-forms above, the three component $A_1^{NS}$ and $A_1^{R}$ one-forms should be dual to the massive vectorial $1^{+-}$ and $1^{--}$ states. Clearly, the $A_1^{R}$ one-form is peculiar. The point is that the $1^{--}$ glueball is much heavier than the ones reported in Table \[glueballs\] (its mass is around 7.5), so there are string modes lighter than this gravity mode. In fact, this is not really a serious problem, as we know that the string modes are not separated from the gravity ones, and this mixing in the spectra is expected. For example, the $2^{-+}$ glueball is clearly a string mode. Something interesting happens in the scalar sector. The dilaton[^8] and tachyon have the quantum numbers of the $0^{++}$ glueballs. Nevertheless, in the light glueball spectrum there are three $0^{++}$ states and we do not have any gravity candidate for the dual to $0^{++}$. One possibility is that there is another scalar in the 5d gravity spectrum that we are overlooking. Or, it could be that the third $0^{++}$ glueball is accounted for by a higher mode of the dilaton or the tachyon. We believe this to be unlikely, for the following reason. The mass ratios of the $0^{++}$ and $0^{++}$ to the $0^{++}$ for $SU(8)$ on the lattice are about $1.42$ and $1.42-1.60$ (the uncertainty here comes from the lacking of the continuum mass value for the $0^{++}$). In the $SU(3)$ case they are instead $1.87$ and $2.29$. As a comparison, the $0^{-+}$ to $0^{++}$ ratios in the two cases ($SU(8)$ and $SU(3)$) are $1.42$ and $1.53$; the $2^{++}$ to the $0^{++}$ ones are $1.40$ and $1.46$; the $1^{+-}$ to the $0^{++}$ ones are $1.72$ and $1.81$; the differences in the masses between these $SU(8)$ and $SU(3)$ glueballs are about $10\%$ or less. As such, it is not very likely that the $0^{++}$ in $SU(8)$ could be the same state as the $0^{++}$ in $SU(3)$, as the change in the mass ratio is from $1.42-1.60$ to $1.87$, an increase of at least the $17\%$ or more. Analogously, it is also unlikely that the $0^{++}$ in $SU(8)$ could be the same state as the $0^{++}$ in $SU(3)$: since the mass of the $0^{++}$ in $SU(3)$ is 1.8 times that of the $0^{++}$, we expect the corresponding state in $SU(8)$ to have a mass of order 6.2, well above the $0^{++}$, $0^{++*}$ states measured on the lattice for $SU(8)$. Thus, as observed in [@meyer], it seems that [both]{} the $0^{++}$ and $0^{++*}$ are more likely to be k-string excitations rather than ordinary gravity modes. As such, we are led to conclude that there is no other scalar beyond the dilaton in the gravity sector, dual to light glueballs. That is, there should be no “light” tachyon. To be more precise, the statement is that even if a tachyon exists in the spectrum of the string theory, out of the two scalars that come out from its possible mixing with the dilaton (and the metric, see footnote \[footnotedil\]), one will be dual to a much heavier glueball than the other. We will continue to call “dilaton” the “lighter” of these two modes.[^9] For similar reasons, we conclude that the RR fields which are the possible Type 0 doubles of the ones considered here, should be dual to much heavier glueballs, since there is no sign of any doubling in the lattice spectrum. We put the proposed identification of the glueball spectrum in Table \[identification\],[^10] together with the dimension $\Delta$ of some of the dual operators, as in [@ooguri; @brower]. $J^{PC}\qquad $ Field Operator $\Delta$ ------------------- ---------------------------- -------------------------------------------------------------------------- ---------- $0^{++}\qquad $ dilaton $\ \Phi$ $\qquad Tr F^2\qquad $ 4 $2^{++}\qquad $ metric $\ g$ $\qquad T_{ij}-trace\qquad $ 4 $0^{++}\qquad $ k-string mode $0^{-+}\qquad $ RR axion $\ \chi$ $\qquad Tr F\wedge F\qquad $ 4 $0^{++*}\qquad $ k-string mode $2^{-+}\qquad $ string mode $1^{+-}\qquad $ NSNS one-form $\ A_1^{NS}$ $\qquad d_{abc}F^{a}_{\mu\alpha}F^{b\alpha\beta}F^{c}_{\beta\nu}\qquad $ 6 $1^{--}\qquad $ RR one-form $\ A_1^{R}$ $\qquad d_{abc}F^{a}_{\mu\alpha}F^{b\alpha\beta}F^{c}_{\beta\nu}\qquad $ 6 : Identification of the light glueball modes with string modes. \[identification\] In the table, $T_{\mu\nu}$ is the dimension four stress-energy tensor of YM and $d_{abc}$ is a symmetric coupling. Non-perturbative string spectrum -------------------------------- Apart from the glueballs, the YM spectrum contains other physical degrees of freedom. Also, in string theory there are non perturbative states. In this section we try and match the two spectra as much as possible. We first consider the string theory and then argue to what YM degrees of freedom the various states could correspond to. ### Branes In string theory there are a collection of D-branes and a solitonic NS-brane. As usual, p-branes couple to (p+1)-form potentials. Thus, looking at the spectrum in section \[secgravity\], and considering the dual forms too, we can write down the spectrum of branes in the theory, as in Table \[branes\], together with their proposed YM interpretation, discussed in the following subsections. Brane Field Sector $\qquad $ YM interpretation ------- -------------------------- -------- ------------------------------ D(-1) $\qquad \chi\qquad $ RR $\qquad $ instanton D0 $\qquad A_1^{R}\qquad $ RR $\qquad $ baryon vertex NS0 $\qquad A_1^{NS}\qquad $ NSNS $\qquad $ monopole vertex F1 $\qquad B\qquad $ NSNS $\qquad $chromoelectric flux D1 $\qquad C_2\qquad $ RR $\qquad $chromomagnetic flux D2 $\qquad C_3\qquad $ RR $\qquad $ domain wall D3 $\qquad C_4\qquad $ RR $\qquad $ colors D4 $\qquad C_5\qquad $ RR $\qquad $flavors : Branes in 5d and their proposed YM interpretation. \[branes\] Actually, the identification of the D3 and D4-branes in the dual filed theory is trivially in terms of color and flavor degrees of freedom support, respectively, as stated in section \[secsetup\]. For what concerns the other branes, although the story is similar to the known 10d duals (we will mostly follow the analysis of [@ooguri1; @ps]), their identification brings to interesting implications. ### D(-1)-brane: the small instanton This is well known from the 10d duals [@banks]. The D(-1)-brane looks like an instanton on the world-volume of the D3-branes. It couples to the axion $\chi$, whose WZ coupling with the D3-brane is just $$\int d^4x\ \chi\ F\wedge F\ .$$ Thus, the axion provides the YM $\theta_{YM}$ angle, $\chi(r=\infty)=\theta_{YM}$ in the microscopic lagrangian. We will assume for simplicity that the $\theta_{YM}$ angle is zero in the following. ### D0-brane: the baryon vertex {#d0} The D0-brane is just a particle and couples to the one-form potential $A_1^{R}$. Fundamental strings (F1) can in principle end on the D0-brane. Thus, the D0 is a source for the F1 flux, which is naturally associated with the chromoelectric flux. We propose to interpret the D0-brane as the baryon vertex.[^11] The open strings coming from the flavor D4-branes would end on it, creating a bound state. In 10d models the baryon vertex is always a Dp-brane wrapped on an internal compact manifold of dimension p. Since in the background there is typically a (p-1)-form potential $C_{p-1}\sim N_c$, on the world-volume there is a non-trivial electric field of magnitude $N_c$, which is inconsistent on a compact space unless there are $N_c$ electric sources: the open strings ending on the brane [@witten; @ps]. Crucially, there is also a simple picture of this mechanism from the non-compact five dimensional point of view [@ooguri1]. It relies on the observation that the supergravity action contains a CS term that can be reduced to 5d once we integrate on the compact space. The latter operation gives a factor of $N_c$, so the five dimensional CS coupling contains a factor $$\label{cs5d} S_{CS} \sim N_c \int d^5x\ \epsilon_{ab}\ B^a \wedge dB^b\ ,$$ where $B^a$ is a doublet of (NSNS,RR) two-forms. But the NSNS two-form $B$ is coupled to the fundamental strings, so a generalization of the standard 3d CS picture implies that one can join $N_c$ fundamental strings together. In fact, this configuration is gauge equivalent to a zero NS charge configuration, thus overcoming the obstruction to join fundamental strings because of the NS charge conservation. The same mechanism should hold in the 5d dual to QCD. The gravity fields comprise both the NSNS and the RR two-forms, and the background includes a non-trivial zero-form field strength $F_0$ dual to $F_5$, so proportional to $N_c$ (this was assumed in the setup in section \[secsetup\]). Thus, it is natural to have a CS term for these fields of the form (\[cs5d\]), where the $N_c$ factor would be given by $F_0$. If such a CS coupling exists, the same argument as above would imply that one can form a baryon vertex. Moreover, the D0-branes are point-like instantons on the five dimensional world-volume of the flavor D4-branes. It has recently been discussed, in the context of the D4/D8 model [@sstutti1; @sstutti2; @sstutti3; @sstutti4; @sstutti5], how instantonic configurations in five dimensions reduce to the Skyrme term in four dimensions.[^12] This strengthens the identification of the D0 as the baryon vertex. Unfortunately, our ignorance of the 5d actions prevents a more detailed discussion. ### NS0-brane: the magnetic monopole vertex The one-form potential $A_1^{NS}$ is a magnetic field with respect to the fundamental string. It couples to a NS0-brane, which is just a particle. The D1-brane can in principle end on the NS0-brane. Thus, the NS0 is a source for the D1 flux, which is naturally associated with the chromomagnetic flux. All these considerations point to the identification of the NS0-brane with a magnetic monopole vertex. As it is well known, a possible mechanism of confinement in the YM theory is via monopole condensation. The latter has the consequence that the monopoles feel no potential: the magnetic charge is screened. Thus, the ’t Hooft loop of the chromomagnetic flux must not give the area law. This is discuss in section \[secD1\]. The NS0-brane has been dubbed “magnetic baryon” in [@gk], along the lines of [@ps], as it would be the S-dual of the D0 brane. In fact, since the argument in section \[d0\] is symmetric in the RR and NSNS 2-forms, it would seem that the D1-branes can be attached to the NS0 only in groups of $N_c$ (this is the reason why we dubbed the NS0 “vertex”). Nevertheless, as discussed in section \[secD1\], the monopole screening implies that the D1 are tensionless at the bottom of the space. As such, they are not forced to end on the NS0, which is therefore completely shielded. It would be interesting to understand better the physics of this object. ### F1: the chromoelectric flux This identification is obvious. The fundamental string is just the YM flux tube. The fact that it has a finite tension just follows from the requirement that the warp factor in the geometry (\[metric\]) has a non-vanishing value at the bottom of the space (as we stated in section \[secsetup\]), since then the Wilson loop gives automatically the area law [@wym; @kinar]. In fact, the action for a fundamental string describing a Wilson loop is $$\label{wl} S \sim \int d^2x\, \sqrt{-G}\ ,$$ where $G$ is the determinant of the induced metric on the world-volume. This action is minimized at $r=0$, where $\sqrt{-G} \sim e^{2f(0)}\neq 0$, giving thus a finite tension $\sigma\sim e^{2f(0)}$ and a linear potential in the limit of large separation $L$ of the two test quarks, $V=\sigma L$. We also recall that the Luscher term comes out trivially for this configuration. In fact, four out of the five string world-sheet bosons, representing the fluctuations around the classical configuration, are massless modes to leading order (the boson representing the radius fluctuation has instead a potential proportional to $e^{2f}$ and the world-sheet fermions most probably take mass by the coupling to the RR five-form as it happens in the 10d models). Thus, the first correction to the string energy, given by the sum of the zero-point energies of the world-sheet massless modes and of the ghosts, is precisely $-\frac{(n-2)}{24\pi L}=-\frac{1}{12\pi L}$ ($n$ being the number of massless modes). In $SU(N_c)$ YM the k-strings take values in $Z_{N_c}$ (= center of the group), so one expects $N_c/2$ different strings. As in [@ps], the presence of these k-strings is deduced from the presence of the baryon vertex: the k-string is just a bound state of k ordinary strings; if we put $N_c$ of them together, they can pair produce baryon vertexes; then, the flux tubes can annihilate, showing that a $N_c$-string is equivalent to no string at all, and thus that the strings take charge in $Z_{N_c}$. Note that the flux tube can break if there are dynamical quarks. This is trivial in our case: the D4/anti D4-branes are space-time filling, so the F1 live on them and can in principle break with pair production of quarks. But since there is a chiral condensate and the open string tachyon has a diverging profile in the IR [@ckp], the flavor branes should “effectively” end at finite distance from the bottom of the space, so the electric fluxes are metastable: sitting at $r=0$ they can break only if, via world-sheet instanton transitions, they reach the effective bottom of the D4-branes. Of course, the same is true in the massive quark case, where the radial position of the effective bottom of the branes is a function of the quark mass $m_q$, so the flux tube breaking effect should be exponentially suppressed with $m_q$. ### D1: the chromomagnetic flux {#secD1} This is an obvious identification too. The important point is that the magnetic flux should not be confined. This should be reflected in the vanishing of the tension of the D1-brane at the bottom of the space. Thus, the energy density would be zero: the flux would be dissolved, the D1-brane would “percolate” and basically fill up the space. As it is well known, the tension of the chromomagnetic flux can be derived from the ’t Hooft loop. In 5d, its only difference with respect to the Wilson loop (\[wl\]) is that the D1 DBI action has a dilaton factor multiplying the area spanned by the string, $$S \sim \int d^2x\ e^{-\Phi} \sqrt{-G}\ .$$ The induced metric on the world-volume is the same as the one for the Wilson loop, which gave a finite tension because $\sqrt{-G} \neq 0$. We conclude that, in order to have magnetic screening, that is a zero tension for the magnetic flux given by the D1-brane, it is necessary and sufficient that the dilaton blows up at the bottom of the space $r=0$, $$e^{\Phi} \rightarrow \infty \qquad {\rm for} \qquad r \rightarrow 0\ .$$ Now, in this five dimensional D3 model, the YM coupling must be proportional to the dilaton exponential.[^13] Thus, we are led to conclude that the YM coupling must diverge in the far IR. This feature is also present in some 10d models (for example in [@wym; @min2]). Thus, it seems that string theory has a tendency of predicting a diverging YM coupling in the IR, rather than a flattening to a constant value. Let us close this section with a comment on oblique confinement. At non trivial $\theta_{YM}$, in YM all the charges are confined, apart from the specific cases when $\theta_{YM}=\frac pq$ (with coprime $p,q$), where $(p,q)$-dyons of electric (magnetic) charge $p$ ($q$) condense. It is well known that in ten dimensional Type IIB theory the tension of a $(p,q)$-string, that is a bound state of $p$ F1s and $q$ D1s, obeys the formula $$\label{oblique} \sigma_{(p,q)}=\sigma\, e^{-\Phi} \sqrt{e^{2\Phi}(p-\chi q)^2+q^2}\ ,$$ where $\sigma$ is the fundamental string tension and $\chi$ is the axion [@schwarz2]. This formula would give oblique confinement provided the axion is $\chi=p/q=\theta_{YM}$ at $r=0$ (i.e. it does not change along he RG flow, as indicated by the lattice studies in [@Giusti:2001xh; @Luscher:2004fu]). In fact, for generic $\theta_{YM}= \chi$, in the limit $e^{-\Phi}\rightarrow 0$ for $r \rightarrow 0$, formula (\[oblique\]) gives $\sigma_{(p,q)} = \sigma\, \|p-\chi q\| \neq 0$, that is, all the charges are confined. Instead, for $\theta_{YM}= \chi = p/q$, exactly the $(p,q)$-strings have vanishing tension, that is, the $(p,q)$-dyons are screened. Of course, in five dimensions formula (\[oblique\]) cannot be the complete story, since the string interactions cannot be discarded. At best, the formula can approximate the actual behavior for $p,q \ll N_c$. With a larger amount of strings, the expected behavior should give the “sine” or “Casimir” scaling of the tension of a $p$-string. Nevertheless, we expect a similar structure to (\[oblique\]), with maybe more general functions of $(p-\chi q)$ (for example, a sine), but with the same pattern of dependence on the dilaton, such that the nice behavior described above would not be spoiled. ### D2: the domain wall The properties of the domain walls, that is that the F1s can end on them and that the baryon vertex can dissolve on their world-volume, are trivially satisfied by the D2-branes (the dissolution of the D0 baryon vertex should be viewed by the WZ coupling $\int A_1^{R}\wedge F_2$, that gives a D0 charge when the gauge field $F_2$ is non-trivial on the world-volume of the D2). In fact, the D0/D2 system realizes in a natural way the scenario, advocated in [@gs],[^14] where the domain walls are “made up” of soliton-like excitations that resemble baryon vertexes (in the sense that $N_c$ fundamental strings can attach to them and make them escape from the domain wall).[^15] Finally, as for the D1-brane before, a crucial point is the possible vanishing of the D2-brane tension at the bottom of the space, where the branes are dragged by the requirement of minimizing their energy in the background (\[metric\]). Instead, in pure YM at $N_c=\infty$ there are infinitely many stable vacua [@Witten:1998uka]. Even if they are non-degenerate in energy and there exists only one true vacuum (for $\theta_{YM}=0$), all the unstable vacua become stable at infinite $N_c$. Thus, we expected finite tension domain walls separating these vacua in the model. It is not clear how this scenario could be realized, but a possibility is that some field is non-trivial on the world-volume of these D-branes, making their tension finite. For example, in the case the domain walls are really “made up” of D0s, a non-trivial $A_1^{R}$ would couple to a magnetic field on the D2, possibly giving it a finite tension, along the lines of [@Schmidhuber:1996fy]. Acknowledgments {#acknowledgments .unnumbered} =============== We would like to thank A. Andrianov, F. Bigazzi, R. Casero, A. Celi, R. Enparan, U. Gursoy, E. Kiritsis, L. Martucci, C. Nunez, H. Ooguri, A. Paredes and L. Tagliacozzo for useful discussions. We also thank the referee of IJMPA for pointing out some problems in section \[secgravity\] in a previous version of these notes. This work is partially supported by the European Commission contracts MRTN-CT-2004-005104, MRTN-CT-2004-503369, CYT-FPA-2004-04582-C02-01, CIRIT-GC-2001SGR-00065. [0]{} G. ’t Hooft, [Nucl. Phys. B* ]{}[72]{}, 461 (1974). G. Veneziano, [Nucl. Phys. B ]{}[117]{}, 519 (1976). A. M. Polyakov, [Phys. Lett. B ]{}[103]{}, 207 (1981). A. M. Polyakov, [Int. J. Mod. Phys. A]{} [14]{}, 645 (1999) \[arXiv:hep-th/9809057\]. J. Erlich, E. Katz, D. T. Son and M. A. Stephanov, [Phys. Rev. Lett. ]{} [95]{}, 261602 (2005) \[arXiv:hep-ph/0501128\]. L. Da Rold and A. Pomarol, [Nucl. Phys. B ]{}[721]{}, 79 (2005) \[arXiv:hep-ph/0501218\]. J. M. Maldacena, [Adv. Theor. Math. Phys. ]{}[2]{}, 231 (1998) [\[Int. J. Theor. Phys. ]{}[38]{}, 1113 (1999)\] \[arXiv:hep-th/9711200\]. S. S. Gubser, I. R. Klebanov and A. M. Polyakov, [Phys. Lett. B ]{}[428]{}, 105 (1998) \[arXiv:hep-th/9802109\]. E. Witten, [Adv. Theor. Math. Phys. ]{}[2]{}, 253 (1998) \[arXiv:hep-th/9802150\]. O. Aharony, S. S. Gubser, J. M. Maldacena, H. Ooguri and Y. Oz, [Phys. Rept. ]{} [323]{}, 183 (2000) \[arXiv:hep-th/9905111\]. F. Bigazzi, R. Casero, A. L. Cotrone, E. Kiritsis and A. Paredes, [JHEP ]{}[0510]{}, 012 (2005) \[arXiv:hep-th/0505140\]. R. Casero, E. Kiritsis and A. Paredes, [Nucl. Phys. B ]{}[787]{}, 98 (2007) \[arXiv:hep-th/0702155\]. C. B. Thorn, arXiv:0808.0458 \[hep-th\]. U. Gursoy and E. Kiritsis, [JHEP]{} [0802]{}, 032 (2008) \[arXiv:0707.1324 \[hep-th\]\]. U. Gursoy, E. Kiritsis and F. Nitti, [JHEP ]{}[0802]{}, 019 (2008) \[arXiv:0707.1349 \[hep-th\]\]. E. Witten, [Adv. Theor. Math. Phys. ]{} [2]{}, 505 (1998) \[arXiv:hep-th/9803131\]. Y. Kinar, E. Schreiber and J. Sonnenschein, [Nucl. Phys. B ]{}[566]{}, 103 (2000) \[arXiv:hep-th/9811192\]. I. R. Klebanov and A. A. Tseytlin, [Nucl. Phys. B ]{}[546]{}, 155 (1999) \[arXiv:hep-th/9811035\]. A. Armoni, E. Fuchs and J. Sonnenschein, [JHEP]{} [9906]{} (1999) 027, \[arXiv:hep-th/9903090\]. G. Ferretti, J. Kalkkinen and D. Martelli, [Nucl. Phys. B]{} [555]{}, 135 (1999) \[arXiv:hep-th/9904013\]. K. Ghoroku, [J. Phys. G ]{}[26]{}, 233 (2000) \[arXiv:hep-th/9907143\]. I. R. Klebanov and J. M. Maldacena, [Int. J. Mod. Phys. A]{} [19]{}, 5003 (2004) \[arXiv:hep-th/0409133\]. I. R. Klebanov and A. A. Tseytlin, [JHEP ]{}[9903]{}, 015 (1999) \[arXiv:hep-th/9901101\]. M. S. Costa, [JHEP ]{}[9904]{}, 016 (1999) \[arXiv:hep-th/9903128\]. S. Sugimoto and K. Takahashi, [JHEP ]{}[0404]{}, 051 (2004) \[arXiv:hep-th/0403247\]. A. Karch and E. Katz, [JHEP]{} [0206]{}, 043 (2002) \[arXiv:hep-th/0205236\]. J. H. Schwarz, historical talk at the GGI workshop “String and M theory approaches to particle physics and cosmology”, Florence, June 2007. J. A. Minahan, [JHEP]{} [9904]{}, 007 (1999) \[arXiv:hep-th/9902074\]. R. C. Brower, S. D. Mathur and C. I. Tan, [Nucl. Phys. B ]{}[587]{}, 249 (2000) \[arXiv:hep-th/0003115\]. H. B. Meyer and M. J. Teper, [Phys. Lett. B ]{}[605]{}, 344 (2005) \[arXiv:hep-ph/0409183\]. H. B. Meyer, arXiv:hep-lat/0508002. C. Csaki, H. Ooguri, Y. Oz and J. Terning, [JHEP]{} [9901]{}, 017 (1999) \[arXiv:hep-th/9806021\]. R. L. Jaffe, K. Johnson and Z. Ryzak, [Annals Phys. ]{} [168]{}, 344 (1986). I. R. Klebanov and A. A. Tseytlin, [Nucl. Phys. B]{} [547]{}, 143 (1999) \[arXiv:hep-th/9812089\]. M. R. Garousi, [Nucl. Phys. B]{} [550]{} (1999) 225 \[arXiv:hep-th/9901085\]. I. R. Klebanov, [Phys. Lett. B]{} [466]{}, 166 (1999) \[arXiv:hep-th/9906220\]. R. Grena, S. Lelli, M. Maggiore and A. Rissone, [JHEP]{} [0007]{}, 005 (2000) \[arXiv:hep-th/0005213\]. C. Csaki and M. Reece, [JHEP ]{}[0705]{}, 062 (2007) \[arXiv:hep-ph/0608266\]. E. Kiritsis and F. Nitti, [Nucl. Phys. B]{} [772]{}, 67 (2007) \[arXiv:hep-th/0611344\]. D. J. Gross and H. Ooguri, [Phys. Rev. D]{} [58]{}, 106002 (1998) \[arXiv:hep-th/9805129\]. J. Polchinski and M. J. Strassler, arXiv:hep-th/0003136. T. Banks and M. B. Green, [JHEP ]{}[9805]{}, 002 (1998) \[arXiv:hep-th/9804170\]. E. Witten, [JHEP ]{}[9807]{}, 006 (1998) \[arXiv:hep-th/9805112\]. T. Sakai and S. Sugimoto, [Prog. Theor. Phys. ]{} [113]{}, 843 (2005) \[arXiv:hep-th/0412141\]. K. Nawa, H. Suganuma and T. Kojo, [Phys. Rev. D ]{}[75]{}, 086003 (2007) \[arXiv:hep-th/0612187\]. D. K. Hong, M. Rho, H. U. Yee and P. Yi, [Phys. Rev. D]{} [76]{}, 061901 (2007) \[arXiv:hep-th/0701276\]. H. Hata, T. Sakai, S. Sugimoto and S. Yamato, arXiv:hep-th/0701280. D. K. Hong, M. Rho, H. U. Yee and P. Yi, [JHEP ]{}[0709]{}, 063 (2007) \[arXiv:0705.2632 \[hep-th\]\]. J. H. Schwarz, [Phys. Lett. B ]{}[360]{}, 13 (1995) \[Erratum-ibid. B [364]{}, 252 (1995)\] \[arXiv:hep-th/9508143\]. L. Giusti, G. C. Rossi, M. Testa and G. Veneziano, [Nucl. Phys. B ]{}[628]{}, 234 (2002) \[arXiv:hep-lat/0108009\]. M. Luscher, [Phys. Lett. B ]{}[593]{}, 296 (2004) \[arXiv:hep-th/0404034\]. G. Gabadadze and M. A. Shifman, [Phys. Rev. D ]{}[61]{}, 075014 (2000) \[arXiv:hep-th/9910050\]. E. Witten, [Phys. Rev. Lett. ]{} [81]{}, 2862 (1998) \[arXiv:hep-th/9807109\]. C. Schmidhuber, [Nucl. Phys. B ]{}[467]{}, 146 (1996) \[arXiv:hep-th/9601003\]. [^1]: Nevertheless, solving the 5d gravity equations, although being not a fully consistent procedure, can in principle give important qualitative indications about the sigma model [@wall]. [^2]: As such, it could be worthwhile to include in these models the first subleading corrections coming from the DBI, which possibly would ameliorate the numerical results. [^3]: For obvious reasons, we will not keep track of $2\pi$ and $\alpha'$ factors. [^4]: The idea that the tachyon should be the pion was already present in the early seventies [@schwarz]. [^5]: A precise matching of the masses between lattice and gravity data is not expected even in 10d, being the two computations done in different regimes of parameters [@brower]. [^6]: The energy momentum tensor $T_{\mu\nu}$ is a symmetric and (classically) traceless operator. Under the spatial rotation group $SO(3)$, which is the little group of massive particles, it decomposes as a spin 2 tensor $T_{ij}-trace$, a vector $T_{ti}$ and a scalar $T_{tt}$ (due to the zero trace condition on $T_{\mu\nu}$, the trace of $T_{ij}$ is equal to $T_{tt}$). The vector operator $T_{ti}$ does not create any glueball due to the conservation law for $T_{\mu\nu}$ [@Jaffe:1985qp]. [^7]: In the known ten dimensional models, the tachyon enters the DBI action schematically as $\int k(T) e^{-\Phi}\sqrt{det(g_{\mu\nu}+F_{\mu\nu})}$, where $k(T)\sim 1 \pm \frac14 T +... $ and the two signs correspond to the “electric” and “magnetic” cases [@kt1; @Klebanov:1998yy; @Garousi:1999fu]. In the “dyonic” case, where there are two gauge groups, the tachyon couples to the operator $TrF_1^2-TrF_2^2$ [@Klebanov:1999um], so the quantum numbers are $++$. Unfortunately we do not know the coupling in the non-critical setting. If the tachyon is present in the spectrum and it couples in a way similar to the “electric” ten dimensional case, it would probably mix with the dilaton. Then, one combination would couple to $TrF^2$ (with an abuse of language we are going to refer to it as “dilaton”). Concerning the other combination (which we are going to call “tachyon”), either it does not couple to any local operator and, as advocated in [@Grena:2000xw; @Csaki:2006ji], it is dual to renormalons; or, if it is dual to a local operator, the most probable thing to happen is that the latter gives a P and C-even assignment as in ten dimensions [@Klebanov:1999um]. In any case, we are going to argue that there is no light glueball candidate to be dual to this scalar field. [^8]: What we call dilaton in this section is actually the unique scalar fluctuation diff. invariant, which is a combination of the dilaton and the metric, see for example [@Kiritsis:2006ua].\[footnotedil\] [^9]: As already said, there is also the possibility, advocated in [@Grena:2000xw; @Csaki:2006ji], that the tachyon is dual to renormalons. [^10]: Note that in QCD, besides the $0^{++}$ corresponding to $Tr F^2$, there are two other $0^{++}$ states with corresponding operators of dimension four (the energy density) and six ($f_{abc}F^{a}_{\mu\alpha}F^{b\alpha\beta}F^{c\mu}_{\beta}$), see for example [@brower]. The claim is that these glueballs are heavier than the ones reported in Table \[glueballs\]. [^11]: As pointed out in [@ps], in the 10d model dual to ${\cal N}=1^$ theory, the baryon vertex and the domain wall look like D0 and D2-branes in the non-compact five dimensions. [^12]: From [@ckp] we also learn that the WZ term on the flavor brane should couple to $F_0=\ ^{}F_5\sim N_c$ via $S_{WZ}\sim \int F_0 \cdot \Omega_5$, where $\Omega_5$ gives the five dimensional field theory CS term. [^13]: The presence of the closed string tachyon would make no difference in this respect: the same combination of the dilaton with the tachyon would enter both the D1 tension calculation and the YM coupling, so the latter would still blow up. [^14]: This scenario was ultimately inspired by string theory. [^15]: This was already observed in [@ps].	{ "pile_set_name": "ArXiv" }
--- abstract: 'Bayesian entity resolution merges together multiple, noisy databases and returns the minimal collection of unique individuals represented, together with their true, latent record values. Bayesian methods allow flexible generative models that share power across databases as well as principled quantification of uncertainty for queries of the final, resolved database. However, existing Bayesian methods for entity resolution use Markov monte Carlo method (MCMC) approximations and are too slow to run on modern databases containing millions or billions of records. Instead, we propose applying variational approximations to allow scalable Bayesian inference in these models. We derive a coordinate-ascent approximation for mean-field variational Bayes, qualitatively compare our algorithm to existing methods, note unique challenges for inference that arise from the expected distribution of cluster sizes in entity resolution, and discuss directions for future work in this domain.' author: - \| Tamara Broderick\ University of California, Berkeley\ `[email protected]` - \| Rebecca C. Steorts\ Carnegie Mellon University\ `[email protected]` bibliography: - 'vb\_er.bib' title: Variational Bayes for Merging Noisy Databases --- Introduction ============ Merging records from multiple databases is a problem that emerged from the genetics literature [@newcombe_1959] and is a pressing issue in statistics and computer science in the modern day [@christen_2011]. For instance, human rights organizations collect records of war crimes in the Middle East and Central America and want to estimate the total number of victims [@lum_2013]. The United States Census Bureau wants to estimate minority representation and child poverty in different parts of the country [@datta_2011; @bell_2013]. In each of these examples, individual records are collected in multiple databases. Due to the collection procedure, records are often duplicated within a single database and across databases. Crucially, due to various factors, some records in these databases are corrupted by noise. In any case, an important part of delivering an estimate for any quantity of interest from the merged databases is also returning some uncertainty for that estimate. [@steorts_2014_eb; @steorts_2014_aistats; @steorts_2014_smered; @sadinle_2014] have recently applied a Bayesian statistical paradigm to merging databases by modeling the noisy corruption as a random process; the authors have shown that their approach provides not only desirable uncertainty quantification for a variety of model queries but also flexible generative models to capture the many unique types of records and record relationships that may be present in these databases. However, the MCMC approximations used in these Bayesian analyses do not scale sufficiently to process the large number of records in many modern and complex databases. Thus, we propose a variational Bayes approximation to capture desired uncertainty in posterior Bayesian estimates while simultaneously allowing the processing of much larger and more realistic databases than is possible with these methods. Finally, we elaborate on how these database-merging models pose unique challenges for variational approximations. Background {#sec:background} ========== Entity resolution refers to the merging of multiple databases (often without shared unique identifiers) into a single database of unique entities [@christen_2011]. Special cases of entity resolution include record linkage, which refers to the identification of records across different databases that represent the same entity, and de-duplication, which refers to the identification of records within the same database that represent the same entity. Traditional approaches for entity resolution that link records directly to other records become computationally infeasible as the number of records grows [@christen_2011; @winkler_2006]. Here, we instead take the approach of [@steorts_2014_smered] and imagine each record as representing a latent individual. When we further suppose that some entries may suffer from noisy corruption, entity resolution can be viewed as a clustering problem. The observed data being clustered are the records in each database, and the latent cluster centers are the unobserved, latent individuals. It is common for record fields to be discrete or categorical: e.g., county of residence, race, gender, etc. Like [@steorts_2014_aistats], we focus on categorical data in what follows.[^1] [@blei_2003_latent; @blei_2006_variational] previously demonstrated the scaling advantages of variational Bayesian approximations to posteriors for mixture and admixture modeling, where the observed data are categorically-valued (e.g. words in a vocabulary). There exist unique challenges in our data clustering problem. A particularly important one lies in an assumption inherent in many of the popular Bayesian models for clustering and admixture—such as mixture models, LDA [@blei_2003_latent], Dirichlet processes [@maceachern_1998_estimating], hierarchical Dirichlet processes [@teh_2006_hierarchical], and many more. These models all implicitly assume that any cluster makes up a non-zero proportion of the data that does not change as the data set size increases without bound [@kingman_1978_representation; @broderick_2013_feature]. This assumption, by contrast, is very clearly inappropriate for entity resolution problems. In entity resolution, we expect each cluster to contain perhaps one and at most a handful of records. We expect the number of clusters to grow linearly with the number of records in a given database—though the number of clusters might be constant as more databases are added. We call the problem of modeling this type of clustering behavior, which differs from classical models and assumptions, the small clustering problem, and we discuss it in more detail in . A new generative model for entity resolution {#sec:gen_model} ============================================ Let $D$ represent the number of databases and $R_{d}$ represents the number of records in the $d$th database. All records contain the same $F$ fields, which are categorical. Let field $f$ have $V_{f}$ possible values, or field attributes. Assume for now that every record is complete; that is, for every database and record, there is no missing field. Let $x_{drf}$ be the observed data value in the $f$th field of the $r$th record in the $d$th database. We make the further simplifying assumption that there are $K$ unique latent individuals. Ultimately, we desire a model where $K$ is random, and we learn a posterior distribution over $K$ given the full data $x = \{x_{drf}\}_{d,r,f}$ across all databases, records, and fields. But as a first step we assume $K$ is fixed and known, as in LDA [@blei_2003_latent]. Let $z_{dr}$ be the latent individual for whom (potentially noisy) data $x_{dr\cdot}$ is recorded in the $r$th record of the $d$th database. In other words, we regard each record $x_{dr\cdot}$ as a possibly distorted copy of an ideal latent record for latent individual $z_{dr}$. To capture this idea, let $\beta_{kf\cdot}$ be a discrete noise distribution associated with the $k$th latent individual. That is, $\beta_{kfv},\,v\in\{1,\ldots,V_f\},$ are numbers between zero and one that sum to one across $v$. If there were no noise in the data entry procedure, the probabilities $\beta_{kf\cdot}$ would correspond to a trivial distribution with all of its mass at some true latent value $v^{}_{kf}$ of the $f$th field for the $k$th individual: $\beta_{kfv} = \mbo\{v = v^{}_{kf}\}.\footnote{Here, $\mbo(E)$ is the indicator function for event $E$.}$ In general, there is some noise in the records, and $\beta_{kfv}$ corresponds to a non-trivial noise distribution; however we assume it has a plurality of its mass at the true value. For our generative model, assume that the observed value $x_{drf}$ of the $f$th field in record $r$ in database $d$ is drawn from the noise distribution associated with the latent individual $z_{dr}$ for this record; that is, $$x_{drf} \| \beta_{\cdot f \cdot}, z_{dr} \sim \cat_{V_{f}}(\beta_{z_{dr} f \cdot}),$$ where $\cat_{V_{f}}$ is the categorical distribution over $1,\ldots,V_{f}$ with probabilities given by the distribution parameter. These draws are independent across records and fields, conditional on $\beta$ and $z$. Next, we form a hierarchical Bayesian model by putting priors on both $z$ and $\beta$. For $z$, we assume that the latent individual for any record is drawn uniformly over all latent individuals and independently across records: $$z_{dr} \sim \cat_{K}(K^{-1}\,\bfo_{K}),$$ where $\bfo_{K}$ is the vector of all ones of length $K$. Since $\beta_{kf\cdot}$ is a vector of probabilities, a natural choice of prior for $\beta_{kf\cdot}$ is the Dirichlet distribution on a vector of size $V_{f}$, which we denote by $\dir_{V_{f}}$. Thus, we assume that the $\beta_{kf\cdot}$ vectors are drawn independently according to $$(\beta_{kfv})_{v=1}^{V_{f}} \sim \dir_{V_{f}}(A_{\cdot}),$$ with hyperparameter vector $A_{\cdot} = (A_{1},\ldots,A_{V_{f}})$. Typically, we assume that the $A_v$ are small (near zero) so that the Dirichlet parameter encourages $\beta_{kf\cdot}$ to be peaked around a single value. We typically choose $A_{1} = \cdots = A_{V_{f}}$. Comparison with previous work {#sec:compare} ============================= We briefly review the model of [@steorts_2014_smered], where the authors introduced the basic Bayesian clustering framework for entity resolution and their Split and MErge REcord linkage and De-duplication (SMERED) algorithm. [@steorts_2014_smered] took a fully hierarchical-Bayesian approach, in the special case where all the record fields are categorical and independent. The authors derived an efficient hybrid (Metropolis-within-Gibbs) MCMC algorithm, SMERED. SMERED is able to update most of the latent variables and parameters using Gibbs sampling steps from conjugate conditional distributions. While SMERED updates the assignment of records to latent individuals using a split-merge step, following [@jain_2004], and can run on a health care databases of 60,000 records in 3.5 hours, it does not scale to “large databases."[^2] In terms of scalability, we wish to scale to millions or billions of records in one or multiple databases. For example, the U.S. Census contains approximately 300 million records, while many medical databases at large universities or in the entire country would contain millions or billions of records. Furthermore, while the model of [@steorts_2014_smered] was shown to work very well for entity resolution applications, it is not easily approximated with variational methods due to various deterministic dependencies in the generative model. We show the full model for SMERED in Appendix \[app:notation\], where we also provide a mapping between the SMERED model and our new generative model from . By contrast, we directly demonstrate in how our new generative model, which is inspired by LDA, is readily amenable to variational approximation. While our model has some similarities to LDA, there are also some differences—large and small. For one, the fields do not enjoy the symmetry of words in a bag-of-words model of a document; that is, the fields are ordered and cannot be interchanged. Second, as we do not expect the distribution of individuals to vary wildly by database, we keep the same uniform distribution over latent individuals in each database. By contrast, an important part of LDA is allowing the admixture proportions of topics to vary by document. We now raise a key issue regarding the main distinction between many classical Bayesian models for mixtures and admixtures (such as mixture models, LDA, feature-allocation models [@broderick_2013_feature] including the Indian buffet process [@griffiths_2005_infinite], etc.) and entity resolution. The issue arises from framing entity resolution as a clustering problem. In clustering and other statistical models, it is common to assume that our data are infinitely exchangeable, meaning that for any data set size, we assume that the distribution of our data would not change if the data were observed in a different order. This simple assumption applied to clustering models implies, via the Kingman paintbox [@kingman_1978_representation; @broderick_2013_feature], that every cluster forms some strictly positive proportion of the data, and this proportion does not change as the data grows. In mixture models, these are the mixing proportions; there may be finitely many in a finite mixture model or infinitely many in a Dirichlet process model. In any of these cases, there are two important consequences for our model. First, as the data set size grows, we always observe more data points in a cluster. In fact, the number of observed data in a cluster grows without bound. Second, because the size of every cluster grows to infinity as the data set size grows to infinity, the usual asymptotic theory applies to inferring cluster properties or parameters. Uncertainty about, e.g., a cluster mean typically shrinks to zero in the limit. When clusters are unique individuals in a population, however, it is not natural to assume that more data always eventually means more records of the same individual. Rather, every cluster should be observed a strictly finite number of times. This means that uncertainty about latent individuals cannot shrink to zero (in general). Since the assumptions of the traditional models (such as LDA) are violated in this case, they do not apply. And we must ask: what are natural regularity assumptions in this small clustering domain, what inferences can we draw about clusters in this domain, and what new families of distributions can we apply? A similar issue to what we have dubbed the small clustering problem has previously been identified for infinitely exchangeable graphs by [@lloyd_2012_random]. This problem is also reminiscent of challenges in high-dimensional statistics, where the number of parameters may grow linearly (or much faster) than the data size. Mean-field variational approximation {#sec:mf_approx} ==================================== The generative model specified in yields the following joint distribution for data $x$ and parameters $\beta, z$: $$\label{eq:p_joint} p(\beta, z, x) = \left[ \prod_{k=1}^{K} \prod_{f=1}^{F} \dir_{V_{f}}(\beta_{kf \cdot} \| A_\cdot) \right] \left[ \prod_{d=1}^{D} \prod_{r=1}^{R_{d}} \prod_{f=1}^{F} \beta_{z_{dr} l x_{drf}} \right].$$ Note that the posterior on the parameters, $p(\beta, z\|x)$, is proportional to $p(\beta, z, x)$. As this posterior cannot be solved for in closed form, we must approximate it. Here, we consider a variational approximation $q$ of the following form: $$\label{eq:q_joint} q(\beta, z) = \left[ \prod_{d=1}^{D} \prod_{r=1}^{R_{d}} q(z_{dr} \| \phi_{dr\cdot}) \right] \left[ \prod_{k=1}^{K} \prod_{f=1}^{F} q(\beta_{kf\cdot} \| \lambda_{kf\cdot}) \right],$$ where we have introduced variational parameters $\phi$ and $\lambda$. We further assume $$\label{eq:q_indiv} q(z_{dr} \| \phi_{dr\cdot}) = \cat_{K}(\phi_{dr\cdot}) \quad \textrm{and} \quad q(\beta_{kf\cdot} \| \lambda_{kf\cdot}) = \dir_{V_{f}}(\beta_{kf\cdot} \| \lambda_{kf\cdot}).$$ The variational optimization problem is to minimize the Kullback-Leibler divergence from $q(\beta,z)$ to $p(\beta,z\|x)$: $ \min_{\lambda,\phi} \kl\left(q_{\lambda,\phi}(\beta, z) \vphantom{\textstyle\frac00} \,\middle\\|\, p(\beta,z\|x)\right). $ To clarify that the optimization is over different choices of the distribution $q$, which are indexed by parameters $\lambda$ and $\phi$, we write $q(\beta,z)$ as $q_{\lambda,\phi}(\beta,z)$ above. We derive the following coordinate-ascent steps in the variational parameters $\phi$ and $\lambda$ for the variational optimization problem $\min_{\lambda,\phi} \kl(q_{\lambda,\phi} \\| p)$ in : $$\begin{aligned} \lambda_{kfv} &\leftarrow A_{v} + \sum_{d=1}^{D} \sum_{r=1}^{R_{d}} \phi_{drk} \mbo\{x_{drf} = v\}, \\ \phi_{drk} &\propto_{k} \exp\left\{ \sum_{f=1}^{F} \sum_{v=1}^{V_{f}} \mbo\{x_{drf} = v\} \left[ \psi( \lambda_{kfv} ) - \psi\left( \sum_{u=1}^{V_{f}} \lambda_{kfv} \right) \right] \right\}.\end{aligned}$$ Future directions {#sec:future} ================= Our next step is to compare the algorithm resulting from our new generative model and variational approximation to the existing Bayesian model and resulting MCMC algorithm SMERED of [@steorts_2014_smered]. We anticipate the variational approach will be much faster, however, there may be accuracy tradeoffs. Also, since we have made many simplifying assumptions, we propose incorporation of more realistic assumptions about record fields as in [@sadinle_2014; @steorts_2014_eb]. Moreover, it remains to allow the number of latent individuals to grow with the size of the data, to construct a model that allows posterior inference of this number, and to address the small clustering problem we have posed. We wish to find a solution that addresses this problem not only for entity resolution but more broadly in other domains, where clusters may not be expected to grow without bound as a proportion of the total data. Acknowledgements ================ TB was supported by the Berkeley Fellowship. RCS was supported by NSF grants SES1130706 and DMS1043903 and NIH grant \#1 U24 GM110707-01. Review of SMERED and notational map with new generative model {#app:notation} ============================================================= The independent fields model of [@steorts_2014_smered] assumes the $d$ databases are conditionally independent, given the latent individuals, and that fields are independent within individuals. We use the same notation as the generative model earlier, with $D$ databases, $R_{d}$ records within the $d$th database, and $F$ fields within each record. Then $\bm{x}_{dr}$ is a categorical vector of length $p$. Let $\bm{y}_{k}$ be the latent vector of true field values for the $k$th record, where $k \in \{1,\ldots,K\}$ indexes the latent individuals. The linkage structure is defined as $\bm{\Lambda}=\{\lambda_{dr}\;;\;d=1,\ldots,D\;;\;r=1,\ldots,R_{d}\}$ where $\lambda_{dr}$ is an integer from $1$ to $K$ indicating the latent individual to which the $r$th record in database $d$ refers, i.e., $\bm{x}_{dr}$ is a possibly-distorted measurement of $\bm{y}_{\lambda_{dr}}$. Finally, $\tilde{z}_{drf}$ is $1$ or $0$ according to whether or not the particular field $f$ is distorted in $\bm{x}_{dr}.$ The Bayesian parametric model is $$\begin{aligned} \bm{x}_{drf}\mid\lambda_{dr},\bm{y}_{\lambda_{dr}\ell},z_{drf},\bm{\theta}_f &\stackrel{\text{ind}}{\sim} \begin{cases} \delta_{\bm{y}_{\lambda_{dr} f }}&\text{ if }z_{drf}=0\\ \cat(1,\bm{\theta}_f)&\text{ if }z_{drf}=1 \end{cases}\\ \nonumber \tilde{z}_{drf} &\stackrel{\text{ind}}{\sim}\bern(\tilde{\beta}_f)\\ \bm{y}_{kf}\mid\bm{\theta}_{f} &\stackrel{\text{ind}}{\sim}\cat(1,\bm{\theta}_f)\\ \bm{\theta}_f &\stackrel{\text{ind}}{\sim}\dir(\bm{\mu}_f)\\ \tilde{\beta}_f &\stackrel{\text{ind}}{\sim}\tb(a_f,b_f) \\ \pi({\boldsymbol{\Lambda}}) &\propto 1.\end{aligned}$$ To compare the SMERED generative model to our generative model in , first note that the observed $r$th record in database $d$ is $x_{dr\cdot}$ in both models. The latent individual for the record at $(d,r)$ in SMERED is $\lambda_{dr}$ and in it is $z_{dr}$. In SMERED, the noise distribution for a latent individual is separated into two steps: whether there is noise for a given field ($\tilde{z}_{drf}$) and the distribution of that noise ($\bm{\theta}_{f}$). By contrast, in , $\beta_{kf\cdot}$ captures the full distribution of field values for individual $k$. Also, while SMERED places a distribution $\bm{\theta}_{f}$ on the underlying distribution of field values, such a distribution is implicit in aggregating over $\beta_{kf\cdot}$ in . Likewise, the “true record values” of SMERED’s $\bm{y}_{\lambda_{dr} \cdot}$ are implicit in the distribution $\beta_{z_{dr}f\cdot}$ of . Mean-field variational approximation derivation {#app:deriv} =============================================== Mean-field variational problem ------------------------------ We recall that minimizing the Kullback Leibler divergence (KL) divergence, $$\min_{\lambda,\phi} \kl(q_{\lambda,\phi}(\beta, z) \|\| p(\beta,z\|x)),$$ is equivalent to maximizing $\elbo(\phi, \lambda)$, where $$\begin{aligned} \elbo(\phi, \lambda) &= -\kl(q_{\lambda,\phi}(\beta, z) \|\| p(\beta,z\|x)) + p(x) \\ &= \mbe_{q}[ \log p(\beta,z,x) ] - \mbe_{q}[ \log q(\beta, z; \phi, \lambda) ].\end{aligned}$$ Henceforth, we concentrate on maximizing $\elbo(\phi, \lambda)$ with respect to mean-field approximation parameters $\phi, \lambda$. From the generative model in , we derived the joint distribution of parameters $\beta,z$ and data $x$, $p(\beta,z,x)$, in . We also assume that the approximating distribution $q(\beta,z)$ for the posterior $p(\beta, z \|x)$ takes the form specified in and . Using these equations, we find $$\begin{aligned} \elbo(\phi, \lambda) &= \sum_{k=1}^{K} \sum_{f=1}^{F} \mbe_{q}[ \log \dir_{V_{f}}(\beta_{kf\cdot} \| A) ] \\ & {} + \sum_{d=1}^{D} \sum_{r=1}^{R_{d}} \sum_{f=1}^{F} \sum_{k=1}^{K} \sum_{v=1}^{V_{f}} \mbe_{q}[ \mbo\{z_{dr} = k\} \mbo\{x_{drf} = v\} \log( \beta_{k f v} ) ] \\ & {} - \sum_{d=1}^{D} \sum_{r=1}^{R_{d}} \sum_{k=1}^{K} \mbe_{q}[ \mbo\{z_{dr} = k\} \log ( \phi_{dr k} ) ] \\ & {} - \sum_{k=1}^{K} \sum_{f=1}^{F} \mbe_{q}[ \log \dir_{V_{f}}(\beta_{kf\cdot} \| \lambda_{kf\cdot}) ].\end{aligned}$$ To evaluate these expectations, we recall the definitions of the digamma $\psi$ and trigamma functions $\psi_{1}$: $$\begin{aligned} \psi(x) &= \frac{d }{d x} \log \Gamma(x) \\ \psi_{1}(x) &= \frac{d^{2}}{dx^{2}} \log \Gamma(x) = \frac{d}{dx} \psi(x).\end{aligned}$$ With these functions in hand, we can write $$\begin{aligned} \lefteqn{ \mbe_{q}[ \log \dir_{V_{f}}(\beta_{kf\cdot} \| A) ] } \\ &\quad = \log \Gamma(\sum_{v=1}^{V_{f}} A_{v}) - \sum_{v=1}^{V_{f}} \log \Gamma( A_{v} ) + \sum_{v=1}^{V_{f}} (A_{v} - 1) \left[ \psi( \lambda_{kfv} ) - \psi( \sum_{u=1}^{V_{f}} \lambda_{kfu} ) \right] \\ \lefteqn{ \mbe_{q}[ \mbo\{z_{dr} = k\} \mbo\{x_{drf} = v\} \log( \beta_{k f v} ) ] } \\ &\quad = \phi_{drk} \mbo\{x_{drf} = v\} \left[ \psi( \lambda_{kfv} ) - \psi( \sum_{u=1}^{V_{f}} \lambda_{kfv} ) \right] \\ \lefteqn{ \mbe_{q}[ \mbo\{z_{dr} = k\} \log ( \phi_{dr k} ) ] } \\ &\quad = \phi_{drk} \log ( \phi_{dr k} ) \\ \lefteqn{ \mbe_{q}[ \log \dir_{V_{f}}(\beta_{kf\cdot} \| \lambda_{kf\cdot}) ] } \\ &\quad = \log \Gamma(\sum_{v=1}^{V_{f}} \lambda_{kfv}) - \sum_{v=1}^{V_{f}} \log \Gamma( \lambda_{kfv} ) + \sum_{v=1}^{V_{f}} (\lambda_{kfv} - 1) \left[ \psi( \lambda_{kfv} ) - \psi( \sum_{u=1}^{V_{f}} \lambda_{kfu} ) \right].\end{aligned}$$ Coordinate ascent ----------------- We find a local maximum of the ELBO via coordinate ascent in each dimension of the variational parameters: $\lambda, \phi$. This method is sometimes known as batch variational inference. First we look at $\lambda$; the partial derivative of the ELBO with respect to $\lambda_{kfv}$ is $$\begin{aligned} \frac{\partial}{\partial \lambda_{kfv}} \elbo(\phi, \lambda) &= (A_{v}-1) \psi_{1}(\lambda_{kfv}) + \left[ \sum_{u=1}^{V_{f}} (A_{u} - 1) \right] \psi_{1}( \sum_{u=1}^{V_{f}} \lambda_{kfu} ) \\ & {} + \sum_{d=1}^{D} \sum_{r=1}^{R_{d}} \phi_{drk} \mbo\{x_{drf} = v\} \left[ \psi_{1}(\lambda_{kfv}) - \psi_{1}( \sum_{u=1}^{V_{f}} \lambda_{kfu} ) \right] \\ & {} + \sum_{u: u \ne v} \sum_{d=1}^{D} \sum_{r=1}^{R_{d}} \phi_{drk} \mbo\{x_{drf} = u\} \left[ - \psi_{1}( \sum_{t=1}^{V_{f}} \lambda_{kft} ) \right] \\ & {} - \psi( \sum_{u=1}^{V_{f}} \lambda_{kfu} ) + \psi(\lambda_{kfv}) \\ & {} - \left\{ (\lambda_{kfv} - 1) \cdot \left[ \psi_{1}(\lambda_{kfv}) - \psi_{1}( \sum_{u=1}^{V_{f}} \lambda_{kfu} ) \right] \right\} \\ & {} - \left[ \psi(\lambda_{kfv}) - \psi( \sum_{u=1}^{V_{f}} \lambda_{kfu} ) \right] \\ & {} - \sum_{u: u \ne v} (\lambda_{kfu} - 1) \psi_{1}( \sum_{t=1}^{V_{f}} \lambda_{kft} ) \\ &= \psi_{1}(\lambda_{kfv}) \left[ A_{v} - \lambda_{kfv} + \sum_{d=1}^{D} \sum_{r=1}^{R_{d}} \phi_{drk} \mbo\{x_{drf} = v\} \right] \\ & {} - \psi_{1}( \sum_{u=1}^{V_{f}} \lambda_{kfu} ) \sum_{u=1}^{V_{f}} \left[ A_{u} - \lambda_{kfu} + \sum_{d=1}^{D} \sum_{r=1}^{R_{d}} \phi_{drk} \mbo\{x_{drf} = u\} \right]\end{aligned}$$ This quantity will be zero for $$\lambda_{kfv} \leftarrow A_{v} + \sum_{d=1}^{D} \sum_{r=1}^{R_{d}} \phi_{drk} \mbo\{x_{drf} = v\}.$$ Next we consider $\phi$. The partial derivative of the ELBO with respect to $\phi_{drk}$ is $$\begin{aligned} \frac{\partial}{\partial \phi_{drk}} \elbo(\phi, \lambda) &= \sum_{f=1}^{F} \sum_{v=1}^{V_{f}} \mbo\{x_{drf} = v\} \left[ \psi( \lambda_{kfv} ) - \psi( \sum_{u=1}^{V_{f}} \lambda_{kfu} ) \right] \\ & {} - \log(\phi_{drk}) - 1.\end{aligned}$$ This quantity will be zero, and the $\phi_{drk}$ will sum across $k$ to one,[^3] if $$\phi_{drk} \propto_{k} \exp\left\{ \sum_{f=1}^{F} \sum_{v=1}^{V_{f}} \mbo\{x_{drf} = v\} \left[ \psi( \lambda_{kfv} ) - \psi( \sum_{u=1}^{V_{f}} \lambda_{kfu} ) \right] \right\}.$$ We note that the $\phi_{drk}$ satisfy the constraint $\phi_{drk} > 0$ for all $k$ and so form a proper probability distribution across $k$. [^1]: The nature of data corruption in the database collection process can often lead to interesting and nonstandard noise distributions for data that might, in other contexts, be treated as non-categorical or continuous (date of birth, age, etc.). Text fields—e.g., name of an individual—must also be treated with more care than assigning a single categorical distribution in this context. These considerations, though addressed elsewhere [@sadinle_2014; @steorts_2014_eb], are outside the scope of this note. [^2]: This database is the National Long Term Care Study (NLTCS), a longitudinal study of the health status of elderly Americans <http://www.nltcs.aas.duke.edu/>. The authors ran the NLTCS on three databases of 20,000 records each, for 1 million iterations of their hybrid MCMC, which took 3.5 hours to run. [^3]: This derivation can be completed using the Lagrange method of multipliers.	{ "pile_set_name": "ArXiv" }
--- abstract: \| The restricted (equilateral) four-body problem consists of three bodies of masses $m_{1}$, $m_{2}$ and $m_{3}$ (called primaries) lying in a Lagrangian configuration of the three-body problem, i,e,. they remain fixed at the apices of an equilateral triangle in a rotating coordinate system. A massless fourth body moves under the Newtonian gravitation law due to the three primaries; as in the restricted three-body problem the fourth mass does not affect the motion of the three primaries. In this paper we show a global regularization of binary collisions of the infinitesimal body with two of the primaries.\ \ [Resumen\ ]{} El problema restringido de cuatro cuerpos equilátero consiste de tres masas puntuales $m_{1}$, $m_{2}$, $m_{3}$ (llamadas primarias) que permanecen a todo tiempo en una configuración Lagrangiana del problema de tres cuerpos, es decir; las masas permanecen fijas en los vértices de un triangulo equilátero en un sistema rotatorio. Un cuarto cuerpo de masa infinitesimal se mueve bajo la ley de gravitación universal de Newton que ejercen las tres masas puntuales; como en el caso del problema restringido de tres cuerpos, la cuarta masa no afecta el movimiento de las tres primarias. El objetivo principal de este artículo es mostrar una regularización global de colisiones binarias de la masa infinitesimal con dos de las primarias. Al final se muestra una aplicación de este proceso de regularización. author: - \| Jaime Burgos–García\ Published in Aportaciones Matemáticas\ Memorias de la Sociedad Matemática Mexicana, 45, pp.1-13. [^1] [^2] bibliography: - 'biblio-u1.bib' title: Regularization in the Restricted Four Body Problem --- Keywords: Four–body problem, Hill’s regions, regularization, ejection-collision orbits. AMS Classification: 70F15, 70F16 Introduction ============ Few bodies problems have been studied for long time in celestial mechanics, either as simplified models of more complex planetary systems or as benchmark models where new mathematical theories can be tested. The three–body problem has been source of inspiration and study in Celestial Mechanics since Newton and Euler. In recent years it has been discovered multiple stellar systems such as double stars, triples systems. The restricted three body problem (R3BP) has demonstrated to be a good model of several systems in our solar system such as the Sun-Jupiter-Asteroid system, and with less accuracy the Sun-Earth-Moon system. In analogy with the R3BP, in this paper we study a restricted problem of four bodies consisting of three primaries moving in circular orbits keeping an equilateral triangle configuration and a massless particle moving under the gravitational attraction of the primaries. In the following discussion we focus on the study of the regularizations of binary collisions of the infinitesimal body with two of the primaries by a simple method similar to Birkhoff’s which permit us to study the dynamic of the equations when they present discontinuities . As an application of the transformed equations by the regularization process it can be shown that some families of periodic orbits end up in a homoclinic connection. This last phenomenon can be dynamically explained by the so called “blue sky catastrophe” termination, a rigorous justification of this phenomena can be found in [@BurgosII]. Equations of Motion =================== Consider three point masses, called $\textit{primaries}$, moving in circular periodic orbits around their center of mass under their mutual Newtonian gravitational attraction, forming an equilateral triangle configuration. A third massless particle moving in the same plane is acted upon the attraction of the primaries. The equations of motion of the massless particle referred to a synodic frame with the same origin, where the primaries remain fixed, are: $$\begin{aligned} \bar{x}''-2n\bar{y}'-n^2\bar{x}&=&-k^2\sum_{i=1}^{3}m_{i}\frac{(\bar{x}-\bar{x_{i}})}{\rho_{i}^3}\nonumber\\ \bar{y}''+2n\bar{x}'-n^2\bar{y}&=&-k^2\sum_{i=1}^{3}m_{i}\frac{(\bar{y}-\bar{y_{i}})}{\rho_{i}^3} \label{sistemaconunidades}\end{aligned}$$ where $k^2$ is the gravitational constant, $n$ is the mean motion, $\rho_{i}^{2}=(\bar{x}-\bar{x}_{i})^2+(\bar{y}-\bar{y}_{i})^2$ is the distance of the massless particle to the primaries, $\bar{x}_{i}$, $\bar{y}_{i}$ are the vertices of equilateral triangle formed by the primaries, and ($'$) denotes derivative with respect to time $t^{}$. We choose the orientation of the triangle of masses such that $m_1$ lies along the positive $x$–axis and $m_2$, $m_3$ are located symmetrically with respect to the same axis, see figure \[triangle\]. ![The restricted four-body problem in a synodic system\[triangle\]](figura1){width="3.5in"} The equations of motion can be recast in dimensionless form as follows: Let $L$ denote the length of triangle formed by the primaries, $x=\bar{x}/L$, $y=\bar{y}/L$, $x_i=\bar{x}_i/L$, $y_i=\bar{y}_i/L$, for $i=1,2,3$; $M=m_{1}+m_{2}+m_{3}$ the total mass, and $t=nt^{}$. Then the equations (\[sistemaconunidades\]) become $$\begin{aligned} \ddot{x}-2\dot{y}-x &=&-\sum_{i=1}^{3}\mu_{i}\frac{(x-x_{i})}{r_{i}^3}\nonumber\\ \ddot{y}+2\dot{x}-y &=&-\sum_{i=1}^{3}\mu_{i}\frac{(y-y_{i})}{r_{i}^3} \label{sistemasinunidades}\end{aligned}$$ where we have used Kepler’s third law: $k^2M=n^2L^3$, and the dot ($\dot{}$) represents derivatives with respect to the dimensionless time $t$ and $r_{i}^2=(x-x_{i})^2+(y-y_{i})^2$. The system (\[sistemasinunidades\]) will be defined if we know the vertices of triangle for each value of the masses. In this paper we suppose $\mu:=\mu_{3}=\mu_{2}$ then $\mu_{1}=1-2\mu$, it’s not hard to prove that the vertices of triangle are given as function of the mass parameter $\mu$ by $x_{1}=\sqrt{3}\mu$, $y_{1}=0$, $x_{2}=-\frac{\sqrt{3}(1-2\mu)}{2}$, $y_{2}=-\frac{1}{2}$, $x_{3}=-\frac{\sqrt{3}(1-2\mu)}{2}$, $y_{3}=\frac{1}{2}$. The system (\[sistemasinunidades\]) can be written succinctly as $$\begin{aligned} \ddot{x}-2\dot{y}&=&\Omega_{x} \label{sistemastandar}\\ \ddot{y}+2\dot{x}&=&\Omega_{y}\end{aligned}$$ where $$\label{omega}\Omega(x,y,\mu):=\frac{1}{2}(x^2+y^2)+\sum_{i=1}^{3}\frac{\mu_{i}}{r_{i}}.$$ is the effective potential function. In the Restricted four–body problem (R4BP) there are three limiting cases: 1. If $\mu=0$, we obtain the rotating Kepler’s problem, with $m_{1}=1$ at the origin of coordinates. 2. If $\mu=1/2$, we obtain the circular restricted three body problem, with two equal masses $m_{2}=m_{3}=1/2$. 3. If $\mu=1/3$, we obtain the symmetric case with three masses equal to $1/3$. It will be useful to write the system (\[sistemastandar\]) using complex notation. Let $z=x+\textit{i}y$, then $$\label{sistemacomplejo}\ddot{z}+2\textit{i}\dot{z}=2\frac{\partial\Omega}{\partial\bar{z}}$$ with $$\Omega(z,\bar{z},\mu)=\frac{1}{2}\vert z\vert^2+U(z,\bar{z},\mu)$$ where the gravitational potential is $$U(z,\bar{z},\mu)=\sum_{i=1}^{3}\frac{\mu_{i}}{\vert z-z_{i}\vert}$$ and $r_{i}=\vert z-z_{i}\vert$, $i=1,2,3.$ are the distances to the primaries. System (\[sistemacomplejo\]) has the Jacobian first integral $$2\Omega(z,\bar{z},\mu)-\vert\dot{z}\vert^{2}=C$$ If we define $P=p_{x}+\textit{i}p_{y}$, the conjugate momenta of $z$, then system (\[sistemastandar\]) can be recast as a Hamiltonian system with Hamiltonian $$\begin{aligned} H&=&\frac{1}{2}\vert P\vert^2+Im(z\overline{P})-U(z,\bar{z},\mu)\nonumber\\ &=&\frac{1}{2}(p^{2}_{x}+p^{2}_{y})+(yp_{x}-xp_{y})-U(x,y,\mu).\label{hamiltoniano}\end{aligned}$$ The relationship with the Jacobian integral is $H=-C/2$. The phase space of (\[hamiltoniano\]) is defined as $$\Delta=\{(z,P)\in\mathbb{C}\times\mathbb{C}\vert z\ne z_{i}, i=1,2,3\},$$ with collisions occurring at $z=z_{i}$, $i=1,2,3$. In the restricted three-body problem there exist five equilibrium points for all values of the masses of the primaries but in this restricted four-body problem the number of equilibrium points depends on the particular values of the masses. A complete discussion of the equilibrium points and bifurcations can be found in [@Del], [@MeyerCC], [@Lea], [@PapaII], [@Simo]. Regularization ============== Where the solutions of the R4BP have binary collisions with any of the primaries the Hamiltonian (\[hamiltoniano\]) is not defined for these solutions, so we have to remove such singularities in the system. The so called regularization process is a technique that enable us to remove singularities of differential equations, therefore we want to apply this technique to the R4BP to study the system when the solutions are near to collision with the primaries. The regularization process is a standar procedure and it can be found in [@Sz] and [@Gia], however, we are going to explain it briefly in the present problem.\ \ First, we perform a translation from the center of mass, namely $z=u+\sqrt{3}\mu-\sqrt{3}/2$, where $u=x_{2}+\textit{i}x_{2}$. The positions of the primaries in these new coordinates become $u_{1}=\frac{\sqrt{3}}{2}$, $u_{2}=-\frac{\textit{i}}{2}$, $u_{3}=\frac{\textit{i}}{2}$. In these coordinates the Hamiltonian is written as $$\label{uhamiltonian}H=\frac{1}{2}\vert U\vert^2+Im((u+\sqrt{3}\mu-\sqrt{3}/2)\overline{U})-V(u,\bar{u},\mu)$$ where $U=P$ is the complex conjugate momenta de $u$. We will denote by $f^{}(w)$ the derivative with respect to complex variable $w$, $f(w)$ represents a complex valued analytic map. The following lemma shows how to complete a point transformation given by an analytic function $u=f(w)$ to a canonical transformation. We will chose later the mapping $f(w)$ to eliminate the singularities due to the primaries. Let $u=f(w)$ be a transformation point, then the transformation of the conjugate momenta $U=W/\overline{f^{}(w)}$ yields a canonical transformation whenever $f^{}(w)\ne0$ $\textit{Proof}$: The mapping $(u,U)\rightarrow(w,W)$ is canonical if $$Re(\overline{U}du)=Re(\overline{W}dw)$$ But $$\overline{U}du=\bar{U}f^{}(u)du=\frac{\overline{W}}{f^{}(u)}f^{}(u)du=\overline{W}dw$$ form which the result follows. $\Box$ The regularization process starts with a canonical transformation followed by a re–parametrization of time on a fixed energy level $H=-C/2$. Let $f(w)$ be as above, this transformation must satisfy the hypothesis of the previous lemma. We perform the following scaling of time $$\label{escalamientotemporal} \frac{d\tau}{dt}=\frac{1}{\vert f^{}(w)\vert^{2}}$$ Now we need to transform the Hamiltonian (\[uhamiltonian\]) to the new variables $$\label{hamiltonianow}H=\frac{1}{2}\frac{\vert W\vert^{2}}{\vert f^{}(w)\vert^{2}}+Im\left((f(w)+\sqrt{3}\mu-\sqrt{3}/2)\frac{\overline{W}}{f^{}(w)}\right)-V(f(w),\overline{f(w)},\mu)$$ Next perform Poincare’s trick to re-parametrize solutions according to (\[escalamientotemporal\]) $\overline{H}=\vert f^{}(w)\vert^{2}(H+C/2)$. Observe that the energy level $H=-C/2$ is carried on to the level $\overline{H}=0$, explicitly $$\label{hamiltonianobarra}\overline{H}=\frac{1}{2}\vert W\vert^{2}+Im((f(w)+\sqrt{3}\mu-\sqrt{3}/2)\overline{f^{}(w)W})-\vert f^{}(w)\vert^{2}V(w,\bar{w},\mu)$$ $$+\vert f^{}(w)\vert^{2}(C/2)$$ where $$V(w,\bar{w},\mu)=\sum_{i=1}^{3}\frac{\mu_{i}}{\vert f(w)-w_{i}\vert}$$ and $w_{i}$ denotes the position of the primaries. Now we must choose the transformation $f(w)$ according to the conditions mentioned above and keeping in mind the singularities that we want to remove. Note that if we want to remove a single collision with any of the primaries, we can apply the Levi–Civita transformation [@Sz] to remove such singularity, however, the interesting problem is to remove simultaneously two or more singularities. It is not hard to see that the equations of the R4BP have the property that if $z(t)$ is a solution (in complex notation) then $\bar{z}(-t)$ is also a solution, in other words, we have symmetry of the solutions with respect to the $x-$axis as in the R3BP. This symmetry of the equations tells us that a collision with the primary $m_{2}$ (respectively $m_{3}$) implies necessarily a collision with the primary $m_{3}$ (respectively $m_{2}$), therefore a simultaneous regularization with the primaries $m_{3}$ and $m_{3}$ is needed. If we want to perform a simultaneous regularization in this case, first, we must note the importance of making the regularized equations as simple as possible in order to simplify the calculations and to save time in the integration of the equations. Therefore, we choose a transformation $f(w)$ similar to the Birkhoff’s transformation [@Birk] $$\label{transformacion} u=f(w)=\frac{1}{2}\left(w-\frac{1}{4w}\right)$$ It’s easy to prove that $f(w)$ has the following properties $$\begin{aligned} u_{i}&=&f(u_{i}),\quad i=2,3.\label{fixed}\\ f^{}(w)&=&\frac{1}{2}\frac{(w-u_{2})(w-u_{3})}{w^{2}}\label{fprima}\end{aligned}$$ In particular $$\vert f^{}(w)\vert^{2}=\frac{1}{4}\frac{(\vert w-u_{2}\vert^{2})(\vert w-u_{3}\vert^{2})}{\vert w\vert^{4}}\label{deriv}$$ and $$f^{}(u_{i})=0,\quad i=2,3.$$ Observe that the positions of the primaries $m_{2}$ and $m_{3}$ remain fixed under the transformation. In fact, the following properties are the key to remove the singularities $w=w_{i}$, $i=2,3.$ $$f(w)-u_{i}=\frac{1}{2}\left(\frac{(w-u_{i})^{2}}{w}\right)$$ and $$f(w)-u_{1}=\frac{1}{2}\left(\frac{(w-a_{1})(w-a_{2})}{w}\right)$$ where $a_{1}=1+\sqrt{3}/2$, $a_{2}=-1+\sqrt{3}/2$. We must check that the Hamiltonian (\[hamiltonianobarra\]) is free of singularities due to collisions with the primaries $m_{2}$ and $m_{3}$, observe that these points are contained in the term $\vert f^{}(w)\vert^{2}V(w,\bar{w},\mu)$, a straightforward calculation using (\[deriv\]) shows $$\vert f^{}(w)\vert^{2}V(w,\bar{w},\mu)=$$ $$\frac{1}{2\vert w\vert^{3}}\left(\frac{(1-2\mu)\vert w-u_{2}\vert^{2}\vert w-u_{3}\vert^{2}}{\vert w-a_{1}\vert \vert w-a_{2}\vert}+\mu(\vert w-u_{2}\vert^{2}+\vert w-u_{3}\vert^{2})\right)$$ Observe that we have removed the singularities due to the primaries $m_{2}$ and $m_{3}$, however, we have introduced new singularities though, $w=a_{1}$, $w=a_{2}$ and $w=0$. We are going to study these new singularities. The origin of the new system $w=0$ is mapped under (\[transformacion\]) to infinity in $u-$space, so it corresponds to escapes and it is not of interest for us. Let us analyze the remaining singularities $w=a_{1}$ and $w=a_{2}$. First we want to describe the number of pre-images of a point $u$ under the transformation $f(w)$, we need to solve the equation given by (\[transformacion\]), or to find the roots of the quadratic polynomial $p(w)=w^{2}-2uw-1/4$. Note that given $u$, we have two roots or pre-images counting multiplicities. We recall the next proposition: Let $w_{0}$ be a root of the polynomial $p(w)$, if $p'(w_{0})=p(w_{0})=0$ but $p''(w_{0})\ne0$ then $w_{0}$ is a root with multiplicity 2. If $p''(w_{0})=p'(w_{0})=p(w_{0})=0$ but $p'''(w_{0})\ne0$ then $w_{0}$ is a root with multiplicity 3 etc. It’s easy to see that $p'(w)=2w-2u$ and $p''(w)=2$ then $p'(w)=0$ $\Leftrightarrow$ $w=u$. Now if we evaluate this root in the polynomial $p(w)$ we see that $p(u)=-u^{2}-1/4$ and $p(u)=0$ $\Leftrightarrow$ $w=u_{i}$ $i=2,3.$ therefore $p(u_{i})=p'(u_{i})=0$ but $p''(u_{i})\ne0$. In conclusion we have proved the following Let $u\in\mathbb{C}$ be a complex number, if $u=u_{i}$, $i=2,3.$ then the number of pre-images is one, if $u\ne u_{i}$, $i=1,2,3.$ then the number of pre-images is two. $\Box$ This shows that the number of pre-images of the positions of the primaries $m_{2}$ and $m_{3}$ is exactly one. Actually for $u=u_{i}$, $i=2,3$, we have $$p(w)=(w-u_{i})^{2}$$ therefore the pre-images of each $u_{i}$, $i=2,3$ are they self. The pre-images of the primary $u_{1}=$ are exactly $a_{1}$ and $a_{2}$, then the new singularities correspond to the singularity $u_{1}$ in the $u$-space however we are not interested in removing this singularity, see figure (\[hillregions\]). Instead, we have performed a global regularization of the singularities due to collisions with the primaries $m_{2}$ and $m_{3}$. The phase space where the Hamiltonian (\[hamiltonianobarra\]) is regular is given by $$\overline{\Delta}=\{(w,W)\in\mathbb{C}\times\mathbb{C}\vert w\notin\{0,a_{1},a_{2}\}, i=1,2,3\}$$ Since the Hamiltonian (\[hamiltonianobarra\]) contains only quadratic modulus, its partial derivatives are continuous throughout the region $\overline{\Delta}$. Hill’s Regions of the Regularized Hamiltonian ============================================= The relation given by the first integral $\vert\dot{z}\vert^{2}=2\Omega(z,\bar{z},\mu)-C$ implies $2\Omega(z,\bar{z},\mu)-C\ge 0$ or $\Omega(z,\bar{z},\mu)\ge C/2$, this inequality places a constraint on the position variable $z$ for each values of $\mu$ and $C$, if $z$ satisfies this condition, then there is a solution through that point $z$ for that values of $C$ and $\mu$ (see for example [@MeyerHDS]). The sets where the inequality $\Omega(z,\bar{z},\mu)\ge C/2$ holds are called the Hill’s regions, in regularized variables the former inequality becomes (see [@Sz])$$\vert f^{}(w)\vert^{2}(\Omega(w,\bar{w},\mu)-C/2)\ge0$$ This inequality defines the Hill’s regions in regularized variables whenever $w\ne0$, $w\ne a_{1}$ and $w\ne a_{2}$. Explicitly these regions are defined by the expression $$\label{regioneshill}\frac{1}{2}\vert f^{}(w)\vert^{2}\vert f(w)+\sqrt{3}\mu-\frac{\sqrt{3}}{2}\vert^{2}+\vert f^{}(w)\vert^{2}V(w,\bar{w},\mu)-\vert f^{}(w)\vert^{2}\frac{C}{2}\ge0$$ The next figure shows the Hill’s regions in the $u-$space and the $w-$space for several values of the Jacobi constant $C$ and for the equal masses case, i.e. $\mu=1/3$. At the reference value $C_{1}=3.35804$ there exist three critical points of the potential $\Omega$ in the $u$-space and the Hill’s regions are very similar, however at the origin of the $w$-space new regions appear around the singularities $a_{1}$ and $0$, see figure (\[hillregions\]). If we increase the value of $C_{1}$ the Hill’s regions become disconnected in both spaces and the new regions around the singularities in the $w$-space increase their size, at this point it is clear the correspondences between the $u$ and $w$ spaces given by the transformation (\[transformacion\]) discussed in the section (3). Finally, if we decrease the value of $C_{1}$ we find that the whole Hill’s region is now connected, see figure (\[hillregions\]). The positions of the primaries are marked by small circles and the singularities $w=0$, $w=a_{1}$ and $w=a_{2}$ are marked by black points. ![Hill’s regions (shaded areas) in the $u$-space (left column), in the $w$-space (center column) and magnifications of the origin in the $w$-space (right column). From top to bottom: $C=C_{1}$, $C=C_{1}+0.2$, $C=C_{1}-0.2$[]{data-label="hillregions"}](figura2){height="18.0cm" width="4.8in"} An application of the regularization process ============================================ In this section we show an application of the regularized equations of the R4BP. We recall that Routh’s criterion for linear stability of the Lagrangian configuration states that the masses of primaries must satisfy the inequality $$\frac{m_{1}m_{2}+m_{2}m_{3}+m_{3}m_{1}}{m_{1}+m_{2}+m_{3}}<\frac{1}{27}.$$ When the three masses are such that $m_{2}=m_{3}:=\mu$ and $m_{1}+m_{2}+m_{3}=1$, the inequality is satisfied in the interval $\mu\in[0,0.0190636...)$. In the paper [@Burgos] we can find a numerical exploration of families periodic orbits of the R4BP for values of the masses satisfying the Routh’s criterion; in that work, there are nine families of periodic orbits and some of them contain ejection–collision orbits with the primaries $m_{1}$ and $m_{2}$, we are going to explain briefly how these orbits were obtained. Suppose that we have calculated a periodic orbit, this periodic orbit lies on a surface defined by the Jacobian first integral and therefore it has a well defined value of the constant $C$, if we use the analytical continuation method [@HenI], [@Sz] we can follow the evolution of this orbits as the value of the constant $C$ varies continuously, in this evolution, the periodic orbit can reach collisions with any of the primaries, if we want to follow the orbit beyond these collisions we need to use regularized equations. When a ejection–collision orbit is reached, we say that the periodic orbit finishes one phase because after this collision the orbit changes its behavior, for instance from direct to retrograde. We refer to the reader to the references to see a complete discussion on families of periodic orbits. In the following we show some ejection–collision orbits in the R4BP and we explain where these orbits can be found.\ \ In the second phase of the family $f$, all of the orbits are near to collision with the primaries $m_{2}$ and $m_{3}$, however, these collisions are never reached but the regularized equations are needed in the numerical calculations to follow the evolution of this phase and to state the “Blue sky catastrophe" termination of this family, see figure \[phasesf\]. ![The evolution of the second phase of family $f$ .\[phasesf\]](fasesf){height="8.0cm" width="4.8in"} In the family $j$ we find two ejection–collision orbits, one of them is at the beginning of the first phase and the second one is at the end of the second phase, see figure \[colisionesj\]. ![Ejection–collision orbits of the family $j$, in the first phase (right) and in the second phase (left).\[colisionesj\]](colisionesj){height="8.0cm" width="4.8in"} In the evolution of the first phase of the family $r_{2}$, we find similar orbits to the family $f$, but in this case a ejection–collision orbit appears before the “Blue sky catastrophe" termination of this family, see figure \[fasesr2\]. ![The evolution of the first phase of family $r_{2}$ .\[fasesr2\]](fasesr2){height="8.0cm" width="4.8in"} Finally, in the family $j_{2}$ we find two ejection–collision orbits, in the figure \[colisionesj2\] we can see that these orbits are very similar to the ones of the family $j$. ![Ejection–collision orbits of the family $j_{2}$, in the first phase (left) and in the second phase (right).\[colisionesj2\]](colisionesj2){height="8.0cm" width="4.8in"} [99]{} A. Deprit, R. Broucke; Régularisation du probléme restreint plan de trois corps par représentations conformes. Icarus 2, 207 (1963). C. Simó ; Relative equilibrium solutions in the four body problem. Cel. Mech. 165-184(1978). D.G. Birkhoff; Sur le probléme restreint des trois corps. Ann. Scuola Superiore de Pisa 4, 267 (1935); also, Birkhoff, D. G., “Collected Mathematical Papers”, Vol.2., p. 466. Am. Math. Soc., New York, 1950. E.S.G. Leandro; On the central configurations of the planar restricted four-body problem. J. Differential Equations.226 p.323-351 (2006). J. Burgos, J. Delgado; Periodic orbits in the restricted four-body problem with two equal masses. Astrophysics & Space Science, 345 Issue 2, pp.247. J. Burgos, J. Delgado; On the “blue sky catastrophe” termination in the restricted four-body problem. Celestial Mechanics and Dynamical Astronomy, 117, Issue 2, pp.113-136. J. Delgado, M. Álvarez-Ramirez; Central Configurations of the symmetric restricted four-body problem. Cel. Mech. and Dynam. Astr.87 p.371-381 (2003). G. Giacaglia; Regularization of the restricted problem of four bodies. The Astronomical Journal.27 No. 5. (1967). J. Henrard; Proof of a conjeture of E. Strömgren. Cel. Mech.7 p.449-457 (1973). K. Meyer; Bifurcation of a central configuration. Cel. Mech.40(3-4) p.273-282 (1987). K. Meyer; Introduction to Hamiltonian Dynamical Systems and the N-body problem. Springer Verlag. (2009). K.E. Papadakis, A.N. Baltagiannis; Families of periodic orbits in the restricted four-body problem. Astrophys. Space Sci. DOI10.1007/s 10509-011-0778-7 (2011). K.E. Papadakis, A.N. Baltagiannis; Equilibrium points and their stability in the restricted four-body problem. Int. J. Bifurc. Chaos doi:IJBC-D-10-00401(2011). M. Álvarez-Ramirez, C. Vidal; Dynamical aspects of an equilateral restricted four-body problem. Math. Probl. Eng. doi:10.1155/2009/181360(2009). M. Hénon; Exploration numérique du probléme restreint I. Masses égales Orbites périodiques Ann. Astrophysics. (1965). M. Hénon; Exploration numérique du probléme restreint II. Masses égales stabilité des orbites périodiques Ann. Astrophysics.28 p.992-1007 (1965). M. Hénon; Generating families in the restricted three body problem. Springer Verlag. (1997). P. Pedersen; Librationspunkte im restringierten vierkoerperproblem. Dan. Mat. Fys. Medd. 1-80(1944). T. Levi-Civita; Sur le régularisation du probléme des trois corps. Acta Math.42. 99-144 (1920). V. Szebehely; Theory of orbits. Academic Press, New York (1967). Departamento de Matemáticas UAM–Iztapalapa. Av. San Rafael Atlixco 186, Col. Vicentina, C.P. 09340, México, D.F. e–mail: [email protected], [^1]: Departamento de Matemáticas UAM–Iztapalapa. Av. San Rafael Atlixco 186, Col. Vicentina, C.P. 09340, México, D.F. e–mail: [email protected] [^2]: Research article	{ "pile_set_name": "ArXiv" }
--- abstract: \| I consider a truncation of low-energy string theory which contains two $U(1)$ gauge fields. After making some general comments on the theory, I describe a previously-obtained instanton for the pair creation of black holes when both gauge fields are non-zero, and obtain the pair creation rate by calculating its action. This calculation agrees qualitatively with the earlier calculation of the pair creation rate for black holes in Einstein-Maxwell theory. That is, the pair creation is strongly suppressed in realizable circumstances, and it reduces to the Schwinger result in the point-particle limit. The pair creation of non-extreme black holes is enhanced over that of extreme black holes by $e^{{\cal A}_{bh}/4}$. author: - \| Simon F. Ross\ [Department of Applied Mathematics and Theoretical Physics]{}\ [University of Cambridge, Silver St., Cambridge CB3 9EW]{}\ [email protected] date: \| \ DAMTP/R-95/47 title: 'Pair creation rate for $U(1)^2$ black holes' --- =8.8 truein =6.2 truein Introduction {#intro} ============ The study of black hole pair creation is of considerable interest for the exploration of quantum gravity. Like black hole evaporation, it represents a truly quantum gravitational process, being classically completely forbidden. At the same time, it is easy to achieve a physical understanding of what is happening; there is a strong analogy with the creation of particle-antiparticle pairs in quantum field theory, as can be seen from the fact that the black hole pair creation rate reduces to the particle-antiparticle rate in the limit of small black holes. If we trust this analogy to particle physics, the pair creation rate should depend on the number of accessible states for the black hole, so we can find out how many states the black holes should have by studying pair creation (although we can’t find out what those states are). Black hole pair creation involves topology change, and this suggests that including the effects of topology change will be important to a proper understanding of quantum gravity. Because of the topology change, black hole pair creation is studied in the path-integral approach to quantum gravity, by finding a suitable instanton (that is, a solution of the classical equations of motion with Euclidean signature) which describes the transition from the given initial to final data. Most of the work to date has focussed on the pair creation of charged black holes in a background electromagnetic field, both in Einstein-Maxwell theory [@ggs] and in a generalisation of this theory which includes a dilaton [@dgkt; @dggh], whose action is $$\label{Faction} I = - \frac{1}{16\pi} \int_M ( R - 2 \partial^{\mu}\phi \partial_{\mu} \phi -e^{-2a\phi} F^2 ) -\frac{1}{ 8\pi}\oint_{\partial M} (K-K_0).$$ The present paper is concerned with a different generalisation of Einstein-Maxwell theory, to include two $U(1)$ gauge fields and a dilaton, called the $U(1)^2$ theory. This is a somewhat more typical example of a low-energy effective theory arising from superstring theory, as the compactification of the extra dimensions will typically give an effective theory with a large number of $U(1)$ gauge fields. As I will argue in Sec. \[dsec\], the most appropriate effective action for this theory is [@Kalloshcens]: $$\label{Kaction} I_{SO(4)} = -\frac{1}{16 \pi} \int_M ( R - 2 \partial^{\mu}\phi \partial_{\mu} \phi -(e^{2\phi} \tilde{F}^2 + e^{-2\phi} G^2)) -\frac{1}{ 8\pi}\oint_{\partial M} (K-K_0),$$ where $\tilde{F}_{\mu\nu}$ and $G_{\mu\nu}$ are the two $U(1)$ gauge fields. This theory is a consistent truncation of low-energy heterotic string theory [@Kalloshcens]. One of the advantages of this truncation is that it includes the Einstein-Maxwell theory as a special case, when $\phi=0$ and $\tilde{F}_{\mu\nu} = G_{\mu\nu}$. That is, Einstein-Maxwell is also a consistent truncation of string theory. The truncation (\[Kaction\]) can also be derived from the $SO(4)$ version of $N=4$ supergravity [@cremmer]. It also includes the action (\[Faction\]) with $a=1$, when one of the gauge fields vanishes. There are two duality symmetries in the $U(1)^2$ theory; one of them is a generalisation of the usual electric-magnetic duality, while the other is trivial on the Einstein-Maxwell solutions. Charged black hole solutions of (\[Kaction\]) were found by Gibbons [@Gary]. These solutions include the Reissner-Nordström metrics when the two gauge charges are equal, so the Reissner-Nordström solutions correspond to dyonic solutions of this theory. The duality symmetries and the black hole solutions are reviewed in Sec. \[dsec\]. An instanton describing pair creation of charged black holes in background fields in this theory was obtained in [@2u1], and is reviewed in Sec. \[Esol\]. This instanton is obtained from a generalisation of the Ernst solution of Einstein-Maxwell theory so that the black holes have two gauge charges and there are two corresponding background fields. The instanton is very similar to the Ernst instanton, but the presence of two background fields introduces some interesting complications. In particular, the black holes are not spherically symmetric in the extremal limit in this case, unlike the Einstein-Maxwell case [@dggh]. The main aim in this paper is to calculate the pair creation rate given by this instanton, which will allow us to extend the conclusions of [@dggh; @entarea] to this case. The amplitude for pair creation in the instanton approximation to the path integral is given by $e^{-I}$, where $I$ is the action of the instanton. The pair creation rate will thus be given by $e^{-I_b}$, where $I_b$ is the action of the “bounce”, an instanton–anti-instanton pair. Sec. \[so4sec\] is thus dedicated to the calculation of the action for the bounce. We find that the pair creation of non-extreme black holes is enhanced over that for extreme black holes by $e^{{\cal A}_{bh}/4}$, from which we conclude that the non-extreme black holes have $e^{{\cal A}_{bh}/4}$ more states than the extreme ones. That is, we conclude that the number of states is given by $e^{S_{bh}}$. In the point-particle limit, where the black holes are small on the scale set by the acceleration, the pair creation rate reduces to the Schwinger result. That is, to leading order, the pair creation rate for the black holes is the same as that for particles of the same mass and charges. In summary, the results of this calculation of the pair creation rate are essentially those of the calculation for the Einstein-Maxwell theory in [@dggh; @entarea]; this might not seem surprising, as the Einstein-Maxwell theory is included as a special case, but there is much more freedom in the $U(1)^2$ theory, so it is a non-trivial result. Properties of the theory {#dsec} ======================== In [@Kalloshcens], two actions were given for a low-energy theory with two $U(1)$ gauge fields and a dilaton, $$\label{Kaction1} I_{SO(4)} = -\frac{1}{16\pi} \int_M ( R - 2 \partial^{\mu}\phi \partial_{\mu} \phi -(e^{2\phi} \tilde{F}^2 + e^{-2\phi} G^2)) -\frac{1}{ 8\pi}\oint_{\partial M} (K-K_0)$$ and $$\label{Kaction2} I_{SU(4)} = -\frac{1}{16\pi} \int_M ( R - 2 \partial^{\mu}\phi \partial_{\mu} \phi - e^{-2\phi}( F^2 + G^2)) -\frac{1}{ 8\pi}\oint_{\partial M} (K-K_0).$$ These can be regarded as arising from the $SO(4)$ and $SU(4)$ versions of $N=4$ supergravity respectively [@cremmer]. If we take $$\label{su4toso4} \tilde{F}_{\mu\nu} = \frac{1}{2} e^{-2\phi} \epsilon_{\mu\nu\rho\sigma} F^{\rho\sigma},$$ It is easy to see that (\[Kaction1\]) and (\[Kaction2\]) give the same equations of motion, but the values of these two actions are different. To calculate the pair creation rate, we need to know which of these actions we should take. If we consider the Einstein-Maxwell case, where $\tilde{F}_{\mu\nu} = G_{\mu\nu} = \frac{1}{\sqrt{2}} {\cal F}_{\mu\nu}$ (say), and $\phi=0$, (\[Kaction1\]) reduces to the usual Einstein-Maxwell action, with Maxwell field ${\cal F}_{\mu\nu}$, while in (\[Kaction2\]), the two gauge field terms cancel. I therefore think that, since we use the Einstein-Maxwell action in the calculation of the pair creation rate in the Einstein-Maxwell case, we should use (\[Kaction1\]) to calculate the pair creation rate in this case. In [@2u1], where the pair creation instanton was obtained, the solutions were written in terms of $F$ and $G$. Since I will use (\[Kaction1\]) to calculate the pair creation rate, I will instead write them here in terms of $\tilde{F}$ and $G$. One interesting feature of the $U(1)^2$ theory is that it has two distinct duality symmetries. The equations of motion of this theory are invariant under a duality transformation, $$\label{dualtransf} F_{\mu\nu} \rightarrow \tilde{F}_{\mu\nu} \equiv \frac{1}{2} e^{-2\phi} \epsilon_{\mu\nu\rho\sigma} F^{\rho\sigma},$$ $$\label{dualt2} G_{\mu\nu} \rightarrow \tilde{G}_{\mu\nu} \equiv \frac{1}{2} e^{-2\phi} \epsilon_{\mu\nu\rho\sigma} G^{\rho\sigma},\, \phi \rightarrow -\phi,$$ which is analogous to the ordinary electric-magnetic duality transformation of Einstein-Maxwell theory. The equations of motion and the action (\[Kaction2\]) are also invariant under the interchange of the two gauge fields, $F_{\mu\nu} \leftrightarrow G_{\mu\nu}$. If we combine these, we find that the equations of motion and the action (\[Kaction1\]) are invariant under the “duality” $$\label{otherdual} F_{\mu\nu} \to \tilde{G}_{\mu\nu}, G_{\mu\nu} \to \tilde{F}_{\mu\nu}, \phi \to -\phi.$$ If we think of $\tilde{F}$ as the field variable rather than $F$, this “duality” just interchanges the two gauge fields and reverses the sign of the dilaton. On the Einstein-Maxwell solutions, for which $\tilde{F}_{\mu\nu} = G_{\mu\nu}$ and $\phi=0$, (\[otherdual\]) is a trivial transformation. In general, we will consider solutions for which the transformation (\[otherdual\]) just corresponds to an interchange of the parameters of the solution. These solutions will be said to have a manifest duality symmetry. The action (\[Kaction1\]) is invariant under this manifest duality symmetry. The charged black hole solutions of the $U(1)^2$ theory are [@Gary]: $$ds^2 = -\lambda dt^2 + \lambda^{-1} dr^2 + R^2 d\Omega, \label{Kbhole1}$$ $$e^{2\phi} = e^{2\phi_0} \frac{r+\Sigma}{r-\Sigma}, \label{Kbhole2}$$ $$\tilde{F} =Q e^{-\phi_0}\sin \theta d\theta \wedge d\varphi,\; G =Pe^{\phi_0}\sin \theta d\theta \wedge d\varphi, \label{Kbhole3}$$ where $$\label{Klambda} \lambda = \frac{(r-r_+)(r-r_-)}{R^2}, \; R^2 = r^2 -\Sigma^2,$$ and [@stringbh] $$\label{Kparam} r_{\pm} = M \pm \sqrt{M^2+\Sigma^2-P^2-Q^2},\; \Sigma = \frac{P^2- Q^2}{2M}.$$ There is a curvature singularity at $r=\|\Sigma\|$. The physical degrees of freedom are $P, Q, M$ and $\phi_0$; $M$ is the mass of the black hole, and $e^{-\phi_0} Q$ and $e^{\phi_0} P$ are its gauge charges. Note that both the gauge fields are magnetic, when we write the solutions this way. One can also obtain a solution with two electric fields, but I will restrict attention to the magnetic case. We could keep the asymptotic value of the dilaton $\phi_0$ as a free parameter, but I will instead fix it by requiring that the dilaton match to an appropriate background value at infinity. The solution has a manifest duality symmetry, as the solution is unchanged when $$\label{bhsymm} \tilde{F} \leftrightarrow G,\; \phi \leftrightarrow -\phi,$$ and $$Q \leftrightarrow P,\; \Sigma \leftrightarrow -\Sigma,\; \phi_0 \leftrightarrow - \phi_0.$$ The pair creation instanton {#Esol} =========================== The pair creation of black holes is described by an instanton, that is, a solution of the classical equations of motion with Euclidean signature, which acts as a saddle-point in the path integral. The solution which gives the instanton in the $U(1)^2$ theory, which I will refer to as the $U(1)^2$ Ernst solution, was obtained in [@2u1]. It is a generalisation of the Ernst solution of Einstein-Maxwell theory [@Ernst]. Like the Ernst solution, it describes a pair of oppositely-charged black holes undergoing uniform acceleration under the influence of background electromagnetic fields. It asymptotically approaches an analogue of the Melvin solution [@melvin], which describes the background fields, which I will refer to as the $U(1)^2$ Melvin solution. The $U(1)^2$ Melvin solution is $$\label{dualMelvin} ds^2 = \Lambda \Psi[-dt^2+d\rho^2+dz^2] + \frac{\rho^2 d\varphi^2}{\Lambda \Psi},$$ $$\label{dMgauge} e^{-2\phi} = \frac{\Lambda}{\Psi},\, A_\varphi = -\frac{\widehat{B}_M\rho^2}{2\Lambda},\, B_\varphi = -\frac{\widehat{E}_M\rho^2}{2\Psi},$$ $$G_{\mu\nu} = \partial_{[\mu} A_{\nu]},\, \tilde{F}_{\mu\nu} = \partial_{[\mu} B_{\nu]},$$ $$\Lambda = 1+ \frac{1}{2} \widehat{B}_M^2 \rho^2,\, \Psi = 1+\frac{1}{2} \widehat{E}_M^2 \rho^2.$$ This solution has a manifest duality symmetry under $$\tilde{F} \leftrightarrow G,\, \phi \leftrightarrow -\phi, \mbox{ and } \widehat{B}_M \leftrightarrow \widehat{E}_M.$$ It represents a pair of magnetic fields which are essentially uniform near the axis $\rho=0$, with field strengths given by $\widehat{E}_M$ and $\widehat{B}_M$. The fields depart from uniformity away from the axis because the field energy curves the spacetime. However, in practice we cannot construct such strong fields, so the physically interesting part of this solution is the region near the axis. The $U(1)^2$ Ernst solution is $$\begin{aligned} ds^2 &=& \frac{\Lambda\Psi}{A^2(x-y)^2}[F(x)(G(y)dt^2-G^{-1}(y) dy^2) \label{Ernst} \\ && +F(y)G^{-1}(x) dx^2]+ \frac{F(y)G(x)}{\Lambda\Psi A^2(x-y)^2} d\varphi^2, \nonumber\end{aligned}$$ $$\label{Ernstf} e^{-2\phi} = e^{-2\phi_0} \frac{\Lambda}{\Psi} \left(\frac{1+\Sigma A y}{1- \Sigma Ay}\right) \left(\frac{1-\Sigma Ax}{1+\Sigma A x} \right),$$ $$A_{\varphi} = -\frac{e^{\phi_0}}{B\Lambda}\left(1+\frac{B\beta x}{1- \Sigma Ax}\right)+k,$$ $$B_{\varphi} = -\frac{e^{-\phi_0}}{E \Psi} \left(1+ \frac{E \alpha x}{1 + \Sigma Ax}\right)+k',$$ $$G_{\mu\nu} = \partial_{[\mu} A_{\nu]},\, \tilde{F}_{\mu\nu} = \partial_{[\mu} B_{\nu]},$$ where $$\begin{aligned} \Lambda &=& \left(1+\frac{B\beta x}{1-\Sigma Ax}\right)^2 \\ && +\frac{B^2(1- x^2-r_+ A x^3)(1+r_- A x)(1-\Sigma Ay)^2}{2A^2(x-y)^2(1-\Sigma A x)^2}, \nonumber\end{aligned}$$ $$\begin{aligned} \Psi &=& \left(1+\frac{E\alpha x}{1+\Sigma Ax}\right)^2 \\ &&+ \frac{E^2(1- x^2-r_+ A x^3)(1+r_- A x)(1+\Sigma Ay)^2}{2A^2 (x-y)^2(1+\Sigma A x)^2}, \nonumber\end{aligned}$$ $$\label{Cmfnsf} F(\xi) = 1-\Sigma^2 A^2 \xi^2,$$ $$\label{Cmfnsg} G(\xi) = \frac{(1-\xi^2-r_+ A \xi^3)(1+r_- A\xi)}{(1- \Sigma^2 A^2 \xi^2)},$$ and $$\begin{aligned} \label{alpha} \alpha^2 &=& \frac{1}{2} (r_+ -\Sigma)(r_- -\Sigma) +\frac{1}{2} A^2 \Sigma^3 (r_- -\Sigma)\\ &=& Q^2 + \frac{1}{2} A^2 \Sigma^3 (r_- - \Sigma), \nonumber\end{aligned}$$ $$\begin{aligned} \label{beta} \beta^2 &=& \frac{1}{2} (r_+ + \Sigma)(r_- + \Sigma) - \frac{1}{2} A^2 \Sigma^3 (r_- + \Sigma) \\ &=& P^2 -\frac{1}{2} A^2 \Sigma^3 (r_- + \Sigma). \nonumber\end{aligned}$$ As we will see below, this solution represents a pair of oppositely-charged black holes accelerating away from each other in a background field, although the coordinate system used here only includes one of the black holes. The black holes carry two magnetic gauge charges, and the background consists of two magnetic fields, which reduce to the fields in the $U(1)^2$ Melvin solution if we go to infinity along the axis of symmetry. The constants $\phi_0, k$, and $k'$ will be chosen so that the solution at infinity agrees with (\[dualMelvin\]). For $r_+ A < 2/(3\sqrt{3})$, the function $G(\xi)$ has four real roots, which I denote in ascending order by $\xi_1,\, \xi_2,\, \xi_3,\, \xi_4$. It is convenient to define another function $H(\xi) = G(\xi) F(\xi)$, so that I may write $$H(\xi) = -(r_+A)(r_-A)(\xi-\xi_1)(\xi-\xi_2)(\xi-\xi_3)(\xi-\xi_4).$$ I restrict the parameters so that $\xi_1 = -1/r_- A$ and $\xi_1 \leq \xi_2 \leq \xi_3 < \xi_4$. The surface $y = \xi_0 \equiv -1/\|\Sigma\| A$ is singular; this is the singular surface inside the black hole, that is, the singular surface at $r=\|\Sigma\|$ in the black hole solutions (\[Kbhole1\]). As $r_- \geq \|\Sigma\|,\, \xi_1 \geq \xi_0$. The surfaces $y=\xi_1,\, y=\xi_2$ are the inner and outer black hole horizons, and $y=\xi_3$ is the acceleration horizon for an observer comoving with the black hole. The coordinates $(x,\varphi)$ are angular coordinates which cover two-spheres around the black hole, except when $y=\xi_3$. So that the metric has the appropriate signature, $x$ is restricted to the range $\xi_3 \leq x \leq \xi_4$ in which $G(x)$ is positive. At $x=\{\xi_3 ,\, \xi_4\}$, the norm of $\partial/\partial \varphi$ vanishes, so these points are interpreted as the poles of the two-spheres; that is, the axis of symmetry is $x=\xi_3,\xi_4$, with $x=\xi_3$ pointing at infinity, and $x=\xi_4$ pointing at the other black hole. There is a divergence in the metric at $x=y$, which is interpreted as the point at infinity, so $y$ is restricted to the range $\xi_0 < y < x$. Spatial infinity is reached only along the axis, that is, when $y=x=\xi_3$, and null or timelike infinity when $y=x\neq \xi_3$ [@ashtekar]. This solution has a manifest duality symmetry under $$\label {Ernstsym} \tilde{F} \leftrightarrow G,\, \phi \leftrightarrow -\phi,$$ and $$Q \leftrightarrow P,\, \Sigma \leftrightarrow -\Sigma,\, B \leftrightarrow E,\, k \leftrightarrow k',\, \phi_0 \leftrightarrow -\phi_0.$$ As in the Ernst solution [@Ernst], the background fields provide the force necessary to accelerate the black holes. To eliminate the nodal singularities in this metric at $x=\xi_3$ and $x=\xi_4$ simultaneously, $A$ must be chosen so that[^1] $$\label{constraint2} G'(\xi_3) \Lambda(\xi_4) \Psi(\xi_4) = -G'(\xi_4) \Lambda(\xi_3) \Psi(\xi_3)$$ and we must take $\Delta \varphi=4\pi L^2 / G'(\xi_3)$, where I have introduced $L^2 = \Lambda(\xi_3) \Psi(\xi_3)$. In the limit $r_+ A \ll 1$, (\[constraint2\]) reduces to Newton’s law, $MA \approx BP + EQ$, and in general it determines the acceleration of the black holes in terms of the other parameters. If I set $r_+ = r_- =0$, (\[Ernst\]) becomes $$\begin{aligned} \label{accmel} ds^2 &= & \frac{\Lambda\Psi}{A^2(x-y)^2}[(1-y^2)dt^2-(1- y^2)^{-1} dy^2\\ && +(1-x^2)^{-1} dx^2] + \frac{1-x^2}{\Lambda\Psi A^2(x-y)^2} d\varphi^2, \nonumber\end{aligned}$$ where $$\Lambda = 1+ \frac{1}{2} B^2 \frac{1-x^2}{A^2(x-y)^2},$$ and $$\Psi = 1+ \frac{1}{2} E^2 \frac{1-x^2}{A^2(x-y)^2}.$$ This is just the $U(1)^2$ Melvin solution (\[dualMelvin\]) in non-standard coordinates [@2u1]. That is, the $U(1)^2$ Melvin solution is a special case of the $U(1)^2$ Ernst solution, where the black hole parameters are set to zero. The $U(1)^2$ Ernst solution (\[Ernst\]) also approaches (\[dualMelvin\]) at large spacelike distances, that is, when we go to infinity along the axis. Spatial infinity corresponds to $x, y \rightarrow \xi_3$, and in this limit it is convenient to use the change of coordinates given in [@dggh], $$\label{inftransf} x - \xi_3 = \frac{4 F(\xi_3) L^2}{G'(\xi_3)A^2}\frac{\rho^2}{(\rho^2+\zeta^2)^2},$$ $$\xi_3 -y = \frac{4 F(\xi_3)L^2}{G'(\xi_3) A^2} \frac{\zeta^2}{(\rho^2+\zeta^2)^2},$$ $$t= \frac{2 \eta}{G'(\xi_3)},\, \varphi = \frac{2L^2 \tilde{\varphi}}{G'(\xi_3)}.$$ For large $\rho^2+\zeta^2$, the $U(1)^2$ Ernst solution in these coordinates reduces to $$ds^2 \rightarrow \tilde{\Lambda}\tilde{\Psi} (-\zeta^2 d\eta^2 + d\zeta^2 + d\rho^2) + \frac{\rho^2 d\tilde{\varphi}^2}{\tilde{\Lambda}\tilde{\Psi}},$$ where $$\tilde{\Lambda} = (1+\frac{1}{2} \widehat{B}_E^2 \rho^2) \mbox{ with } \widehat{B}_E^2 =\frac{B^2 G'^2(\xi_3)}{4L^2\Lambda(\xi_3)},$$ and $$\tilde{\Psi} = (1+\frac{1}{2} \widehat{E}_E^2 \rho^2) \mbox{ with } \widehat{E}_E^2 = \frac{E^2 G'^2(\xi_3)}{4L^2\Psi(\xi_3)}.$$ If I now set $\hat{t} = \zeta \sinh \eta, z = \zeta \cosh \eta$, we once again regain (\[dualMelvin\]). For large $\rho^2+\zeta^2$, the dilaton and gauge fields tend to $$e^{-2\phi} \rightarrow e^{-2\phi_0} \frac{\Lambda(\xi_3)}{\Psi(\xi_3)} \frac{\tilde{\Lambda}}{\tilde{\Psi}},$$ $$A_{\tilde{\varphi}} \rightarrow e^{\phi_0} \frac{\Psi(\xi_3)^{1/2}}{\Lambda(\xi_3)^{1/2}} \frac{\hat{B}\rho^2}{ 2\tilde{\Lambda}},\, B_{\tilde{\varphi}} \rightarrow e^{-\phi_0} \frac{\Lambda(\xi_3)^{1/2}}{\Psi(\xi_3)^{1/2}} \frac{\hat{E} \rho^2}{2\tilde{\Psi}},$$ so if I set $e^{2\phi_0}=\Lambda(\xi_3)/\Psi(\xi_3)$, I recover (\[dMgauge\]) in this limit. I will take this to define $\phi_0$ in general. Thus, I recover the $U(1)^2$ Melvin solution at large spacelike distances, and this allows me to identify the physical strength of the background fields in the Ernst solution as $\widehat{E}_E$ and $\widehat{B}_E$. We can also calculate the physical charges on the black hole by integrating the field tensors over two-spheres surrounding the black holes. We find $$\widehat{P} = \frac{1}{4\pi} \int G = \frac{\Lambda(\xi_3) \Psi(\xi_3)^{3/2}}{ G'(\xi_3) \Lambda(\xi_4)^{1/2}} \frac{\beta (\xi_4 - \xi_3)}{ (1-\Sigma A \xi_4) (1 - \Sigma A \xi_3)} \label{physP}$$ and $$\widehat{Q} = \frac{1}{4\pi} \int \tilde{F} = \frac{\Psi(\xi_3) \Lambda(\xi_3)^{3/2}}{ G'(\xi_3) \Psi(\xi_4)^{1/2}} \frac{\alpha (\xi_4 -\xi_3)}{ (1+\Sigma A \xi_4) (1+\Sigma A \xi_3)}, \label{physQ}$$ where $\alpha$ and $\beta$ are given by (\[alpha\],\[beta\]). The solution (\[Ernst\]) describes two black holes accelerating away from each other, propelled by the background fields. Now we take the Euclidean section obtained by taking $\tau = it$ in (\[Ernst\]). Half the Euclidean section gives an instanton describing black hole pair production [@gwg; @garstrom]. There are three possible instantons: one describing pair production of non-extreme black holes, one describing pair production of extreme black holes, with $\xi_1=\xi_2$, and another special case when $\xi_2 = \xi_3$. We will not consider this last here, as it does not describe black hole pair production (see [@robb] for more details of this case). Let us first consider the non-extreme or wormhole instantons, [ i.e.]{}, $\xi_1<\xi_2< \xi_3$. In the Euclidean section, we must restrict $y$ to $\xi_2 \leq y \leq \xi_3$ to obtain a positive definite metric, and $(y,\tau)$ are now also coordinates on a two-sphere, except when $x=\xi_3$. We must impose another condition on the parameters to eliminate the possible conical singularities at the black hole horizon $y=\xi_2$ and the acceleration horizon $y=\xi_3$ simultaneously. Namely, the period of $\tau$ must be taken to be $\Delta \tau = 4\pi/\|G'(\xi_2)\|$, and we must set $$\|G'(\xi_2)\| = \|G'(\xi_3)\|,$$ where $G(\xi)$ is given by (\[Cmfnsg\]). This condition is satisfied by setting $$\left(\frac{\xi_2^2-\xi_0^2}{\xi_3^2-\xi_0^2}\right) \left( \frac{\xi_3-\xi_1}{\xi_2-\xi_1} \right)= \frac{\xi_4-\xi_2}{\xi_4 - \xi_3}. \label{instcond}$$ This condition provides a further restriction on the black hole parameters, which may be thought of as determining the mass of the black hole in terms of its charges. More precisely, we can solve it for $r_-A$ in terms of $r_+A$ and $\Sigma A$. The whole Euclidean section is a bounce, that is, an instanton–anti-instanton pair joined along a spacelike slice. The topology of the bounce is $S^2 \times S^2 - \{pt\}$, where the removed point is $x=y=\xi_3$. For the extremal instantons, when $\xi_1 = \xi_2$, we must take $\Delta \tau = 4\pi/\|G'(\xi_3)\|$ to ensure regularity at the acceleration horizon. The black hole event horizon is at infinite distance in all spatial directions, so we do not have to worry about a conical singularity there. The range of $y$ in the Euclidean section is now $\xi_2 < y \leq \xi_3$, so that $(y,\tau)$ are now polar coordinates on an $R^2$, except when $x=\xi_3$. The extremal bounce has topology $S^2 \times R^2 - \{pt\}$, and the instanton can be interpreted as creating a pair of extremal black holes, with infinitely long throats. However, I have found that, unlike the case with one $U(1)$ gauge field [@dggh], the extremal solutions do not become spherically symmetric near the event horizon, and therefore do not approach the static black hole solutions at this internal infinity. This can be most easily seen by computing the intrinsic curvature scalar $^2 R$ for the black hole horizon itself when the black holes are extremal, and calculating its numerical values at some typical horizon positions. I will omit the rather unilluminating formula for $^2 R$, and simply state that one finds that the curvature is larger at the poles than at the equator of the two-sphere. Since the horizon is not a round two-sphere, the solution cannot be spherically symmetric. The point of this is that, unlike the case with one gauge field, even the extremal black holes are accelerating in some sense. It would be interesting to see if this could be extended to a Kaluza-Klein theory with two gauge fields, as in the usual Kaluza-Klein theory there is a well-defined sense in which the extremal black holes move on geodesics, and are thus not accelerating. Another difference that it is worth highlighting is that, even once the no-strut condition (\[constraint2\]) and either (\[instcond\]) or $\xi_1 =\xi_2$ have been satisfied, there are still four parameters in the solution, the two charges $\widehat{Q}$ and $\widehat{P}$ of the black hole and the background field strengths $\widehat{B}_E$ and $\widehat{E}_E$. This means we have a lot more freedom than in the Einstein-Maxwell case, where we only had two parameters once the regularity constraints were satisfied. In particular, if $\widehat{Q}$ and $\widehat{P}$ have opposite signs, it is possible to take large values of $\widehat{B}_E$ and $\widehat{E}_E$ without producing very large accelerations. This implies that, unlike the case with one $U(1)$ gauge field [@dggh], there does not seem to be any universal bound on $\widehat{Q} \widehat{E}_E$ or $\widehat{P} \widehat{B}_E$. The pair creation rate {#so4sec} ====================== Having described the pair creation instanton, I now turn to the calculation of the pair creation rate. The principal results are that the pair creation rate for non-extreme black holes is enhanced over that for extreme black holes by $e^{{\cal A}_{bh}/4}$ (as in the Einstein-Maxwell case [@dggh]), the pair creation rate is always suppressed, and it reduces to the Schwinger result in the limit of small black holes. The $U(1)^2$ Ernst metric reduces to the Ernst metric when $\phi=0$, and to the dilaton Ernst metric when either $\tilde{F}$ or $G$ vanishes, and so I can check the calculation by showing that it agrees with the results of [@entarea; @dggh] in these cases. The amplitude for pair creation in a background field is given by the path integral $$\Psi = \int d[g] d[A] d[B] e^{-I}, \label{pint}$$ where the action $I$ in the path integral is the action (\[Kaction\]), and the integral is over all metrics and gauge fields which interpolate between the background fields at infinity and a spacelike slice which contains the pair of black holes. If there is an appropriate instanton, we assume that $\Psi$ will be approximately $\Psi \approx e^{-I}$, where $I$ is now the action of the instanton. The pair creation rate $\Gamma$ is given by the modulus squared of this amplitude, so it will be approximately $\Gamma \approx e^{-I_b}$, where $I_b$ is the action of the bounce. For the pair creation of black holes, the Euclidean sections of the solutions discussed in Sec. \[Esol\] are the bounces, so the calculation of the pair creation rates reduces to the problem of the calculation of the actions of these bounces. The simplest way to evaluate the action is by a Hamiltonian decomposition, following the techniques given in [@haho]. Since the solutions we are interested in are stationary, if the Euclidean section was of the form $\Sigma \times S^1$, where the $S^1$ factor represents the time direction, the action would just be given by $I = \beta H$, where $H$ is the Hamiltonian and $\beta = \Delta \tau$ is the period in imaginary time. However, the time-translation Killing vector has fixed points at the black hole event horizon and the acceleration horizon, so by doing this we have neglected a contribution from a neighbourhood of each horizon. Including the contributions from these corners, the total Euclidean action is (in the non-extreme case) $$\label{impeq} I = \beta H -\frac{1}{ 4} (\Delta {\cal A} + {\cal A}_{bh}),$$ where ${\cal A}_{bh}$ is the area of the black hole horizon, and $\Delta {\cal A}$ is the difference in area of the acceleration horizon between the solution and the background [@haho; @entarea]. In the extreme case, the term proportional to ${\cal A}_{bh}$ is absent, as the black hole event horizon is not part of the Euclidean section. The Hamiltonian $H$, which is only defined with respect to the background spacetime, can be expressed as [@haho] $$\label{hamil} H = \int_{\Sigma} N{\cal H} - \frac{1}{ 8\pi} \int_{S^\infty} N({}^2 K - {}^2K_0),$$ where $N$ is the lapse, ${\cal H}$ is the Hamiltonian constraint, ${}^2 K$ is the trace of the two dimensional extrinsic curvature of the boundary near infinity, and ${}^2 K_0$ is the analogous quantity for the background spacetime. On solutions, the constraint vanishes, and so the only non-zero contribution comes from the gravitational surface term. To calculate this surface term, we need to introduce a boundary near infinity, and calculate its extrinsic curvature in the instanton and the background solution. To ensure that the boundary used in both calculations is the same, I need to match the intrinsic features of the boundary; that is, the induced metric, the gauge field, and the value of the dilaton on the boundary. I take the boundary in the $U(1)^2$ Ernst solution to be $$\label{bern} x = \xi_3 + \epsilon_E \chi, \ \ y = \xi_3 + \epsilon_E (\chi-1),$$ where $0 \leq \chi \leq 1$, and make the coordinate transformations $$\label{changei} \varphi = \frac{ 2L^2 }{ G'(\xi_3)} \varphi',\ t = \frac{2 }{ G'(\xi_3)} t',$$ and I assume that the boundary in the $U(1)^2$ Melvin solution lies at $$\label{melbdry} x = -1 + \epsilon_M \chi[1 + \epsilon_E f(\chi)],$$ $$y = -1 +\epsilon_M (\chi-1) [1+\epsilon_E g(\chi)]$$ in the accelerated coordinate system (\[accmel\]). Other choices for the boundary in the $U(1)^2$ Melvin solution may be possible, but this is the only choice that I have been able to explicitly carry out. We evaluate all quantities to second nontrivial order in $\epsilon_E$, as higher-order terms will not affect the result in the limit $\epsilon_E \to 0$. For the $U(1)^2$ Ernst metric, the induced metric on the boundary is $$\begin{aligned} {}^{(2)} ds^2 &=& \frac{ L^2 F(\xi_3) }{ A^2 \epsilon_E G'(\xi_3)} \left\{ - \frac{ \lambda \psi d \chi^2 }{ \chi (\chi-1)} \left[ 1+\epsilon_E (2 \chi -1) \frac{F'(\xi_3) }{ F(\xi_3)} \right] \right. \\ \nonumber &&+ \left. \frac{4 \chi }{\lambda \psi} \left[1 + \epsilon_E \chi \frac{H''(\xi_3) }{ 2 H'(\xi_3)} -\epsilon_E \frac{F'(\xi_3) }{ F(\xi_3)} \right] d\varphi'^2 \right\}, \label{emetric}\end{aligned}$$ where $$\label{lambda} \lambda = 1 + \frac{ 2 L^2 \widehat{B}_E^2 F(\xi_3)\chi}{ A^2 \epsilon_E G'(\xi_3)} \left( 1 + \frac{1}{2} \epsilon_E \chi \frac{H''(\xi_3)}{H'(\xi_3)} + \frac{2 \Sigma A \epsilon_E}{1-\Sigma A \xi_3} \right)$$ and $$\label{psi} \psi = 1 + \frac{ 2 L^2 \widehat{E}_E^2 F(\xi_3) \chi}{ A^2 \epsilon_E G'(\xi_3)} \left( 1 + \frac{1}{2} \epsilon_E \chi \frac{H''(\xi_3)}{H'(\xi_3)} - \frac{2 \Sigma A \epsilon_E}{1 + \Sigma A \xi_3} \right).$$ The gauge potentials on the boundary for the $U(1)^2$ Ernst solution are $$\label{fgauge} A_{\varphi'} = \frac{ L e^{\phi_0}}{\Lambda(\xi_3) \widehat{B}_E} \left[ 1 - \frac{A^2 \epsilon_E G'(\xi_3)}{2 L^2 F(\xi_3) \widehat{B}_E^2 \chi} \right]$$ and $$\label{ggauge} B_{\varphi'} = \frac{ L e^{-\phi_0}}{\Psi(\xi_3) \widehat{E}_E} \left[ 1 - \frac{A^2 \epsilon_E G'(\xi_3)}{2 L^2 F(\xi_3) \widehat{E}_E^2 \chi} \right].$$ The dilaton at the boundary is $$\label{dilb} e^{-2\phi} = e^{-2\phi_0} \frac{\Lambda(\xi_3) \lambda}{\Psi(\xi_3) \psi} \left( 1 -\frac{2 \Sigma A \epsilon_E}{1 - \Sigma^2 A^2 \xi_3^2} \right).$$ For the $U(1)^2$ Melvin solution, the induced metric on the boundary is $$\begin{aligned} \label{bmetM} {}^{(2)} ds^2 &=& \frac{-\Lambda \Psi }{ 2 \chi (\chi-1) \bar{A}^2 \epsilon_M} \left\{ 1 - \epsilon_E (\chi-1) f(\chi) + \epsilon_E \chi g(\chi) \right. \\ \nonumber && -2 \epsilon_E \chi (\chi-1) [f'(\chi) -g'(\chi)] - \left. 2 \epsilon_E [\chi f(\chi) - (\chi-1) g(\chi)] \right\} d\chi^2 \\ \nonumber && +\frac{2 \chi }{\Lambda \Psi \bar{A}^2 \epsilon_M}\left\{ 1 -\frac{1}{2} \epsilon_M \chi + \epsilon_E f(\chi) -2 \epsilon_E [\chi f(\chi) - (\chi-1) g(\chi)] \right\} d\varphi^2,\end{aligned}$$ where $$\begin{aligned} \label{lambdaM} \Lambda &=& 1+\frac{ \widehat{B}_M^2 \chi }{ \bar{A}^2 \epsilon_M} \left\{ 1 - \frac{1}{ 2} \epsilon_M \chi + \epsilon_E f(\chi) \right. \\ \nonumber &&- \left. 2 \epsilon_E [\chi f(\chi) - (\chi-1) g(\chi)]\right\}\end{aligned}$$ and $$\begin{aligned} \label{psiM} \Psi &=& 1+\frac{ \widehat{E}_M^2 \chi }{\bar{A}^2 \epsilon_M} \left\{ 1 - \frac{1}{ 2} \epsilon_M \chi + \epsilon_E f(\chi) \right. \\ \nonumber &&- \left. 2 \epsilon_E [\chi f(\chi) - (\chi-1) g(\chi)]\right\} .\end{aligned}$$ The gauge potentials on the boundary in $U(1)^2$ Melvin are $$\label{mgauge} A_{\varphi} = \frac{1}{ \widehat{B}_M} \left[ 1 - \frac{ \bar{A}^2 \epsilon_M}{\widehat{B}_M^2 \chi} \right]$$ and $$\label{mgauge2} B_{\varphi} = \frac{1}{ \widehat{E}_M} \left[ 1 - \frac{ \bar{A}^2 \epsilon_M}{\widehat{E}_M^2 \chi} \right],$$ and the dilaton at the boundary in $U(1)^2$ Melvin is $$\label{mdil} e^{-2 \phi} = \frac{\Lambda}{\Psi},$$ where $\Lambda$ is given by (\[lambdaM\]) and $\Psi$ is given by (\[psiM\]). I fix the remaining coordinate freedom by taking $$\label{abar2} \bar{A}^2 = -\frac{G'(\xi_3) }{ 2 L^2 F(\xi_3)}\frac{H'(\xi_3) }{ H''(\xi_3)}A^2,$$ and write $$\label{expang} e^{\phi_0} = \frac{\Lambda(\xi_3)^{1/2}}{\Psi(\xi_3)^{1/2}} \left(1 - \gamma \epsilon_E \right), \widehat{B}_M = \widehat{B}_E \left(1+\alpha \epsilon_E \right), \widehat{E}_M = \widehat{E}_E \left(1 + \beta \epsilon_E \right).$$ I then find that the intrinsic metric, gauge potentials and dilaton on the boundary can all be matched by taking $$\label{subl1} \epsilon_M = -\frac{H''(\xi_3)}{H'(\xi_3)} \epsilon_E,$$ $$\label{subl2} f(\chi) = \frac{F'(\xi_3) }{ F(\xi_3)} (4\chi -3),\ \ g(\chi) = \frac{F'(\xi_3) }{ F(\xi_3)} (4\chi-1),$$ and $$\label{subl3} \gamma = \alpha = -\beta = \frac{\Sigma A}{1-\Sigma^2 A^2 \xi_3^2}.$$ Note that the lapse function is also matched by these conditions. For the $U(1)^2$ Ernst metric, the lapse function at the boundary is given by $$N = \left[\frac{4 L^2 F(\xi_3)(1-\chi)\lambda \psi}{ A^2 \epsilon_E G'(\xi_3)}\right]^{\frac{1}{2}} \left[ 1 + \frac{1}{4} \epsilon_E (\chi-1) \frac{H''(\xi_3) }{ H'(\xi_3)} + \frac{1}{2} \epsilon_E \frac{F'(\xi_3) }{ F(\xi_3)} \right], \label{lapseE}$$ While the lapse function for the $U(1)^2$ Melvin metric is $$\begin{aligned} \label{lapseM} N &=& \left[\frac{2(1-\chi)\Lambda \Psi}{ \bar{A}^2 \epsilon_M}\right]^{\frac{1}{2}} \left\{ 1 - \frac{1}{4} \epsilon_M (\chi-1)+ \frac{1}{2} \epsilon_E g(\chi) \right. \\ \nonumber && \left. - \epsilon_E [\chi f(\chi) - (\chi-1) g(\chi)] \phantom{\frac{1}{2}} \right\},\end{aligned}$$ so we see that the matching conditions (\[abar2\]-\[subl3\]) make (\[lapseE\]) and (\[lapseM\]) equal as well. The extrinsic curvature of this boundary embedded in the $U(1)^2$ Ernst solution is $$\begin{aligned} \label{excurv2} {}^2 K &=& \frac{A \epsilon_E^{1/2} G'(\xi_3)^{1/2} }{ L F(\xi_3)^{1/2} \lambda \psi} \left[ 1+ \frac{1}{4} \epsilon_E \frac{H''(\xi_3) }{ H'(\xi_3)} (4 \chi -3) \right. \\ \nonumber &&- \left.\frac{1 }{2}\epsilon_E \frac{F'(\xi_3) }{ F(\xi_3)} (4 \chi -3) \right],\end{aligned}$$ while the extrinsic curvature of the boundary embedded in the $U(1)^2$ Melvin solution is $$\begin{aligned} \label{excurvM2} {}^2 K_0 &=& \frac{\bar{A} \epsilon_M^{1/2} \sqrt{2} }{ \Lambda \Psi} \left[ 1 -\frac{1}{4} \epsilon_M (4\chi-3) \right. \\ \nonumber &&- \left. \frac{1 }{ 2} \epsilon_E \frac{F'(\xi_3) }{ F(\xi_3)} (24\chi -13) \right].\end{aligned}$$ Using the matching conditions (\[abar2\]-\[subl3\]), one may now evaluate $$\label{ecdiff} {}^2 K - {}^2 K_0 = \frac{5 A \epsilon_E^{3/2} G'(\xi_3)^{1/2} }{ L F(\xi_3)^{1/2} \lambda \psi} \frac{F'(\xi_3) }{ F(\xi_3)} (2\chi -1).$$ Therefore, taking the limit $\epsilon_E \rightarrow 0$, the Hamiltonian is $$\label{hameval} H_E = -\frac{1}{4} \int_0^1 d \chi N \sqrt{h} (^2 K - {}^2 K_0) = - \frac{5 L^2 F'(\xi_3) }{ A^2 G'(\xi_3)} \int_0^1 d\chi (2 \chi -1) =0.$$ The action is thus given by $$I = -\frac{1}{4} (\Delta {\cal A} + {\cal A}_{bh} ) \label{keyeq}$$ when the black holes are non-extremal, and by $$I = -\frac{1}{4} \Delta {\cal A} \label{keyeqq}$$ if the black holes are extremal. Note that the action in the non-extreme case is less than the action in the extreme case by $-\frac{1}{4} {\cal A}_{bh}$, and thus the pair creation rate for non-extreme black holes in enhanced by $e^{{\cal A}_{bh}/4}$ over that for extreme black holes. A natural interpretation of this result is that the non-extreme black holes have $e^{{\cal A}_{bh}/4}$ more states than the extreme ones, as in the Einstein-Maxwell case [@entarea]. As the difference in entropy between the non-extreme and extreme solutions is also $\frac{1}{4} {\cal A}_{bh}$ [@entarea], this suggests that the entropy is a reliable guide to the number of states, that is, the number of states of a black hole $\sim e^{S_{bh}}$. I now proceed to calculate the right hand side of (\[keyeq\]) and (\[keyeqq\]). The area of the black hole horizon is $$\label{bhhoriz} {\cal A}_{bh} = \int_{y=\xi_2} \sqrt{g_{xx} g_{\varphi\varphi}} dx d\varphi = \frac{4\pi F(\xi_2) L^2}{A^2 G'(\xi_3)} \frac{(\xi_4-\xi_3)}{(\xi_3-\xi_2)(\xi_4-\xi_2)}.$$ Again, in the calculation of the difference in area of the acceleration horizon, I need to introduce a boundary, in this case a circle, at large distances, and match the intrinsic features of this boundary. The area of the acceleration horizon in the $U(1)^2$ Ernst spacetime up to a large circle at $x = \xi_3 + \epsilon_E$ is $$\label{eaccel} {\cal A}_E = \int_{y=\xi_3} \sqrt{g_{xx} g_{\varphi\varphi}} dx d\varphi = - \frac{4 \pi L^2 F(\xi_3)}{A^2 G'(\xi_3) (\xi_4 - \xi_3)} + \pi \rho_E^2,$$ where $\rho_E^2 = 4 F(\xi_3) L^2/[G'(\xi_3) A^2 \epsilon_E]$. The area of the acceleration horizon in the $U(1)^2$ Melvin spacetime inside a circle at $\rho = \rho_M$ is ${\cal A}_M = \pi \rho_M^2$. Now I need to match the proper length of the boundary, the integral of both gauge potentials around the boundary, and the value of $\phi$ at the boundary. The proper length of the boundary in the $U(1)^2$ Ernst solution is $$\begin{aligned} \label{lengE} l_E &= &\frac{4\pi}{\widehat{E}_E \widehat{B}_E \rho_E} \left[ 1 - \frac{F(\xi_3) L^2}{G'(\xi_3) A^2} \frac{H''(\xi_3) }{H'(\xi_3)} \frac{1}{\rho_E^2} + \frac{2 F(\xi_3) L^2}{ G'(\xi_3) A^2} \frac{F'(\xi_3)}{F(\xi_3)} \frac{1}{\rho_E^2} \right. \\ \nonumber && \left. - \frac{1}{\widehat{E}_E^2 \rho_E^2} - \frac{1}{\widehat{B}_E^2 \rho_E^2} \right],\end{aligned}$$ while the proper length of the boundary in the $U(1)^2$ Melvin solution is $$\label{lengM} l_M = \frac{4 \pi}{\widehat{E}_M \widehat{B}_M \rho_M} \left( 1 -\frac{1}{\widehat{E}_M^2 \rho_M^2} - \frac{1}{ \widehat{B}_M^2 \rho_M^2} \right).$$ The integral of the gauge potentials around the boundary are, in the $U(1)^2$ Ernst solution, $$\label{pot1E} \oint A_\varphi d\varphi = \frac{2 \pi}{\widehat{B}_E} \frac{ e^{\phi_0} \Psi(\xi_3)^{1/2}}{\Lambda(\xi_3)^{1/2}} \left( 1 - \frac{2}{\widehat{B}_E^2 \rho_E^2} \right)$$ and $$\label{pot2E} \oint B_\varphi d\varphi = \frac{2 \pi}{\widehat{E}_E} \frac{ e^{-\phi_0} \Lambda(\xi_3)^{1/2}}{\Psi(\xi_3)^{1/2}} \left( 1 - \frac{2}{\widehat{E}_E^2 \rho_E^2} \right).$$ while in the $U(1)^2$ Melvin solution, they are $$\label{pot1M} \oint A_\varphi d\varphi = \frac{2 \pi}{\widehat{B}_M} \left( 1 - \frac{2}{\widehat{B}_M^2 \rho_M^2} \right)$$ and $$\label{pot2M} \oint B_\varphi d\varphi = \frac{2 \pi}{\widehat{E}_M} \left( 1 - \frac{2}{\widehat{E}_M^2 \rho_M^2} \right).$$ The dilaton at the boundary is $$\begin{aligned} \label{dilatE} e^{-2\phi} &=& e^{-2\phi_0} \frac{\Lambda(\xi_3) \widehat{B}_E^2}{\Psi(\xi_3) \widehat{E}_E^2} \left[ 1 + \frac{2 \Sigma A}{1-\Sigma^2 A^2 \xi_3^2} \frac{4 F(\xi_3) L^2}{G'(\xi_3) A^2 \rho_E^2} \right. \\ \nonumber && \left. + \frac{2}{\widehat{B}_E^2 \rho_E^2} - \frac{2}{\widehat{E}_E^2 \rho_E^2} \right]\end{aligned}$$ in the $U(1)^2$ Ernst solution, and $$\label{dilatM} e^{-2\phi} = \frac{\widehat{B}_M^2}{\widehat{E}_M^2} \left( 1 + \frac{2}{\widehat{B}_M^2 \rho_M^2} - \frac{2}{\widehat{E}_M^2 \rho_M^2} \right)$$ in the $U(1)^2$ Melvin solution. Now $e^{\phi_0}, \widehat{B}_M$ and $\widehat{E}_M$ are given by (\[expang\]) and (\[subl3\]), and we may see that we can match the proper length of the boundary, the integrals of the gauge fields and the dilaton if we also take $$\label{rho} \rho_M = \rho_E \left\{ 1+ \frac{1}{\rho_E^2} \frac{F(\xi_3) L^2}{G'(\xi_3) A^2} \left[ \frac{H''(\xi_3)}{H'(\xi_3)} - \frac{2 F'(\xi_3)}{F(\xi_3)} \right] \right\}.$$ This implies that the difference in horizon area is $$\begin{aligned} \label{diffha} \Delta {\cal A} &=& -\frac{4 \pi L^2 F(\xi_3)}{G'(\xi_3) A^2} \left[ \frac{1}{\xi_4 - \xi_3} + \frac{H''(\xi_3)}{2H'(\xi_3)} - \frac{F'(\xi_3)}{F(\xi_3)} \right] \\ \nonumber &=& - \frac{4 \pi L^2 F(\xi_3)}{G'(\xi_3) A^2} \left[ \frac{(\xi_2-\xi_1)}{(\xi_3-\xi_2)(\xi_3-\xi_1)} + \frac{2}{(\xi_3 -\xi_1)} -\frac{F'(\xi_3)}{F(\xi_3)} \right].\end{aligned}$$ For the extreme case, we therefore have $$- \frac{1}{4} \Delta {\cal A} = \frac{2 \pi L^2 F(\xi_3)}{ G'(\xi_3) A^2} \left[ \frac{1}{\xi_3- \xi_1} - \frac{F'(\xi_3)}{2 F(\xi_3)} \right],$$ while for the non-extreme case, we have $$\begin{aligned} -\frac{1}{4} (\Delta {\cal A} + {\cal A}_{bh}) &=& \frac{\pi L^2 F(\xi_3)}{G'(\xi_3) A^2} \left[ \frac{2}{\xi_3- \xi_1} - \frac{F'(\xi_3)}{ F(\xi_3)} \right. \\ \nonumber && \left. + \frac{(\xi_2-\xi_1)}{(\xi_3-\xi_2)(\xi_3-\xi_1)} - \frac{ F(\xi_2) (\xi_4-\xi_3)}{F(\xi_3) (\xi_3-\xi_2) (\xi_4 - \xi_2)} \right] \\ \nonumber &=& \frac{2 \pi L^2 F(\xi_3)}{ G'(\xi_3) A^2} \left[ \frac{1}{\xi_3- \xi_1} - \frac{F'(\xi_3)}{2 F(\xi_3)} \right],\end{aligned}$$ where I have used the instanton condition (\[instcond\]) to cancel the last two terms. Thus, I deduce that the action must be $$\label{actionres} I_{b} = \frac{2 \pi L^2 F(\xi_3)}{ G'(\xi_3) A^2} \left[ \frac{1}{\xi_3- \xi_1} - \frac{F'(\xi_3)}{2 F(\xi_3)} \right]$$ in both cases. This answer agrees with the action of the Ernst solution found in [@ggs; @entarea] when $\Sigma =0$ (which implies $F(\xi)=1$), as it should. It also reduces to the answer for the action of the dilaton Ernst solution found in [@dggh; @entarea] when either $Q=0$ or $P=0$. Thus, this result is consistent with the previously-obtained results. The point-particle limit is $r_+ A \ll 1$, as the black hole becomes small on the scale set by the acceleration in this limit. In this limit, both the extreme and non-extreme black holes satisfy $r_+ \approx r_-$ [@2u1]. When $r_+ A \ll 1$, the action reduces to $$I_b \approx \frac{\pi r_-}{A} \approx \frac{\pi M^2}{BP + EQ}, \label{lowact}$$ where I have used Newton’s law in the second step. The pair creation rate is $e^{-I_b}$, so we recover the Schwinger result (generalised to the case of two gauge fields) in this limit, as we would expect. That is, we find that small black holes are pair created at the same rate (to leading order) as we would expect for some hypothetical particles carrying the same mass and charges. In particular, the pair creation rate will be very small for realistic fields, as we must have $M> M_{pl}$ for this semi-classical approximation to be valid. Because of the number of parameters involved, it is difficult to say anything more about the general behaviour of this action, but the qualitative agreement with [@entarea] is remarkable, given the much more complicated nature of this solution, and the presence of twice as many free parameters. Acknowledgements ================ I am happy to acknowledge many helpful conversations with friends and colleagues, particularly Stephen Hawking, who I also thank for his comments on an earlier draft of this paper. I thank the Association of Commonwealth Universities and the Natural Sciences and Engineering Research Council of Canada for financial support. [99]{} D. Garfinkle, S.B. Giddings and A. Strominger, Phys. Rev. D [49]{}, 958 (1994). H.F. Dowker, J.P. Gauntlett, D.A. Kastor and J. Traschen, Phys. Rev. D [49]{}, 2909 (1994). H.F. Dowker, J.P. Gauntlett, S.B. Giddings and G.T. Horowitz, Phys. Rev. D [50]{}, 2662 (1994). R. Kallosh, A. Linde, T. Ortín, A. Peet, and A. Van Proeyen, Phys. Rev. D [46]{}, 5278 (1992). E. Cremmer, J. Scherk, and S. Ferrara, Phys. Lett. [74B]{}, 61 (1978). G.W. Gibbons, Nucl. Phys. [B 207]{}, 337 (1982); G.W. Gibbons and K. Maeda, [ibid.]{} [298]{}, 741 (1988). S.F. Ross, Phys. Rev. D [49]{}, 6599 (1994). S.W. Hawking, G.T. Horowitz and S.F. Ross, Phys. Rev. D [51]{}, 4302 (1995). D. Garfinkle, G.T. Horowitz, and A. Strominger, Phys. Rev. D [43]{}, 3140 (1990), erratum D [45]{}, 3888(E) (1992). F.J. Ernst, J. Math. Phys. [17]{}, 515 (1976). M.A. Melvin, Phys. Lett. [8]{}, 65 (1964). A. Ashtekar and T. Dray, Comm. Math. Phys. [79]{}, 581 (1981); T. Dray, Gen. Rel. Grav. [14]{}, 109 (1982). G.W. Gibbons, in [Fields and Geometry 1986]{}, Proceedings of the 22nd Karpacz Winter School of Theoretical Physics, Karpacz, Poland, edited by A. Jadczyk (World Scientific, Singapore, 1986). D. Garfinkle and A. Strominger, Phys. Lett. [B256]{}, 146 (1991). R.B. Mann and S.F. Ross, Phys. Rev. D [52]{}, 2254 (1995). S.W. Hawking and G. Horowitz, “The gravitational Hamiltonian, action, entropy and surface terms”, DAMTP/R/94-52, gr-qc/9501014. [^1]: Note that $\Lambda(\xi_i) \equiv \Lambda(x=\xi_i)$ and $\Psi(\xi_i) \equiv \Psi(x=\xi_i)$ are constants.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Motivated by the celebrated Schoen-Yau-Gromov-Lawson surgery theory on metrics of positive scalar curvature, we construct a double manifold associated with a minimal isoparametric hypersurface in the unit sphere. The resulting double manifold carries a metric of positive scalar curvature and an isoparametric foliation as well. To investigate the topology of the double manifolds, we use K-theory and the representation of the Clifford algebra for the FKM-type, and determine completely the isotropy subgroups of singular orbits for homogeneous case.' address: - 'School of Mathematical Sciences, Laboratory of Mathematics and Complex Systems, Beijing Normal University, Beijing 100875, China' - 'Department of Mathematics, Hangzhou Normal University, Hangzhou, Zhejiang 310036, China; School of Mathematical Sciences, Peking University, Beijing 100871, China' - 'School of Mathematical Sciences, Laboratory of Mathematics and Complex Systems, Beijing Normal University, Beijing 100875, China' author: - Zizhou Tang - Yuquan Xie - Wenjiao Yan title: 'Schoen-Yau-Gromov-Lawson theory and isoparametric foliations' --- Dedicated to Professor Yuanlong Xin on his 70th birthday. Introduction ============ One of the simplest invariants of a Riemannian manifold is its scalar curvature function. Here, we say an $n$-dimensional manifold $M$ carries a metric of positive scalar curvature $R_M$ if $R_M\geq 0$ and $R_M(p) > 0$ for some point $p\in M $. Then a natural question to raise is “Which manifolds admit Riemannian metrics of positive scalar curvature?" In recent decades, this subject has been the focus of lively research. The first important contribution to this subject was made by A. Lichnerowicz ([@Lic]) in 1962, who showed that a compact spin manifold with non-vanishing $\widehat{A}$-genus cannot carry a Riemannian metric of positive scalar curvature. N. Hitchin ([@Hit]) generalized this result. More precisely, he used a ring homomorphism $\alpha$, constructed by J. Milnor, from $\Omega_^{\text{spin}}$, the spin cobordism ring, to $KO^{-}(\text{pt})$, and proved that $\alpha(M)$ vanishes if $M$ carries a metric of positive scalar curvature. When $\dim M \equiv 0~ (\mod 4)$, $\alpha(M)$ can be identified with $\widehat A(M)$ (up to a factor), so this recovers the result of Lichnerowicz. One surprising and beautiful result of this study was that half of the exotic spheres in dimensions $8k+1$ and $8k+2$ cannot carry metrics of positive scalar curvature. Another remarkable step toward answering the question above was made when Schoen and Yau ([@SY]), and independently, Gromov and Lawson ([@GL1]) established the following “surgery theorem" on metrics of positive scalar curvature: Theorem [Let $X$ be a compact manifold which carries a Riemannian metric of positive scalar curvature. Then any manifold which can be obtained from $X$ by performing surgeries in codimension $\geq3$ also carries a metric of positive scalar curvature. ]{} Inspired by Schoen-Yau-Gromov-Lawson’s surgery theory, we will construct a new manifold with rich geometrical properties from a Riemannian manifold with an embedding hypersurface. In particular, we implement this construction on a unit sphere with a minimal isoparametric hypersurface, finding that the new manifold admits not only complicated topology, but also a metric of positive scalar curvature. Moreover the isoparametric foliation is kept. Details of the construction are given in the following. Given a compact, connected manifold $X^n$ ($n \geq 3$) without boundary. Let $Y^{n-1}\hookrightarrow X^n$ be a connected embedding hypersurface with a trivial normal bundle, and $\pi_0(X- Y) \neq 0$, i.e., the complement of $Y$ in $X$ is not connected. Then $Y^{n-1}$ separates $X^n$ into two components, say $X^{n}_{+}$ and $X^n_{-}$, with the same boundary $Y^{n-1}$. (The assumption $\pi_0(X- Y) \neq 0$ is necessary. For example, $T^2-S^1$, removing a latitude circle from the torus, is connected.) Since $Y$ has a trivial normal bundle in $X$, we can choose a unit normal vector field $\xi$ on $Y$, which is an interior normal direction with respect to $X_+$. Define a continuous function $r: X^n \longrightarrow \mathbb{R}$ $$x \mapsto \left\{ \begin{array}{ll} ~~\mathrm{dist}(x, Y), & \textrm{if $x \in X_+$}\\ -\mathrm{dist}(x, Y), & \textrm{if $x \in X_-$} \end{array} \right.$$ where $\mathrm{dist}(x, Y)$ means the distance from $x$ to $Y$. Clearly, $X_+$ ($X_-$) is just the subset that $r\geq 0$ (resp. $r\leq 0$). Let $Y_r:=\{x \in X\|~ r(x)=r\}$ for $\|r\|$ so small that $Y_r$ is still an embedding hypersurface. We extend $\xi$ to a unit vector field in a neighborhood of $Y$ such that $\xi$ is normal to $Y_r$. From now on, without loss of generality, we only deal with $X^n_+$. Concerning with the Riemannian product space $X^n_+\times \mathbb{R}$ with coordinates $(x, t)$, for a small number $\bar{r}>0$, we can define a hypersurface $M^n$ in $X^n_+\times \mathbb{R}$ as [@GL1] $$M^n := \{(x,t)\in X^n_+\times \mathbb{R}~:~(r(x), t)\in \gamma, ~r(x)\leq \bar{r}\}$$ where $\gamma$ is a curve in the $(r,t)$-plane as pictured below: \[N\] ![image](N){height="46mm"} $$N:~unit~``exterior"~normal~vector~field~on~M,~~\sin\theta:=\langle N, \xi\rangle$$ The curve $\gamma$ begins at one end with a vertical line segment $t=0$, $r_1\leq r\leq \bar{r}$, and ends with a horizontal line segment $r=r_\infty >0$, with $r_\infty$ small enough as we will require. Now fix a point $q=(x,t)\in M$ corresponding to $(r, t)\in \gamma$. Choose an orthonormal basis $e_1, e_2,...,e_{n-1}$ of $T_xY_r$ such that the shape operator $A_{\xi}$ is expressed as $A_{\xi}e_i=\mu_i(r)e_i$ for $i=1,...,n-1$. Then the associated principal curvatures of $M$ at $q$ are of the form $\lambda_i=\mu_i(r) \sin\theta$ for $i=1,...,n-1$. As observed by Gromov-Lawson ([@GL1]), the tangent vector of $M\cap (l\times \mathbb{R})$ is also a principal direction for the second fundamental form of $M$ in $X^n_+\times \mathbb{R}$, where $l$ is a geodesic ray in $X_+$ emanating from $Y$. We denote this tangent vector by $e_n$. Thus the $n$-th principal curvature at $q$ is $\lambda_n:= k$, the (nonnegative) curvature of $\gamma$ at $(r, t)$. Look at the Gauss equation: $$K_{ij}^M=K_{ij}^{X\times\mathbb{R}} + \lambda_i\lambda_j,\qquad1\leq i,j\leq n,$$ where $K_{ij}^M$ is the sectional curvature of $M$ of the plane $e_i\wedge e_j$, and $K_{ij}^{X\times\mathbb{R}}$ is the corresponding sectional curvature of $X_+\times\mathbb{R}$. Since the metric of $X_+\times\mathbb{R}$ is the product metric, we see: $$\begin{aligned} &&K_{ij}^{X\times\mathbb{R}}=K_{ij}^X, \qquad1\leq i,j\leq n-1,\\ &&K_{n,j}^{X\times\mathbb{R}}=K_{\xi,j}^X\cos^2\theta,\end{aligned}$$ where $K^X$ is the sectional curvature of $X_+$. It follows immediately that the scalar curvature of $M$ with the induced metric can be expressed as: $$\label{scalar} R_M=\sum_{i\neq j}^nK_{ij}^M =R_X+2A\sin^2\theta+2kH(r)\sin\theta$$ where $$A:=\sum_{i< j\leq n-1}\mu_i(r)\mu_j(r) -Ric^X(\xi,\xi);\quad H(r)=\sum_{i=1}^{n-1}\mu_i(r), \emph{ ~the ~mean~ curvature~ of}~ Y_r.$$ \[scalar1\] Since $\gamma$ ends with a horizontal line segment, $M$ has the standard product metric as $t$ goes to infinity. This guarantees that we can glue $X_+$(resp.$X_-$) smoothly in metric with a copy of itself along $Y$ to get a new manifold called the double of $X_+$ (resp.$X_-$), and denoted by $D(X_+)$(resp.$D(X_-)$). The double of a manifold with boundary, as a topological concept, appeared in 1930’s. Gromov-Lawson ([@GL2]) studied the geometric property of the double of a manifold, and showed an interesting theorem which states that if $X$ carries a metric of positive scalar curvature, and $Y$ is a minimal hypersurface, then the double manifold $D(X_+)$(resp.$D(X_-)$) also carries a metric of positive scalar curvature. But in their construction of the double manifold, they “bent" too much near the boundary of $X_+$, inducing some singularities or creases. However in our method, an explicit construction of $D(X_+)$(resp.$D(X_-)$) with satisfactory properties is given. \[scalar2\] Formula (\[scalar\]) is the expression of the scalar curvature of $M$ in $X_+\times\mathbb{R}$. It also holds in $X_-\times\mathbb{R}$ although $\xi$ is the exterior normal vector field of $X_-$. Gromov and Lawson([@GL1]) studied the scalar curvature of $M$. Their formula is expressed in form of the estimate of principal curvatures, while ours is an explicit expression. Their main result on surgery is of course correct although in their formula $(1)$ ([@GL1]) they lost a factor $2$. In addition, in $(1')$ they missed one item related to the second fundamental form of the submanifold. But this mistake would result in the missing of the item $H(r)$ in our formula (\[scalar\]), which is, however, essential for our research. Rosenberg and Stolz([@RS]) modified Gromov-Lawson’s expression, but they also lost the principal curvatures or the second fundamental form of the submanifold. From now on, we will be concerned with $X^n=S^n(1)$. Suppose that $Y^{n-1}$ is a compact minimal isoparametric hypersurface, i.e., a compact hypersurface with vanishing mean curvature and constant principal curvatures $\mu_1, \mu_2,...,\mu_{n-1}$ (cf. [@CR]). In fact, in every isoparametric family in the unit sphere, there does exist one and only one minimal hypersurface (cf. [@GT1]). Recall an elegant result of Münzner ([@Mu80]) that the number $g$ of distinct principal curvatures must be $1$, $2$, $3$, $4$ or $6$; and the multiplicities $m_i$ $(i=1, 2,..., g)$ of distinct principal curvatures satisfy $m_k=m_{k+2}$ (subscripts $\mod ~g$). We will denote them by $m_+$ and $m_-$, respectively. More precisely, as it is well known that every isoparametric hypersurface in the unit sphere corresponds to an isoparametric function $f$ with image $[-1, 1]$. Denote the focal submanifolds by $M_+:=f^{-1}(1)$ and $M_-:=f^{-1}(-1)$ so that $codim(M_+)=m_++1$, $codim(M_-)=m_-+1$. One of the main results of the present paper is: \[isop1\] Let $Y^{n-1}$ be a compact minimal isoparametric hypersurface in $S^n(1)$, $n\geq 3$, which separates $S^n$ into $S^n_+$ and $S^n_-$. Then each of doubles $D(S^n_+)$ and $D(S^n_-)$ admits a metric of positive scalar curvature. Moreover, there is still an isoparametric foliation in $D(S^n_+)$ (or $D(S^n_-)$). As a direct result, we get the $KO$-characteristic numbers $$\alpha(D(S^n_+))=0,~~~~~~\alpha(D(S^n_-))=0.$$ Furthermore, we have: \[pi-manifold\] $D(S^n_+)$ (resp.$D(S^n_-)$) is a $\pi$-manifold, i.e., a stably parallelizable manifold. In particular, it is an orientable, spin manifold with vanishing Stiefel-Whitney classes and Pontrjagin classes. It is worth pointing out that the condition “$D(S^n_+)$ is stably parallelizable" does not imply the conclusion $\alpha(D(S^n_+))=0$. For instance, as Kervaire-Milnor ( Theorem 3.1 in [@KM]) proved, every homotopy sphere is a $\pi$-manifold. But as we stated before, there do exist some $8k+1$ and $8k+2$ dimensional exotic spheres with non-vanishing $KO$-characteristic number $\alpha$. For isoparametric hypersurfaces in unit spheres, taking the different values of $g$ into account, we know: When $g = 1$, an isoparametric hypersurface must be a great or small sphere. Thus the double construction is trivial, namely, $D(S^n_+) \cong S^n$. When $g = 2$, an isoparametric hypersurface must be a standard product of two spheres $S^k(r)\times S^{n-k-1}(s)$ with $r^2+s^2=1$. Thus $D(S^n_+) \cong S^k \times S^{n-k}$ or $S^{k+1} \times S^{n-k-1}$. When $g = 3$, E. Cartan has classified the isoparametric hypersurfaces. In fact, they are all homogeneous (see, for example, [@CR]). When $g = 4$, except for the unknown case $(m_+, m_-)=(7, 8)$ (or $(8, 7)$), all the isoparametric hypersurfaces are either FKM-type or homogeneous (cf. [@CCJ], [@Chi]). When $g = 6$, all the isoparametric hypersurfaces must be homogeneous (see, for example, [@Miy09]). Given all these classifications, in order to study the properties of the double manifold $D(S^n_+)$, it suffices to consider the cases that $Y$ is either homogeneous or of the FKM-type, except for the case $(g, m_+, m_-)=(4, 7, 8)$. Therefore, we divide our research into two parts, one is on the homogeneous case, and the other is on the FKM-type. We begin by recalling a well known result that homogeneous hypersurfaces in $S^n$ are isoparametric since they have constant principal curvatures. They have been characterized as principal orbits of the isotropy representation of rank two symmetric spaces, and are classified completely by Hsiang and Lawson (cf. [@HL], [@TT]). From the corresponding cohomogeneity one action on $S^n$ with a certain slice representation of the normal disc, we derive a cohomogeneity one action on $D(S^n_+)$. In terms of the isotropy subgroup $K_0$ of the principal orbit and $K_\pm$ of the singular orbits (focal submanifolds) $M_\pm$, we classify $D(S^n_+)$ in Section $3$ with respect to the homogeneous hypersurface $Y$. In particular, we investigate $D(S^4_+)$ in the case $(g,m_+,m_-)=(3,1,1)$, finding an interesting phenomenon that $D(S^4_+)\cong S^2\times S^2/\sigma$, where $\sigma$ is an involution different from that of the oriented Grassmannian $G_2(\mathbb{R}^4)\cong S^2\times S^2/\sim $. Next, we turn to the FKM-type. For every orthogonal representation of the Clifford algebra $\mathcal{C}_{m-1}$ on $\mathbb{R}^l$, Ferus, Karcher and Münzner ([@FKM]) constructed a homogeneous polynomial $F$ on $\mathbb{R}^{2l}$. The level hypersurfaces of $F\|_{S^{2l-1}}$ are isoparametric in $S^{2l-1}$ with $g=4$ and multiplicities of distinct principal curvatures $(m_+,m_-,m_+,m_-)=(m,l-m-1,m,l-m-1)$. If $m\not \equiv 0 $ $(\mod~4)$, $F$ is determined by $m$ and $l$ up to a rigid motion of $S^{2l-1}$; if, however $m\equiv 0$ $(\mod~4)$, there are inequivalent representations of $\mathcal{C}_{m-1}$ on $\mathbb{R}^l$ parameterized by an integer $q$, the index of the representation (cf. [@Wan]). In the second case, denote by $M_+(m,l,q)$, $M_-(m,l,q)$ the corresponding focal submanifolds, respectively. According to [@FKM], $M_+$ has a trivial normal bundle, while $M_-$ is diffeomorphic to an $S^{l-1}$ bundle over $S^m$. Thus $D(S^{2l-1}_+)\cong M_+\times S^{m+1}$. As for $D(S^{2l-1}_-)$, a delicate calculation of the topology on a sphere bundle over $M_-$ leads to the following \[FKM\] Given an odd prime $p$. If $q_1\not \equiv \pm q_2$ $(\mod~p)$, then $D(S^n_-)(m,l,q_1)$ and $D(S^n_-)(m,l,q_2)$ have different homotopy types. As we claimed in Proposition \[pi-manifold\], $D(S^n_+)$ is a $\pi$-manifold with vanishing Stiefel-Whitney classes and Pontrjagin classes. Without the aid of characteristic classes, it is usually not easy to distinguish the homeomorphism classes. Our Theorem \[FKM\] is established by using $\mod~p$ cohomology operators. Geometry of the double manifold $D(S^n_+)$ ========================================== This section will be devoted to the proof of Theorem \[isop1\]. We prefer to prove this result by making use of fundamental properties of isoparametric hypersurfaces and some straightforward verifications. Let $Y^{n-1}$ be a minimal isoparametric hypersurface in the unit sphere $S^n(1)$. It is well known that $Y$ is a level hypersurface with vanishing mean curvature of an isoparametric function $f$ on $S^n(1)$. By an isoparametric function on $S^n(1)$, we mean a function $f: S^n(1)\rightarrow \mathbb{R}$ satisfying: $$\label{ab} \left\{ \begin{array}{ll} \|\nabla f\|^2= b(f)\\ \quad\triangle f=a(f) \end{array}\right.$$ where $\nabla f$ is the gradient of $f$, $\triangle f$ is the Laplacian of $f$, $b$ is a smooth function on $\mathbb{R}$, and $a$ is a continuous function on $\mathbb{R}$ (see [@Tho], for an excellent survey). We require that the isoparametric function is proper (cf. [@GT2]) so that both focal submanifolds have codimensions greater than $1$. Recall that an isoparametric hypersurface $Y^{n-1}$ in $S^n(1)$ has constant principal curvatures, which we denote by $\mu_1(0), \mu_2(0),...,\mu_{n-1}(0)$ as before corresponding to the unit normal vector field $\xi=\frac{\nabla f}{\|\nabla f\|}$. A key reason for choosing $Y^{n-1}$ to be minimal isoparametric is that, as we will see, its induced metric from $S^n(1)$ has positive scalar curvature. By Gauss equation, for a closed minimal hypersurface $N$ in a unit sphere $S^n(1)$, $$S=(n-1)(n-2)-R_N$$ where $S$ is square of the length of the second fundamental form. If, in addition, $N$ is a minimal isoparametric hypersurface on $S^n$, Peng and Terng ([@PT]) asserted that: $$S=(g-1)(n-1),$$ which implies $R_N\geq 0$, and “=" is achieved if and only if $(m_+, m_-)=(1,1)$ since $n-1=\frac{g}{2}(m_++m_-)$. It follows immediately that the minimal isoparametric hypersurface $Y$ has $R_Y\geq 0$, $H(0)={\displaystyle\sum_{i=1}^{n-1}\mu_i(0)}=0$, and $S=\displaystyle\sum_{i=1}^{n-1}\mu_i^2(0)=(g-1)(n-1)$, which imply that $$2\sum_{i< j}^{n-1}\mu_i\mu_j\|_Y=H(0)^2-\sum_{i=1}^{n-1}\mu_i^2\|_Y=-(g-1)(n-1).$$ By definition in formula $(\ref{scalar})$, we see $A=\displaystyle\sum_{i< j\leq n-1}\mu_i\mu_j -(n-1)$. In order to simplify the calculation of $R_M$, we set $$a(r):=2\sum_{i<j}^{n-1}\mu_i\mu_j\|_{Y_r}-2\sum_{i<j}^{n-1}\mu_i\mu_j\|_{Y}=2\sum_{i<j}^{n-1}\mu_i\mu_j\|_{Y_r}+(g-1)(n-1).$$ Since $$\lim_{r\rightarrow 0}2\sum_{i<j}^{n-1}\mu_i\mu_j\|_{Y_r}=\lim_{r\rightarrow 0}\Big(H(r)^2-\sum_{i=1}^{n-1}\mu_i^2\|_{Y_r}\Big)=-(g-1)(n-1),$$ we have $$\label{a(r)} \lim_{r \to 0}a(r)=0$$ In fact, according to the Bochner-Weitzenböck formula: $$\frac{1}{2}\triangle\|\nabla f\|^2=\|Hess f\|^2+\langle\nabla f, \nabla(\triangle f)\rangle + Ric(\nabla f, \nabla f),$$ by virtue of the expression of Hessian of $f$ (cf. [@GTY]): $$Hess f=diag\Big(-\sqrt{b(f)}\mu_1,\cdots,-\sqrt{b(f)}\mu_{n-1}, b'(f)/2\Big),$$ with $b(f)=g^2(1-f^2)$ in formula (\[ab\]) by the famous Cartan-Münzner equations, we obtain $$\label{sum of mu} \sum_{i=1}^{n-1}\mu_i^2\|_{Y_r}=\frac{(n-1)(g-1)-cf+(n-1)f^2}{1-f^2}\quad \emph{with}\quad c=\frac{g^2(m_--m_+)}{2}.$$ Hence we can express $H(r)$ explicitly as: $$\label{Hr} H(r)=(n-1)\frac{f\|_{Y_r}}{\sqrt{1-f^2\|_{Y_r}}}-\frac{c}{g\sqrt{1-f^2\|_{Y_r}}}.$$ It follows immediately that $H(0)=0$ and $H(r)>0$ for any $r>0$. Consequently, from the definition of $a(r)$ and (\[sum of mu\]), (\[Hr\]), it follows that $$\begin{aligned} \label{exact a(r)} &&a(r)\\ &=&~H^2(r)-\sum_{i=1}^{n-1}\mu_i^2\|_{Y_r}+(g-1)(n-1)\nonumber\\ &=&~(g-1)(n-1)+\frac{1}{1-f^2}\{(n-1)^2f^2-\frac{2c}{g}(n-1)f+\frac{c^2}{g^2}-(n-1)f^2+cf\nonumber\\ & &\quad-(g-1)(n-1)\},\nonumber\end{aligned}$$ Substituting all these equalities in $(\ref{scalar})$, we get immediately $$\label{scalarYr} R_{M}\|_{Y_r}=n(n-1)\cos^2\theta+(n-g-1)(n-1)\sin^2\theta+a(r)\sin^2\theta+2kH(r)\sin\theta$$ with $H(r)$ and $a(r)$ in (\[Hr\]), (\[exact a(r)\]), respectively. Since we have the dimension relation $n-1=\frac{g}{2}(m_++m_-)$, it suffices to analyze the following two cases for our destination. $(A)$: $(m_+, m_-)=(1, 1)$. This is just the case that $n-g-1=0$. Since $a(r)$ is identically $0$ in this case, (\[scalarYr\]) becomes $$R_M=n(n-1)\cos^2\theta+2kH(r)\sin\theta.$$ By controlling the “bending" angle of the curve $\gamma$, we can assume $0\leq k\leq \frac{1}{2}$ so that $R_M\|_{Y_r}=n(n-1)\cos^2\theta+2kH(r)\sin\theta \geq 0,$ and “=" is achieved if and only if $r=0$. $(B)$: $Max\{m_+, m_-\}\geq 2$. In this case, $n-g-1>0$. For $\theta \in [0,\frac{\pi}{2}]$, it is easily seen that $$Min\{n(n-1)\cos^2\theta+(n-g-1)(n-1)\sin^2\theta\}=(n-g-1)(n-1),$$ thus by (\[scalarYr\]), $$R_M\geq (n-g-1)(n-1) + a(r)\sin^2\theta + 2kH(r)\sin\theta.$$ With the same assumption on $k$ as in case $(A)$, $R_M$ has a positive lower bound. Up to now, we changed only the metric near the minimal isoparametric hypersurface $Y$ along the curve $\gamma$ into a product metric while preserving the positive scalar curvature, as desired. In this way, gluing two copies of $S^n_+$, we get the double manifold of positive scalar curvature. More importantly, there is still an isoparametric foliation on $D(S^n_+)$, remaining the same with that in $S^n(1)$ as $r\geq r_1$. In a neighborhood of $Y$ with diameter $2r_1$, the principal curvatures turn out to be $\mu_1\cos\theta, \mu_2\cos\theta,...,\mu_{n-1}\cos\theta$. The proof is now complete.$\Box$ Topology of the double manifold $D(S^n_+)$ ========================================== First of all, we compute the cohomology groups. \[homology\] Let the ring of coefficient $R=\mathbb{Z}$ if $M_+$ and $M_-$ are both orientable and $R=\mathbb{Z}_2$ otherwise. Then $$\left\{ \begin{array}{ll} H^0(D(S^n_+))=R & \textrm{}\\ H^1(D(S^n_+))=H^1(M_+) & \textrm{}\\ H^q(D(S^n_+))=H^{q-1}(M_-) \oplus H^q(M_+) & \textrm{ for $2 \leq q \leq n-2$ }\\ H^{n-1}(D(S^n_+))=H^{n-2}(M_-)\\ H^n(D(S^n_+))=R \end{array} \right.$$ For $D(S^n_-)$, analogous identities hold.$\Box$ \[ori\] By Morse theory, we see that if $m_+>1$ (resp. $m_->1$), then $M_-$ (resp. $M_+$) is orientable. In fact, we define a spherical distance function on the focal submanifold $M_-$. $$\begin{aligned} &&L_p:~M_-\longrightarrow \mathbb{R}\\ &&\qquad\quad x\mapsto \cos^{-1}\langle p, x\rangle\end{aligned}$$ where $p$ belongs to the complement of $M_{\pm}$ in $S^n$. The Morse index theorem states that the index of $L_p$ at a non-degenerate critical point $x$ equals the number of focal points (counting multiplicities ) of $(M_-, x)$ on the shortest geodesic segment from $p$ to $x$. Immediately, we obtain, for example when $g=4$, the index of non-degenerate critical points are $0$, $m_+$, $m_++m_-$ and $2m_++m_-$, respectively. Consequently, we have the cell decomposition $M_-=S^{m_+}\bigcup e^{m_++m_-}\bigcup e^{2m_++m_-}$. Thus if $m_+>1$, $M_-$ is simply connected. Similar results hold for other values of $g$. In order to prove Proposition \[homology\], we recall a topological theorem of Münzner (cf. [@Mu80]) stated as Theorem Let $N$ be a compact connected hypersurface in $S^{n}$ such that: $(a)$ $S^{n}$ is divided into two manifolds $B_+$ and $B_-$ with the same boundary $N$. $(b)$ $B_+$ $(resp.~ B_-)$ has the structure of a disc bundle over a compact manifold $M_+$ $(resp.~ M_-)$ of dimension $n-1-m_+$ $(resp.~ n-1-m_-)$. Let the ring of coefficient $R=\mathbb{Z}$ if $M_+$ and $M_-$ are both orientable and $R=\mathbb{Z}_2$, otherwise. Let $\nu=m_+ + m_-$. Then $$H^q(M_\pm) = \left\{ \begin{array}{ll} R, & \textrm{for $q\equiv 0$ $(\mod ~\nu)$, $0 \leq q < n-1$}\\ R, & \textrm{for $q\equiv m_\mp$ $(\mod ~\nu)$, $0 \leq q < n-1$}\\ 0, & \textrm{otherwise } \end{array} \right.$$ Further, $$H^q(N) = \left\{ \begin{array}{ll} R, & \textrm{for $q=0, n-1$}\\ H^q(M_+) \oplus H^q(M_-), & \textrm{for $1 \leq q \leq n-2$} \end{array} \right.$$ $\Box$ To complete the proof of Proposition \[homology\], we observe that a minimal isoparametric hypersurface $Y$ in $S^n$ satisfies the hypotheses of the previous theorem, getting the cohomology groups $H^q(M_\pm)$, equivalently, $H^q(S^n_\pm)$. Finally, by the Mayer-Vietoris sequence of $(D(S^n_+), S^n_+, S^n_+)$, we arrive at the conclusion immediately.$\Box$ Next, we give a proof of Proposition 1.1 [ $D(S^n_+)$ (resp.$D(S^n_-)$) is a $\pi$-manifold, i.e., a stably parallelizable manifold. In particular, it is an orientable, spin manifold with vanishing Stiefel-Whitney classes and Pontrjagin classes. ]{} Suppose we are now given a (minimal) isoparametric hypersurface in $S^n$. As Münzner asserted (cf. [@CR] p.283), $S^n_+$ has the structure of a differential disc bundle over $M_+$. In fact, it is the normal disc bundle over $M_+$. More precisely, we have $$\begin{aligned} \label{normal disc bundle} && B^{m_++1} \hookrightarrow S^n_+ =B(\nu_+)\\ &&\qquad\qquad\quad \downarrow \pi\nonumber\\ &&\qquad\quad\quad\quad M_+\nonumber\end{aligned}$$ where $\nu_+$ is the normal bundle over $M_+$, $B^{m_++1}$ is the fiber disc. Since $S^n_+$ has a metric, we can define a homeomorphism as: $$\begin{aligned} &&B^n_1\sqcup_{id}B^n_2 \longrightarrow S(\nu_+\oplus\textbf{1})\\ &&\qquad \quad e \longmapsto \left\{ \begin{array}{ll} (e, \sqrt{1-\|e\|^2}), & \textrm{for $e \in B^n_1$}\\ (e, -\sqrt{1-\|e\|^2}), & \textrm{for $e \in B^n_2$} \end{array} \right.\end{aligned}$$ where $B^n_1$, $B^n_2$ are two copies of $S^n_+=B(\nu_+)$, $S(\nu_+\oplus \textbf{1})$ is a sphere bundle of the Whitney sum $\nu_+\oplus \textbf{1}$, here $\textbf{1}$ is a trivial line bundle over $M_+$. As a result, we get a new bundle $$\begin{aligned} && S^{m_++1} \hookrightarrow D(S^n_+)\\ &&\qquad\qquad\quad \downarrow \rho\\ &&\qquad\quad\quad\quad M_+\end{aligned}$$ and a correspondence $D(S^n_+)\cong S(\nu_+\oplus \textbf{1})$. It follows immediately that $$\begin{aligned} T(D(S^n_+))\oplus \textbf{1} &\cong& T(S(\nu_+\oplus \textbf{1}))\oplus \textbf{1} \\ &\cong& \rho^TM_+\oplus \rho^(\nu_+\oplus\textbf{1})\\ &\cong& \rho^{j^_+}TS^n\oplus\textbf{1}\cong (\textbf{n+1})\end{aligned}$$ where $j_+: M_+\rightarrow S^n$ is an inclusion. In other words, $D(S^n_+)$ is stably parallelizable, i.e., a $\pi$-manifold. This completes the proof.$\Box$ As indicated in Introduction, we will be mainly concerned with the minimal isoparametric hypersurface $Y$ in the following two cases: homogeneous hypersurface and FKM-type. Homogeneous hypersurface ------------------------ Let $Y$ be a homogeneous hypersurface in $S^n$, as Hsiang and Lawson([@HL]) showed: $Y$ can be characterized as a principal orbit of the isotropy representation of some rank two symmetric space $U/K$. To begin with, we provide a brief description of the corresponding rank two symmetric spaces. Again, let $g$ be the number of distinct principal curvatures of the homogeneous hypersurface $Y$. As mentioned before, $g$ can only be $1$, $2$, $3$, $4$ or $6$. When $g = 1$, $Y$ is a hypersphere in $S^n$, the corresponding rank two symmetric space is $$(S^1\times SO(n + 1))/SO(n) = S^1\times S^n.$$ When $g = 2$, $Y$ is a Riemannian product of two spheres $S^k(r)\times S^{n-k-1}(s)$ with $r^2+s^2=1$, $1\leq k\leq n-2$, the corresponding rank two symmetric space is $$(SO(k + 2)\times SO(n- k + 1))/(SO(k + 1)\times SO(n-k)) = S^{k+1}\times S^{n-k}.$$ When $g = 3$, $Y$ is congruent to a tube of constant radius around the Veronese embedding of real projective plane $\mathbb{R}P^2$ into $S^4$, or complex projective plane $\mathbb{C}P^2$ into $S^7$, or quaternionic projective plane $\mathbb{H}P^2$ into $S^{13}$, or Cayley projective plane $\mathbb{O}P^2$ into $S^{25}$. The corresponding rank two symmetric spaces are $$SU(3)/SO(3);~~ SU(3)\times SU(3)/SU(3);~~ SU(6)/Sp(3);~~ E_6/F_4.$$ When $g = 4$, $Y$ is a principal orbit of the isotropy representation of $$SO(5)\times SO(5)/SO(5);~~ SO(10)/U(5);~~ E_6/T\cdot Spin(10);$$ or of two-plane Grassmannians $$SO(k + 2)/SO(k)\times SO(2)~~(k\geq 3);$$ $$SU(k + 2)/S(U(k)\times U(2))~~(k \geq 3);$$ $$Sp(k + 2)/Sp(k)\times Sp(2)~~(k\geq 2).$$ When $g = 6$, $Y$ is a principal orbit of the isotropy representation of $$G_2/SO(4)~~ or ~~G_2\times G_2/G_2.$$ Now let $G$ be a closed subgroup of the isometry group of $S^n$ acting on $S^n$ with cohomogeneity one. We equip the orbit space $S^n/G$ with the quotient topology relative to the canonical projection $S^n\rightarrow S^n/G$. Since $n>1$, $S^n$ is simply connected and compact, for topological reasons $S^n/G$ must be homeomorphic to $[-1, 1]$ and each singular orbit has codimension greater than one. We denote the singular orbits corresponding to $\pm 1$ by $M_{\pm}$, and their isotropy subgroups by $K_{\pm}$, respectively. Naturally, there is a diffeomorphism $M_{\pm}\cong G/K_{\pm}$. The other orbits are congruent to each other, and they are all principal orbits. It makes sense to fix an orbit $Y$ corresponding to a certain value in $(-1,1)$ so that $Y$ is minimal, and denote its isotropy subgroup by $K_0$. The existence of such $Y$ is clear. Similarly, $Y\cong G/K_0$. Based on the bundle structure of $S^n_+$ over $M_-$ and the following group action of $K_{\pm}$: $$\begin{aligned} && K_{\pm}\times(G\times B^{m_++1}_{\pm})\longrightarrow G\times B^{m_++1}_{\pm}\\ &&\qquad\qquad\quad~~(k,g,x)\longmapsto (gk^{-1}, k\bullet x)\end{aligned}$$ where $\bullet$ is a slice representation (for details, see for example [@Bre]), we decompose $S^n$ into $$S^n=G\times_{K_+}B^{m_++1}_+\cup_{Y}G\times_{K_-}B^{m_-+1}_-.$$ Next, by gluing two copies of $S^n_+$, we define a new action of the isotropy group $K_+$ on $G\times S^{m_++1}$ : $$\begin{aligned} &&K_+\times(G\times S^{m_++1})\longrightarrow G\times S^{m_++1}\\ &&\qquad(k,g,(x,t))\longmapsto (gk^{-1},k\star(x,t))\end{aligned}$$ where $k\star(x,t):=(k\bullet x,t)$, $t=\pm\sqrt{1-\|x\|^2}$, $(x,t)\in S^{m_++1}$. Consequently, we have the diffeomorphism: $$D(S^n_+)\cong G\times_{K_+}B^{m_++1}\cup_{Y}G\times_{K_+}B^{m_++1}\cong G\times S^{m_++1}/K_+.$$ In conclusion, $D(S^n_+)$ can be determined by $K_+$ and its action on $G\times S^{m_++1}$. Similarly, we can also express $D(S^n_-)$ in this way. A series of delicate calculations lead us to a complete list of the isotropy subgroups $K_0$ and $K_{\pm}$ of homogeneous hypersurfaces and focal submanifolds as follows. To the best of our knowledge, the determinations of $K_+$ and $K_-$ have not previously appeared in the literature. The main difficult occurred in the calculation of exceptional Lie groups. Homogeneous (isoparametric) hypersurfaces in the unit sphere. g $(m_{+}, m_-)$ (U, K) $K_0$ $K_+$ $K_-$ --- ---------------- --------------------------------- ------------------------------ ------------------------ ------------------------- 1 $n-1$ ($S^1\times SO(n+1)$, $SO(n)$) $SO(n-1)$ $SO(n)$ $SO(n)$ $n\geq 2$ 2 $(p, q)$ ($SO(p+2)\times SO(q+2)$, $SO(p)\times SO(q)$ $SO(p+1)\times SO(q)$ $SO(p)\times SO(q+1)$ $SO(p+1)\times SO(q+1)$) $ p,q\geq 1$ 3 (1, 1) $(SU(3), SO(3))$ $\mathbb{Z}_2+ \mathbb{Z}_2$ $S(O(2)\times O(1))$ $S(O(1)\times O(2))$ 3 (2, 2) $(SU(3)\times SU(3), SU(3))$ $T^2$ $S(U(2)\times U(1))$ $S(U(1)\times U(2))$ 3 (4, 4) $(SU(6), Sp(3))$ $Sp(1)^3$ $Sp(2)\times Sp(1)$ $Sp(2)\times Sp(1)$ 3 (8, 8) $(E_6, F_4)$ $Spin(8)$ $Spin(9)$ $Spin(9)$ 4 (2, 2) $(SO(5)\times SO(5), SO(5))$ $T^2$ $SO(2)\times SO(3)$ $U(2)$ 4 (4, 5) $(SO(10), U(5))$ $SU(2)^2\times U(1)$ $Sp(2)\times U(1)$ $SU(2)\times U(3)$ 4 (6, 9) $(E_6, T\cdot Spin(10))$ $U(1)\cdot Spin(6)$ $U(1)\cdot Spin(7) $ $S^1\cdot SU(5)$ 4 (1, m-2) $(SO(m+2), SO(m)\times SO(2))$ $SO(m-2)\times \mathbb{Z}_2$ $SO(m-2)\times SO(2)$ $O(m-1)$ $m\geq 3$ 4 (2, 2m-3) $(SU(m+2), S(U(m)\times U(2)))$ $S(U(m-2)\times T^2)$ $S(U(m-2)\times U(2))$ $S(U(m-1)\times T^2)$ $m\geq 3$ 4 (4, 4m-5) $(Sp(m+2), Sp(m)\times Sp(2))$ $Sp(m-2)\times Sp(1)^2$ $Sp(m-2)\times Sp(2)$ $Sp(m-1)\times Sp(1)^2$ $m\geq 2$ 6 (1, 1) $(G_2, SO(4))$ $\mathbb{Z}_2+ \mathbb{Z}_2$ $O(2)$ $O(2)$ 6 (2, 2) $(G_2\times G_2, G_2)$ $T^2$ $U(2)$ $U(2)$ In the following, we first illustrate the calculations of $K_0, K_+, K_-$ in the case of the symmetric pair $(E_6, T\cdot Spin(10))$ with $(g,m_+,m_-)=(4,6,9)$, and then give an example of the case $(SU(3), SO(3))$ with $(g,m_+,m_-)=(3,1,1)$. \[469\] The calculation of $K_+, K_-$ of the symmetric pair $(E_6, T\cdot Spin(10))$ with $(g,m_+,m_-)=(4,6,9)$. At the beginning, we introduce some notations and operations on the division Cayley algebra ${\mathbb{O}}$, which is generated by $\{e_0=1, e_1,\cdots,e_7\}$ and satisfies 1. For $i>0$, $e_i^2=-1$; 2. For $i,j>0$, $i\neq j$, $e_ie_j=-e_je_i$; 3. $e_1e_2=e_4$; 4. If $e_ie_j=e_k$ for some $i,j,k>0$, then $e_{i+1}e_{j+1}=e_{k+1}$ and $e_{2i}e_{2j}=e_{2k}$ (subscripts $\mod~7$). Let $M_3({\mathbb{O}})$ be the set of $3\times 3$ matrices with entries in ${\mathbb{O}}$, ${\mathcal{H}_3}$ the set of Hermitian matrices in $M_3({\mathbb{O}})$, namely, $${\mathcal{H}_3}=\{ X\in M_3({\mathbb{O}}) \| {\hspace{10pt}}^t\overline{X}=X \},$$ where the conjugate of any element $x=\sum_{i=0}^7x_ie_i\in {\mathbb{O}}$ is defined by $$\overline{x}=x_0e_0-\sum_{i=1}^7x_ie_i.$$ In the following, we will always denote an element $X\in{\mathcal{H}_3}$ of the form $$X=X(\xi,x)=\left( \begin{array}{ccc} \xi_1 & x_3 & \overline{x}_2 \\ \overline{x}_3 & \xi_2 & x_1 \\ x_2 & \overline{x}_1 & \xi_3 \\ \end{array} \right), {\hspace{10pt}}\mbox{for}~~ \xi_i\in {\mathbb{R}},~ x_i\in {\mathbb{O}}$$ by $$X=\xi_1 E_1+\xi_2 E_2+\xi_3 E_3+F_1(x_1)+F_2(x_2)+F_3(x_3).$$ The Jordan product, as a basic operation in ${\mathcal{H}_3}$, is a multiplication defined by $$X\circ Y=\frac{1}{2}(XY+YX), {\hspace{10pt}}\mathrm{for}~ X,Y\in {\mathcal{H}_3}.$$ Usually, $({\mathcal{H}_3}, \circ)$ is called the exceptional Jordan algebra. Moreover, the trace ${\mathrm{Tr}\hspace{1pt}(X)} $, the inner product $(X,Y)$ and the determinant $\det X$ can be defined respectively by $$\begin{aligned} &&{\mathrm{Tr}\hspace{1pt}(X)}=\xi_1+\xi_2+\xi_3, {\hspace{10pt}}\mathrm{for}~ X=X(\xi,x),\\ &&(X,Y)={\mathrm{Tr}\hspace{1pt}(X\circ Y)},\\ && \det X=\xi_1\xi_2\xi_3+ \operatorname{Re}(x_1x_2x_3)-\xi_1\|x_1\|^2-\xi_2\|x_2\|^2-\xi_3\|x_3\|^2.\end{aligned}$$ Let ${\mathcal{H}_3}^{{\mathbb{C}}}=\{X_1+\sqrt{-1}X_2~\|~ X_1,X_2\in {\mathcal{H}_3}\}$ be the complexification of Jordan algebra ${\mathcal{H}_3}$. In the same manner, we have the Jordan product, the trace, the ${\mathbb{C}}$-linear form $(~,~)$ and the determinant in ${\mathcal{H}_3}^{{\mathbb{C}}}$. A Hermitian inner product $\langle~,~ \rangle$ on ${\mathcal{H}_3}^{{\mathbb{C}}}$ is given by $$\langle X,Y\rangle=(X,\tau Y),{\hspace{10pt}}\mbox{for}~X,Y\in {\mathcal{H}_3}^{{\mathbb{C}}},$$ where $\tau$ is the complex conjugate of ${\mathcal{H}_3}^{{\mathbb{C}}}$. With all these notations, an equivalent definition of the group $E_6$ can be given by (cf. [@Yok]) $$E_6 = \{ \alpha\in GL({\mathcal{H}_3}^{{\mathbb{C}}},{\mathbb{C}}) ~\|~ \det (\alpha X)=\det X, \langle\alpha X,\alpha Y\rangle=\langle X,Y\rangle\}.$$ Set $${\mathcal{SH}_3}=\{ A\in M_3({\mathbb{O}})\|{\hspace{10pt}}^t\overline{A}=-A,{\mathrm{Tr}\hspace{1pt}(A)}=0 \}.$$ As above, we denote an element $A\in {\mathcal{SH}_3}$ of the form $$A=\left( \begin{array}{ccc} a_1 & x_3 & -\overline{x}_2 \\ -\overline{x}_3 & a_2 & x_1 \\ x_2 & -\overline{x}_1 & a_3 \\ \end{array} \right), {\hspace{10pt}}a_i,x_i\in {\mathbb{O}}, \overline{a}_i=-a_i, a_1+a_2+a_3=0.$$ by $$A=a_1 E_1+a_2 E_2+a_3 E_3+A_1(x_1)+A_2(x_2)+A_3(x_3).$$ Notice that $[{\mathcal{SH}_3},{\mathcal{SH}_3}]={\mathcal{SH}_3}$, $[{\mathcal{SH}_3},{\mathcal{H}_3}]={\mathcal{H}_3}$. Thus any $A\in {\mathcal{SH}_3}$ induces a map $\tilde{A}:{\mathcal{H}_3}\rightarrow {\mathcal{H}_3}$ expressed as $$\tilde{A}(X)=\frac{1}{2}[A,X].$$ Let ${\mathfrak{t}}'$ be the subalgebra of $gl({\mathcal{H}_3})$ generated by $\{\tilde{A}\|A\in {\mathcal{SH}_3}\}$. Then ${\mathfrak{t}}'$ is isomorphic to the (compact) Lie algebra of $F_4$. Furthermore, observing that any $X\in {\mathcal{H}_3}$ also induces a map $\tilde{X}:{\mathcal{H}_3}\rightarrow {\mathcal{H}_3}$ defined by $$\tilde{X}(Y)=X\circ Y, {\hspace{10pt}}Y\in {\mathcal{H}_3},$$ we set ${\mathfrak{p}}'=\{\tilde{X}\| X\in {\mathcal{H}_3}, {\mathrm{Tr}\hspace{1pt}(X)}=0\}$, then the Lie algebra ${\mathfrak{t}}'+\sqrt{-1}{\mathfrak{p}}'$ is just the (compact) Lie algebra of $E_6$, denoted by $\mathfrak{e}_6$. In the following discussions, we will omit the symbol ‘$\sim$’ for simplicity. Let $\mathfrak{d}_4$ be the subalgebra of ${\mathfrak{t}}'$ generated by $\{ \Sigma a_i E_i \| a_i\in{\mathbb{O}}, \overline{a}_i=-a_i, \Sigma a_i=0\}$. Then for any $D \in \mathfrak{d}_4$, $X=X(\xi,x)$, $$D \left( \begin{array}{ccc} \xi_1 & x_3 & \overline{x}_2 \\ \overline{x}_3 & \xi_2 & x_1 \\ x_2 & \overline{x}_1 & \xi_3 \\ \end{array} \right)=\left( \begin{array}{ccc} 0 & D_3x_3 & \overline{D_2x}_2 \\ \overline{D_3x}_3 & 0 & D_1x_1 \\ D_2x_2 & \overline{D_1x}_1 & 0 \\ \end{array} \right),$$ where $D_1,D_2,D_3$ are elements of Lie algebra $so(8)$ and satisfy the principle of triality: $$(D_1 x)y+x(D_2y)=\overline{D_3(\overline{xy})}, {\hspace{10pt}}x,y\in {\mathbb{O}},$$ which implies that $D_2,D_3$ are uniquely determined by $D_1$. Hence the map defined by $D\mapsto D_1$ is an isomorphism from $\mathfrak{d}_4$ to $so(8)$. Furthermore, setting $$\begin{aligned} &&\mathfrak{D}_i= \{ A_i(x)~\|~ x\in {\mathbb{O}}\}, {\hspace{10pt}}i=1,2,3,\\ &&\mathfrak{R}_i= \{ F_i(x)~\|~ x\in {\mathbb{O}}\}, {\hspace{10pt}}i=1,2,3,\\ &&\mathfrak{R}_0= \{ \Sigma \xi_iE_i ~\|~ \xi_i\in {\mathbb{R}}, \Sigma\xi_i=0 \},\end{aligned}$$ we can decompose the Lie algebra $\mathfrak{e}_6$ as $$\mathfrak{e}_6=\mathfrak{d}_4+\mathfrak{D}_1+\mathfrak{D}_2+\mathfrak{D}_3 +\sqrt{-1}\mathfrak{R}_0+\sqrt{-1}\mathfrak{R}_1+\sqrt{-1}\mathfrak{R}_2+\sqrt{-1}\mathfrak{R}_3.$$ Since there is a transformation $\sigma$ of ${\mathcal{H}_3}^{{\mathbb{C}}}$ expressed as $$\sigma \left( \begin{array}{ccc} \xi_1 & x_3 & \overline{x}_2 \\ \overline{x}_3 & \xi_2 & x_1 \\ x_2 & \overline{x}_1 & \xi_3 \\ \end{array} \right)=\left( \begin{array}{ccc} \xi_1 & -x_3 & -\overline{x}_2 \\ -\overline{x}_3 & \xi_2 & x_1 \\ -x_2 & \overline{x}_1 & \xi_3 \\ \end{array} \right),$$ ( obviously, $\sigma^2=\emph{id}$ ), an involution $\gamma$ of $E_6$ can be naturally defined by $\gamma(\alpha)=\sigma \alpha \sigma, $ for $\alpha\in E_6$. Thus the decomposition of $\mathfrak{e}_6$ corresponding to $\gamma$ can be written as $\mathfrak{e}_6=\mathfrak{t}+\mathfrak{p}$, where $$\begin{aligned} \mathfrak{t}&=&\{\delta\in e_6~~\| ~~\sigma \delta =~~\delta \sigma\}=\mathfrak{d}_4+\mathfrak{D}_1+\sqrt{-1}\mathfrak{R}_0+\sqrt{-1}\mathfrak{R}_1,\\ \mathfrak{p}&=&\{\delta\in e_6~~\| ~~ \sigma \delta =-\delta \sigma\}=\mathfrak{D}_2+\mathfrak{D}_3+\sqrt{-1}\mathfrak{R}_2+\sqrt{-1}\mathfrak{R}_3.\end{aligned}$$ Choosing a maximal Abelian subspace of $\mathfrak{p}$ as $\mathfrak{h}=\{A_2(\lambda_1e_0)+\sqrt{-1}F_2(\lambda_2e_1)~\|~\lambda_1,\lambda_2\in {\mathbb{R}}\}$, and denoting by $\Delta$ the set of restricted positive roots with respect to $\mathfrak{h}$, we have $$\label{e6} \mathfrak{e}_6=\mathfrak{m}+\mathfrak{h}+\sum_{\lambda\in \Delta}\{\mathfrak{t}_{\lambda}+\mathfrak{p}_{\lambda}\},$$ where $$\begin{aligned} &&\mathfrak{m}=\{A\in \mathfrak{t}~\| {\hspace{10pt}}[A,H]=0, \mbox{for~} H\in \mathfrak{h}\},\\ &&\mathfrak{t}_{\lambda}= \{ A\in \mathfrak{t}~\| {\hspace{10pt}}ad(H)^2A=-\lambda(H)^2A, \mbox{for~} H\in \mathfrak{h} \},\\ &&\mathfrak{p}_{\lambda}= \{ A\in \mathfrak{p}~\| {\hspace{10pt}}ad(H)^2A=-\lambda(H)^2A, \mbox{for~} H\in \mathfrak{h} \}.\end{aligned}$$ Set $\widetilde{e}_i=e_1e_i$, for $i>1$ and $G_{ij}=E_{ij}-E_{ji}$, for $i,j=0,1,\cdots,7$, where $E_{ij}$ is the matrix with $(i,j)$ entry $1$ and all others $0$. By a direct computation, we can express $\mathfrak{m}$ explicitly as $$\mathfrak{m}=\mbox{span}\{\sqrt{-1}(E_1-2E_2+E_3), D~\|~ D_2=G_{ij}, i,j>1\}\cong so(6)\oplus {\mathbb{R}}.$$ Moreover, we calculate $\mathfrak{t}_{\lambda}$ and $\mathfrak{p}_{\lambda}$ in (\[e6\]) with respect to the root system $\Delta$, and list them in the following table. $\mathrm{dimension}$ $\lambda\in \Delta$ basis of $\mathfrak{t}_{\lambda}$ basis of $\mathfrak{p}_{\lambda}$ ---------------------- ------------------------------------ ----------------------------------------------- ----------------------------------------------- $6$ $\lambda_1$ $D\in \mathfrak{d}_4 : D_2=G_{0i}, i>1 $ $ A_2(e_i) :i>1$ $6$ $\lambda_2$ $D\in \mathfrak{d}_4 : D_2=G_{1i}, i>1$ $ \sqrt{-1}F_2(e_i): i>1$ $1$ $\lambda_1-\lambda_2$ $D+\sqrt{-1}(E_1-E_3) : D_2=G_{01}$ $ A_2(e_1)+\sqrt{-1}F_2(e_0)$ $1$ $\lambda_1+\lambda_2$ $D-\sqrt{-1}(E_1-E_3) : D_2=G_{01}$ $ A_2(e_1)-\sqrt{-1}F_2(e_0)$ $A_1(e_0)-\sqrt{-1}F_1(e_1)$ $ A_3(e_0)-\sqrt{-1}F_3(e_1)$ $8$ $\frac{1}{2}(\lambda_1-\lambda_2)$ $A_1(e_1)+\sqrt{-1}F_1(e_0)$ $ A_3(e_1)+\sqrt{-1}F_3(e_0)$ $A_1(e_i)+\sqrt{-1}F_1(\widetilde{e}_i): i>1$ $ A_3(e_i)-\sqrt{-1}F_3(\widetilde{e}_i):i>1$ $A_1(e_0)+\sqrt{-1}F_1(e_1)$ $ A_3(e_0)+\sqrt{-1}F_3(e_1)$ $8$ $\frac{1}{2}(\lambda_1+\lambda_2)$ $A_1(e_1)-\sqrt{-1}F_1(e_0)$ $ A_3(e_1)-\sqrt{-1}F_3(e_0)$ $A_1(e_i)-\sqrt{-1}F_1(\widetilde{e}_i):i>1$ $ A_3(e_i)+\sqrt{-1}F_3(\widetilde{e}_i):i>1$ Let $K=\{\alpha\in E_6 ~\|~ \alpha\sigma=\sigma\alpha\}$, which acts on $\mathfrak{p}$ by the adjoint representation. The orbits can only be of the following three types: $1^{\circ}$ If $H_0\in \mathfrak{h}$ with $\lambda_1(H_0)\cdot\lambda_2(H_0)\neq 0$ and $\lambda_1(H_0)\neq \pm \lambda_2(H_0)$, the Lie algebra of the isotropy subgroup $~K_0$ at $H_0$ is $\mathfrak{m} $ and $ K_0\cong U(1)\cdot Spin(6). $ $2^{\circ}$ If $H_+\in \mathfrak{h}$ with either $\lambda_1(H_+)= 0$ or $\lambda_2(H_+)= 0$. Without loss of generality, assume $\lambda_2(H_+)= 0$. According to the previous table, the Lie algebra of the isotropy subgroup $K_+$ is $\mathfrak{m}\oplus \mathfrak{k}_{\lambda_2}\cong so(7)\oplus {\mathbb{R}}. $ Then it is not difficult to see that $$K_+\cong U(1)\cdot Spin(7).$$ $3^{\circ}$ If $H_-\in \mathfrak{h}$ with either $\lambda_1(H_-)=\lambda_2(H_-)$ or $\lambda_1(H_-)=-\lambda_2(H_-)$. Without loss of generality, assume $\lambda_1(H_-)=\lambda_2(H_-)$. According to the previous table, the Lie algebra $\mathfrak{t}_-$ of the isotropy subgroup $K_-$ is given by $$\mathfrak{t}_-=\mathfrak{m}+\mathfrak{t}_{\mu}+\mathfrak{t}_{\mu/2}, {\hspace{10pt}}\mbox{for~} \mu=\lambda_1-\lambda_2.$$ Let $V^{10}$ be a $10$-dimensional vector space defined by $$V^{10}=\bigg\{\left( \begin{array}{ccc} 0 & 0 & 0 \\ 0 & \xi & x \\ 0 & \overline{x} & -\tau(\xi) \\ \end{array} \right) {\hspace{10pt}}\|{\hspace{10pt}}\xi\in {\mathbb{C}}, x\in {\mathbb{O}}\bigg\}\subset{\mathcal{H}_3}^{{\mathbb{C}}}.$$ It is well known that $K\cong T^1\cdot Spin(10)$, and the representation $\phi:Spin(10)\rightarrow SO(V^{10})$ is just the vector representation. We finally introduce a complex structure $J$ on $V^{10}$ as follows, $$J\left( \begin{array}{ccc} 0 & 0 & 0 \\ 0 & \xi & x \\ 0 & \overline{x} & -\tau(\xi) \\ \end{array} \right)= \left( \begin{array}{ccc} 0 & 0 & 0 \\ 0 & -\sqrt{-1}\xi & x \cdot e_1 \\ 0 & \overline{x\cdot e_1} & -\sqrt{-1}\tau(\xi) \\ \end{array} \right), {\hspace{10pt}}\mbox{for}~~\xi\in {\mathbb{C}}, x\in {\mathbb{O}}.$$ By a direct computation, we find that elements of the subalgebra $[\mathfrak{t}_-,\mathfrak{t}_-]\subset so(10)$ commute with the complex structure $J$, specifically, $[\mathfrak{t}_-,\mathfrak{t}_-]\cong su(5)$. Moreover, the center $\mathfrak{c(t_-)}\cong {\mathbb{R}}$ is not contained in $so(10)$. Therefore, we can conclude $$\mathfrak{t}_-=\mathfrak{c(t_-)}\oplus [\mathfrak{t}_-,\mathfrak{t}_-]\cong {\mathbb{R}}\oplus su(5),$$ and via the representation $\phi$ the corresponding isotropy subgroup is $$K_-\cong S^1\cdot SU(5),$$ where $S^1$ is a group generated by the center $\mathfrak{c(t_-)}$. \[311\] An explicit description of $D(S^4_+)$ with $(g,m_+, m_-)=(3,1,1)$. Firstly, recall a result of E. Cartan that the isoparametric hypersurface in this case must be a tube of constant radius over a standard Veronese embedding of $\mathbb{R}P^2$ into $S^4$. Let $\nu$ be the normal bundle of $\mathbb{R}P^2\hookrightarrow S^4$, which is non-orientable since $T\mathbb{RP}^2 \oplus \nu = \textbf{4}$, a $4$-dimensional trivial bundle. Let $\eta$ be the Hopf line bundle over $\mathbb{R}P^2$. It is well known that $T\mathbb{R}P^2\oplus\textbf{1}=3\eta$. Thus $3\eta\oplus\nu=T\mathbb{R}P^2\oplus\textbf{1}\oplus \nu=\textbf{5}$. Hence $4\eta\oplus\nu$=$\textbf{5}\oplus\eta$. Assertion 1: $4\eta\cong \textbf{4}$. It follows at once that $\nu\oplus \textbf{4}=\eta\oplus\textbf{5}$. Then we deduce by obstruction theory that $\nu\oplus\textbf{1}=\eta\oplus\textbf{2}$, and thus $D(S^4_+)\cong S(\nu\oplus \textbf{1})\cong S(\eta\oplus \textbf{2})$. Furthermore, we show Assertion 2: $D(S^4_+)\cong S^2\times S^2/\sigma$, where $\sigma$ is an involution. Proof of Assertion 2:\[assertion 2\] Again, let $\eta$ be a Hopf line bundle over $\mathbb{R}P^n$, $E(\eta)$ be the total space of $\eta$, then $$\begin{aligned} \label{eta} &&E(\eta)\cong S^n\times \mathbb{R}\Big/(x,t)\sim(-x,-t)\\ &&\qquad\quad \downarrow\nonumber\\ &&\qquad S^n/x\sim-x=\mathbb{R}P^n\nonumber\end{aligned}$$ where $x\in S^n$, $t\in \mathbb{R}$. This interpretation deduces that for $x\in S^n$, $(t_1,...,t_p)\in \mathbb{R}^p$, and $(s_1,...,s_q)\in \mathbb{R}^q$, $$\label{p+qeta} S^n\times \mathbb{R}^{p+q}\Big/(x,t_1,...,t_p,s_1,...,s_q)\sim(-x,t_1,...,t_p,-s_1,...,-s_q)\cong E(\textbf{p}\oplus q\eta)$$ In particular, $$D(S^4_+)\cong S(\eta\oplus\textbf{2})\cong S^2\times S^2\Big/(x,y_1,y_2,y_3)\sim(-x,-y_1,y_2,y_3).$$ where $x\in S^2$, $(y_1,y_2,y_3)\in S^2$.$\Box$ Assertion $1$ should be well known. However, we would like to give an interesting and simple proof. Proof of Assertion 1: By (\[p+qeta\]), it suffices to define a pointwise isomorphism $\Phi$ $$\begin{aligned} \label{assertion 1} && S^2\times \mathbb{R}^4 \Big/ (x, t)\sim (-x, -t) ~~\xrightarrow{\Phi}~~ S^2/\mathbb{Z}_2\times \mathbb{R}^4\\ &&\qquad\qquad\qquad\qquad \downarrow \qquad\qquad\qquad\qquad \downarrow\nonumber\\ &&\qquad\qquad\qquad\quad \mathbb{R}P^2\qquad\quad \xrightarrow{\emph{id}}\qquad \mathbb{R}P^2\nonumber\end{aligned}$$ where $\mathbb{R}^4$ is identified with the quaternions $\mathbb{H}$, and $x\in S^2=\{x\in \mathbb{H}~\|~\|x\|=1, Re~x=0\}$, $t\in\mathbb{H}$. Define $\Phi (x,t):= (x, xt)$. Obviously, $\Phi$ is well-defined, and for a fixed $x$, it is a linear isomorphism.$\Box$ It is worth remarking that $D(S^4_+)$ is not diffeomorphic to the oriented Grassmannian $$G_2(\mathbb{R}^4)\cong S^2\times S^2/(x,y)\sim(-x,-y),$$ where $x, y\in S^2$. To show this remark, firstly, recall that (cf. [@Tan]) $$H^(G_2(\mathbb{R}^4);\mathbb{Z}_2)\cong \mathbb{Z}_2[a_1,a_2]\Big/a^3_1=0, a^2_1a_2+a^2_2=0,$$ where $a_1\in H^1(G_2(\mathbb{R}^4);\mathbb{Z}_2)$, $a_2\in H^2(G_2(\mathbb{R}^4);\mathbb{Z}_2)$. By this cohomology ring structure, it is not difficult to conclude the total Stiefel-Whitney class $$W(G_2(\mathbb{R}^4))=1+w^2_1.$$ This means that the first Stiefel-Whitney class vanishes, and the second Stiefel-Whitney class $w_2(G_2(\mathbb{R}^4))\neq 0$. In other words, $G_2(\mathbb{R}^4)$ is an orientable manifold, but not spin, while $D(S^4_+)$ is spin as mentioned in Proposition \[pi-manifold\]. FKM-type -------- In this subsection, we investigate the FKM-type isoparametric hypersurfaces in spheres with four distinct principal curvatures (cf.* [@FKM] and [@CCJ]). According to [@FKM], for a symmetric Clifford system $\{P_0,...,P_m\}$ on $\mathbb{R}^{2l}$, i.e., $P_i$’s are symmetric matrices satisfying $P_iP_j+P_jP_i=2\delta_{ij}I_{2l}$, there is a homogeneous polynomial on $\mathbb{R}^{2l}$ defined by $$\begin{aligned} \label{FKM isop. poly.} && \quad F: \mathbb{R}^{2l} \rightarrow \mathbb{R}\\ && F(z) = \|z\|^4 - 2\displaystyle\sum_{i = 0}^{m}{\langle P_iz,z\rangle^2}.\nonumber\end{aligned}$$ It can be shown that if $l-m-1>0$, then the level sets of the restriction $f=F\|_{S^{2l-1}}$ constitute a family of isoparametric hypersurfaces with $g=4$ distinct principal curvatures with multiplicities $m_+=m$, $m_-=l-m-1$. The focal submanifolds are $M_+=f^{-1}(1)$, $M_-=f^{-1}(-1)$, with codimensions $m+1$ and $l-m$ in $S^{2l-1}$, respectively. Clearly, the $+1$ eigenspace of $P_0$, say $E_+(P_0)$, is invariant under the transformations $E_1=P_1P_2$,..., $E_{m-1}=P_1P_m$. As usual, let $\delta(m)$ be the dimension of the irreducible Clifford algebra $\mathcal{C}_{m-1}$-modules (e.g., $\delta(4)=4$, $\delta(8)=8$, $\delta(m+8)=16\delta(m)$). Then $l=k\delta(m)$, for some positive integer $k$. As is known in representation theory, when $m\equiv 0$ $(\mod~4)$, there exist exactly two irreducible $\mathcal{C}_{m-1}$-modules $\Delta^+_m$ and $\Delta^-_m$ distinguished by $E_1E_2\cdots E_{m-1}=Id~~or-Id.$ If we write $E_+(P_0)=a\Delta^+_m\oplus b\Delta^-_m$ as $\mathcal{C}_{m-1}$-modules, then $$tr(P_0P_1\cdots P_m) = 2q\delta(m),$$ where $q=a-b.$ On the other hand, noticing $k=a+b$, we see $$q\equiv k~(\mod~2).$$ By [@Wan], two symmetric Clifford systems with the same index $q$ give rise to equivalent isoparametric functions on $S^{2l-1}$. Therefore, when $m\equiv 0$ $(\mod~4)$, $f$ is determined by $m$, $l=k\delta(m)$, as well as $q$ up to a rigid motion of $S^{2l-1}$. However, when $m\not \equiv 0 $ $(\mod~4)$, $q$ is always zero, so $f$ is determined only by $m$ and $l$ up to a rigid motion of $S^{2l-1}$. In the rest of this subsection, we focus on the case when $m\equiv 0$ $(\mod~4)$. First, we denote the corresponding isoparametric hypersurface by $M(m,l,q)$, and focal submanifolds by $M_{\pm}(m,l,q)$. Next, recall the following conclusions shown in ([@FKM]): $(a)$ The normal bundle $\nu_+$ is trivial, in particular, $M(m,l,q)$ is diffeomorphic to the product $M_+(m,l,q)\times S^m$; $(b)$ $M_-(m,l,q)$ is diffeomorphic to an $S^{l-1}$ bundle over $S^m$. Therefore, the property $(a)$, together with the proof of Proposition \[pi-manifold\], implies that $D(S^{2l-1}_+)\cong D(B(\nu_+))\cong S(\nu_+\oplus \textbf{1}) \cong M_+\times S^{m+1}$. In the remaining part, we will be concerned with the topology of $D(S^{2l-1}_-)$. On this account, we prove Theorem \[FKM\] in two steps. Step 1: As stated before, $D(S^{2l-1}_-)\cong D(B(\nu_-))\cong S(\nu_-\oplus \textbf{1})$ is the total space of a sphere bundle. Set $\zeta:=\nu_-\oplus\textbf{1}$, then $D(S^{2l-1}_-)=S(\zeta)$, where $\zeta$ is a vector bundle over $M_-$ with disc bundle $B(\zeta)$ and sphere bundle $S(\zeta)$, respectively. $$\begin{aligned} \label{zeta} &&\zeta: \mathbb{R}^{l-m+1} \hookrightarrow E(\zeta)\supset B(\zeta)\supset S(\zeta) = D(B(\nu_-))\\ &&\qquad\qquad\qquad\quad \downarrow \pi\nonumber\\ &&\qquad\qquad\quad\quad M^{m+l-1}_-\nonumber\end{aligned}$$ As mentioned in $(b)$ above, $M_-(m,l,q)$ is diffeomorphic to an $S^{l-1}$ bundle over $S^m$, that is: $M_-(m,l,q)\cong S(\xi)$, where $\xi$ is a vector bundle over $S^m$ so that $$\begin{aligned} \label{M-} && S^{l-1} \hookrightarrow M_-=S(\xi)\\ &&\qquad\qquad \downarrow \pi_1\nonumber\\ &&\qquad\qquad S^m\nonumber\end{aligned}$$ \[3.1\] The Pontrjagin class $$p_{\frac{m}{4}}(\zeta)=-\pi^_1p_{\frac{m}{4}}(\xi)=-q\cdot\beta(m)\cdot (\frac{m}{2}-1)!\cdot \pi_1^\gamma,$$ where $\beta(m)=\left\{\begin{array}{c c} 1, & m\equiv 0~ (\mod~8) \\ 2, & m\equiv 4~ (\mod~8) \end{array}\right.$, $\gamma \in H^m(S^m; \mathbb{Z})$ is a suitable generator. Denoting the total Pontrjagin class of an arbitrary vector bundle $\eta$ by $P(\eta)=1+p_1(\eta)+p_2(\eta)+...$, from [@MS], we know that $P(\eta\oplus\textbf{1})=P(\eta)$. Firstly, since $\nu_-\oplus TM_-\cong TS^{2l-1}\|_{M_-}$ is stably parallelizable (in fact it is trivial for the dimension reason), we have $P(\nu_-\oplus TM_-)=1$. On the other hand, since $m\geq 4$, $l=k\delta(m)\equiv 0~(\mod~4)$, thus $l-1$ can not be divided by $4$. By reason of rank of the sphere bundle $S(\xi)$, we can deduce that $P(M_-):=P(TM_-)=1+p_{\frac{m}{4}}(M_-)$. Consequently, since in this case the cohomology of $M_-$ with coefficients $\mathbb{Z}$ has no torsion ( Proposition \[homology\] ), it follows from $P(\zeta)=P(\nu_-\oplus \textbf{1})=P(\nu_-)$ that $$\label{p4m} p_{\frac{m}{4}}(\zeta)=-p_{\frac{m}{4}}(M_-).$$ Next, for the tangent bundle of $M_-$, we have $TM_-\oplus\textbf{1}\cong \pi_1^TS^m\oplus\pi_1^\xi$. As a direct result, $TM_-\oplus\textbf{2}=\textbf{(m+1)}\oplus\pi_1^\xi$, which implies that $p_{\frac{m}{4}}(M_-)=p_{\frac{m}{4}}(\pi^_1\xi)$. Since $l-1>m$, there exists a section of the sphere bundle $S(\xi)$, thus its Euler class $e$ vanishes. Hence, from the Gysin cohomology sequence with coefficients $\mathbb{Z}$ associated with $S(\xi)$: $$\label{Gysin} \rightarrow H^i(S^m)\xrightarrow{e}H^{i+l}(S^m)\xrightarrow{\pi_1^}H^{i+l}(M_-)\rightarrow H^{i+1}(S^m)\rightarrow\cdots,$$ we deduce that $$\label{pi1 isomorphism} \pi^_1:~~H^m(S^m; \mathbb{Z})\longrightarrow H^m(M_-; \mathbb{Z})~\qquad~~~~is~ an ~isomorphism.$$ Thus by (\[p4m\]) $$\label{p4m xi} p_{\frac{m}{4}}(\zeta)=-\pi^_1p_{\frac{m}{4}}(\xi)~\quad under~ the~ isomorphism ~\pi^_1.$$ At the mean time, $\xi-rank~\xi \in \widetilde{KO}(S^m)$, which will be abbreviated as $\xi$. Let us consider the complexification homomorphism $$\begin{aligned} \label{c} && \mathbb{C}:~\widetilde{KO}(S^m)~\rightarrow~\widetilde{K}(S^m)\\ &&\qquad\quad \quad\xi ~~\mapsto~~~\xi\otimes \mathbb{C}\nonumber\\ &&\qquad\quad \quad 1 ~~\mapsto~~~\left\{\begin{array}{c c} 1, & m\equiv 0~ (\mod~8) \\ 2, & m\equiv 4~ (\mod~8) \end{array}\right.\nonumber\end{aligned}$$ Recall a well known result that the Chern character $Ch:~\widetilde{K}(S^m)\longrightarrow H^m(S^m; \mathbb{Z})$ is an isomorphism for even $m$, namely, the top Chern class of the generator of $\widetilde{K}(S^m)$ is equal to a generator of $H^m(S^m; \mathbb{Z})$ multiplied by $(\frac{m}{2}-1)!$. By the isomorphisms $\xi\cong a\Delta^++b\Delta ^-\cong (a-b)\Delta^++b(\Delta^++\Delta ^-)\cong q\Delta^++\textbf{b}$, (since $\Delta^++\Delta ^-$ is trivial) where $\Delta^+\in \widetilde{KO}(S^m)$ is a generator, we finally arrive at $$\label{delta} p_{\frac{m}{4}}(\xi)= q\cdot\beta(m)\cdot (\frac{m}{2}-1)!\cdot \gamma \in H^m(S^m; \mathbb{Z})\cong \mathbb{Z}, $$ where $\gamma\in H^m(S^m; \mathbb{Z})$ is a suitable generator, and $\beta(m)=\left\{\begin{array}{c c} 1, & m\equiv 0~ (\mod~8) \\ 2, & m\equiv 4~ (\mod~8) \end{array}\right.$ In summary, $p_{\frac{m}{4}}(\zeta)=-\pi^_1p_{\frac{m}{4}}(\xi)=-q\cdot\beta(m)\cdot (\frac{m}{2}-1)!\cdot \pi_1^\gamma$. Step 2: We mainly make use of Wu Square $\mathcal{P}^1$ (cf. [@MS] ). Let $p:=2r+1\geq 3$ be an odd prime, and $m=2(p-1)=4r\equiv 0~(\mod~4)$. We will need the following fundamental criterion from number theory. Wilson’s Theorem: [$p$ is a primeif and only if$(p-1)!\equiv-1~(\mod~p)$. ]{} We recall the Wu Squares with coefficients $\mathbb{Z}_p$, which are generalized from Steenord Squares with coefficients $\mathbb{Z}_2$ by Wu Wen-Tsün: $$\mathcal{P}^i:~~H^j(X; \mathbb{Z}_p)\rightarrow H^{j+4ri}(X; \mathbb{Z}_p).$$ For $(B(\zeta), S(\zeta))\supset B(\zeta) \supset S(\zeta)$, one has the following commutative diagram of the cohomology sequences with coefficients $\mathbb{Z}_p$: $$\begin{aligned} \label{delta Pi} && \quad H^j(B(\zeta), S(\zeta))~~\rightarrow ~~ H^j(B(\zeta))~~ \rightarrow~~ H^j(S(\zeta))~~\xrightarrow{\delta}~~ H^{j+1}(B(\zeta), S(\zeta))\rightarrow \cdots\\ &&\qquad\quad\downarrow \mathcal{P}^i\qquad\qquad\qquad \downarrow \mathcal{P}^i\qquad\qquad\quad \downarrow \mathcal{P}^i\qquad\qquad\quad \downarrow \mathcal{P}^i\nonumber\\ &&H^{j+4ri}(B(\zeta), S(\zeta))\rightarrow H^{j+4ri}(B(\zeta))\rightarrow H^{j+4ri}(S(\zeta))\xrightarrow{\delta} H^{j+1+4ri}(B(\zeta), S(\zeta))\rightarrow\cdots\nonumber\end{aligned}$$ in particular, $\mathcal{P}^i$ satisfies $\delta\mathcal{P}^i=\mathcal{P}^i\delta$. Define $q_i(\zeta):=\Phi^{-1}\cdot \mathcal{P}^i\cdot \Phi(1)$, where $1\in H^0(M_-; \mathbb{Z}_p)$, and $\Phi$ is the Thom isomorphism $$\Phi: H^i(M_-; \mathbb{Z}) \longrightarrow H^{i+l-m+1}(B(\zeta), S(\zeta); \mathbb{Z}).$$ By Wu Theorem (cf. [@MS]), $q_i(\zeta)$ can be expressed in form of a combination of Pontrjagin classes $p_1(\zeta)$, $p_2(\zeta)$,..., $p_{ri}(\zeta)$. Observing $p_0(\zeta)=1$, $p_1(\zeta)=0$,...,$p_{r-1}(\zeta)=0$, $p_r(\zeta)\neq 0$, we want to represent $q_1(\zeta)$ by $p_r(\zeta)$. By Newton’s identities and Lemma \[3.1\], $$\begin{aligned} &&q_1(\zeta)\\ &\equiv& (-1)^{r+1}\cdot r\cdot p_r(\zeta)~(\mod~p)\\ &\equiv& (-1)^{\frac{m}{4}+1}\cdot r\cdot (-1)\cdot q \cdot \beta(m)\cdot (\frac{m}{2}-1)!\pi_1^(\gamma)~ (\mod~p)\\ &\equiv& (-1)^{\frac{m}{4}}\cdot q\cdot \beta(m)\cdot\frac{p-1}{2}\cdot(p-2)!\pi_1^(\gamma)~ (\mod~p)\end{aligned}$$ When $m\equiv 4 ~(\mod~8)$, a direct application of Wilson Theorem gives $q_1(\zeta)\equiv (-1)^{\frac{m}{4}+1}\cdot q\cdot \pi_1^(\gamma)~ (\mod ~p)$; When $m\equiv 0 ~(\mod~8)$, by Wilson Theorem, it is not difficult to show $\frac{1}{2}(p-1)!\equiv \frac{1}{2}(p-1)$ $(\mod~p)$, which yields that $q_1(\zeta)\equiv (-1)^{\frac{m}{4}}\cdot q\cdot\frac{p-1}{2}\cdot \pi_1^(\gamma)~ (\mod ~p)$. In summary, $$\label{p1 phi} q_1(\zeta)\equiv\left\{ \begin{array}{c c} (-1)^{\frac{m}{4}}\cdot q\cdot\frac{p-1}{2}\pi_1^(\gamma)~ (\mod ~p), & m\equiv 0 ~(\mod~8)\\ (-1)^{\frac{m}{4}+1}\cdot q\cdot \pi_1^(\gamma)~ (\mod ~p),& m\equiv 4 ~(\mod~8) \end{array}\right.$$ Fix $j=l-m$, $i=1$ in (\[delta Pi\]). Since $m<l-m<l-1$, $H^{l-m}(B)=0$, $H^l(B)=0$, so both $\delta$’s are injective: $$\begin{aligned} &&H^{l-m}(S) \autorightarrow{$\delta$}{inj.} H^{l-m+1}(B,S) \autoleftarrow {$\Phi$}{$\cong$} H^0(M_-)\cong \mathbb{Z}_p\\ &&\quad\downarrow\mathcal{P}^1\qquad\qquad\quad~~ \downarrow\mathcal{P}^1\nonumber\\ &&~~~H^{l}(S) ~~\autorightarrow{$\delta$}{inj.}~~ H^{l+1}(B,S) \autoleftarrow {$\Phi$}{$\cong$} H^m(M_-)\cong \mathbb{Z}_p\nonumber\end{aligned}$$ On the other hand, by Gysin cohomology sequence of $\zeta$ with coefficient $\mathbb{Z}_p$: $$\label{Gysin zeta} \rightarrow H^k(M_-)\xrightarrow{\cup e}H^{k+l-m+1}(M_-)\xrightarrow{\pi^}H^{k+l-m+1}(S(\zeta))\rightarrow H^{k+1}(M_-)\rightarrow\cdots$$ we get $H^{l-m}(S(\zeta))\cong \mathbb{Z}_p$, $H^l(S(\zeta))\cong \mathbb{Z}_p$, thus $\delta$ are isomorphisms since they are injective as we stated above. At last, by the definition $q_1(\zeta):=\Phi^{-1}\cdot \mathcal{P}^1\cdot \Phi(1)$, different values of $q$ give rise to different Wu squares $\mathcal{P}^1$, as we desired. The proof of Theorem \[FKM\] is now complete! The authors would like to thank Professor Hui Ma for her valuable conversations concerning the calculations on exceptional Lie groups. They also want to express their gratitude to Professor Kefeng Liu and the referee for their interest and helpful comments. [123]{} G. E. Bredon, Introduction to compact transformation groups, Academic Press, INC, Orlando, Florida, 1972. T. E. Cecil, Q. S. Chi, and G. R. Jensen, Isoparametric hypersurfaces with four principal curvatures, Ann. Math. 166* (2007), 1–76. Q. S. Chi, Isoparametric hypersurfaces with four principal curvatures, III, preprint, 2011, arXiv:1104.3249v3. T. E. Cecil and P. T. Ryan, Tight and taut immersions of manifolds, Research Notes in Math. 107, Pitman, London, 1985. D. Ferus, H. Karcher, and H. F. M[ü]{}nzner, Cliffordalgebren und neue isoparametrische Hyperflächen, Math. Z. 177 (1981), 479–502. For an English version, see arXiv: 1112.2780. M. Gromov and H. B. Lawson, Jr., The classification of simply connected manifolds of positive scalar curvature, Ann. of Math. 111 (1980), 423-434. M. Gromov and H. B. Lawson, Jr., Positive scalar curvature and the Dirac operator on complete Riemannian manifolds, Publ. Math. I.H.E.S., 58 (1983), 83-196 . J. Q. Ge and Z. Z. Tang, Geometry of isoparametric hypersurfaces in Riemannian manifolds, preprint, 2010, arXiv:1006.2577. J. Q. Ge and Z. Z. Tang, Isoparametric functions and exotic spheres, J. Reine Angew. Math. DOI: 10.1515/crelle-2012-0005, March 2012. J. Q. Ge, Z. Z. Tang and W. J. Yan, A filtration for isoparametric hypersurfaces in Riemannian manifolds, preprint, 2011, arXiv:1102.1126. N. Hitchin, Harmonic spinors, Adv. in Math. 14 (1974), 1-55. W. Y. Hsiang and H. B. Lawson, Minimal submanifolds of low cohomogenity, J. Diff. Geom. 5 (1971), 1–38. M. A. Kervaire and J. W. Milnor, Groups of homotopy spheres. I, Ann. of Math. 77 (1963), 504–537. A. Lichnerowicz, Spineurs harmoniques, C.R. Acad. Sci. Paris, Ser. A-B, 257 (1963), 7-9. R. Miyaoka, Isoparametric hypersurfaces with (g,m)=(6,2), to appear in Ann. Math. J. W. Milnor and J. D. Stasheff, Characteristic classes, Princeton University Press, Princeton, NJ, 1974. H.F. M[ü]{}nzner, Isoparametric hyperflächen in sphären, I and II, Math. Ann. 251 (1980), 57–71 and 256 (1981), 215–232. C. K. Peng and C. L. Terng, Minimal hypersurfaces of spheres with constant scalar curvature, Annals of Math. Studies, No.103, Princeton University Press, 1983, 177–198. J. Rosenberg and S. Stolz, Metrics of positive scalar curvature and connections with surgery, Surveys on Surgery Theory, Vol. 2, Annals of Math. Studies, No.149, Princeton University Press, 2001. R. Schoen and S. T. Yau, On the structure of manifolds with positive scalar curvature, Manuscripta Math. 28 (1979), 159–183. Z. Z. Tang, Nonexistence of weakly almost complex structures on Grassmannians, Proc. Amer. Math. Soc. 121 (1994), 1267–1270. G. Thorbergsson, A survey on isoparametric hypersurfaces and their generalizations, In Handbook of diff. geom., Vol. I, North - Holland, Amsterdam, 963 - 995, 2000. R. Takagi and T. Takahashi, On the principal curvatures of homogeneous hypersurfaces in a sphere, Differential Geometry, in honor of K. Yano, Kinokuniya, Tokyo, 1972. Q. M. Wang, On the topology of Clifford isoparametric hypersurfaces, J. Diff. Geom. 27 (1988), 55–66. I. Yokota, Exceptional Lie groups, preprint, 2009, arXiv:0902.0431v1.	{ "pile_set_name": "ArXiv" }
--- abstract: \| We study incentive compatible mechanisms for Combinatorial Auctions where the bidders have submodular (or XOS) valuations and are budget-constrained. Our objective is to maximize the liquid welfare, a notion of efficiency for budget-constrained bidders introduced by Dobzinski and Paes Leme (2014). We show that some of the known truthful mechanisms that best-approximate the social welfare for Combinatorial Auctions with submodular bidders through demand query oracles can be adapted, so that they retain truthfulness and achieve asymptotically the same approximation guarantees for the liquid welfare. More specifically, for the problem of optimizing the liquid welfare in Combinatorial Auctions with submodular bidders, we obtain a universally truthful randomized $O(\log m)$-approximate mechanism, where $m$ is the number of items, by adapting the mechanism of Krysta and V[ö]{}cking (2012). Additionally, motivated by large market assumptions often used in mechanism design, we introduce a notion of competitive markets and show that in such markets, liquid welfare can be approximated within a constant factor by a randomized universally truthful mechanism. Finally, in the Bayesian setting, we obtain a truthful $O(1)$-approximate mechanism for the case where bidder valuations are generated as independent samples from a known distribution, by adapting the results of Feldman, Gravin and Lucier (2014). author: - 'Dimitris Fotakis[^1]' - 'Kyriakos Lotidis[^2]' - 'Chara Podimata[^3]' bibliography: - 'refs.bib' title: A Bridge between Liquid and Social Welfare in Combinatorial Auctions with Submodular Bidders --- Introduction {#sec:intro} ============ Imagine that you are a social planner wanting to auction-off the seats of a local stadium at an extremely wealthy neighborhood (i.e., people have no budget constraints for the seats) for a big concert. As a social planner, your goal is to allocate the seats in a way that maximizes (or, at least, approximates as closely as possible) the happiness of the people interested in these seats. However, different people have different seat preferences; some people are happy with two consecutive seats anywhere in the stadium, and some might want a whole row. Phrased in mechanism design language, this is a Combinatorial Auction, where you seek to optimize the social welfare by a truthful mechanism. Combinatorial Auctions, like the one above, appear in many AI-centric contexts (e.g., spectrum auctions, network routing auctions [@HS01], airport time-slot auctions [@RSB82], etc.) and have been a central topic in the study of Multi-Agent Systems. They have also experienced a recent interest in the AI community with works employing ML algorithms to overcome standard complexity problems (e.g., [@BL18; @BLS18]). As if this problem was not hard enough to solve, imagine that you find out two unfortunate events; the stadium is in fact at a working-middle class neighborhood (i.e., people do have budget constraints) and your boss is concerned about the effect of these budget constraints on the potential revenue. Now, the objective function should balance between the willingness and the ability of the people to pay for their seats. Motivated by usual discrepancies between the auction participants’ ability and willingness to pay, @DPL14 introduced the notion of liquid welfare, which is the minimum of an agent’s budget and valuation for a bundle of goods. As such, maximizing the liquid welfare achieves a reasonable compromise between social efficiency and potential for revenue extraction (which is constrained by the budgets). [Problem Definition.]{} More formally, a Combinatorial Auction (CA) consists of a set $U$ of $m$ items to be allocated to $n$ bidders. Each bidder $i$ has a valuation function $v_i : 2^U \to \reals_{\geq 0}$. Valuation functions, $v$, are assumed to be non-decreasing, i.e., $v(S) \leq v(T)$, for all $S \subseteq T \subseteq U$, and normalized $v(\emptyset) = 0$. For the objective of social welfare (SW), the goal is to compute a partitioning $\mathcal{S} = (S_1, \ldots, S_n)$ of the set of items, $U$, that maximizes $v(\mathcal{S}) = \sum_{i=1}^n v_i(S_i)$. For the objective of liquid welfare (LW), we assume that each bidder $i$ also has a budget $B_i \in \reals_{\geq 0}$ and the liquid welfare that can be extracted from agent $i$ for each set of items $S \subseteq U$ is $\bar{v}_i(S) = \min\{ v_i(S), B_i \}$[^4]. Under this objective, the goal is to compute a partitioning $\mathcal{S} = (S_1, \ldots, S_n)$ of $U$ that maximizes $\bar{v}(\mathcal{S}) = \sum_{i=1}^n \bar{v}_i(S_i)$. We focus on CAs with submodular or XOS bidders. A set function $v: 2^U \to \reals_{\geq 0}$ is submodular if for every $S, T \subseteq U$, $v(S) + v(T) \geq v(S \cap T) + v(S \cup T)$ and subadditive if $v(S) + v(T) \geq v(S \cup T)$. A set function $v$ is XOS (a.k.a. fractionally subadditive, see [@F09]) if there exist additive functions $w_k : 2^U \to \reals_{\geq 0}$ such that for every $S \subseteq U$, $v(S) = \max_{k} \{ w_k(S) \}$. The class of submodular functions is a proper subset of the class of XOS functions, which in turn is a proper subset of the class of subadditive functions. Since bidder valuations have exponential size, a polynomial (in $m$ and $n$) algorithm must have oracle access to them. A value query specifies a set $S \subseteq U$ and receives the value $v(S)$. A demand query, denoted by $\DQ(v, U, \vec{p})$, specifies a valuation function $v$, a set $U$ of available items and a price $p_j$ for each available item $j \in U$, and receives the set (or bundle) $S \subseteq U$ maximizing $v(S) - \sum_{j \in S} p_j$, i.e., the set of available items that maximizes bidder’s utility at these prices. For brevity, we often write $p(S) = \sum_{j \in S} p_j$ to denote the price of a bundle $S$. Demand queries are strictly more powerful than value queries. Value queries can be simulated by polynomially many demand queries, and in terms of communication cost, demand queries are exponentially stronger than value queries [@BN09]. Our mechanisms are polynomial-time, given access to demand oracles, which in general can be NP-hard to compute. Previous Work on Social Welfare ------------------------------- Truthful maximization of SW in CAs with submodular or XOS bidders has been a central problem in Algorithmic Mechanism Design, with many powerful results. Due to space restrictions, we only discuss results most relevant to our work. While discussing previous work below, we assume XOS bidders and polynomial-time randomized truthful mechanisms that approximate the SW, by accessing valuations through demand queries, unless mentioned otherwise. In the worst-case setting, where we do not make any further assumptions on bidder valuations, @DNS06 presented the first truthful mechanism with a non-trivial approximation guarantee of $O(\log^2 m)$. @D07 improved the approximation ratio to $O(\log m \log \log m)$ for the more general class of subadditive valuations. Subsequently, @KV12 provided an elegant randomized online mechanism that achieves an approximation ratio of $O(\log m)$ for XOS valuations. @D16 broke the logarithmic barrier for XOS valuations, by providing an approximation guarantee of $O(\sqrt{ \log m })$. We highlight that accessing valuations through demand queries is essential for these strong positive results. @D11 proved that any truthful mechanism for submodular CAs with approximation ratio better than $m^{\frac{1}{2}-{\varepsilon}}$ must use exponentially many value queries. In the Bayesian setting, bidder valuations are drawn as independent samples from a known distribution. @FGL14 showed how to obtain item prices that provide a constant approximation ratio for XOS valuations. These results were significantly extended and strengthened in the recent work of @DFKL17, and a (truly) polynomial algorithm was provided as well. Intuition, Main Ideas, and Contribution --------------------------------------- Our aim is to extend these results to the objective of LW. To this end, we exploit the fact that most of the mechanisms above follow a simple pattern: first, by exploring either part of the instance in [@KV12] or the knowledge about the valuation distribution in [@FGL14], the mechanism computes appropriate (a.k.a. supporting) prices for all items. Then, these prices are “posted” to the bidders, who arrive one-by-one and select their utility-maximizing bundle, through a demand query, from the set of available items (see Algorithm \[algo:pp\]). The technical intuition behind the high level approach above is nicely explained in [@D16 Section 1.2]. Let ${\mathcal{O}}= (O_1, \ldots, O_n)$ be an optimal solution for the SW (in fact, any constant factor approximation suffices). The supporting price of item $i$ in $\mathcal{O}$ is $q_j = w_k(\{ j \})$, where $w_k$ is the additive valuation determining the value $v_i(O_i)$ (recall that valuation functions are XOS). Intuitively, $q_j$ is how much item $j$ contributes to the social welfare of $\mathcal{O}$. Then, a price of $p_j = q_j/2$ for each item $j$ is appropriate in the sense that a constant approximation to $v(\mathcal{O})$ can be obtained by letting the bidders arrive one-by-one, in an arbitrary order, and allocating to each bidder $i$ her utility maximizing bundle, chosen from the set of available items by a demand query (see [@D16 Lemma 4.2]). Hence, approximating the SW by demand queries boils down to computing such prices $p_j$. In the Bayesian setting, prices $p_j$ can be obtained by drawing $n$ samples from the valuation distribution and computing the expected contribution of each item $j$ to a constant factor approximation of the optimal allocation (see Section 3 and Lemma 3.4 in [@FGL14]). Similarly, the idea of estimating the contribution of the items would work under some market uniformity assumption, as the one introduced in Definition \[def:competitive\]. In the worst-case setting, if we assume integral and polynomially-bounded valuations (i.e., that $\max_{i} \{ v_i(U) \} \leq m^d$, for some constant $d$), a uniform price for all items selected at random from $1, 2, 4, 8, \ldots, 2^{d \log m}$ results in an logarithmic approximation ratio. @KV12 show how to estimate supporting prices online, by combining binary search and randomized rounding. Importantly, as long as each bidder does not affect the prices offered to her, this general approach results in (randomized, universally) truthful mechanisms. Towards extending the above approach and results to the LW, our first observation (Lemma \[lem:val-liqval\]) is that if a valuation function $v$ is submodular (resp. XOS), then the corresponding liquid valuation function $\bar{v} = \min\{ v, B \}$ is also submodular (resp. XOS). Then, one can directly use the mechanisms of e.g., [@KV12; @D16; @FGL14] with valuation functions $\bar{v} = \min\{ v, B \}$ and demand queries of the form: $\DQ(\min\{v, B\}, U, \vec{p})$ (i.e., wrt. the liquid valuation of the bidders) and obtain the same approximation guarantees but now for the LW. However, the resulting mechanisms are no longer truthful; bidders still seek to maximize their utility (i.e., value minus price) from the bundle that they get, subject to their budget constraint, rather than their liquid utility (i.e., liquid value minus price). Specifically, given a set of items $U$ available at prices $p_j$, $j \in U$, a budget-constrained bidder $i$ wants to receive the bundle $S_i = \arg\max_{S \subseteq U} \{ v_i(S) - p(S)\,\|\,p(S) \leq B_i \}$, and might not be happy with the bundle $S'_i = \arg\max_{S \subseteq U} \{ \bar{v}_i(S) - p(S) \}$ computed by the demand query for the liquid valuation[^5]. To restore truthfulness, we replace demand queries with budget-constrained demand queries. A budget-constrained demand query, denoted by $\BCDQ(v, U, \vec{p}, B)$, specifies a valuation function $v$, a set of available items $U$, a price $p_j$ for each $j \in U$ and a budget $B$, and receives the set $S \subseteq U$ maximizing $v(S) - p(S)$, subject to $p(S) \leq B$, i.e., the set of available items that maximizes bidder’s utility subject to her budget constraint. To establish the approximation ratio, we first observe that the fact that liquid valuations are XOS suffices for estimating supporting prices, as in previous work on the SW. Additionally, we show that the bundles allocated by $\BCDQ(v, U, \vec{p}, B)$ approximately satisfy the efficiency guarantees on the liquid welfare and the liquid utility of the allocated bundles (see Lemma \[lem:bfdo\]). Specifically, we observe that the approximation guarantees of mechanisms for the SW mostly follow from the fact that a demand query $\DQ(v, U, \vec{p})$ guarantees that for the allocated bundle $S$ and for any $T \subseteq U$, $v(S) - p(S) \geq v(T) - p(T)$, and $v(S) \geq v(T) - p(T)$ . In Lemma \[lem:bfdo\], we show that a budget-constrained demand query, $\BCDQ(v, U, \vec{p}, B)$, guarantees that for the allocated bundle $S$ and any $T \subseteq U$, $2 \bar{v}(S) - p(S) \geq \bar{v}(T) - p(T)$, and $\bar{v}(S) \geq \bar{v}(T) - p(T)$. Using this property, we can prove the equivalent of [@D16 Lemma 4.2]and also the approximation guarantees of the mechanisms in @KV12 [@FGL14] but for the LW. #### Contribution. Formalizing the intuition above, we obtain a randomized universally truthful mechanism that approximates the LW within a factor of $O(\log m)$ (Section \[sec:worst-case\]), and a posted-price mechanism that approximates the LW within a constant factor when bidder valuations are drawn as independent samples from a known distribution (Section \[sec:stochastic\]). Both mechanisms assume XOS bidder valuations; the former is based on the mechanism of @KV12 and the latter on the mechanism of @FGL14. Motivated by large market assumptions often used in Algorithmic Mechanism Design (see e.g., [@BCIMS05; @EFV17; @LX17] and the references therein), we introduce a competitive market assumption in Section \[sec:lcma\]. Competitive Markets are closer to practice, since they stand in between the stochastic and the worst-case settings, in terms of the assumptions made. The main idea is that when there is an abundance of bidders, even if we remove a random half of them, the optimal LW does not decrease by much. Then, computing supporting prices for all items based on a randomly chosen half of the bidders, and offering these prices through budget-constrained demand queries to the other half, yields a universally truthful mechanism that approximates LW within a constant factor (Theorem \[thm:cm-apx\]). Conceptually, in this work, we present a general approach through which known truthful approximations to the SW, that access valuations through demand queries, can be adapted so that they retain truthfulness and achieve similar approximation guarantees for the LW. The important properties required are that liquid valuation functions $\bar{v}$ belong to the same class as valuation functions $v$ (proven for submodular, XOS and subadditive valuations), and that the efficiency guarantees of budget-constrained demand queries on liquid welfare and liquid utility are similar to the corresponding efficiency guarantees of standard demand queries for liquid valuations (proven for all classes of valuations functions). Indeed, applying this approach to the mechanism of @D16, we obtain a universally truthful mechanism that approximates the LW for CAs with XOS bidders within a factor of $O(\sqrt{\log m})$ (the details are omitted due to space constraints). Similarly, we can take advantage of the improved results of @DFKL17 in the Bayesian setting. All the missing proofs can be found in the full version of the paper on [@FLP18]. Previous Work on Liquid Welfare ------------------------------- Liquid welfare was introduced as an efficiency measure for auctions when bidders are budget constrained in [@DPL14] (since it was known that getting any non-trivial approximation for the SW in these cases is impossible) and it corresponds to the optimal revenue an omniscient seller could extract from the set of the bidders, had he known their valuations and their budgets. Moreover, @DPL14 proved a $O( \log n )$ (resp. $(\log^2 n)$)-approximation to the optimal LW for the case of a single divisible item and submodular (resp. subadditive) bidders. @DPL14 and @LX15 proved that the optimal LW can be approximated truthfully within constant factor for a single divisible good, additive bidder valuations and public budgets. Closer to our setting, @LX17 provided a truthful mechanism that achieves a constant factor approximation to the LW for multi-item auctions with divisible items, under a large market assumption. Under similar large market assumptions, @EFV17 obtained mechanisms that approximate the optimal revenue within a constant factor for multi-unit online auctions with divisible and indivisible items, and a mechanism that achieves a constant approximation to the optimal LW for general valuations over divisible items. However, prior to our work, there was no work on approximating the LW in CAs (in fact, that was one of the open problems in [@DPL14]). Our work is remotely related to the literature of budget feasible mechanism design, a topic introduced by @S10 and studied in e.g., [@DPS11; @CGL11; @BCGL12; @BH16; @WLL18]. Budget feasible mechanism design focuses on payment optimization in reverse auctions, a setting almost orthogonal to the setting we consider in this work. Notation and Preliminaries {#sec:model} ========================== The problem and most of the terminology and the notation are introduced in Section \[sec:intro\]. In this section, we introduce some additional notation required for the technical part. We use $\operatorname{\mathbb{E}}[X]$ to denote the expectation of a random variable $X$ and $\operatorname{\mathbb{P}}[E]$ to denote the probability of an event $E$. Let ${\text{OPT}}$ (resp. ${\overline{\text{OPT}}}$) denote the optimal SW (resp. LW)[^6]. For some $\rho > 1$, which may depend on $n$ and $m$, we say that a mechanism is $\rho$-approximate for the optimal SW (resp. LW) if it produces an allocation ${\mathcal{S}}$ with $\rho \cdot v({\mathcal{S}}) \geq {\text{OPT}}$ (resp. $\rho \cdot {\bar{v}}({\mathcal{S}}) \geq {\overline{\text{OPT}}}$). Let a social choice function $f: {\bar{V}}^n \to A$, which maps the set of liquid valuation functions of the bidders, ${\bar{V}}: V \times B$, to an allocation, $A$, and a payment scheme $q = (q_1, \dots, q_n)$ for this allocation. A deterministic mechanism is defined by the pair $(f,q)$. Our mechanisms in this work are going to be randomized, i.e., they are probability distributions over deterministic mechanisms. The incentives desiderata for randomized mechanisms are usually either universal truthfulness (when for all the deterministic mechanisms, the bidders’ dominant strategy is truthfulness) or truthfulness in expectation [@DFK10; @DRY11] (when bidders’ expected utilities are maximized under truthful reporting of their private information). In this work, we are focusing on the former, stronger notion; the one of universal truthfulness, under the bidders’ budget constraints. \[def:truth-budgets\] Let $(\tilde{f}, \tilde{q})$ be a randomized mechanism over a set of deterministic mechanisms $\{(f^1, q^1 ), \dots, (f^k, q^k ) \}$. Mechanism $(\tilde{f}, \tilde{q})$ is universally truthful if for all $i \in [n], \kappa \in [k]$ and for any $v_i'$ and any $B_i'$, such that $q^\kappa(v_i', v_{-i}) \leq B_i'$ and $q^\kappa(v_i, v_{-i}) \leq B_i$, it holds that: v\_i(f\^(v\_i,B\_i, v\_[-i]{},B\_[-i]{})) - q\^(v\_i,B\_i, v\_[-i]{}, B\_[-i]{}) v\_i(f\^(v\_i’, B\_i’, v\_[-i]{},B\_[-i]{})) - q\^(v\_i’, B\_i’, v\_[-i]{}, B\_[-i]{}) Approach {#sec:approach} ======== First, we show that if the bidder valuations are submodular (resp. XOS, subadditive), then their liquid valuations are submodular (resp. XOS, subadditive) as well. \[lem:val-liqval\] Let $v$ be a non-negative monotone submodular (resp. XOS, subadditive) function. Then, for any $B \in \reals_{\geq 0}$, ${\bar{v}}= \min\{ v, B \}$ is also monotone submodular (resp. XOS, subadditive). In Algorithm \[algo:pp\], we present a universally truthful (since the prices offered to each bidder do not depend on her declaration and demand queries maximize bidders’ utility) mechanism, which is a simplified version of the mechanism in [@KV12] for approximating SW in CAs. Since for the LW, bidders have budgets, we replace the demand queries $\DQ(v_i, U_i, \vec{p}^{(i)})$ in line 4 with budget constrained demand queries $\BCDQ(v_i, U_i, \vec{p}^{(i)}, B_i)$[^7]. As explained in Section 1.2, Algorithm \[algo:pp\] with $\BCDQ$s remains universally truthful for budget-constrained bidders. Fix an ordering $\pi$ of bidders and set $U_1 = U$. Set initial prices for the items: ${\vec{p}}^{(1)} = (p_1^{(1)}, \dots, p_m^{(1)})$. Let $S_i = \DQ(v_i, U_i, {\vec{p}}^{(i)},)$ With probability $q$, allocate $S_i$ to $i$ and set $U_{i+1} = U_i \setminus S_i$. Otherwise, set $U_{i+1} = U_i$. Update item prices to ${\vec{p}}^{(i+1)} = ( p_1^{(i+1)}, \dots, p_m^{(i+1)})$. For budget-constrained bidders, Algorithm \[algo:pp\] with $\BCDQ$s in line 4, is universally truthful. The lemma follows directly from Definition \[def:truth-budgets\]. Nevertheless, universal truthfulness is not our sole desideratum; in each of the three settings analyzed in the following sections, we show why mechanisms similar in spirit to Algorithm \[algo:pp\] with $\BCDQ$s, yield good approximation guarantees for the LW. Before the setting-specific analyses, we relate the efficiency of $\BCDQ$ to the efficiency of standard $\DQ$s for liquid valutions. \[lem:bfdo\] Let $S \subseteq U$ be the set allocated by the BCDQ for a bidder with valuation $v$ and budget $B$. Then, for every other $T \subseteq U$, the following hold: 1. \[lem:prop1\] ${\bar{v}}(S) \geq {\bar{v}}(T) - p(T)$ 2. \[lem:prop2\] $2{\bar{v}}(S) - p(S) \geq {\bar{v}}(T) - p(T)$. We will prove each claim of the lemma separately. For claim \[lem:prop1\], notice that if $p(T) > B$, then the right hand side of the inequality will be negative and thus, the inequality trivially holds. So, we will focus on the case where $p(T) \leq B$. We distinguish the following cases: 1. (${\bar{v}}(S) = v(S)$ and ${\bar{v}}(T) = v(T)$) Hence, $B \geq v(T)$. Bundle $T$ was considered at the time of the query and yet, the query returned set $S$. Thus: ${\bar{v}}(S) \geq {\bar{v}}(S) - p(S) = v(S) - p(S) \geq v(T) - p(T) = {\bar{v}}(T) - p(T)$. 2. (${\bar{v}}(S) = B$ and ${\bar{v}}(T) = B$) Then, the inequality trivially holds since: $B \geq B - p(T)$ and prices are non-negative. 3. (${\bar{v}}(S) = B$ and ${\bar{v}}(T) = v(T)$) The inequality holds since: $B \geq B - p(T) \geq v(T) - p(T) = {\bar{v}}(T) - p(T)$. 4. (${\bar{v}}(S) = v(S)$ and ${\bar{v}}(T) = B$) Hence, $B \leq v(T)$. Bundle $T$ was considered at the time of the query and yet, the query returned set $S$. Thus, ${\bar{v}}(S) \geq {\bar{v}}(S) - p(S) = v(S) - p(S) \geq v(T) - p(T) \geq B - p(T) = {\bar{v}}(T) - p(T)$. This concludes our proof for claim \[lem:prop1\]. For claim \[lem:prop2\], notice that since $S$ is the set received from the BCDQ, then it is affordable: ${\bar{v}}(S) \geq p(S)$. Adding this inequality to the inequality of claim \[lem:prop1\], we have that: $2{\bar{v}}(S) - p(S) \geq {\bar{v}}(T) - p(T)$. Worst-Case Setting {#sec:worst-case} ================== For the worst-case instances, adapting appropriately our Core Mechanism, we present Algorithm \[algo:kv12\] (based again, on the mechanism of [@KV12]). The only difference is that budget-constrained bidders in Algorithm \[algo:kv12\] are restricted to using $\BCDQ$s, instead of $\DQ$s, thus making the mechanism universally truthful (see Section \[sec:approach\]). Resembling the analysis of [@KV12], we show that for $1/q = \Theta(\log m)$, Algorithm \[algo:kv12\] achieves an approximation ratio of $O(\log m)$ for the LW. First, we note that parameter[^8] $L$ is selected so that there exists only one bidder whose liquid valuation for $U$ (weakly) exceeds it. Fix an ordering $\pi$ of bidders and set $U_1 = U$. Set initial prices $p_1^{(1)} = \cdots = p_m^{(1)} = \frac{L}{4m}$. Let $S_i = \BCDQ(v_i, U_i, {\vec{p}}^{(i)}, B_i)$ With probability $q$, allocate $R_i = S_i$ to $i$ and set $U_{i+1} = U_i \setminus S_i$. Otherwise, set $U_{i+1} = U_i, R_i = \emptyset$. Update prices $\forall j \in S_i$: $p_j^{(i+1)} =2 p_j^{(i)}$. \[thm:lw-wc\] Algorithm \[algo:kv12\] is universally truthful and for $q = 1/\Theta(\log m)$, achieves an approximation ratio of $O(\log m)$ for the LW. We present a series of Lemmas that will lead us naturally to the proof of the Theorem. Let ${\mathcal{S}}= (S_1, \dots, S_n)$ and ${\mathcal{R}}= (R_1, \dots, R_n)$ the provisional and the final allocation of Algorithm \[algo:kv12\] respectively. First, we provide two useful bounds on ${\bar{v}}({\mathcal{S}})$. We find it important to also discuss an overselling variant of Algorithm \[algo:kv12\]. In the Overselling variant, allow us to assume that for Step 5 of Algorithm \[algo:kv12\], $q=1$ (i.e., $S_i$ is allocated to bidder $i$ with certainty) and $U_{i+1} = U_i = U$ (thus the name of the variant). The Overselling variant allocates at most $k=\log(4m)+2$ copies of each item and collects a liquid welfare within a constant factor of the optimal LW. To see that, observe that for $q=1$, after allocating $k-1$ copies of some item $j$, $j$’s price becomes $\frac{L}{4m} 2^{\log(4m)+1} = 2L$. Then, there is only one agent with liquid valuation larger than $L$ who can get a copy of $j$. \[lem:lb-1\] Let $p_j$ denote the final price of each item $j$. Then, for any sets $U_1, \dots, U_n \subseteq U$ of items available when the bidders arrive, Algorithm \[algo:kv12\] with $q = 1$ satisfies ${\bar{v}}({\mathcal{S}}) \geq \sum_{j \in U} p_j - \frac{L}{4}$. Since bidders are individually rational and do not violate their budget constraints, for every bidder $i$ it holds that $B_i \geq \sum_{j \in S_i} p_j^{(i)}$ and $v_i(S_i) \geq \sum_{j \in S_i} p_j^{(i)}$. The rest of the proof is identical to the proof of [@KV12 Lemma 2] for $b=1$. Specifically, let $\ell_j^{(i)}$ be the number of copies of item $j$ allocated just before bidder $i$ arrives, and let $\ell_j$ be the total number of copies of item $j$ allocated by Algorithm \[algo:kv12\] with $q = 1$. Then, using the fact that $p_j =L \cdot \frac{2^{\ell_j}}{4m}$: $$\begin{aligned} {\bar{v}}({\mathcal{S}}) & \geq \sum_{i=1}^n \sum_{j \in S_i} p_j^{(i)} = \frac{L}{4m} \sum_{i=1}^n \sum_{j \in S_i} 2^{\ell_j^{(i)}} = \frac{L}{4m} \sum_{j \in U} (2^{\ell_j} - 1) = \sum_{j \in U} p_j - \frac{L}{4}\end{aligned}$$ \[lem:lb-2\] For sets $U_1 = \dots = U_n \subseteq U$, the Overselling variant of Algorithm \[algo:kv12\] with $q = 1$ satisfies ${\bar{v}}({\mathcal{S}}) \geq {\overline{\text{OPT}}}- \sum_{j \in U} p_j$. Let ${\mathcal{O}}= (O_1, \dots, O_n)$ be the optimal allocation. From Lemma \[lem:bfdo\], we get that for each bidder $i$, ${\bar{v}}(S_i) \geq {\bar{v}}(O_i) - \sum_{j \in O_i} p_j^{(i)} \geq {\bar{v}}(O_i) - \sum_{j \in O_i} p_j$, where we use that the final price of each item is the largest one. Summing over all bidders, we have that ${\bar{v}}({\mathcal{S}}) \geq {\bar{v}}({\mathcal{O}}) - \sum_{i =1}^n \sum_{j \in O_i} p_j \geq {\overline{\text{OPT}}}- \sum_{j \in U}p_j$, where the last inequality uses the fact that the optimal solution is feasible and thus, each item is allocated at most once in ${\mathcal{O}}$. \[lem:oversell-apx\] The Overselling variant of Algorithm \[algo:kv12\] with $q = 1$ allocates at most $\log(4m)+2$ copies of each item and computes an allocation ${\mathcal{S}}$ with liquid welfare ${\bar{v}}({\mathcal{S}}) \geq \frac{3}{8} {\overline{\text{OPT}}}$. Follows directly from Lemma \[lem:lb-1\], Lemma \[lem:lb-2\] and the fact that ${\overline{\text{OPT}}}\geq L$. Of course, the allocation ${\mathcal{S}}$ in Lemma \[lem:oversell-apx\] is infeasible, since it allocates a logarithmic number of copies of each item. The remedy is to use an allocation probability $q = 1/\Theta(\log m)$. For such values of $q$, we can plug in the proof of [@KV12 Lemma 6] (which just uses that the valuation functions are fractionally subadditive) and show that for each agent $i$ and for all $A \subseteq U$, $\operatorname{\mathbb{E}}[{\bar{v}}_i(A \cap U_i)] \geq {\bar{v}}_i(A)/2$. We are now ready to conclude the proof of Theorem \[thm:lw-wc\]. For Algorithm \[algo:kv12\] with $q^{-1} = 4(\log(4m) + 1)$, it holds that $\operatorname{\mathbb{E}}[{\bar{v}}({\mathcal{S}})] \geq {\overline{\text{OPT}}}/8$ and $\operatorname{\mathbb{E}}[{\bar{v}}({\mathcal{R}})] \geq q\,{\overline{\text{OPT}}}/8$. Let ${\mathcal{O}}= ( O_1, \ldots, O_n )$ be the optimal allocation. For each bidder $i$, Lemma \[lem:bfdo\] implies that the response $S_i$ of $\BCDQ$ satisfies ${\bar{v}}_i(S_i) \geq {\bar{v}}_i(O_i \cap U_i) - \sum_{j \in O_i \cap U_i}p_j^{(i)}$, for any $U_i$ resulted from the outcome of the random coin flips. Therefore, $\operatorname{\mathbb{E}}[{\bar{v}}_i(S_i) ] \geq \operatorname{\mathbb{E}}[ {\bar{v}}_i(O_i \cap U_i) ] - \operatorname{\mathbb{E}}[\sum_{j \in O_i \cap U_i} p_j^{(i)}]$. By the choice of $q$, for any bidder $i$, $\operatorname{\mathbb{E}}[ {\bar{v}}_i ( O_i \cap U_i )] \geq {\bar{v}}_i(O_i)/2$. Then, working with the expectations as in the proofs of Lemma \[lem:lb-1\], Lemma \[lem:lb-2\] and Lemma \[lem:oversell-apx\], we can show that $\operatorname{\mathbb{E}}[{\bar{v}}({\mathcal{S}})] \geq {\overline{\text{OPT}}}/8$. Finally, one can use linearity of expectation and show that $\operatorname{\mathbb{E}}[{\bar{v}}({\mathcal{R}})] = q \operatorname{\mathbb{E}}[{\bar{v}}({\mathcal{S}})]$. The details are omitted, due to lack of space, and can be found in [@KV12 Lemma 4]. Competitive Market {#sec:lcma} ================== @BCIMS05 were the first ones to define a budget dominance parameter that corresponded to the ratio of the maximum budget of all the bidders to the value of the optimum SW in the context of multi-unit auctions with budget-constrained bidders. More recently, @EFV17 and @LX17 used similar notions of budget dominance[^9] (termed large market assumptions) as a means to achieve constant factor approximation to the LW in multi-unit auctions and auctions with divisible items respectively. However, for the case of indivisible items, it is clear that the definition of a large market used in the previous cases, becomes almost void (see Appendix \[appendix:lcma\] for a discussion). Below, we first introduce our definition of Competitive Markets for indivisible goods and then, show how one can obtain a constant factor approximation of the optimal LW, when bidders have XOS liquid valuations. \[def:competitive\] Let $0 \leq {\varepsilon}< 2$ and a constant $0 \leq \delta \leq 1/2$. A market is called $({\varepsilon}, \delta)$ - Competitive Market, if for any randomly removed set of bidders, ${\mathbb{S}}$, with cardinality $n/2$, then for the remaining set of bidders, ${\mathbb{T}}$, it holds that: $$\operatorname{\mathbb{P}}\!\left[{\overline{\text{OPT}}}_{{\mathbb{T}}} \geq \left(1 - \frac{{\varepsilon}}{2}\right) \cdot {\overline{\text{OPT}}}\right] \geq 1 - \delta$$ where by ${\overline{\text{OPT}}}_{{\mathbb{T}}}$ we denote the optimal LW achieved by bidders in set ${\mathbb{T}}$. In an $({\varepsilon},\delta)$ - Competitive Market, let ${\mathbb{S}}\subseteq [n]$ be randomly chosen s.t. $\|{\mathbb{S}}\| = \frac{n}{2}$ and let ${\mathbb{T}}= [n] \setminus {\mathbb{S}}$. Then: $$\begin{aligned} \operatorname{\mathbb{P}}\!&\left[\left\{{\overline{\text{OPT}}}_{{\mathbb{T}}} \geq \left(1-\tfrac{{\varepsilon}}{2}\right){\overline{\text{OPT}}}\right\}\cap \left\{{\overline{\text{OPT}}}_{{\mathbb{S}}} \geq \left(1- \tfrac{{\varepsilon}}{2}\right){\overline{\text{OPT}}}\right\}\right] \\ &\geq 1 - 2\delta\end{aligned}$$ Let $X_{{\mathbb{S}}}$ the event that ${\overline{\text{OPT}}}_{{\mathbb{S}}} \geq \left(1-\frac{{\varepsilon}}{2}\right){\overline{\text{OPT}}}$ and $X_{{\mathbb{T}}}$ the event that ${\overline{\text{OPT}}}_{{\mathbb{T}}} \geq \left(1-\frac{{\varepsilon}}{2}\right){\overline{\text{OPT}}}$. Then, we have: $$\operatorname{\mathbb{P}}\left[ X_{{\mathbb{S}}} \cap X_{{\mathbb{T}}} \right] = 1 - \operatorname{\mathbb{P}}\left[ \overline{X_{{\mathbb{S}}}} \cup \overline{X_{{\mathbb{T}}}} \right] \geq 1 - 2\delta$$ where the inequality follows from the Union Bound. We are now ready to state our Competitive Market mechanism that will be used for approximating the optimal LW. We note here that the greedy algorithm ${\mathcal{A}}$ is due to @LLN06. Divide the bidders into sets ${\mathbb{S}}, {\mathbb{T}}$ uniformly at random, s.t., $\|{\mathbb{S}}\| = \frac{n}{2} = \|{\mathbb{T}}\|$. Run the greedy algorithm ${\mathcal{A}}$ for bidders in ${\mathbb{S}}$ and denote the solution obtained by ${\mathcal{A}}^{{\mathbb{S}}}$. Set $p_j = \frac{1}{2\beta}{\bar{v}}\left( {\mathcal{A}}_j^{{\mathbb{S}}} \right)$, where $\beta > 1$ is a constant Fix an internal ordering of bidders in ${\mathbb{T}}$, $\pi$, and set $U_1 = U$. Let $S_i = \BCDQ(v_i, U_i, {\vec{p}})$. Set $U_{i+1} = U_i \setminus S_i$. As usual, we denote ${\mathcal{S}}= (S_1, \dots, S_n)$ the final allocation from mechanism presented in Algorithm \[algo:lcma\]. Valuations of bidders are XOS (and so are the liquid valuations (Lemma \[lem:val-liqval\])); let $a_i$ be the maximizing clause of $S_i$ in the liquid valuation ${\bar{v}}_i$ of bidder $i$. Since $a_i$’s are additive, for each bidder $i$ and $j \in S_i$ let $q_j = a_i (\{j\})$. Notice that $\sum_{i \in [n]} {\bar{v}}(S_i) = \sum_{j \in \cup_{i \in [n]} S_i}q_j$. We denote by ${\overline{\text{OPT}}}_{{\mathbb{T}}} = \sum_{j \in U} q_j^{{\mathbb{T}}}$, where $q_j^{{\mathbb{T}}}$ is the contribution of item $j$ in ${\overline{\text{OPT}}}_{{\mathbb{T}}}$. We divide the set of all items $U$ into two sets; the set of competitive items, denoted by ${\mathcal{C}}$ and the set of non-competitive items, denoted by ${\overline{\mathcal{C}}}= {\mathcal{M}}\setminus {\mathcal{C}}$. The following lemma upper bounds the contribution of non-competitive items in the optimal solution. \[lem:contr-non-compet\] Let ${\mathcal{C}}= \{j \big\| q_j^{{\mathbb{T}}} > \frac{{\bar{v}}({\mathcal{A}}_j^{{\mathbb{S}}})}{\beta}\}$ for constant $\beta >1$. Then, $\sum_{j \in {\overline{\mathcal{C}}}} q_j^{{\mathbb{T}}} \leq \frac{{\varepsilon}}{2(\beta - 1)}{\overline{\text{OPT}}}$ and $\sum_{j \in {\mathcal{C}}}q_j^{{\mathbb{T}}} \geq \frac{\beta(2-{\varepsilon})-2}{2(\beta - 1)}{\overline{\text{OPT}}}$. From Definition \[def:competitive\], it holds with constant probability (w.c.p) that: $${\overline{\text{OPT}}}\geq \sum_{j \in {\mathcal{C}}}q_j^{{\mathbb{T}}} + \sum _{j \in {\overline{\mathcal{C}}}}q_j^{{\mathbb{T}}} = \sum _{j \in U}q_j^{{\mathbb{T}}} \geq \left(1-\frac{{\varepsilon}}{2}\right) \cdot {\overline{\text{OPT}}}$$ Let ${\mathbb{S}}_{{\overline{\mathcal{C}}}} \subseteq {\mathbb{S}}$ be the set of the bidders that are allocated the non-competitive items from the greedy algorithm ${\mathcal{A}}$ when running on set ${\mathbb{S}}$. Then, in the augmented set ${\mathbb{T}}\cup {\mathbb{S}}_{{\overline{\mathcal{C}}}}$, there exists an allocation ${\mathcal{Q}}$[^10] with liquid valuation, $$\label{eq:lv1} {\bar{v}}({\mathcal{Q}}) \geq \sum_{j \in {\mathcal{C}}} q_j^{{\mathbb{T}}} + \sum_{j \in {\overline{\mathcal{C}}}} {\bar{v}}\left({\mathcal{A}}_j^{{\mathbb{S}}}\right)$$ and therefore we have w.c.p: $$\begin{aligned} {\overline{\text{OPT}}}&\geq {\bar{v}}({\mathcal{Q}}) \geq \sum_{j \in {\mathcal{C}}} q_j^{{\mathbb{T}}} + \sum_{j \in {\overline{\mathcal{C}}}} {\bar{v}}\left({\mathcal{A}}_j^{{\mathbb{S}}}\right) \geq \sum_{j \in {\mathcal{C}}}q_j^{{\mathbb{T}}} + \beta \sum _{j \in {\overline{\mathcal{C}}}}q_j^{{\mathbb{T}}} \\ &\geq \left(1 -\frac{{\varepsilon}}{2}\right){\overline{\text{OPT}}}+ (\beta -1)\sum _{j \in {\overline{\mathcal{C}}}}q_j^{{\mathbb{T}}} \end{aligned}$$ Re-arranging the latter and using the fact that $$\sum_{j \in {\mathcal{C}}} q_j + \frac{{\varepsilon}}{2(\beta - 1)} {\overline{\text{OPT}}}\geq \sum_{j \in U}q_j^{{\mathbb{T}}} \geq \left(1-\frac{{\varepsilon}}{2}\right){\overline{\text{OPT}}}$$ then, for the items in ${\mathcal{C}}$ it holds w.c.p that: $\sum_{j \in {\mathcal{C}}}q_j^{{\mathbb{T}}} \geq \frac{\beta(2-{\varepsilon})-2}{2(\beta - 1)}{\overline{\text{OPT}}}$. In the next Lemma, we prove a lower bound on the contribution of competitive items to the solution obtained by the greedy algorithm, with respect to ${\overline{\text{OPT}}}$. \[lem:bound-alg\] $\sum_{j \in {\mathcal{C}}} {\bar{v}}\left( {\mathcal{A}}_j^{{\mathbb{S}}} \right) \geq \frac{2 (\beta - 1) - {\varepsilon}\cdot (3\beta -1)}{4(\beta-1)}{\overline{\text{OPT}}}$. Combining Inequality and Lemma \[lem:contr-non-compet\] we get that $\sum_{j \in {\overline{\mathcal{C}}}} {\bar{v}}\left( {\mathcal{A}}_j^{{\mathbb{S}}} \right) \leq \frac{\beta {\varepsilon}}{2 (\beta-1)} {\overline{\text{OPT}}}$. Algorithm ${\mathcal{A}}$ provides a $2$-approximation to the optimal LW of set ${\mathbb{S}}$ [@LLN06], so w.c.p we have: $$\sum_{j \in {\mathcal{C}}} {\bar{v}}\left( {\mathcal{A}}_j^{{\mathbb{S}}} \right) + \sum_{j \in {\overline{\mathcal{C}}}} {\bar{v}}\left( {\mathcal{A}}_j^{{\mathbb{S}}} \right) \geq \frac{1}{2} {\overline{\text{OPT}}}_{{\mathbb{S}}} \geq \frac{1 - \frac{{\varepsilon}}{2}}{2} {\overline{\text{OPT}}}$$ Combining the last two equations, we get the result. \[thm:cm-apx\] The CM Algorithm is universally truthful and achieves, in expectation, a constant approximation to the optimal LW, i.e., $$\operatorname{\mathbb{E}}\left[ {\bar{v}}\left( {\mathcal{S}}\right) \right] \geq (1 - 2\delta) \cdot \frac{2(\beta -1)-{\varepsilon}\cdot(3\beta-1)}{16\beta(\beta-1)}{\overline{\text{OPT}}}$$ Since the bidders that control the prices being posted belong to set ${\mathbb{S}}$ and they never get any item, it is their (weakly) dominant strategy to report their valuations and their budgets truthfully. Furthermore, the bidders that are buying under the said posted prices belong to set ${\mathbb{T}}$ and they make BCDQs, which we showed to be truthful. Finally, the bidders are uniformly at random* split to sets ${\mathbb{S}}$ and ${\mathbb{T}}$. For each item $j \in {\mathcal{C}}$ we have $q_j^{{\mathbb{T}}} > {\bar{v}}({\mathcal{A}}_j^{{\mathbb{S}}})/\beta$. Therefore, there exists an allocation for bidders in ${\mathbb{T}}$ and items in ${\mathcal{C}}$ that is supported by prices $p_1, \dots, p_m$, where $p_j ={\bar{v}}({\mathcal{A}}_j^{{\mathbb{S}}})/\beta$. Thus, a modification of [@D16 Lemma 4.2] implies that if we we set $p_j' = p_j/2$, for each $j \in {\mathcal{C}}$, and run a fixed price auction in ${\mathbb{T}}$ with prices $p_1',\dots,p_m'$, we get that ${\bar{v}}({\mathcal{S}}) \geq \sum_{j \in {\mathcal{C}}} p_j/4$. Using the latter, along with the prices of the items, we have that $${\bar{v}}({\mathcal{S}}) = \frac{1}{4\beta}\sum_{j \in {\mathcal{C}}} {\bar{v}}({\mathcal{A}}_j^{{\mathbb{S}}})\geq \frac{2(\beta -1)-{\varepsilon}(3\beta-1)}{16\beta(\beta-1)} {\overline{\text{OPT}}}$$ where the last inequality is due to Lemma \[lem:bound-alg\]. Thus, we conclude that $\operatorname{\mathbb{E}}\left[{\bar{v}}\left( {\mathcal{S}}\right)\right] \geq \left(1-2\delta\right)\frac{2(\beta -1)-{\varepsilon}\cdot (3\beta-1)}{16\beta(\beta-1)}{\overline{\text{OPT}}}$. Bayesian Setting {#sec:stochastic} ================ The Bayesian Setting offers a great middle ground between the unstructured worst-case instances and the very structured Competitive Markets. In this setting, let ${\vec{v}}= (v_1, \dots, v_n)$ be a profile of bidder valuations and ${\mathbf{B}}= (B_1, \dots, B_n)$ a profile of bidder budgets. Assume that the bidders’ valuations are drawn independently from distributions ${\mathcal{V}}_1, \dots, {\mathcal{V}}_n$ and the budgets from distributions ${\mathcal{B}}_1, \dots, {\mathcal{B}}_n$. For simplicity, let us assume that their liquid valuations are drawn independently from distributions ${\mathcal{D}}_1, \dots, {\mathcal{D}}_n$. We will denote by ${\mathcal{D}}= {\mathcal{D}}_1 \times \dots \times {\mathcal{D}}_n$ the product distribution where liquid valuations profiles, ${\mathbf{\bar{v}}}= ({\bar{v}}_1, \dots, {\bar{v}}_n)$, are independently drawn from. We are going to show that the results presented in @FGL14 can be extended for budget-constrained bidders. Specifically, we are going to show that, if liquid valuations are fractionally subadditive, then we can create appropriate prices such that, when presented to the bidders in a posted-price mechanism and bidders are making BCDQs, then we can obtain universally truthful constant-factor approximation mechanisms for the LW in Bayesian CAs. Our Lemma \[lem:like3.4\] establishes the existence of such appropriately scaled prices. The key component activating our results is that instead of reasoning about the utility* achieved from the bundle purchased by bidder $i$ (as received by the BCDQ), we instead have to use Lemma \[lem:bfdo\]. We also note that using our techniques one could even achieve the better approximation guarantees presented by @DFKL17. Their analysis is significantly more complex, however, and we omit it in the interest of space. \[thm:anon-price\] Let distribution ${\mathcal{D}}$ over XOS liquid valuation profiles be given via a sample access to ${\mathcal{D}}$. Suppose that for every ${\mathbf{\bar{v}}}\sim {\mathcal{D}}$, we have: black-box access to a LW maximization algorithm, ${\text{ALG}}$[^11] an XOS value query oracle (for liquid valuations sampled from ${\mathcal{D}}$)[^12]. Then, for any $\epsilon > 0$, we can compute item prices in ${\text{POLY}}(m,n,1/\epsilon)$ time such that, for any bidder arrival order, the expected liquid welfare of the posted price mechanism is at least $\frac{1}{4}\operatorname{\mathbb{E}}_{{\mathbf{\bar{v}}}\sim {\mathcal{D}}} [{\mathbf{\bar{v}}}({\text{ALG}}({\mathbf{\bar{v}}}))] - \epsilon$, where by ${\text{ALG}}({\mathbf{\bar{v}}})$ we denote the solution produced by algorithm ${\text{ALG}}$. \[lem:like3.4\] Given a distribution ${\mathcal{D}}$ over XOS liquid valuations, let ${\vec{p}}$ be the price vector s.t. $p_j = \frac{1}{2} \operatorname{\mathbb{E}}_{{\mathbf{\bar{v}}}\sim {\mathcal{D}}} [{\text{LW}}_j({\mathbf{\bar{v}}})]$. Let ${\vec{p}}'$ be any price vector such that $\|p_j' - p_j\| < \delta$ for all $j \in [m]$. Then, for any arrival order, $\pi$, bidders buying bundles by making BCDQs under prices ${\vec{p}}'$ results in expected liquid welfare at least $\frac{1}{4} \operatorname{\mathbb{E}}_{{\mathbf{\bar{v}}}\sim {\mathcal{D}}} [{\mathbf{\bar{v}}}({\text{ALG}}({\mathbf{\bar{v}}}))] - \frac{m\delta}{2}$. Conclusion {#sec:concl} ========== In real-life auctions, bidders are always* constrained by budgets, which we tend to overlook due to the technical difficulties that they add. The role of budgets in welfare/revenue optimization is amplified in CAs, where bidders have richer valuations and hence, studying budgeted CAs is a step towards bridging the gap between the theory on truthful mechanism design for CAs and constraints faced in practice. In this work, we showed how the liquid welfare can be approximated in CAs where bidders are budget-constrained in three settings: worst-case, Competitive Markets and stochastic. The most meaningful question that arises from our work (apart, of course, from the ever existent one of lowering the approximation guarantee in worst-case instances) is related to the competitive markets. We conjecture that the condition that we provide can be made even weaker, and leave it to future research. Finally, our results can also be used to extend a variety of already known results in CAs without budgets, to CAs with budget-constrained bidders. For example, Lemma \[lem:bfdo\] (with some changes in the constants of [@DK17]) implies a constant factor approximation for best response dynamics in XOS CAs with budgeted bidders, that apply after a single round of bid updates. Acknowledgments =============== We are thankful to Yiling Chen for her most valuable feedback in an initial draft of this work, as well as the anonymous reviewers for their helpful comments and suggestions. Part of this work was done while Chara Podimata was visiting Yahoo Research, NY. This work is partially supported by the National Science Foundation under grant CCF-1718549 and the Harvard Data Science Initiative. Any opinions, findings, conclusions, or recommendations expressed here are those of the authors alone. [^1]: Yahoo! Research, New York, NY USA and National Technical University of Athens, Athens, Greece, [^2]: National Technical University of Athens, Athens, Greece, [^3]: Harvard University, Cambridge, MA, USA, [^4]: Slightly abusing the terminology, we refer to ${\bar{v}}_i(S)$ as agent $i$’s liquid valuation. [^5]: For a concrete example, consider a bidder with budget $B = 2$ and two items $a$ and $b$ available at prices $p_a = 2$ and $p_b = 1$. Assume that the bidder’s valuation function is $v(\{ a \}) = v(\{ a, b \}) = 10$, $v(\{ b \}) = 2$ (and therefore, her liquid valuation is $\bar{v}(\{ a \}) =\bar{v}(\{ b \}) = \bar{v}(\{ a, b \}) = 2$). The bidder wants to get item $a$ at price $2$, which gives her utility $8$. However, the demand query for her liquid valuation function $\bar{v}$ allocates item $b$, which gives her a utility of $1$. Clearly, in this example, the bidder would have incentive to misreport her preferences to the demand query oracle. [^6]: The instance is clear from the context. [^7]: In all our mechanisms, if budgets are larger than the valuations of the allocated bundles, the mechanism with $\BCDQ$ behaves identically to the mechanism with $\DQ$ (i.e., revenue and SW are not affected by the change in the objective.) [^8]: $L$ can be computed with standard techniques, as explained in [@KV12]. [^9]: Namely, $\forall i \in [n]: B_i \leq {\overline{\text{OPT}}}/(mc)$, with $c$ a large constant. [^10]: Allocation ${\mathcal{Q}}$ is realized by allocating all items in ${\mathcal{C}}$ to bidders in ${\mathbb{T}}$ that also had them in the ${\overline{\text{OPT}}}_{{\mathbb{T}}}$ allocation and all items in ${\overline{\mathcal{C}}}$ to the bidders in ${\mathbb{S}}_{{\overline{\mathcal{C}}}}$ that had them in the allocation of the greedy ${\mathcal{A}}$. The claim is completed by submodularity. [^11]: ${\text{ALG}}$ can be any algorithm that provides a $O(1)$-approximation to the optimal LW, since we do not care about incentives (access to ${\text{ALG}}$ will only happen for ghost samples). [^12]: An XOS value oracle takes as input a set $T$ and returns the corresponding additive representative function for the set $T$, i.e., an additive function $A_i(\cdot)$, such that (i) ${\bar{v}}_i(S) \geq A_i(\hat{S})$ for any $\hat{S} \subset [m]$ and (ii) ${\bar{v}}_i(T) = A_i(T)$.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We consider convex combinations of finite-valued almost periodic sequences (mainly substitution sequences) and put them as potentials of one-dimensional tight-binding models. We prove that these sequences are almost periodic. We call such combinations [hybrid quasicrystals]{} and these studies are related to the minimality, under the shift on both coordinates, of the product space of the respective (minimal) hulls. We observe a rich variety of behaviors on the quantum dynamical transport ranging from localization to transport.' author: - \| Túlio O. Carvalho\ [Departamento de Matemática – UEL, CP 6001, Londrina, PR, 86051-970 Brazil]{}\ and\ César R. de Oliveira[^1]\ [Departamento de Matemática – UFSCar, São Carlos, SP, 13560-970 Brazil]{} title: 'Hybrid Quasicrystals, Transport and Localization in Products of Minimal Sets' --- Introduction ============ The study of transport in one-dimensional aperiodic lattices may be modeled by the nearest-neighbors tight-binding Hamiltonian (Schrödinger operator) in $l^2(\Z)$ $$\label{hamiltonian} (H\psi)_n=\psi_{n+1}+\psi_{n-1}+\lambda V_n\psi_n,$$ with $\lambda>0$ and potentials $V=(V_n)_{n\in\Z}$ generated by aperiodic sequences. In many circumstances the potentials are real-valued functions of sequences on a finite set $\A$, called [alphabet]{}; these are models of one-dimensional quasicrystals [@BQ]. Quantum interferences may lead to localization of the solutions of the corresponding Schrödinger equation $$i\frac{\partial}{\partial t}\psi(t)= H\psi(t),$$ as in case of (random) Bernoulli potentials [@GdeB], but also to ballistic motion, mainly related to periodic potentials. Among the characterizations of localization and transport we single out the second moment of the position operator $$\label{moments} m_2(T):= \sum_{n=-\infty}^ \infty \|n-n_0\|^2 \|\psi_n(T)\|^2,$$ usually with initial condition concentrated on a single site $n_0$. For a large class of potentials the moment $m_2(T)\le C T^2$ (at least for $T>1$) and if $m_2(T)\approx C T^2$ holds we have the definition of [ballistic motion]{}. Localization will be characterized by a bounded function $m_2(T)\le C$, $\forall T$; lack of localization is usually referred to as [delocalization]{} or [transport]{}. Half the way between these extremes are the [anomalous transport]{}, that is, $$m_2(T)\approx C T^\beta \quad\text{with}\quad0<\beta<2,$$ which are usually accompanied by singular continuous spectrum of the operator $H$. Important examples of such anomalous behavior are the above cited models of quasicrystals, among which the most prominent are the (primitive) substitution sequences [@BQ; @Q], for instance, Fibonacci, Thue-Morse and Period Doubling sequences. The Schrödinger operators whose potentials are generated by these sequences have singular continuous spectrum of zero Lebesgue measure (see [@DReview] and references therein). A widespread spectral point of view makes the association of singular continuous spectrum with anomalous transport, absolutely continuous to ballistic motion and point spectrum of the Schrödinger operator with localization, even though there are known exceptions, namely of operators with purely point spectra showing transport. Even rank one perturbation (a very localized one) can exchange point and singular continuous spectra [@SW], and the latter surely implies transport (any continuous spectrum does, as a consequence of RAGE theorem). What about unlocalized perturbations, i.e., those spread over the whole lattice? Certainly this becomes a too huge class of problems to be reasonably dealt with. However, there is a special type of such perturbations we think it is worth considering and may be of some (experimental) relevance in the near future. A particular model of quasicrystal, as a substitution sequence, is an almost periodic sequence that grows up from a seed (i.e., an initial condition) and a specific “growing rule.” If one has control of the growing technique, one could grow a quasicrystal in one direction following one such rule, and in a perpendicular direction following another rule. This hybridization creates a potential which is a linear (convex) combinations of the original ones. This type of long range perturbations of the potentials can also be considered from the theoretical point of view, the sequence spaces are two-dimensional and have been considered before [@vuil]. The potentials we shall consider are constructed as follows. Given two [parent]{} potentials $v=(v_n)_{n\in\Z}$, $u=(u_n)_{n\in\Z}$ and $0\le \kappa\le 1$, the [hybrid potential]{} is $$\mathcal{I}_\kappa(v,u):=\kappa v+(1-\kappa)u = \left(\kappa v_n+(1-\kappa)u_n\right)_{n\in\Z}.$$ Experience with random potentials indicates that if one of them is random then this characteristic will prevail with respect to localization. If both potentials are periodic, then the resulting one will also be periodic with period given by their least common multiple. So, in these extreme cases again, localization and ballistic motion, respectively, are persistent. Note that the number of values a hybrid potential assumes is in general larger than the number of values of each of its components; e.g., if both $v,u$ take values in $\{0,1\}$, then $\mathcal{I}_\kappa(v,u)$ will generally assume all values in $\{0,\kappa,(1-\kappa),1\}$. This work is an initial study on this proposal, and we will limit ourselves to almost periodic potentials taking a finite number of values (notably, substitution sequences). We present theoretical results on minimality, and data for the moment $m_2(T)$ from numerical time evolution simulations. Section \[sectBRSS\] review briefly some aspects of finitely valued sequences. In Section \[sectMinimality\] we address the question about minimality of the product of minimal sets, giving a sufficient condition for it. In Section \[sectNumerics\] we report some outcomes of numerical simulations of the moment $m_2(T)$ for the hybrid potentials, closing in the final section with our concluding remarks. Summary on Sequences and Substitutions {#sectBRSS} ====================================== We denote by $\A^$ (resp. $\A^\Z$) the sets of finite (resp. bi-infinite) words with letters in the finite set $\A$ (called [alphabet]{}), which can be considered a subset of the real numbers. The metric on $\A^\Z$ is $$d(a,b) = \begin{cases} 0, \text{ if } \forall n\in \Z, a_n=b_n \\ \frac{1}{2^n}, \text{ where } n=\min\{\|j\| : a_j\neq b_j\} \end{cases} .$$ A dynamics on this set is the (left) shift $(\sigma (v))_n=v_{n+1}$. Recall that a sequence $v\in\A^\Z$ is almost periodic iff its hull (the bar indicates the closure of the set) $$\Omega(v) := \overline{\{\sigma^j(v): j\in\Z\}}$$is minimal, that is, the hull of any sequence in $\Omega(v)$ coincides with $\Omega(v)$. The set $\mathcal{O}(v)=\{\sigma^j(v): j\in\Z\}$ is the orbit of $v$. By Tychonov theorem $\A^\Z$ is compact and so is every hull as above. The minimality is an important property of the hull of (primitive) substitution sequences (see ahead), as well as the existence of a unique ergodic measure, and up to now rigorous and numerical studies have revealed just one dynamical behavior in each minimal component (usually rigorous results are restricted to generic or full measure sets). So, as a first step in the study of such new class of systems, in this work we address the problem of minimality of the hull of hybrid sequences in case their respective parent potentials are almost periodic. Setting a product metric, the dynamics with respect to which one considers minimality, is on the product space of the hulls of the two parent potentials, and is generated by the (natural) product shift $$\sigma(u,v)_n=(u_{n+1},v_{n+1}) . \label{prodshift}$$ We use the same notation for the shift in two and one-dimensional sequences. In order to investigate the minimality of the product spaces it turns out to be important hybridizing not only of $v$ and $u$, but also of elements of their one-dimensional orbits; namely, to consider $\mathcal{I}_\kappa(v,\sigma^j(u))$, for each $j\in\Z$. A finite word $w$ is indexed $a_0a_1\cdots a_{\|w\|-1}$, $a_i\in \A$, where $\|w\|$ denotes the length of $w$. Given a set of infinite words $X$, the [language]{} of $X$, ${\cal L}(X)$, is the set of finite words occurring in some $w\in X$. Let us describe some substitution rules which generate sequences of interest for this work; details can be found in [@BQ; @Q]. Given a finite alphabet $\A$ a substitution is a map $\xi:\A\to \A^ $. Its iterations are defined by concatenation, that is, $\xi(abc):=\xi(a)\xi(b)\xi(c)$, $\xi^{n+1}(a):=\xi(\xi^n(a))$, $n\ge1$. A substitution is [primitive]{} if there exists $k\in\N$ so that for every $a\in\A$ the word $\xi^k(a)$ contains all letters of $\A$. All substitutions in this work are primitive (see [@LdO] for some nonprimitive substitutions as potentials of Schrödinger operators). A fixed point of a substitution is a sequence $u\in \A^\N$ such that $\xi(u)=u$. In order to exist, it must be the case that $u_0$ is the first letter of $\xi(u_0)$. It is known that if $\xi$ is primitive, there is some $l$ such that $\xi^l$ has a fixed point [@Q], so it is no loss to assume $\xi$ has a fixed point. Fibonacci ([Fcc]{}), Period Doubling ([PD]{}) and Thue-Morse ([TM]{}) substitution sequences are constructed with an alphabet of two letters $\{a,b\}$ through the rules $$a\mapsto ab, \;b\mapsto a\; {\rm(Fcc), \;\;\;}\;\;\; a\mapsto ab, \;b\mapsto ba \;\;\;{\rm(TM),}$$ $$a\mapsto ab, \;b\mapsto aa\; {\rm(PD).}$$ Beginning with $a$ (the seed) and applying successively the substitution rules (the growing rules), aperiodic sequences are obtained; e.g., the Thue-Morse sequence is given by $$abbabaabbaababba\cdots$$ The paper folding ([PF]{}) sequence can be obtained with an alphabet of four letters $\{1,2,3,4\}$, the substitution $$1\mapsto 12, \;\;\;2\mapsto 32,\;\;\; 3\mapsto 14,\;\;\; 4\mapsto 34,$$ (the seed is $1$) and then applying the literal map $1,2\mapsto a$ and $3,4\mapsto b$. We then use these substitution sequences to define our potentials $V$; we take $V_n=-1$ if the $n$-th letter of the sequence is $a$ and $V_n=1$ in case it is $b$. There are standard ways of extending such substitution potentials for negative values of $n$ [@BG; @HKS]. We do not have to deal with this issue in numerical simulations because we take a large finite sample of $N$ sites, using the initial wavefunction concentrated on position $N/2$, i.e., $\psi_n(t=0)=\delta_{N/2,n}$, $n\ge0$. This is the procedure we use to construct almost periodic substitution potentials $V$. It is known that the spectrum of the operator (\[hamiltonian\]) with finite-valued aperiodic and almost periodic potentials has no absolutely continuous component (primitive substitutions are included) [@HKS; @K]; although from a rigorous point of view the lack/presence of eigenvalues in cases of primitive substitution sequences is an open question, as already remarked, no strong evidence of the presence of eigenvalues and localization was found yet. Given a substitution $\xi$ over a finite alphabet $\A$, denote by $M_\xi$ its substitution matrix, i.e., $M_\xi=a_{w,w'}$, where $a_{w,w'}$ is the number of occurrences of the letter $w'$ in $\xi(w)$. $\xi$ is a [Pisot substitution]{} if the dominant eigenvalue of $M_\xi$ has modulus greater than one, while all the other eigenvalues have absolute values strictly less than one. For example, the matrix substitution for [TM]{} and [Fcc]{} substitution are $$M_{\text{\sc TM}}=\left( \begin{array}{cc} 1 & 1 \\ 1 & 1\end{array}\right)\quad\text{and}\quad M_{\text{\sc Fcc}}=\left( \begin{array}{cc}1 & 1 \\ 1 & 0\end{array}\right),$$ whose dominant eigenvalues are $2$ and $(1+\sqrt{5})/2$, respectively. The dominant eigenvalue of the [PD]{} substitution is $2$, but it is not Pisot, since the other eigenvalue is $-1$. [PF]{} is not Pisot either. Minimality of Hybrid Hulls {#sectMinimality} ========================== Let $v$ and $u$ denote almost periodic sequences and $\Omega(v)$, $\Omega(u)$ be their respective hulls. In the product space $\Omega(v)\times \Omega(u)$ we have the shift defined by $\sigma(x,y)=(\sigma(x),\sigma(y))$. This dynamics does not imply that the product space is minimal if the parent hulls $\Omega(v)$ and $\Omega(u)$ are minimal. The orbit of a point $(x,y)$ is $\O(x,y)=\{\sigma^n(x,y) : n\in\Z\}$. For each $\kappa$, there is a correspondence between elements of this product space and hybrid sequences $\mathcal{I}_\kappa(v_l,u_l)$, $v_l\in \Omega(v)$, $u_l\in \Omega(u)$. The potential is a real-valued function on one such sequence. It is of interest to know whether the potential is almost-periodic. We present in this section some results concerning minimality on the product space of minimal subsets of $\A^\Z$. Given $\epsilon>0$, the subset of integer numbers $$\{n\in\Z :\ d(\sigma^n(x),x)<\epsilon \}$$ is called the set of $\epsilon$-periods of $x\in X\subset \A^\Z$. When $X$ is minimal, the above set is [syndetic]{}, i.e., there is an integer $m$ such that any interval $[n,n+m]\subset \Z$ intersects it. Recall that $x$ is almost periodic iff that set is syndetic for all $\epsilon>0$; in this case of finite-valued sequences this is equivalent to the fact that every finite word in $x$ appears with bounded gaps. There is an alternative view of periods in terms of words, or equivalently the cylinder sets generated by them. If $a\in\A^\Z$, let $R_a(w)$ denote the set of integers $n$ such that $$a_na_{n+1}\cdots a_{n+\|w\|-1}=w .$$ Thus $R_a(w)$ is the set of integers $n$ for which $w$ is a [prefix]{} of $\sigma^na$. It can be ordered $$R_a(w)= \{\alpha_i,i\in\Z:\alpha_i<\alpha_{i+1}\}$$ for some arbitrary choice of $\alpha_0$. The minimality of $X$ is equivalent to the fact that for each finite word $w$ that occurs in $X$ there is an integer $m(w)$ so that $\alpha_{k+1}-\alpha_k<m(w)$, $\forall k$ (i.e., $w$ occurs with bounded gaps). Similarly, for $b\in Y\subset \A^\Z$, $Y$ minimal, let $R_b(u)=\{\beta_j,\ j\in\Z:\beta_j<\beta_{j+1}\}$ (the notation should be clear). In the product space $X\times Y$ we seek a description of the possible minimal sets under the shift and metric $$D((a,b),(c,d)):=d_X(a,c)+d_Y(b,d),\quad a,c\in X,\quad b,d\in Y.$$ The existence of these minimal sets is a consequence of $X\times Y$ compactness and Zorn’s Lemma. Picture $X\times Y$ as the orbit closure of the union of $(a,\sigma^n(b))$, $n\in \Z$. If we represent the sequence $a$ along a horizontal lattice $(\cdot,0)\subset \Z^2$ and $\sigma^r(b)$ along vertical lattices, each attached to the corresponding horizontal position $(r,0)$, the orbit $(a,\sigma^rb)$ is the left translation of the horizontal line $(\cdot,0)$. Analogously the orbit of $(\sigma^k(a),b)$ may be followed by pulling horizontally the line at $(\cdot,k)$. We begin to address the question about minimality of $\overline{\O(a,b)}$ by asking if, as one sits at different positions along the horizontal axis, one sees the same pair of finite words $u$, $w$ upwards and to the right respectively, infinitely often. While this certainly happens at each $\beta_n$ and $\alpha_n$ alternatively upwards and to the right, one is interested in these words appearing at the same time [and]{} with bounded gaps. If $X\subset \A^\Z$ and $Y\subset \B^\Z$ are minimal sets, then $X\times Y$ decomposes into finitely many minimal sets. \[teo1\] Suppose on the contrary that we had infinitely many invariant sets $M_i$. Choose a point in each $M_i$ and an open set $U_i$ containing it but with no intersection with $M_j$, $j>i$. Complete this cover with $V_i=M_i\setminus U_i$ (recall that $U_i$ is also closed since we are dealing with product of cylinders). From the cover of $U_i$ and $V_i$’s we cannot extract a finite subcover, but $X \times Y$ is compact. Theorem 1.17 in [@furst] yields a point whose orbit closure is a minimal set. We can show that this holds for every point in $X\times Y$, whenever $X$ and $Y$ are minimal sets. \[propOrbMin\] Suppose $X,Y\subset \A^\Z$ are minimal. Given a point $z\in X\times Y$, its orbit closure $\overline{\O(z)}$ is minimal. Pick a point $(a,\sigma^jb)$ from a minimal set $M\subset X\times Y$. Then, for any $\epsilon>0$, there exists a sequence $(n_p)_{p\in \N}$, $n_p \nearrow \infty$, $\|n_{p+1}-n_p\|$ bounded such that $$D((a,\sigma^j(b)),\sigma^{n_p}(a,\sigma^j(b)))< \epsilon .$$ Now pick a point $(\sigma^ka,\sigma^lb)\in X\times Y$. For any $n_p>h=\max\{\|l-j\|,\|k\|\}$ $$\begin{split} & D(\sigma^k(a,\sigma^{l-k}(b)),\sigma^{n_p+k}(a,\sigma^{l-k}(b)))= \\ &\; d_X(\sigma^k a,\sigma^{n_p+k}a)+d_Y(\sigma^lb,\sigma^{n_p+l}b) \\ &\; \le 2^k d_X(a,\sigma^{n_p}a)+ 2^{\|l-j\|}d_Y(\sigma^j b,\sigma^{n_p+j}b) <2^{h+1}\epsilon \end{split}$$ and this can be made arbitrarily small. Since any $z\in X\times Y$ belongs to the closure of the orbit of some $(\sigma^ka,\sigma^lb)$, the proposition is proved. This proves the assertion in the abstract If $X,Y\subset \A^\Z$ are minimal sets, then a sequence $z\in X\times Y$ is almost periodic, as well as any sequence obtained from it by some real-valued function defined on $X\times Y$. In what follows, unless stated on the contrary, we assume that $X$ and $Y$ are minimal sets. Now the question is to characterize when the product $X\times Y$ is minimal. Suppose there exists a sequence $n_k\nearrow \infty$ so that $\sigma^{n_k}a\rightarrow a^$ and $\{\sigma^{n_k+l}b: n_k\}$, for some $l$ fixed, is dense in $Y$. Then $X\times Y$ is minimal. Our hypothesis asserts that $(a^,Y)$ is contained in the orbit closure $\overline{\O(a,\sigma^l(b))}$. Due to Proposition \[propOrbMin\], it is enough to show that there is a dense orbit in $X\times Y$. Given $(x,y)\in X\times Y$ and $\epsilon>0$ $$\begin{split} &D((\sigma^na,\sigma^{n+l}b),(x,y)) \\ &\ \leq D((\sigma^na,\sigma^{n+l}(b)),(\sigma^j(a^),y))+D((\sigma^j(a^),y),(x,y)) \ . \end{split}$$ Since $X$ is minimal, the second term may be made less than $\epsilon/3$, and this fixes $j$. Now $$D((\sigma^na,\sigma^{n+l}(b)),(\sigma^j(a^),y)) =d(\sigma^n(a),\sigma^j(a^))+d(\sigma^{n+l}(b),y) .$$ We note that along the subsequence $n=n_k+j$ the first term is less than $\epsilon/3$ for every $n_k$ sufficiently large. The set of points $\sigma^{n_k+l}b$, $n_k>N$, is dense in $Y$, $N$ a fixed arbitrary integer. Therefore, if $z=\sigma^{-j}y$, there exists $n_k$ so that $d(\sigma^{n_k+l}b,z)<{\epsilon'}$. Choosing $\epsilon'$ small enough yields $d(\sigma^{n_k+(j+l)}b,y)<\epsilon/3$, for some $n_k>N$. Therefore, if $X\times Y$ is not minimal, then for every convergent sequence $\sigma^{n_k}(a)\rightarrow a^$ one has that $\sigma^{n_k+l}(b)$ is not dense, for any $l$. We know that the dynamics of a map on the torus ${\mathbb T}^d$: $T({\mathbf\theta})={\mathbf\theta}+{\mathbf\alpha}\pmod 1$ is ergodic when ${\mathbf \alpha}$ is rationally independent. By coding this dynamics with a partition along each circle, we get a symbolic sequence which is semi-conjugate to the original dynamics [@alessandri]. This is an example of a product of two sequences spaces which is minimal. We say that $(w,u)$ is a [prefix]{} of $(a,b)$ if $w$ is a prefix of $a$ and $u$ is a prefix of $b$. It is easy to characterize lack of minimality in terms of the language of $X\times Y$. Indeed, if ${\cal L}(a,b)$ denote the set of words of $(a,b)\in X\times Y$, we have ${\cal L}(a,b)\subset {\cal L}(a)\times {\cal L}(b)$. Hence, if $X\times Y$ is not minimal, for each invariant set $M\subset X\times Y$, there are words $r\in {\cal L}(a)$ and $s\in {\cal L}(b)$ such that $(r,s)$ does not occur in $M$. If $a$ is an almost periodic sequence, note that ${\cal L}(a)={\cal L}(\Omega(a))$. By hull of a substitution we understand the hull of any of its fixed points. For primitive substitution sequences we get a simple criterion for minimality of the product of their hulls. Recall that this case is our choice of prototypes of hybrid quasicrystals. The argument comes from the proof of a result Hansel in [@han] related to Cobham’s Theorem (see also [@cob; @dur]). Recall that two positive numbers $\theta$ and $\vartheta$ are multiplicatively independent if the equation $\theta^l=\vartheta^k$ holds only for $l=k=0$. \[thmMultIndepMin\] Let $\xi$ and $\zeta$ be two primitive substitutions on the (finite) alphabets $\A$ and $\B$, respectively, and denote by $X$ and $Y$ their respective hulls under the shift. If $M_\xi$ and $M_\zeta$ have multiplicatively independent dominant eigenvalues, then $X\times Y$ is minimal. Let $a=\xi(a)=(a_j)_{j\in\Z}$ and $b=\zeta(b)=(b_j)_{j\in\Z}$ be fixed points of the corresponding substitutions and $M=\overline{\O(a,b)}$. If $X\times Y$ is not minimal, then $M\ne X\times Y$ and there is a finite word $(r,s)$ in $X\times Y$ that does not occur in $M$. Thus, for any $r_0,r_1,s_0,s_1$, $\|r_0\|=\|s_0\|$, $(r_0rr_1,s_0ss_1)$ does not occur in the orbit of $(a,b)$ either. Since the substitutions are primitive, there is a $k$ so that for all $n\geq k$ the words $\xi^n(w)$, $w\in\A$, contain $r$, and $\zeta^n(u)$, $u\in \B$, contain $s$. We choose $r_0$ and $r_1$ so that $\xi^n(w)=r_0rr_1$ above. Then $s_0$ and $s_1$ are chosen so that $\zeta^n(u)$ contains $s_0ss_1$. We conclude that for any pair $(w,u)\in \A\times\B$ there is some $n_0$ (which may be taken [big]{}) so that $$(\xi^{n_0}(w),\zeta^{n_0}(u)) \text{ is not a prefix of } \sigma^k(a,\sigma^{-l}b) \text{ for every } k \text{ and some } l. \label{one}$$ Let $\theta$ and $\vartheta$ be the dominant eigenvalues of $M_\xi$ and $M_\zeta$ respectively. Consider the subsets of $\N$ $$\begin{aligned} & E(X)=\{\|\xi(a_0a_1\cdots a_m)\|, m>0\}, \\ & E(Y)=\{\|\zeta(b_0b_1\cdots b_m)\|, m>0\} .\end{aligned}$$ $E(X)$ contains some of the positions where the words $\xi^j(w),\ \forall j, \forall w\in\A$, occur in $a$. By Lemma 2 in [@han], for large enough $m$, these positions are the integer numbers closest to $a\theta^{pj}+b$ for some integer $p>0$, and real $a>0,b$. The same holds for $E(Y)$, in that it contains integers closest to $a'\vartheta^{qj}+b'$, for some integer $q>0$, and real $a'>0,b'$. But $\theta^p$ and $\vartheta^q$ are multiplicatively independent, so the set of ratios $$\frac{a\theta^{pj}+b}{a'\vartheta^{qj}+b'}$$ is dense in $\R^+$. Therefore, it must be the case that the intersections $E(X)\cap E(Y)$ and $E(X)\cap \{E(Y)-l\}$, for any $l$, are not empty. This contradicts (\[one\]). In the case of Pisot substitutions, the proof is simpler in that the $E(X)$ will contain integer numbers close to $\theta^{pj}$, while $E(Y)$ contains integers close to $\vartheta^{qj}$, for integer $j>0$. If $\theta$ and $\vartheta$ are multiplicatively independent, the same argument follows. \[corolExemMin\] If $\Omega_{\text{\sc TM}}$, $\Omega_{\text{\sc Fcc}}$ and $\Omega_{\text{\sc PD}}$ are the hulls of the indicated substitution sequences, then the products $\Omega_{\text{\sc TM}}\times\Omega_{\text{\sc Fcc}}$ and $\Omega_{\text{\sc PD}}\times\Omega_{\text{\sc Fcc}}$ are minimal. One can investigate whether some product spaces generated by constant length substitutions are not minimal in a case by case analysis. For instance if $\xi$ denotes the Thue-Morse substitution and $\eta$ is the period-doubling substitution, then contained in the above described $X\times Y$, one has the following subset $\Omega$. Let $\B$ denote the four letter alphabet: $\{(a,a),(a,b),(b,a),(b,b)\}$. On $\cal B$ we define the substitution $$\zeta(x,y):=(\xi(x),\eta(y)) .$$ Explicitly, $\zeta(a,a)=(ab,ab)=(a,a)(b,b)$, $\zeta(a,b)=(ab,aa)=(a,a)(b,a)$, $\zeta(b,a)=(ba,ab)=(b,a)(a,b)$ and $\zeta(b,b)=(ba,aa)=(b,a)(a,a)$. This substitution $\zeta$ can be shown to be primitive with two fixed points. The fixed points belong to the same hull, since the languages of the fixed points of a primitive substitution coincide. In a four letter alphabet, Berstel has considered a substitution isomorphic to $\zeta$ when constructing square-free words [@bers]. Let $u=abbabaab\cdots$ be one fixed point of the Thue-Morse substitution and $w=abaaabab\cdots$ be the aperiodic fixed point of the period-doubling substitution. We can see that $\Omega$ is an invariant minimal subset strictly contained in $X\times Y$ by noticing that, while $(abba,baaa)$ is a prefix of $(u,\sigma w)$, it does not occur in any point of the orbit $\sigma^j(u,w)$. Similarly, we have analyzed the substitutions defined in an eight letter alphabet by the product of period doubling and Rudin-Shapiro, and the product of period doubling and paper-folding. These substitutions are [semi-primitive]{}, in the sense of [@BG], see also [@ltww] where semi-primitiveness is shown for Rudin-Shapiro substitution. There is a sub-alphabet, with six letters, in which they are primitive. These substitutions also have two fixed points. The same argument on the location of the letter $b$ in the period doubling substitution leads to more than one invariant set in $X\times Y$. Numerics of the Moment {#sectNumerics} ====================== In this section we report some numerical simulations of the moment $m_2(T)$ as a function of time $T$ for some hybrid quasicrystals. Basis sets were usually of size $2^{14}$, and the time evolution was done by integrating the Schrödinger equation using a sympletic integrator, as described in [@cestu]. The emphasis will be on hybrid substitution quasicrystals. It is expected that different minimal sets present different behavior of $m_2(T)$ and, with respect to numerics, this is the working setting accepted here. In these numerical experiments we have mostly fixed $\kappa=1/2$, but exceptions are explicitly mentioned. We also set $\lambda=1$ (preliminary results indicate that the qualitative behavior is independent of $\lambda\ne0$). The guide to the simulations was based on two properties used in Theorem \[thmMultIndepMin\], that is, the multiplicatively independent dominant eigenvalues of their substitution matrices. ![The moment as function of time ($\log-\log$ scale) for the combination of [TM]{} and [Fcc]{} substitution sequences. The sequence [Fcc]{} was kept fixed, while [TM]{} was shifted by $0,1,\cdots,5$ in order to explore different elements of the product of their hulls. \[figTMFcc\]](tmfcc.eps){width="4.0in" height="2.5in"} First consider the hybridizing of [TM]{} and [Fcc]{}. The results are summarized in Figure \[figTMFcc\]. Different elements of the product of the hulls are obtained by keeping one sequence fixed and shifting the other before the combination. Although both sequences individually generate transport (for [TM]{} see the dashed line in Figure \[figTMTM\]), when combined we have got only one behavior, in accordance with Theorem \[thmMultIndepMin\] and Corollary \[corolExemMin\], since the hybrid hull is minimal in this case. This gives an example of numerical dynamical localization in an almost periodic sequence. As a complement to such simulations we have also considered $\kappa=0.2$ and $0.8$, and localization was always found; again the minimality seems to be the important property. Another possibility we have investigated is when the two involved substitutions have multiplicative dependent dominant eigenvalues. The extreme case is for equal eigenvalues and we have selected this situation by hybridizing a substitution with shifts of itself. Figure \[figTMTM\] presents the results of these simulations for [TM]{} sequence; transport was found in all cases, although with different exponents $\beta$, indicating the presence of more than one minimal component in the product of the hulls; so there are different hybrid quasicrystals in this case. In Figure \[figTMTM\] the dashed line is for the original [TM]{} sequence. We have noted three distinct behaviors, with the dashed line as a border between them: for some shift values the moment follows the dashed line ($\beta\approx \beta_{\text{\sc TM}}=1.8$), others present a range of $0<\beta< \beta_{\text{\sc TM}}$ values, while others with near ballistic behavior (i.e., $\beta> 1.9$); that is, if $\beta>\beta_{\text{\sc TM}}$ then it is near the maximum possible value. We add that for the combination of [Fcc]{} with itself similar results were obtained (not shown), that is, transport prevails and different exponent values of $\beta$ were found; however, without a case near the ballistic motion. ![The moment as function of time ($\log-\log$ scale) for the combination of [TM]{} with itself shifted. The dashed line is for the original [TM]{} (no shift at all). \[figTMTM\]](tmtm.eps){width="4.0in" height="2.5in"} The same procedure was applied to the [PD]{} substitution. If no shift is applied to the sequences, then the original sequence is obtained and it cannot be considered a hybrid quasicrystal, although it is embedded in the product space. Except for this case, where $\beta_{\text{\sc PD}}\approx 1.78$, all other simulations clearly indicate a motion near the ballistic one (no figure is shown). It appears that the self-product of period-doubling substitution contains only two minimal components. We have also combined almost periodic substitution sequences with periodic ones (with periods up to 32), and quite distinct behaviors were found. A periodic sequence is also almost periodic and its hull has finitely many elements (as many as its period). We have hybridized [PD]{}, [TM]{}, [PF]{} and [Fcc]{} with periodic sequences and, depending on the choice of the period, for some cases we have found transport, with different values of $\beta$, but in some other periods we got localization. Figure \[figpfper\] shows some instances of [PF]{} combined with periodic sequences. Such long range perturbations have shown a rich range of possibilities. ![The moment as function of time ($\log-\log$ scale) for the combination of [PF]{} with periodic sequences. The periods were 4 (dashed), 16 (line), 7 and 10 (localization).\[figpfper\]](pfper.eps){width="4.0in" height="2.5in"} Conclusions =========== In this work we considered hybrid quasicrystals, defined by the convex combination two parent finitely valued almost periodic sequences, as new models of one-dimensional quasicrystals. Hybridization of substitution sequences was given special attention. We investigated in some generality the minimality of product spaces $X\times Y$, when both $X$ and $Y$ are minimal, and Section \[sectMinimality\] presented a sufficient condition for primitive substitutions, which is the multiplicative independence of the eigenvalues of their substitution matrices. Minimality is well known when the metric on the sequence space is given by the sup-norm [@pete], but requires extra work in the setting of finitely valued sequences. Some hybrid potentials were inserted into Schrödinger equation and the time evolution of concentrated initial conditions numerically investigated; the interest was in localization and transport in such structures. In order to classify our numerical results we have adopted the pragmatic position that elements in the same minimal set should generate similar time evolutions. This was confirmed in cases our analytical results proved minimality for the product of minimal hulls, and suggested the presence of more than one minimal component in other cases. The figures presented in Section \[sectNumerics\] illustrate these behaviors. The hybridization with periodic sequences was also numerically considered. The numerical results suggest a rigorous investigation of localization in some hybrid quasicrystals. This could be accomplished by proving that the Lyapounov exponent $\gamma$ in these sequences is uniformly positive, that is, the existence of $c>0$ such that $\gamma>c>0$. To our knowledge, this result would be relevant since minimal sequences generated by primitive substitutions have been shown to have zero Lyapounov exponent by the following reasoning. Recall that ${\cal L}(\Omega)$ denotes the language of the minimal subshift $\Omega$, and let $[v]$ be the cylinder set defined by the word $v$: $$[v]\equiv \{\omega\in \Omega: \omega_1\cdots \omega_{\|v\|}=v\}$$ Let $\nu$ be a $\sigma$-invariant probability on $(\Omega,\sigma)$ and $n\in \N$ and ${\cal L}_n(\Omega)$ the set of words of length $n$ occurring in $\Omega$. Define $$\eta_\nu(n)=\min\{\nu([w]):w\in L_n(\Omega)\}$$ Boshernitzan’s condition, first considered in subshifts related to interval exchange transformations [@bosh], may be written as $$\limsup_{n\rightarrow \infty} n\eta_\nu(n)>0 \ .$$ It is proven that the family of ergodic operators $(H_\omega)_{\omega\in \Omega}$, when $\Omega$ is a minimal subshift satisfying Boshernitzan’s condition, have zero Lyapounov exponent everywhere in their spectrum, which is a Cantor set of zero Lebesgue measure [@dalebos]. We can see that Boshernitzan’s condition does not hold in hybrid quasicrystals. For an almost periodic hybrid sequence $z\in \Omega$, the complexity $p_z(n)$, which counts the number of words of length $n$ in $z$, is at least of the order of $n^2$, because the complexity of each component in a hybrid sequence is at least of order $n$. On the other hand, the measure of any cylinder must be inversely proportional to the complexity, because $$p_z(n)\min_{v\in {\cal L}_n(\Omega)} \nu([v])< \sum_{v\in {\cal L}_n(\Omega)} \nu([v]) = 1 \ .$$ for any probability $\nu$. This leaves one important theoretical question, to pursue the possibility of Anderson localization in minimal hybrid quasicrystals. [90]{} P. Alessandri, [Codages de rotations et basses complexités]{}, Ph.D. Thesis, University Aix-Marseille 2, (1996). F. Axel, and D. Gratias (eds.), [Beyond Quasicrystals]{}, Les Editions de Physique, Springer-Verlag, Berlin (1995). J. Berstel, Sur la construction de mots sans carrè, [Sem. Th. des Nombres de Bordeaux]{}, [8]{}, (1979-1980). M. Boshernitzan, A condition for minimal interval exchange maps to be uniquely ergodic, [Duke Math. J.]{} [52]{}: 723–752 (1985). A. Bovier and J.-M. Ghez, Spectral properties of one-dimensional Schrödinger operators with potentials generated by substitutions, [Commun. Math. Phys.]{} [158]{}: 45–66 (1993). Erratum: [Commun. Math. Phys.]{} [166]{}: 431–432 (1994). T. O. Carvalho and C. R. de Oliveira, Spectra and transport in almost periodic dimers, [J. Stat. Phys.]{} [107]{}: 1015-1030 (2002). A. Cobham, On the base dependence of sets of numbers recognizable by finite automata, [Math. Syst. Th.]{} [6]{}: 164–192 (1972). D. Damanik, Strictly ergodic subshifts and associated operators, to appear in Simon Festschrift, [Proceedings of Symposia in Pure Mathematics]{}, American Mathematical Society, Providence. mp-arc 05-309. D. Damanik, D. Lenz, A condition of Boshernitzan and uniform convergence in the multiplicative ergodic theorem, [Duke Math. J.]{} [133]{}: 95–123 (2006). F. Durand, A Generalization of Cobham’s Theorem, [Th. Comp. Syst.]{} [31]{}: 169-185 (1998). H. Furstenberg, [Recurrence in Ergodic Theory and Combinatorial Number Theory]{}, Princeton Univ. Press, Princeton (1981). F. Germinet and S. De Bièvre, Dynamical localization for discrete and continuous random Schrödinger operators, [Commun. Math. Phys.]{} [194]{}: 323–341 (1998). G. Hansel, Systèmes de numération indépendants et syndéticité, [Th. Comp. Science]{} [204]{}: 119–130 (1998). A. Hof, O. Knill and B. Simon, Singular Continuous Spectrum for Palindromic Schrödinger Operators, [Commun. Math. Phys.]{} [174]{}: 149–159 (1995). S. Kotani, Jacobi matrices with random potential taking finitely many values, [Rev. Math. Phys.]{} [1]{}: 129–133 (1989). M. V. Lima and C. R. de Oliveira, Uniform Cantor Singular Continuous Spectrum for Nonprimitive Schrödinger Operators, [J. Stat. Phys.]{} [112]{}: 357–374 (2003). Q.-H. Liu, B. Tan, Z.-X. Wen and J. Wu, Measure zero spectrum of a class of Schrödinger operators, [J. Stat. Phys.]{} [106]{}: 681–691 (2002). M. Queffélec, [Substitution Dynamical Systems – Spectral Analysis]{}, Lect. N. Math. [1294]{}, Springer-Verlag, Berlin (1987). K. Petersen, [Ergodic Theory]{} , Cambridge Univ. Press, Cambridge (1983). B. Simon and T. Wolf, Singular Continuous Spectrum under rank one Perturbations and Localization for Random Hamiltonians, [Commun. Pure Appl. Math.]{} [39]{}: 75–90 (1986). L. Vuillon, Combinatoire des motifs d’une suite sturmienne bidimensionnelle, [Theor. Comp. Sc.]{} [209*]{}: 261–285 (1998). [^1]: CRdO thanks the partial support by CNPq.	{ "pile_set_name": "ArXiv" }
--- abstract: 'This study proposes a model for multimodal, imitative interaction of agents, based on frameworks of predictive coding and active inference, using a variational Bayes recurrent neural network. The model dynamically predicts visual sensation and proprioception simultaneously through generative processes by associating both modalities. It also updates the internal state and generates actions by maximizing the lower bound. A key feature of the model is that the complexity of each modality, as well as of the entire network can be regulated independently. We hypothesize that regulation of complexity offers a common perspective over two distinct properties of embodied agents: coordination of multimodalities and strength of agent intention or belief in social interactions. We evaluate the hypotheses by conducting experiments on imitative human-robot interactions in two different scenarios using the model. First, regulation of complexity was changed between the vision module and the proprioception module during learning. The results showed that complexity of the vision module should be more strongly regulated than that of proprioception because of its greater randomness. Second, the strength of complexity regulation of the whole network in the robot was varied during test imitation after learning. We found that this affects human-robot interactions significantly. With weaker regulation of complexity, the robot tends to move more egocentrically, without adapting to the human counterpart. On the other hand, with stronger regulation, the robot tends to follow its human counterpart by adapting its internal state. Our study concludes that the strength with which complexity is regulated significantly affects the nature of dynamic interactions between different modalities and between individual agents in a social setting.' author: - \| Wataru Ohata\ Cognitive Neurorobotics Research Unit\ Okinawa Institute of Science and Technology\ Okinawa, Japan\ `[email protected]`\ Jun Tani[^1]\ Cognitive Neurorobotics Research Unit\ Okinawa Institute of Science and Technology\ Okinawa, Japan\ `[email protected]`\ bibliography: - 'references.bib' title: 'Investigation of Multimodal and Agential Interactions in Human-Robot Imitation, based on frameworks of Predictive Coding and Active Inference[^2]' --- Introduction ============ Predictive coding [@rao1999; @tani1999learning; @lee2003; @friston2005theory; @hohwy2013predictive; @clark2015; @friston2018] and active inference [@friston2009; @friston2010action; @baltieri2017; @buckley2017; @pezzulo2018hierarchical; @oliver2019active] have attracted increasing attention from a variety of research disciplines, such as neuroscience, cognitive science, and psychology, for their potential to provide a unified theoretical framework explaining perception and action generation. They attempt to describe how organisms perceive sensory information and generate actions to act on the environment, so as to minimize surprise. Typically, in the formulation of predictive coding, perceptions are modeled using generative models in which the logarithm of marginal likelihood is bounded by the evidence lower bound as: $$\begin{aligned} \ln p_\theta(\boldsymbol{X})&\geq\underbrace{\int q_\phi(\boldsymbol{z}\|\boldsymbol{X})\ln\frac{p_\theta(\boldsymbol{X},\boldsymbol{z})}{q_\phi(\boldsymbol{z}\|\boldsymbol{X})}d\boldsymbol{z}}_{\rm Evidence\ lower\ bound}\\ \label{elbo} &=\underbrace{\mathbb{E}_{q_\phi(\boldsymbol{z\|\boldsymbol{X}})}[\ln p_\theta(\boldsymbol{X}\|\boldsymbol{z})]}_{\rm Accuracy}-\underbrace{D_{\rm KL}[q_\phi(\boldsymbol{z}\|\boldsymbol{X})\Vert p(\boldsymbol{z})]}_{\rm Complexity}\end{aligned}$$ where $\boldsymbol{X}$ is the observation, $\boldsymbol{z}$ is the latent variable, $q_\phi(\boldsymbol{z}\|\boldsymbol{X})$ is the approximated posterior, and $\theta$ and $\phi$ are the parameters of the model. The first term of the right side of equation (\[elbo\]) is denominated [accuracy]{}, which is the expectation of the model with respect to the approximated posterior. The second term on the right side is termed [complexity]{}, which is the Kullback-Leibler divergence (KL divergence) between the approximated posterior and the prior. The accuracy term represents the reconstruction of the observation with the posterior, and the complexity term serves to regularize the model. The model is optimized by maximizing the lower bound, which is equivalent to minimizing [free energy]{} proposed by [@friston2005theory]. Recently, [@ahmadi2019novel] proposed a network model, the so-called Predictive coding-inspired Variational Recurrent Neural Network (PV-RNN), in which the evidence lower bound is maximized in the processes of perception, action generation, and learning. They introduced a hyper parameter $w$ called the meta-prior, which weights the regulation of the complexity term in the evidence lower bound. They found that a model trained under weaker regulation of the complexity term by setting the meta-prior to a larger value, develops more deterministic dynamics with higher estimated precision, whereas one with stronger regulation by setting the meta-prior to a smaller value does so with more probabilistic dynamics with lower estimated precision. It was found that adequate regulation of complexity by adjusting the meta-prior, contributes greatly to a model’s ability to generalize in learning. The current study hypothesizes that regulating the complexity term with different strengths should affect cognitive functions of agents in at least two distinct ways, especially for those interacting with others in social settings using multimodal sensation. First, regulating the complexity term with an adequate weighting may enhance effectiveness in coordinating multimodal perceptions. In general our experience is not comprised of a single modality, but of many, such as vision, audio perception, and tactile sensation, and so forth. How those multimodal perceptions can be adequately coordinated for association is not a trivial problem because intrinsic complexity and randomness in spatio-temporal pattern differ in each modality. In this situation, we speculate that such differences between different modalities can be arbitrated by regulating the complexity term of each with an appropriate strength. Second, regulating the complexity term differently after learning may affect the strength of agent intention or belief in the social interaction. Here, acting with strong intention in social interactions means that an agent acts egocentrically toward others or with volatility, whereas acting with a weak intention means to act passively by following intentions of others. In this regard, [@chame2019cognitive] showed that when the PV-RNN is trained under weaker regulation of the complexity term, the network tends to behave egocentric with less adaptation to sensory inputs, whereas under stronger regulation of the complexity term, the network tends to behave more passively with adaptation to sensor inputs. However, such characteristics of the network, once developed through learning under particular conditions for regulating the complexity term, cannot be changed after learning. In social interactions, agents are better able to act differently, depending on the social context at a given moment. Sometimes they tend to preserve their prior intention for acting perversely, and at other times they change it more easily by adapting to intentions of others. The current study examines such possibilities for modulating behavioral characteristics by dynamically changing parameters to regulate the complexity term during social interaction after learning. For purposes of evaluating these two hypotheses, we conducted a set of robotics experiments by extending the PV-RNN to facilitate a model of multimodal, agential interaction using visuo-proprioceptive sensory inputs. The model network employed in the current study is comprised of a multi-layered PV-RNN with a branching structure, in which each branch is responsible for one modality. These are connected through an associative module. The model is a generative model and predicts incoming visual perception and proprioception simultaneously through a generative process. This corresponds to the prior generation. The resultant prediction error for each sensory modality is back-propagated through time [@rumelhart1985learning] and through each module to the associative module by which the internal state in each module is modulated in the direction of maximization of the evidence lower bound. This corresponds to the posterior inference. The network is trained through supervised learning by maximizing the evidence lower bound. One thing to note is that the model can have different values of the meta-prior in each layer of the PV-RNN in each modality. This means that the model is capable of regulating complexity with different weighting in a layer- and modality-specific manner. First, we investigated how assigning different values of the meta-prior to the proprioception and vision modules affects learning and the resultant imitation performance with multimodal sensation of visuo-proprioceptive patterns. Our results suggest that regulating complexity more in the vision module than in the proprioception module facilitates better imitation performance in multi-modal sensation after the learning, when visual sensory information contains more randomness than proprioceptive information. Second, we examined how shifting a set of meta-prior values in the whole network to larger or smaller values than those used in the learning phase affects the characteristics of imitative behavior. We found that the network that strongly regulated the complexity term by setting smaller values of the meta-prior tended to follow human movement patterns by adapting its internal states. On the other hand, the network that weakly regulated the complexity term by setting larger values of the meta-prior tended to generate more egocentric/self-centered movement patterns with less sensitivity to changes or fluctuations in the human movement patterns by adapting its internal state less. The current paper presents a detailed analysis of the underlying mechanisms accounting for these observed phenomena. The next section describes the employed model in greater detail. Model ===== Model overview {#model overview} -------------- This subsection describes briefly how we can model multimodal imitative interaction of agents perceiving visuo-proprioceptive sensory inputs by using concepts of predictive coding and active inference. Among various types of imitation, synchronized imitation is considered in the current study by virtue of its simplicity. In synchronized imitation, the agent is required to imitate target patterns demonstrated by the human counterpart by predicting them on the basis of prior learning. Although target patterns to imitate are structurally the same as previously learned patterns, they could involve marginal variations, as in speed, amplitude, and shape. The process of synchronized imitation can be achieved by means of iterative cycling of prediction of sensory inputs during the demonstration, generation of corresponding movement, and updating the latent state of the network using the resultant sensory prediction error. To generate movement, one step, look-ahead prediction of proprioception is fed to an inverse model [@kawato1987hierarchical], which is often implemented by a PID feedback controller in robots. The inverse model or PID feedback controller computes an optimal motor torque as the motor command for minimizing the error between the predicted proprioception (the target joint angles) and the actual proprioception (the actual joint angles). This process of generating adequate motor commands for achieving the expected sensory inputs in terms of the proprioception can be considered a primitive type of active inference [@friston2010action; @friston2011action]. Updating the latent state can be performed using a scheme called error regression [@tani1999learning; @ito2004line; @hwang2018dealing; @ahmadi2019novel], by which the process of sensory perception assumed in a predictive coding framework can be performed. Now we look at how the PV-RNN [@ahmadi2019novel] can be used to implement the model for multimodal imitative interaction of robot agents receiving visuo-proprioceptive sensory inputs based on the aforementioned frameworks of predictive coding and active inference. Figure \[model\] shows the overall system view, consisting of a PV-RNN, a robot, and a human counterpart. The human demonstrates movement patterns to the robot both visually and kinesthetically, guiding the robot’s posture by wearing a motion capture suit. PV-RNN is considered a generative model, formulated in a continuous spatio-temporal domain, using a variational Bayes framework. A PV-RNN inherits the concept of a Multiple Timescale Recurrent Neural Network (MTRNN)[@yamashita2008], which is characterized by its architecture, which allocates different timescale dynamics to different layers. Higher layers are endowed with slower timescale dynamics and lower layers with faster dynamics, as inspired by recent cognitive neuroscience evidence [@newell2001time; @huys2004multiple; @smith2006interacting; @kording2007dynamics]. It has been shown that introduction of multiple timescale dynamics can enhance abstraction and generalization in learning by extracting action-primitive hierarchies or chunking structures from observed multimodal sensory inputs [@yamashita2008; @choi2018predictive; @hwang2018dealing]. PV-RNN approximates the posterior at each time step to minimize the reconstruction error in previous steps through variational inference, which is implemented by means of the error regression scheme. Each sensory module for proprioception and vision was modeled with a multi-layered PV-RNN and modules were connected with an associative module, which also consists of a PV-RNN (Figure \[model\]). The associative module generates the top-down prior, conditioned by the current latent state in this module at each time step. The top-down prior is then fed to both the proprioception and vision modules. Each sensory module also generates the top-down prior at each time step conditioned by the previous latent state of the module computed using top-down information provided by the associative module, by which the prediction of sensory inputs, proprioception and vision at the subsequent time step is generated. Here, we note that the actual pixel visual image is compressed into a low-dimensional vector value using a CNN-type [@lecun1989backpropagation; @lecun1998gradient] encoder-decoder, which is connected to the bottom layer of the vision PV-RNN module for the purpose of reducing the computational cost. By feeding the prediction of the next proprioception to the motor controller as the target, the corresponding motor commands for generating expected movements can be generated. When the sensory outcome is observed at each sensory module after the movement, the prediction error is computed. Errors generated in both the proprioceptive and visual modules are back-propagated through time [@werbos1974beyond; @rumelhart1985learning], as well as through the layers from those sensory modules to the associative module, where the latent state at each layer in each module is updated so as to maximize the lower bound in terms of the posterior inference using the error regression scheme (more details in Section \[error-regression\]). This cycle of prediction with the conditional prior, generating the corresponding movement, and updating the latent state in terms of the posterior inference using the error, is repeated at each small $\Delta t$ of the sensory sampling, where robot agents predict future movement patterns of the human counterpart while updating their own intentions or briefs represented by the latent state for possible adaptation to spontaneous changes or fluctuations. Finally, learning of demonstrated movement patterns can be done by performing mostly the same cycling process, but by updating connectivity weights without generating actual movement. In this process, the prediction error for the target visuo-proprioceptive sequence patterns sampled in the training exemplar is computed offline without generating movement. The obtained error is used to update the connectivity weights, as well as the latent states, at each time step in the whole network for maximizing the lower bound. Learning is performed in an end-to-end manner across all layers and modules throughout the whole network because the whole network is differentiable. ![The overall system configuration consists of the PV-RNN, the robot, and the human counterpart, who wears a motion capture suit. The PV-RNN consists of three modules: the associative module, a vision module, and a propriopcetion module. Each module is comprised of layer(s) of PV-RNN. $p$ and $q$ represent prior and posterior in each layer, and blue arrows illustrate how the top-down intention or brief in terms of the prior, propagates downward through the whole network to predict both sensory modalities. Red arrows illustrate how the bottom-up prediction error in proprioception and vision propagates upward through the whole network to infer the posterior. The PV-RNN layers in the vision module predict the latent dynamics of visual patterns $\boldsymbol{l}_t$, which are fed to the decoder to generate visual patterns $\boldsymbol{v}_t$ during learning. In error-regression, perceived visual patterns are embedded with the encoder to latent patterns with which the prediction error of visual patterns $e_t^l$ is computed. The proprioception module directly predicts the joint-angle configuration of the robot $\boldsymbol{p}_t$, which is fed to the motor control system to generate the motor command $\boldsymbol{m}_t$. The prediction error in proprioception $e_t^p$ is obtained by measuring the actual joint-angle $\bar{\boldsymbol{p}}_t$. The model is optimized by minimizing the prediction error and KL divergence between the approximated posterior and prior, which is weighted with the meta-prior.[]{data-label="model"}](model.jpg){width="100.00000%"} Derivation of evidence lower bound ---------------------------------- PV-RNN is a generative and inference model based on graphical representation (Figure \[er\]). It is comprised of deterministic variables $\boldsymbol{d}$, i.e. following Dirac delta distributions, and stochastic variables $\boldsymbol{z}$. The model includes prior and infers posterior by variational inference. We modified the original PV-RNN at three points with respect to dependencies of variables. First, in our model, there are no connections between the output of the network $\boldsymbol{x}$ and $\boldsymbol{z}$, which exist in the original PV-RNN. This is for simplification of the model, and it was confirmed that removing these connections did not hinder network performance. Second, the current network does not have connections from the lower layer to the higher layer, which the original network does have. This modification is intended to separate more clearly the information flow between top-down generative prediction and bottom-up error propagation. Last, diagonal connections from the higher layer during the previous time-step to the lower layer during the succeeding time-step are changed to vertical connections during the same time-step. Following the derivation of the evidence lower bound in [@ahmadi2019novel], the evidence lower bound of the proposed visuo-proprioceptive model is derived as $$\begin{split} \ln(\boldsymbol{p}_{1:T}, \boldsymbol{v}_{1:T}\|\boldsymbol{d}_0^{})&\geq\sum_{t=1}^T\left(\mathbb{E}_{q^a, q^p}[\ln P(\boldsymbol{p}_t\|\boldsymbol{d}_t^{p, 1})]+\mathbb{E}_{q^a, q^v}[\ln P(\boldsymbol{v}_t\|\boldsymbol{d}_t^{v, 1})]\right.\\ &-\sum_{l\in A}D_{\rm KL}[q(\boldsymbol{z}_t^{l}\|\boldsymbol{d}_{t-1}^{l}, e_{t:T}^p, e_{t:T}^v)\Vert p(\boldsymbol{z}_t^{l}\|\boldsymbol{d}_{t-1}^{l})]\\ &-\sum_{l\in P}D_{\rm KL}[q(\boldsymbol{z}_t^{l}\|\boldsymbol{d}_{t-1}^{l}, e_{t:T}^p)\Vert p(\boldsymbol{z}_t^{l}\|\boldsymbol{d}_{t-1}^{l})]\\ &\left.-\sum_{l\in V}D_{\rm KL}[q(\boldsymbol{z}_t^{l}\|\boldsymbol{d}_{t-1}^{l}, e_{t:T}^{v})\Vert p(\boldsymbol{z}_t^{l}\|\boldsymbol{d}_{t-1}^{l})]\right) \end{split}$$ where A, P, and V represent the associative module, the proprioception module, and the vision module, respectively, and $l$ indicates a layer in each module. $\boldsymbol{p}_{1:T}$ and $\boldsymbol{v}_{1:T}$ are time series propriocetive and visual patterns. $\boldsymbol{d}_0^$ represents $\boldsymbol{d}$ in all the layers at time-step 0. $\mathbb{E}_{q^a, q^p}$ denotes the expectation over all the distributions of $\boldsymbol{z}$ in the associative module and the proprioception module, and $\mathbb{E}_{q^a, q^v}$ denotes expectation over all distributions of $\boldsymbol{z}$ in the associative module and the vision module. $\boldsymbol{d}_t^{p,1}$ is the deterministic valuable in the lowest layer of the proprioception module at time-step $t$, and $\boldsymbol{d}_t^{v, q}$ is that in the lowest layer of the vision module. $\boldsymbol{z}_t^{l^a}$, $\boldsymbol{z}_t^{l}$ are the stochastic variables in $l$th layer in the associative module, in the proprioception module, and in the vision module at time-step $t$, respectively. $\boldsymbol{d}_{t-1}^{l}$ represents the activities of deterministic variables in $l$th layer in the associative module, in the proprioception module, and in the vision module at time-step $t$, respectively. $e^p_{t:T}$ and $e^v_{t:T}$ are the prediction error between the predicted patterns and the target patterns at time-step from $t$ to $T$ in proprioception and vision, respectively. By introducing the meta-prior, which weighs the KL divergence between the approximated posterior and the prior, the evidence lower bound of the model is defined as $$\begin{split} \mathcal{L}_w&:=\sum_{t=1}^T\left(\underbrace{\mathbb{E}_{q^a, q^p}[\ln P(\boldsymbol{p}_t\|\boldsymbol{d}_t^{p, 1})]+\mathbb{E}_{q^a, q^v}[\ln P(\boldsymbol{v}_t\|\boldsymbol{d}_t^{v, 1})]}_{\rm accuracy}\right.\\ &-\sum_{l\in A}w^{l}\underbrace{D_{\rm KL}[q(\boldsymbol{z}_t^{l}\|\boldsymbol{d}_{t-1}^{l}, e_{t:T}^p, e_{t:T}^v)\Vert p(\boldsymbol{z}_t^{l}\|\boldsymbol{d}_{t-1}^{l}))]}_{\rm complexity\ in\ associative\ module}\\ &-\sum_{l\in P}w^{l}\underbrace{D_{\rm KL}[q(\boldsymbol{z}_t^{l}\|\boldsymbol{d}_{t-1}^{l}, e_{t:T}^p)\Vert p(\boldsymbol{z}_t^{l}\|\boldsymbol{d}_{t-1}^{l})]}_{\rm complexity\ in\ proprioception\,module}\\ &\left.-\sum_{l\in V}w^{l}\underbrace{D_{\rm KL}[q(\boldsymbol{z}_t^{l}\|\boldsymbol{d}_{t-1}^{l}, e_{t:T}^{v})\Vert p(\boldsymbol{z}_t^{l}\|\boldsymbol{d}_{t-1}^{l})]}_{\rm complexity\ in\ vision\ module}\right) \end{split}$$ where $w^{l}$ represents the meta-priors in the $l$th layer in the associative module, the proprioception module, and the vision module, respectively. The parameters of the model are optimized by maximizing the lower bound, which corresponds to minimizing the free energy. The associative module ---------------------- The associative module is comprised of a PV-RNN. Since we adopted MTRNN in PV-RNN, its computations are as follows. [\_t\^[a,l]{}=]{} \_[dd]{}\^[a,ll]{}\_[t-1]{}\^[a, l]{}+\_[dz]{}\^[a,ll]{}\^[a,l]{}\_t+\^[a,l]{} & [if top layer]{}\ \_[dd]{}\^[a,ll]{}\_[t-1]{}\^[a, l]{}+\_[dd]{}\^[a,ll+1]{}\_t\^[a,l+1]{}+\_[dz]{}\^[a,l]{}\^[a,l]{}\_t+\^[a,l]{} & [otherwise]{} $$\begin{aligned} \boldsymbol{h}_t^{a,l}&=\left(1-\frac{1}{\tau^{a,l}}\right)\boldsymbol{h}_{t-1}^{a,l}+\frac{1}{\tau^{a,l}}\boldsymbol{u}_{t^l}\\ \boldsymbol{d}_t^{a,l}&=\tanh\left(\boldsymbol{h}_t^{a,l}\right)\end{aligned}$$ where $\boldsymbol{u}_t^{a,l}$ is the sum of inputs to $l$th layer of the associative module. $\boldsymbol{W}_{dd}^{a,ll}$, $\boldsymbol{W}_{dz}^{a,ll}$, and $\boldsymbol{W}_{dd}^{a,ll+1}$ are weight matrices for recurrent connections, the stochastic variable $\boldsymbol{z}$, and the input from the higher layer, respectively. $\boldsymbol{b}^{a,l}$ is the bias in the $l$th layer in the associative module, and $\tau^{a,l}$ is the time constant for MTRNN computation in the $l$th layer in the associative module. $\tanh$ is the activation function. The stochastic variable $\boldsymbol{z}$ is assumed to follow a multivariate Gaussian distribution with diagonal covariant matrix, and the deterministic variable $\boldsymbol{d}$ predicts mean $\boldsymbol{\mu}$ and variance $\boldsymbol{\sigma}$ of the distribution. That is, for computation of prior, $$\begin{aligned} &p(\boldsymbol{z}_t^{p,a,l}\|\boldsymbol{d}_{t-1}^{a,l})=\mathcal{N}(\boldsymbol{z}_t^{p,a,l};\boldsymbol{\mu}_t^{p,a,l},\boldsymbol{\sigma}_t^{p,a,l})\\ &\boldsymbol{\mu}_t^{p,a,l}=\tanh(\boldsymbol{W}_{\mu d}^{a,l}\boldsymbol{d}_{t-1}^{a,l} + \boldsymbol{b}_\mu^{a,l})\\ &\boldsymbol{\sigma}_t^{p,a,l}=\exp(\boldsymbol{W}_{\sigma d}^{a,l}\boldsymbol{d}_{t-1}^{a,l}+\boldsymbol{b}_\sigma^{a,l})\\ &\boldsymbol{z}_t^{p,a,l}=\boldsymbol{\mu}_t^{p,a,l}+\boldsymbol{\sigma}_t^{p,a,l}\boldsymbol{\epsilon}\end{aligned}$$ where $\boldsymbol{\mu}_t^{p,a,l}$ and $\boldsymbol{\sigma}_t^{p,a,l}$ are the mean and variance for the prior distribution of $\boldsymbol{z}_t^{p,a,l}$ at time-step $t$ in $l$th layer in the associative module. $\boldsymbol{W}_{\mu d}^{a,l}$ and $\boldsymbol{W}_{\sigma d}^{a,l}$ are the weight matrices for $\boldsymbol{d}^{a,l}_{t-1}$. $\boldsymbol{b}_{\mu}^{a,l}$ and $\boldsymbol{b}_{\sigma}^{a,l}$ are the biases for each computation. $\tanh$ in computation of mean is used for stability of optimization, and $\exp$ in $\boldsymbol{\sigma}$ is for variance to be positive. $\boldsymbol{\epsilon}$ is sampled from $\mathcal{N}(\boldsymbol{0},\boldsymbol{I})$. To approximate the posterior, PV-RNN has adaptive variables $\boldsymbol{a}$ which are specific to time-step and sequence. $\boldsymbol{a}$ is optimized during learning with prediction error through backpropagation through time (BPTT). By considering $\boldsymbol{a}$, the computations of posterior are $$\begin{aligned} &q(\boldsymbol{z}_t^{q,a,l}\|\boldsymbol{d}_{t-1}^{a,l}, e_{t:T}^p, e_{t:T}^v)=\mathcal{N}(\boldsymbol{z}_t^{q,a,l};\boldsymbol{\mu}_t^{q,a,l},\boldsymbol{\sigma}_t^{q,a,l})\\ &\boldsymbol{\mu}_t^{q,a,l}=\tanh(\boldsymbol{W}_{\mu d}^{a,l}\boldsymbol{d}_{t-1}^{a,l} + \boldsymbol{a}_{\mu,t}^{a,l} + \boldsymbol{b}_\mu^{a,l})\\ &\boldsymbol{\sigma}_t^{q,a,l}=\exp(\boldsymbol{W}_{\sigma d}^{a,l}\boldsymbol{d}_{t-1}^{a,l}+\boldsymbol{a}_{\sigma,t}^{a,l}+\boldsymbol{b}_\sigma^{a,l})\\ &\boldsymbol{z}_t^{q,a,l}=\boldsymbol{\mu}_t^{q,a,l}+\boldsymbol{\sigma}_t^{q,a,l}\boldsymbol{\epsilon}\end{aligned}$$ where $\boldsymbol{\mu}_t^{q,a,l}$ and $\boldsymbol{\sigma}_t^{q,a,l}$ are mean and variance for the posterior distribution of $\boldsymbol{z}_t^{q,a,l}$ at time-step $t$ in $l$th layer in the associative module. Note that the weight matrices for $\boldsymbol{d}$ are different from those used to compute the prior. The proprioception module ------------------------- Proprioceptive patterns are directly generated from the PV-RNN. The highest layer in the proprioception module receives the input from the lowest layer in the associative layer, and its computations are [\_t\^[p,l]{}=]{} \_[dd]{}\^[pa]{}\_t\^[a,1]{}+\_[dd]{}\^[p,ll]{}\_[t-1]{}\^[p,l]{}+\_[dz]{}\^[p,l]{}\_t\^[p,l]{}+\^[p,l]{} & [if top layer]{}\ \_[dd]{}\^[p,ll]{}\_[t-1]{}\^[p,l]{}+\_[dz]{}\^[p,l]{}\_t\^[p,l]{}+\^[p,l]{} & [otherwise]{} $$\begin{aligned} \boldsymbol{h}_t^{p,l}&=\left(1-\frac{1}{\tau^{p,l}}\right)\boldsymbol{h}_{t-1}^{p,l}+\frac{1}{\tau^{p,l}}\boldsymbol{u}_t^{p,l}\\ \boldsymbol{d}_t^{p,l}&=\tanh\left(\boldsymbol{h}_t^{p,l}\right)\end{aligned}$$ Proprioceptive patters are generated from the lowest layer of the proprioception module. $$\boldsymbol{p}_t=\tanh\left(\boldsymbol{W}^p\boldsymbol{d}_t^{p,1}+\boldsymbol{b}^p\right)$$ The vision module ----------------- For the vision module, a special adjustment was made. Although introducing recurrent connections in convolutional neural networks (CNN) offers great performance[@xingjian2015convolutional] for modelling dynamic visual patterns, it is computationally expensive. Our model is developed to work in physical robots in real time in the future. Therefore, an architecture that is computationally less intensive was needed. In this vision module, PV-RNN layers predict latent dynamics of visual patterns, and the predicted latent dynamics are fed to decoders comprised of static CNNs to generate images. Furthermore, during the error-regression scheme described below, visual patterns are perceived as latent dynamics through an encoder. Then, the prediction error is computed as the discrepancy between the predicted latent patterns and the perceived latent patterns, which is computed in a relatively small dimension, reducing the computational burden. Similar to the proprioception module, the highest layer of the vision module receives input from the lowest layer of the associative layer, and its computations are [\_t\^[v,l]{}=]{} \_[dd]{}\^[va]{}\_t\^[a,1]{}+\_[dd]{}\^[v,ll]{}\_[t-1]{}\^[p,l]{}+\_[dz]{}\^[v,l]{}\_t\^[v,l]{}+\^[v,l]{} & [if top layer]{}\ \_[dd]{}\^[v,ll]{}\_[t-1]{}\^[v,l]{}+\_[zd]{}\^[v,l]{}\_t\^[v,l]{}+\^[v,l]{} & [otherwise]{} $$\begin{aligned} \boldsymbol{h}_t^{v,l}&=\left(1-\frac{1}{\tau^{v,l}}\right)\boldsymbol{h}_{t-1}^{v,l}+\frac{1}{\tau^{v,l}}\boldsymbol{u}_t^{v,l}\\ \boldsymbol{d}_t^{v,l}&=\tanh\left(\boldsymbol{h}_t^{v,l}\right)\end{aligned}$$ Then the lowest layer of the PV-RNNs predicts the latent dynamics $\boldsymbol{l}$, and the visual patterns $\boldsymbol{v}$ are generated through a decoder. $$\begin{aligned} &\boldsymbol{l}_t=\tanh\left(\boldsymbol{W}^l\boldsymbol{d}_t^{v,1}+\boldsymbol{b}^l\right)\\ &\boldsymbol{v}_t={\rm decoder}(\boldsymbol{l}_t)\end{aligned}$$ In error-regression, the targets of latent dynamics $\bar{\boldsymbol{l}}$ of visual patters $\bar{\boldsymbol{v}}$ are computed through an encoder. $$\bar{\boldsymbol{l}}_t={\rm encoder}(\bar{\boldsymbol{v}}_t)$$ To improve the generalization capability of the encoder and decoder, [CoordConv]{} architecture[@liu2018intriguing] was introduced in them. Learning process ---------------- The visuo-proprioceptive model is trained by maximizing the evidence lower bound. Thus, given $T$ time-step length of proprioceptive patterns $\boldsymbol{p}_{1:T}$ and visual patterns $\boldsymbol{v}_{1:T}$, the cost function to be minimized is defined as $$\begin{split} {\rm cost}:=&\sum_{t=1}^{T}\left(\frac{1}{R^p}\left\Vert \boldsymbol{p}_t-\bar{\boldsymbol{p}_t}\right\Vert^2+\frac{1}{R^v}\left\Vert \boldsymbol{v}_t-\boldsymbol{\bar{v}}_t\right\Vert^2 +\sum_{l\in A}w^{l}\frac{1}{R^{l}}D_{\rm KL}[q(\boldsymbol{z}_t^{l}\|\boldsymbol{d}_{t-1}^{l}, e_{t:T}^p, e_{t:T}^v)\Vert p(\boldsymbol{z}_t^{l}\|\boldsymbol{d}_{t-1}^{l})]\right.\\ &\left.+\sum_{l\in P}w^{l}\frac{1}{R^{l}}D_{\rm KL}[q(\boldsymbol{z}_t^{l}\|\boldsymbol{d}_{t-1}^{l}, e_{t:T}^p)\Vert p(\boldsymbol{z}_t^{l}\|\boldsymbol{d}_{t-1}^{l})] +\sum_{l\in V}w^{l}\frac{1}{R^{l}}D_{\rm KL}[q(\boldsymbol{z}_t^{l}\|\boldsymbol{d}_{t-1}^{l}, e_{t:T}^{v})\Vert p(\boldsymbol{z}_t^{l}\|\boldsymbol{d}_{t-1}^{l})]\right) \end{split}$$ where A, P, and V represent the associative module, the proprioception module, and the vision module. $R^p$ and $R^v$ are the dimensions of proprioceptive patterns and visual patterns to normalize prediction errors, and $R^{l}$ is the dimension of the distributions of $\boldsymbol{z}$ to normalize the KL divergence. Since the prior and posterior distributions are assumed to follow a multivariate Gaussian distribution with diagonal covariant matrix, the KL divergence in the cost function is analytically computed. Given two $n$ dimensional multivariate Gaussian distributions $p(\boldsymbol{z})=\mathcal{N}(\boldsymbol{z};\boldsymbol{\mu}^p,\boldsymbol{\sigma}^p)$ and $q(\boldsymbol{z})=\mathcal{N}(\boldsymbol{z};\boldsymbol{\mu}^q,\boldsymbol{\sigma}^q)$ where $\boldsymbol{\mu}=(\mu_1,\mu_2,...,\mu_n)^T$ and $\boldsymbol{\sigma}=(\sigma_1,\sigma_2,...,\sigma_n)^T$, $$D_{\rm KL}[q(\boldsymbol{z})\Vert p(\boldsymbol{z})]=\sum_{i=1}^n\left(\ln\left(\frac{\sigma^p_i}{\sigma^q_i}\right)+\frac{(\mu^p_i-\mu^q_i)^2+(\sigma^q_i)^2}{2(\sigma^p_i)^2}-\frac{1}{2}\right)$$ The parameters of the model, including $\boldsymbol{a}$, are optimized using BPTT. To perform error-regression, an encoder was also trained separately. Testing method {#error-regression} -------------- [@ahmadi2019novel] proposed a method to test the trained model in a way that is consistent with concepts of predictive coding and active inference. Since a PV-RNN is a generative model. It does not accept any input; thus, it is not straightforward to test with the model with novel patterns. In testing, weights and biases of the networks are fixed, and the only adaptive variable $\boldsymbol{a}$ is updated (Figure \[er\]). The network has a window with a specified time-step that represents the immediate past for perception. Within the window, the posterior is inferred by means of the error-regression. In the error-regression, an optimal posterior is inferred for a given observation by maximizing the evidence lower bound, and the network makes the best prediction for incoming perception. In this process, the adaptive variable $\boldsymbol{a}$ assigned to the time window is updated through BPTT. Optimization is accomplished in an iterative manner, and after a given number of iterations, the window slides one time-step ahead, and the same process occurs during the next time-step. An example of error-regression in real-time applied to a human-robot interactions through proprioception using PV-RNN is found in [@chame2019cognitive]. ![A graphical representation of error-regression. This is an example in which the length of the window is two (the yellow area), and the network has two layers. In error-regression, the weights and the biases of the network are fixed, and only the adaptive variables $\boldsymbol{a}$ are updated. Black arrows represent forward computations, and red arrows indicate how prediction errors are propagated to $\boldsymbol{a}$ by BPTT.[]{data-label="er"}](er.jpg){width="50.00000%"} Experiments =========== Experimental design ------------------- Using the proposed model, imitative interaction experiments between a human and a humanoid robot were conducted. Although human-robot interaction ought to be studied in an online fashion to reflect human behavior in response to robot actions, because of the intensive computation required in the error regression scheme, we could not conduct such experiments online. Therefore, the current study examined only the dynamic response of the model network using recorded sequences of visuo-proprioceptive patterns. For this purpose, the dataset containing human-demonstrated movement patterns in terms of visuo-proprioceptive sequences were collected both for training the network and for later testing of pseudo-synchronized imitative interaction. After training, the model was tested with novel visuo-proprioceptive patterns in error-regression with two different scenarios (Experiment 1 and Experiment 2). In Experiment 1, different values of the meta-prior were assigned to the proprioception module and the vision module to examine how regulating complexity in each modality affects coordination of different modalities. In Experiment 2, a set of meta-prior values in the whole network was shifted to larger or smaller values than those used in the learning phase, in order to examine the effect of regulating complexity on the manner of synchronized imitation. This experimental setting is different from [@chame2019cognitive], in which dynamic characteristics during imitative interaction were compared between different networks trained with different values of the meta-prior. In the following experiments, some parameters that determine the network structure were set as follows. The associative module consisted of a one-layer PV-RNN, and the proprioception module and the vision module consisted of two-layers of PV-RNN, respectively. These PV-RNN layers were characterized by a time scale imposed on MTRNN computation. That is, the higher layer had a larger time constant, producing slow time-scale dynamics, and lower layer had smaller time constants, generating fast time-scale dynamics. Therefore, in this study, the higher layer of the proprioception module and the vision module, which receive input from the associative module, are referred to as the [slow layer]{}, and the lower layer is referred to as the [fast layer]{}. Data preparation ---------------- To obtain a dataset of synchronized vision and proprioception, we used a humanoid robot, Torobo (Tokyo Robotics Inc.) and a motion capture suit (Perception Neuron, Noitom Ltd.). Torobo is a human-sized, torso-type humanoid robot with 16 joint-angles, of which 6 are for each arm and 4 are for the torso and head positions. Human body movements can be mapped to joint-angle trajectories of the robot using the motion capture suit. A human experimenter wearing the suit demonstrated a set of body movements, which were mapped as joint-angle trajectories. This demonstration was also recorded with a camera to obtain corresponding visual patterns. The target sequential movement pattern to be learned by the robot was designed by considering a probabilistic finite state machine that can generate probabilistic sequences of three different primitive movement patterns. Those were (A) waving with both arms three times, (B) rotating the torso to the left with the arms, and (C) rotating the torso to the right with the arms. Primitive pattern A is followed either by primitive pattern B or primitive pattern C with a 50% chance, and primitive patterns B and C are followed by pattern A with 100% chance (Figure \[training\_example\] (A)). One sequence consists of 8 probabilistic transitions of primitive movements, and movement sequences demonstrated by three human participants were recorded for 10 sequences for each. In other words, the dataset was comprised of 30 sequences of visuo-proprioceptive temporal patterns. Recorded visuo-proprioceptive patterns were down-sampled to 3.75 Hz so that one sequence became 400 time-steps. Joint-angle trajectories were normalized to a range between $-1$ and $1$. Vision patterns were further converted into gray scale images and down-sized to $64\times64$ pixels (Figure \[training\_example\] (B)). It is noted that the visual trajectories are far more fluctuated than the proprioceptive ones due to various optical conditions, such as illumination and surface reflectiveness. ![Training data. (A) A diagram of the probabilistic finite state machine. (B) An example of the training dataset. The top row is a part of a joint-angle trajectory. The corresponding labels of primitive patterns are (A, B, and C) indicated above the plots. For simplicity, only 4 joint-angles out of 16 representing the movements are shown. The middle row show the corresponding visual pixel images at each period. The bottom row shows visual trajectories in the latent space embedded by the encoder. For simplicity, only three variables out of 20 are shown.[]{data-label="training_example"}](training_example.jpg){width="100.00000%"} By using the training example, the model is required to extract a probabilistic structure that the primitive pattern of B or C appears with only a 50% chance after every appearance of the primitive A by estimating precision in transitions of primitives by learning. Such learning should be achieved without providing explicit labels for those primitives, but by extracting the underlying chunking and segmentation structure from the continuous sensory signals prepared in the dataset. The PV-RNN can achieve such tasks by using the multiple timescale RNN scheme combined with Bayesian inference approach [@ahmadi2019novel]. Experiment 1: Different meta-priors in different modalities ----------------------------------------------------------- In this experiment, we examined how assigning different meta-priors to the proprioception and vision modules affects the learning process and error-regression performance. Two sets of meta-priors $w_1$ and $w_2$ were assigned to the model (Table \[network configuration\]). $w_1$ has larger values of the meta-prior in the proprioception module than in the vision module, and they were exchanged in $w_2$. Both $w_1$ and $w_2$ have the same value for the meta-prior in the associative module. First, the model was trained with the $w_1$ and $w_2$, and the learning process was examined, with special attention to each component of the lower bound. To facilitate training, Adam optimizer [@kingma2014adam] was utilized with the parameter setting $\alpha=0.001$, $\beta_1=0.9$, and $\beta_2=0.999$. The model was trained 10 times with different random initializations of the parameters of the model for 10,000 epochs, and the mean and standard deviation of the prediction error of proprioception and vision, and the KL divergence of each layer of the model at each epoch were computed. [lccccc]{} & $\mathbb{R}^d$ & $\mathbb{R}^z$ & $\tau$ & $w_1$ & $w_2$\ \ Associative module & 10 & 1 & 15 & 0.0025 & 0.0025\ Proprioception slow layer & 20 & 2 & 8 & 0.005 & 0.0025\ Proprioception fast layer & 30 & 3 & 2 & 0.01 & 0.005\ Vision slow layer & 20 & 2 & 8 & 0.0025 & 0.005\ Vision fast layer & 30 & 3 & 2 & 0.005 & 0.01\ \ Results are summarized in Figure \[ex1\_1\]. In comparing $w_1$ and $w_2$ conditions, even though the prediction errors in the proprioception and vision modules showed similar behavior (Figure \[ex1\_1\] (A), (B)), the KL divergence in each module was optimized differently. Despite different values of the meta-prior assigned to the fast layer of the proprioception module, its KL divergences in $w_1$ and $w_2$ conditions were reduced in exactly the same way (Figure \[ex1\_1\] (E)). This is not the case in the fast layer of the vision module (Figure \[ex1\_1\] (G)). The KL divergence in the slow layer of the proproception module and the slow layer of the vision module showed different values in $w_1$ and $w_2$ settings (Figure \[ex1\_1\] (D),(F)). Interestingly, although the associative module was set to the same value of meta-prior in $w_1$ and $w_2$ conditions, the KL divergence in the $w_2$ setting reached a larger value than that in the $w_1$ setting. This is because the larger value of the meta-prior assigned to the fast layer of the vision module in the $w_2$ condition prevented the vision module from absorbing the fluctuation in the observed visual patterns, which resulted in bottom-up fluctuation from the vision module to the associative module, appearing as a discrepancy between the prior and the posterior in this module. Because visual sensation tends to contain more inherent randomness compared to proprioceptive sensation as described previously, complexity in this modality should be adequately regulated by setting a smaller meta-prior value. Otherwise, the discrepancy that appears in the visual module tends to leap to the higher associative module without being well resolved before. ![The learning process of the model with two different meta-prior settings. (A) The prediction error in proprioception. (B) The prediction error in vision. (C) The KL divergence in the associative module. (D) The KL divergence in the slow layer of the proprioception module. (E) The KL divergence in the fast layer of the proprioception module. (F) The KL divergence in the slow layer of the vision module. (G) The KL divergence in the fast layer of the vision module. The shadows are the standard deviation of 10 trials with different parameter initializations. Note that the values of prediction errors are the sum of the prediction errors at all time-steps and sequences normalized by the dimension of data.[]{data-label="ex1_1"}](exp1_1.jpg){width="100.00000%"} We further tested the trained models in the error-regression scheme. Training of the models stopped after 4,000 epochs. Novel visuo-proprioceptive patterns were prepared for the error-regression, which also comprised the previous primitive body movements A, B, and C, the lengths of which were 400 time-steps. The length of the time window was set at 30, and the number of iterations of optimization by BPTT at each time-step was 30. As in learning, Adam was used to improve optimization with the parameter settings $\alpha=0.2$, $\beta_1=0.9$, and $\beta_2=0.999$. Evaluation of the error-regression examined how much the reconstruction error in each modality and the KL divergence at each subnetwork in the PV-RNN were minimized. That is, at the point when $T'$ time-step window length for the immediate past shifts $t$ times, the adaptive variable $\boldsymbol{a}$ assigned within the window is optimized with the iterative process, and at the last iteration, the reconstruction error and the KL divergence are computed inside the window. Therefore, they are defined as $$\begin{aligned} &{\rm Proprioception\ error}:=\frac{1}{T}\sum_{t=1}^{T}\frac{1}{T'}\sum_{t'=1}^{T'}\frac{1}{R^p}\Vert \boldsymbol{p}_{t'}-\bar{\boldsymbol{p}}_{t'}\Vert^2\\ &{\rm Vision\ error} := \frac{1}{T}\sum_{t=1}^T\frac{1}{T'}\sum_{t'=1}^{T'}\frac{1}{R^l}\Vert \boldsymbol{l}_{t'}-\bar{\boldsymbol{l}}_{t'}\Vert^2\\ &{\rm KLD} := \frac{1}{T}\sum_{t=1}^{T}\frac{1}{T'}\sum_{t'=1}^{T'}\frac{1}{R^z}D_{\rm KL}[q(\boldsymbol{z}_{t'}\|\boldsymbol{d}_{t'-1},e_{t':T})\Vert p(\boldsymbol{z}_{t'}\|\boldsymbol{d}_{t'-1})]\end{aligned}$$ where $t'$ is the time step inside the window. $R^p$ and $R^l$ are the dimension of proprioception and the latent space of vision, respectively. $R^z$ is the dimension of $\boldsymbol{z}$, and the KL divergence is computed for every PV-RNN submodule. Models trained in previous experiments were used. Error-regression was run 10 times with different random number seeds, and the mean and standard deviation of each quantity were computed. In addition, one-step look-ahead prediction error, the discrepancy between the prediction in the next time-step of the current window and the observation, was computed in vision to evaluate the prediction accuracy. Results are summarized in Table \[exp1\_2\]. The prediction error in proprioception was remarkably minimized compared to that in vision, in both conditions $w_1$ and $w_2$. This is because the visual modality contains more noise than the proprioceptive modality. It was also observed that the reconstruction error in vision was smaller for the $w_1$ condition than the $w_2$ condition. Furthermore, the KL divergence in the associative module was reduced more significantly in the $w_1$ condition than the $w_2$ condition. This resulted because the vision module generalized better with noisy visual patterns in the test of error regression in the $w_1$ case than the $w_2$ case by minimizing the complexity term more. Because fluctuation or randomness in visual sensation was well resolved within the vision module in the $w_1$ case, the associative module became relatively free from such fluctuation, as evidenced by the smaller KL divergence observed in the associative module under this condition of minimizing the complexity term. As a result, the one-step, look-ahead prediction was also more accurate in $w_1$ condition. [lccccc]{} & Proprioception & Vision & Associative & Proprioception\ & reconstruction error & reconstruction error & KLD & slow KLD\ \ $w_1$ & $0.017\pm9.5\times10^{-4}$ & $0.12\pm6.9\times10^{-3}$ & $2.0\pm0.091$ & $1.6\pm0.12$\ $w_2$ & $0.011\pm2.9\times10^{-4}$ & $0.19\pm1.5\times10^{-2}$ & $3.2\pm0.15$ & $2.6\pm0.11$\ \ & Proprioception & Vision & Vision & Vision one-step\ & fast KLD & slow KLD & fast KLD & prediction error\ \ $w_1$ & $0.57\pm0.040$ & $8.1\pm0.17$ & $0.50\pm0.048$ & $0.20\pm0.0097$\ $w_2$ & $0.57\pm0.041$ & $1.9\pm0.071$ & $0.54\pm0.048$ & $0.24\pm0.017$ Experiment 2: Different meta-priors of the whole system ------------------------------------------------------- Based on Experiment 1, we further investigated how values of the meta-prior of the whole system affect interactions with a human. The model that was trained for 4,000 epochs in Experiment 1 with the $w_1$ setting was used. Five different meta-prior settings were prepared: from smaller values of the meta-prior setting $W_1$ to the larger setting $W_5$ with a consistent ratio among layers (Table \[w setting in exp2\]). Error-regression with different meta-prior settings was performed with the novel visuo-proprioceptive patterns used in Experiment 1. As in the previous experiment, the error-regression was run 10 times with different random number seeds, and the mean and variance of each quantity introduced in the previous experiment were computed. In this experiment, one-step look-ahead prediction error in proprioception was also measured. [lccccc]{} & $W_1$ & $W_2$ & $W_3$ & $W_4$ & $W_5$\ \ Associative module & $2.5\times10^{-5}$ & $2.5\times10^{-4}$ & $2.5\times10^{-3}$ & $2.5\times10^{-2}$ & $2.5\times10^{-1}$\ Proprioception slow layer & $5.0\times10^{-5}$ & $5.0\times10^{-4}$ & $5.0\times10^{-3}$ & $5.0\times10^{-2}$ & $5.0\times10^{-1}$\ Proprioception fast layer & $1.0\times10^{-4}$ & $1.0\times10^{-3}$ & $1.0\times10^{-2}$ & $1.0\times10^{-1}$ & $1.0$\ Vision slow layer &$2.5\times10^{-5}$ & $2.5\times10^{-4}$ & $2.5\times10^{-3}$ & $2.5\times10^{-2}$ & $2.5\times10^{-1}$\ Vision fast layer & $5.0\times10^{-5}$ & $5.0\times10^{-4}$ & $5.0\times10^{-3}$ & $5.0\times10^{-2}$ & $5.0\times10^{-1}$ The results are summarized in Figure \[exp2-1\]. As a whole, with smaller values of the meta-prior, the reconstruction error was minimized more (Figure \[exp2-1\](A)), and the KL divergence remained large (Figure \[exp2-1\](D)), whereas with larger values of the meta-prior, the KL divergence was minimized more (Figure \[exp2-1\](D)), and the reconstruction error remained large (Figure\[exp2-1\](A)). This tendency can also be seen in the local proprioception module and vision module. In the proprioception module, as the values of meat-prior increased, the reconstruction error in proprioception became large (Figure \[exp2-1\](B)), and the KL divergence became small, both in the slow layer (Figure \[exp2-1\](F)) and in the fast layer (Figure \[exp2-1\](G)). Similarly, in the vision module, as the values of meta-prior increased, the reconstruction error in vision increased (Figure \[exp2-1\](C)), and the KL divergence became small both in the slow layer (Figure \[exp2-1\](H)) and in the fast layer (Figure \[exp2-1\](I)). The KL divergence at the associative module also increased as the values of the meta-prior increased (Figure \[exp2-1\](E)). In addition, with smaller values of the meta-prior, the one-step, look-ahead prediction error was minimized more both in proprioception (Figure \[exp2-1\](J)) and in vision (Figure \[exp2-1\](K)). This is because the KL divergence term in the evidence lower bound was weighted more for minimization than was the reconstruction error term. In this situation, the posterior $q(\boldsymbol{z}_t\|\boldsymbol{d}_{t-1},e_{t:T'})$ at each time step in the error regression window (ER-W) approached its prior $p(\boldsymbol{z}_t\|\boldsymbol{d}_{t-1})$ in the window by modulating the adaptive value $\boldsymbol{a}_t$, which is fed into the computation of the posterior $q(\boldsymbol{z}\|\boldsymbol{d}_{t-1},e_{t:T'})$, while the prior $p(\boldsymbol{z}_t\|\boldsymbol{d}_{t-1})$ was less changed. This means that network dynamics were driven mainly by the prior, while less affected by sensory inputs. Network dynamics become more egocentric by following the prior, which was less modified by weaker regulation of the complexity term (i.e., more weighting for the KL divergence term). On the other hand, with more weighting (i.e., less weighting for the KL divergence term), network dynamics became more adaptive to change or fluctuations of sensory inputs, by freely modulating the posterior in the direction of error minimization without being much constrained by the prior. In this condition, the prior $p(\boldsymbol{z}_t\|\boldsymbol{d}_{t-1})$ at each step in the window also changes because the posterior $q(\boldsymbol{z}_{t-1}\|\boldsymbol{d}_{t-2})$ at the previous step, which is mapped to $p(\boldsymbol{z}_t\|\boldsymbol{d}_{t-1})$ through $\boldsymbol{d}_t$ also changes. ![Reconstruction error, KL divergence minimization, and one-step, look-ahead prediction error in error-regression with five meta-prior settings. (A) Sum of the reconstruction error in proprioception and vision. (B) The reconstruction error in proprioception. (C) The reconstruction error in vision. (D) Sum of the KL divergence at all layers. (E) The KL divergence in the associative module. (F) The KL divergence in the slow layer of the proprioception module. (G) The KL divergence in the fast layer of the proprioception module. (H) The KL divergence in the slow layer of the vision module. (I) The KL divergence of the fast layer of the vision module. (J) One-step, look-ahead prediction error in proprioception. (K) One-step, look-ahead prediction error in vision. Error bars represent the standard deviation of 10 trials with different random seeds. Note that each graph has a different scale.[]{data-label="exp2-1"}](exp2_1.jpg){width="100.00000%"} Figure \[exp2-2\] (A) and (B) shows an example of time-series plots of related neural activities of the proprioception module, comparing these two cases of strong ($W_1$ setting) and weak ($W_5$ setting) regulation of the complexity term. Both cases are computed for a situation observing the same visuo-proprioceptive sequence pattern. With strong regulation of the complexity term (Figure \[exp2-2\] (A) top), the observation (dashed lines) was well-reconstructed in the reconstructed outputs (real lines) inside the ER-W (shaded area) from time-step 120 to 150 due to relatively stronger weighting in the accuracy term compared to $W_5$ setting. The plots after time-step 150 represent future prediction of expecting to encounter the primitive B. From time-steps 150 to 180 (Figure \[exp2-2\] (A) bottom), the robot observed new sensory information where the primitive C instead of the predicted primitive of B was encountered. (Remember that there is a 50% chance of encountering the primitive B or the primitive C.) This new observation was reconstructed inside the ER-W, and based on the inferred posterior during this period, the robot updated the future prediction after time-step 180 as the primitive C to be continued. Because of relatively stronger weighting in the accuracy term, the posterior was inferred to adapt to reality. The prediction was also updated accordingly (Figure \[exp2-2\] (A) bottom). In the case of weak regulation (Figure \[exp2-2\](B)top), the observation was still well reconstructed inside the ER-W. This is because primitive A always follows primitives either B or C so that it is easy to predict primitive A. Therefore, the reconstruction error inside the ER-W was small from the beginning. The plots after time-step 150 represent future predictions of expecting primitive B to be encountered. After observing new sensory information in which primitive C instead of the predicted primitive C was encountered between time-step 150 and 180 (Figure \[exp2-2\](B)bottom), however, the new observation was not reconstructed well inside the ER-W. Due to strong regulation of the KL divergence term (weak regulation of the complexity term), the posterior was forced closer to the prior by ignoring the new observation. Consequently, the inferred posterior did not affect the prior as much as in the $W_1$ setting, which resulted in generation of consistent prediction for future. Actually, the look-ahead prediction made at time-step 150, shown in the top row, and the one made at time-step 180 in the bottom row are almost the same. These observations imply that both the prediction of the future and the reflection of the past become more adaptive to sensory observation in the case of stronger regulation of the complexity term, whereas they become more persistent regardless of the sensory observation in the case of weaker regulation. ![An example of time-series plots of neural activities in the output layer of proprioception module in $W_1$ setting (A) and in $W_5$ setting (B). The reconstruction of the past observation and the future prediction at time-step 150 (top) and at time-step 180 (bottom) are shown. The solid lines represent the neural activities, and the dashed lines represent the observations. The shadow area indicates the error-regression window. For simplicity, only 4 out of 16 joint-angles representing the movements are shown.[]{data-label="exp2-2"}](exp2_2.jpg){width="100.00000%"} Some representative videos related to Experiment 2 can be seen in [video link A](https://youtu.be/WaODCuuVZjA) and in [video link B](https://youtu.be/rXDCOb7Q89I) for $W_1$ condition and $W_5$ condition, respectively. These videos show how prediction of future as well as reflection of past can be performed for each condition. Also, further temporal details during the error regression process can be seen in [video link C](https://youtu.be/jCU96QZzPpA) and in [video link D](https://youtu.be/7pGwITPIFTE) for $W_1$ condition and $W_5$ condition, respectively. In these videos, it can be observed that there are some divergence between the prior and the posterior in terms of mean and variance and that they are dynamically changing inside the ER-W in $W_1$ condition whereas these two profiles become close each other with showing relatively persistent patterns in $W_5$ condition. These observations accord with our analysis described previously. Discussion ========== In this study, we hypothesized that the means of regulating the complexity term in the evidence lower bound in the model contributes greatly to two distinct properties of agents: coordination of multimodal perception and strength of intention or belief in acting with others. We evaluated the hypotheses by building a model for multimodal imitative interaction of agents using visuo-proprioceptive sensory modalities, based on frameworks of predictive coding and active inference, and by conducting experiments on imitative human-robot interactions. We performed experiments in two different scenarios. First, the ratio of regulating the complexity term was changed between the vision module and the proprioception module. Our results show that the complexity term in the vision module should be regulated more than that of the proprioception module. This is because visual sensory and proprioceptive inputs are fundamentally different with respect to their intrinsic randomness. Visual inputs fluctuate more often due to various optical conditions, such as illumination and surface reflectiveness. In this situation, the complexity of information processing in the vision module should be more regulated to achieve better generalization. Second, regulation strength for the complexity term throughout the whole network was varied after learning. Our results demonstrated that this significantly affects the manner of generating human-robot interaction. With weaker regulation of the complexity term, the robot tends to act more egocentrically, without adapting to the other. In contrast, with stronger regulation of the complexity term, the robot tends to follow its human counterpart by adapting its internal state. This result implies that the strength of intention or belief in acting with others can be modulated by adjusting the strength of regulation of the complexity term after the learning phase. In the current study, we evaluated the hypotheses using imitative human-robot interactions. In imitative interactions between a human and a robot, there are two scenarios: the robot following the human’s movements, and the human following the robot’s movements. In our experimental setup, the agent with strong regulation of the complexity term corresponds to the former case, and that with weak regulation to the latter. This may provide new insights into studies of mirror neuron systems [@rizzolatti2014mirror; @kilner2007predictive]. Mirror neurons were first discovered in area F5 of the monkey premotor cortex[@di1992understanding; @gallese1996action], and it was reported that mirror neurons are activated while executing their own actions, as well as observing those performed by others. Because mirror neurons transfer observations of an action to generation of the same action, they could underlie the mechanism of imitative behaviors, which are thought to be the basis of various higher cognitive functions[@aly2015online; @kohler2002hearing; @oztop2006mirror; @oztop2013mirror]. These considerations imply that imitation by mirror neurons can be explained by the current model in the case of strong regulation of the complexity term. However, in everyday social interactions, humans do not just follow others, but they also lead them with their own belief, depending on each contextual or social situation. Psychological studies indicate that turn-taking between following and leading can take place rather spontaneously in various social cognitive behaviors, including conversation [@sacks1978simplest], mother-infant pre-verbal communication [@trevarthen1979communication] and imitation [@nadel2002imitation]. In considering possible underlying mechanisms for turn-taking, some researchers [@ikegami2007turn; @ito2004line] suggest that turn-taking may develop due to potential instability, such as chaos formed in coupled dynamics between two agents in their modeling studies. We could consider a meta-level dynamic of coupling two agents, whereby the value of the meta-prior for regulating the complexity term for each agent counteracts mutually. This could result in autonomous shifts between the leading mode by increasing the meta-prior and the following mode by reducing it. Future studies should examine the aforementioned mechanism for turn-taking by conducting an online experiment of human-robot interactions. However, the computational cost of online error regression for inference of the posterior has been the major bottleneck for conducting such experiments in real time and this is why the current study was limited to using recorded visuo-proprioceptive sequence patterns as the target in testing synchronized imitation, rather than introducing actual, real-time, human-robot interaction. Regarding this problem, some may suggest employing other types of variational models, such as a variational recurrent neural network (VRNN) [@chung2015recurrent], because a VRNN demands far less computation time, since the posterior at each time step can be inferred by simple sequential mapping of inputs using an autoencoder [@kingma2016improved]. However, it is argued that the current scheme for inference of the posterior through iterative computation for optimization should be vital for any embodied cognitive systems that require rapid adaptation of the internal states to the environment. Actually, Ahmadi and Tani [@ahmadi2019novel] showed that PV-RNN can perform better than VRNN [@chung2015recurrent] in an online prediction in dynamically changing environment by inferring the posterior using the error regression scheme. Therefore, future studies are expected to explore possible methods for accelerating online error regression of the model, such as by massive parallelisation for the purpose of conducting real-time, human-robot interactions using the model. Acknowledgments {#acknowledgments .unnumbered} =============== This work was supported by funding from Okinawa Institute of Science and Technology Graduate University. We thank the lab members in the Cognitive Neurorobotics Research Unit. Especially, we are grateful to Ahmadreza Ahmadi and Prasanna Vijayaraghavan for their help in developing the model. We thank Siqing Hou for his help in collecting the dataset. We also thank Dr. Steven Aird for assisting improving the manuscript. [^1]: Corresponding author [^2]: This is a preprint.	{ "pile_set_name": "ArXiv" }
--- abstract: \| Background : Proton-induced nucleon knockout $(p,pN)$ reactions have been successfully used to study the single-particle nature of stable nuclei in normal kinematics with the distorted-wave impulse approximation (DWIA) framework. Recently, these reactions have been applied to rare-isotope beams at intermediate energies in inverse kinematics to study the quenching of spectroscopic factors. Purpose : Our goal is to investigate the effects of various corrections and uncertainties within the standard DWIA formalism on the $(p,pN)$ cross sections. The consistency of the extracted reduction factors between DWIA and other methods is also evaluated. Method : We analyze the $(p,2p)$ and $(p,pn)$ reactions data measured at the R$^3$B/LAND setup at GSI for carbon, nitrogen, and oxygen isotopes in the incident energy range of 300–450 MeV/u. Cross sections and reduction factors are calculated by using the DWIA method. The transverse momentum distribution of the $^{12}$C($p$,$2p$)$^{11}$B reaction is also investigated. Results : We have found that including the nonlocality corrections and the Møller factor affects the cross sections considerably. The proton-neutron asymmetry dependence of reduction factors extracted by the DWIA calculation is very weak and consistent with those given by other reaction methods and ab initio structure calculations. Conclusions : The results found in this work provide a detailed investigation of the DWIA method for $(p,pN)$ reactions at intermediate energies. They also suggest that some higher-order effects, which is essential for an accurate cross-section description at large recoil momentum, is missing in the current DWIA and other reaction models. author: - Nguyen Tri Toan Phuc - Kazuki Yoshida - Kazuyuki Ogata title: ' Toward a reliable description of $\bm{(p,pN)}$ reactions in the distorted-wave impulse approximation' --- Introduction ============ In the last fifty years, nucleon knockout reactions induced by intermediate energy protons of the type $(p,pN)$ in normal (forward) kinematics have been one of the most successful tools for studying the single-particle nature of stable nuclei [@Jac66; @Jac73; @Kit85; @Wak17]. These reactions are sometimes referred to as quasifree scattering due to the dominance of scattering processes between the proton and the knocked-out nucleon. The most widely used theoretical approach to analyze these reactions is the distorted-wave impulse approximation (DWIA) method of Chant and Roos [@Cha77; @Cha83]. A detailed review of this method including its uncertainties and applications to existing forward-kinematics $(p,pN)$ data has been reported in Ref. [@Wak17]. With the availability of radioactive beams at energies up to 450 MeV/u, there has been a renewed interest to elucidate single-particle properties of unstable nuclei by using nucleon knockout reactions on the hydrogen target in inverse kinematics. The DWIA method has first been theoretically applied to inverse kinematics $(p,pN)$ reactions with the scattering wave functions treated with the eikonal approximation [@Aum13; @Oga15] and later in (standard) partial-wave expansion form [@Yos18]. The latter method was then successfully applied in several experimental studies [@Oli17; @Ele19; @Tan19; @Chen19; @Kaw18] carried out at RIKEN, Japan. However, an extensive investigation of the sensitivity of calculated cross sections to various choices of inputs and corrections (as in Ref. [@Wak17]) was not done in these studies. Besides DWIA, other theoretical methods such as the three-body Faddeev equation in the Alt-Grassberger-Sandhas formulation (FAGS) [@Cre08; @Cre19] and the transfer-to-the-continuum (TC) method [@Mor15] have also been used to analyze the $(p,pN)$ experimental data measured in inverse kinematics. Although there is a good consistency between DWIA and TC formalism for the specific $^{15}$C($p$,$pn$)$^{14}$C reaction [@Yos18], a systematic comparison between the DWIA (eikonal and partial-wave form), TC, and FAGS analyses on the actual experimental data is essential to determine the range of applicability for $(p,pN)$ reactions. Among the possible uses of proton-induced nucleon knockout reactions, the study of quenching single-particle strength and its proton-neutron asymmetry dependence is one of the most important subjects. A reduction of 30%–40% with respect to the independent-particle model (IPM) limit in spectroscopic factors (SFs) deduced from $(e,e'p)$ experiments was observed at NIKHEF [@Lap93]. This quenching of the SF, quantified as the reduction factor $R_s$, is due to the lack of short-range (including tensor) and long-range correlations in the IPM and standard shell model (SM) calculations [@Pan97; @Dic04; @Sick07; @Bar09; @Jen11; @Cip15]. Systematic analysis of nucleon removal reactions on light composite targets ($^{9}$Be and $^{12}$C) [@Gade08; @Tos14] suggests a strong dependence on the proton-neutron asymmetry defined as ${\Delta S=S_p-S_n}$ $(\Delta S=S_n-S_p)$ for proton (neutron) removal, where $S_p$ $(S_n)$ is the proton (neutron) separation energies. However, such strong dependence cannot be observed in systematic nucleon-transfer studies [@Lee06; @Kay13; @Fla13; @Fla18; @Xu19]. A weak dependence of $R_s$ on $\Delta S$ is also supported by ab initio coupled-cluster (CC) [@Jen11] and self-consistent Green’s function (SCGF) [@Cip15] calculations. Very recently, a series of $(p,pN)$ measurements for carbon-, nitrogen-, and oxygen-isotope beam with incident energy range of 300–450 MeV/u in inverse kinematics was performed at the R$^3$B/LAND setup at GSI Helmholtzzentrum für Schwerionenforschung in Darmstadt, Germany [@Pan16; @Atar18; @Fer18; @Holl19]. While the measurements in Ref. [@Pan16] are exclusive, those in Refs. [@Atar18; @Fer18; @Holl19] provide the semi-inclusive cross sections and momentum distributions, in which various bound states of the residual nucleus are summed. The eikonal DWIA [@Pan16; @Atar18; @Holl19], FAGS [@Fer18], and TC [@Gom18] methods were applied to these experimental data and gave a generally similar conclusion that the reduction factors depend very weakly on the proton-neutron asymmetry, in contradiction with the much steeper asymmetry found in nucleon removal analysis. However, as was noted in Ref. [@Gom18], the $(p,pN)$ data analyzed by different reaction models exhibit some discrepancies due to choices of inputs and nonrelativistic treatments. These results also slightly underestimate the magnitude of $R_s$ given by ab initio calculations [@Jen11; @Cip15]. To clarify the inconsistencies between these models and give a reliable evaluation of the experimental data, it is of great interest to perform a careful DWIA analysis on the reduction factors with the GSI $(p,pN)$ data [@Pan16; @Atar18; @Fer18; @Holl19]. The content of this paper is as follows. In Sec. \[secformalism\] the formulation of the standard partial-wave DWIA formalism is given. In Sec. \[secresult\] the cross sections and reduction factors are calculated for all published GSI data. The impact of several corrections in the DWIA framework on these observables are investigated. The transverse momentum distributions for the specific $^{12}$C($p$,$2p$)$^{11}$B case is also discussed. Finally, the summary is given in Sec. \[secsum\]. Formalism {#secformalism} ========= The A($p$,$pN$)B knockout reaction is analyzed with the same partial-wave DWIA framework as in Ref. [@Yos18]. Observables with superscript A are evaluated in the A-rest frame while those without the superscript are in the three-body center-of-mass (c.m.) frame, also called the G frame. The transition amplitude for the A($p$,$pN$)B reaction is given by $$\begin{aligned} T_{\bm{K}_0\bm{K}_1\bm{K}_2}^{nljm} &= \Braket{ \chi_{1,\bm{K}_1}^{(-)}\chi_{2,\bm{K}_2}^{(-)} \|t_{pN}\| \chi_{0,\bm{K}_0}^{(+)}\varphi^{nljm} }, \label{eqtrans}\end{aligned}$$ where $\chi_{i,\bm{K}_i} (i=0,1,2)$ are the distorted scattering wave functions of the $p$-A, $p$-B, and $N$-B systems, respectively. $\bm{K}_i$ is the momentum of particle $i$ in the G frame. As has been shown in Eqs.(3.8)-(3.15) of Ref. [@Wak17], when the kinematic coupling term in the exit channel Hamiltonian is approximated, the three-body scattering wave function can be separated into two two-body distorted-wave functions in Eq. (1), where $\bm{K}_i$ can now be interpreted as the relative momentum. The superscripts $(+)$ and $(-)$ specify the outgoing and the incoming boundary conditions of these scattering waves, respectively. The relative single-particle wave function of the $N$-B system bound inside A is denoted as $\varphi^{nljm}$ where $n$, $l$, $j$, and $m$ are the principal quantum number, the orbital angular momentum, the total angular momentum, and its third component, respectively. $t_{pN}$ is a transition operator for the $p$-$N$ scattering, which is sometimes called an effective interaction [@FL85]. The absolute square of its matrix element is proportional to the $pN$ elastic-scattering cross section ${d\sigma_{pN}}/{d\Omega_{pN}}$. Following the same theoretical treatment as in Ref. [@Yos18], upon disregarding the spin-orbit distortion, the momentum distribution (MD) is given by $$\begin{aligned} \frac{d\sigma}{d\bm{K}_\mathrm{B}^\mathrm{A}} = &C_0 \int d\bm{K}_1^\mathrm{A}d\bm{K}_2^\mathrm{A} \delta(E_f^\mathrm{A}-E_i^\mathrm{A}) \delta(\bm{K}_f^\mathrm{A}-\bm{K}_i^\mathrm{A}) \nonumber \\ &\times \frac{E_1 E_2 E_\mathrm{B}}{E_1^\mathrm{A} E_2^\mathrm{A} E_\mathrm{B}^\mathrm{A}} \frac{d\sigma_{pN}}{d\Omega_{pN}} \sum_{m}(2\pi)^2 \lvert \bar{T}_{\bm{K}_0\bm{K}_1\bm{K}_2}^{nljm} \rvert ^2, \label{eq.md}\end{aligned}$$ where $$\begin{aligned} C_0 &\equiv \frac{E_0^\mathrm{A}}{(\hbar c)^2 K_0^\mathrm{A}} \frac{f_{pN}}{(2l+1)}\frac{\hbar^4}{(2\pi)^3 \mu_{pN}^2}.\end{aligned}$$ The factor $f_{pN}$ equals to 1 for $(p,pn)$ and $1/2$ for $(p,2p)$ reactions. $E_i$ and $E^{\rm A}_i$ are the total (relativistic) energies of particle $i$ in the G and A-rest frames, respectively. The G-frame $pN$ scattering cross section in Eq. (\[eq.md\]) is related to the one in the two-nucleon c.m. frame, which we term the t frame, through $$\begin{aligned} \frac{d\sigma_{pN}}{d\Omega_{pN}}= \eta^2\frac{d\sigma_{pN}^\textrm{\,t}}{d\Omega_{pN}^\textrm{\,t}},\end{aligned}$$ where $$\begin{aligned} \eta=\left(\frac{E_1^\mathrm{t} E_2^\mathrm{t} E_0^\mathrm{t} E_N^\mathrm{t}}{E_1 E_2 E_0 E_N}\right)^{1/2}, \label{eq.mol}\end{aligned}$$ is the Møller factor required for the transformation of $t_{pN}$ from the t frame to the G frame in relativistic kinematics [@Mol45; @Ker59]. The total energy of the struck nucleon $E_N$ is determined by the momentum conservation of the two colliding nucleons as [@Yos18] $$\begin{aligned} E_N=\dfrac{\hbar^2}{2\mu_{NB}}\left[\bm{K}_1+\bm{K}_2-\dfrac{(A+1)}{A}\bm{K}_0\right]^2.\end{aligned}$$ The reduced transition amplitude in Eq. (\[eq.md\]) is given by $$\begin{aligned} \bar{T}_{\bm{K}_0\bm{K}_1\bm{K}_2}^{nljm} &= \int d\bm{R}\,\chi_{1,\bm{K}_1}^{(-)}(\bm{R})\, \chi_{2,\bm{K}_2}^{(-)}(\bm{R})\, \chi_{0,\bm{K}_0}^{(+)}(\bm{R})\, \nonumber \\ &\times \varphi^{nljm}(\bm{R}) e^{-i\bm{K}_0\cdot\bm{R}/A}.\end{aligned}$$ For the bound-state wave function $\varphi^{nljm}$ generated by a local potential, the effect of nonlocality is taken into accounted in the interior region by multiplying $\varphi^{nljm}$ by the Perey factor [@Per63] $$\begin{aligned} F_\textrm{PR}(R)=C_\textrm{PR}\left[1-\dfrac{\mu_{N\mathrm{B}}}{2\hbar^2}\beta^2 V_{N\mathrm{B}}(R)\right]^{-1/2},\end{aligned}$$ where the nonlocality range $\beta=0.85$ fm for nucleon [@Per62], and $V_{N\mathrm{B}}$ is the single-particle binding potential. The factor $C_\textrm{PR}$ is chosen so that the modified bound-state wave function is normalized to unity. Similarly, for a scattering wave function obtained from a Dirac phenomenology optical potential (OP), the relativistic velocity-dependent term modifies the wave function by what we refer to as the Darwin factor [@Arn81; @Ham90] $$\begin{aligned} F_\textrm{DW}(R)=\left[\dfrac{E_i+U_S(R)-U_V(R)}{E_i}\right]^{1/2},\end{aligned}$$ where $U_S$ and $U_V$ are the scalar and vector potentials in the Dirac equation, respectively. This Darwin factor is regarded as a kind of nonlocality correction and has been well known to be very important in order to fully take into account relativistic effect in $(e,e'p)$ reactions [@Udi95]. The cylindrical transverse momentum distributions (TMD) are obtained from the MD as $$\begin{aligned} \frac{d\sigma}{dK_{\mathrm{B}b}^\mathrm{A}} &= 2\pi\int dK_{\mathrm{B}z}^\mathrm{A} K_{\mathrm{B}b}^\mathrm{A} \frac{d\sigma}{d\bm{K}_\mathrm{B}^\mathrm{A}}.\end{aligned}$$ The integrated single-particle cross section is then calculated from the TMD as $$\begin{aligned} \sigma_\textrm{sp} &= \int \frac{d\sigma}{dK_{\mathrm{B}b}^\mathrm{A}}dK_{\mathrm{B}b}^\mathrm{A}. \label{eq.sp}\end{aligned}$$ Results and discussion \[secresult\] ==================================== In this section, we applied the DWIA framework described in the preceding section to the GSI $(p,pN)$ data [@Pan16; @Atar18; @Fer18; @Holl19] to evaluate the single-particle cross section and deduce the reduction factor. The impact of nonlocality corrections, the Møller factor, and energy dependence of final-state OP on the $R_s$ are clarified. Numerical inputs {#subsecinput} ---------------- We perform DWIA calculations for all 18 published $(p,pN)$ cases of the R$^3$B/LAND setup [@Pan16; @Atar18; @Fer18; @Holl19]. In addition, we also perform calculations with several other choices of inputs to estimate the theoretical uncertainty on reduction factors. All calculations in this work adopt the relativistic treatment of the kinematics, which is essential to reproduce the correct MD [@Yos18]. The nonlocality correction is taken into account in our calculation through the use of the Perey factor for the single-particle bound state and Darwin factor for distorted waves. For the distorting potential of the $p$-A, $p$-B, and $N$-B systems, we use the EDAD2 parameter set of the Dirac phenomenology [@Coo93]. Calculations using the EDAD1, EDAD3 [@Coo93], and the “democratic” EDAD [@Coo09] Dirac OP sets give a difference of 10% at most, which is consistent with the one observed in normal kinematics [@Wak17]. The single-particle wave function of the struck nucleon is obtained from a Woods-Saxon potential with central and spin-orbit components defined in the same manner as in Refs. [@Gade08; @Gom18]. For both components, a diffuseness $a=0.7$ fm is used for all the cases. The radius parameter is adjusted following the prescription $\langle r^2 \rangle=[A/(A-1)]\langle r^2 \rangle_\textrm{HF}$, where $\langle r^2 \rangle_\textrm{HF}$ is the single-particle mean square radius of the Hartree-Fock calculation with the Skyrme SkX interaction [@Bro98]. The depth $V_\textrm{so}=6$ MeV is fixed for the spin-orbit term while the central one is adjusted to reproduce the experimental separation energies. This choice of binding potential gives a difference in the $(p,pN)$ cross section within 10% compared with the one used in the $(e,e'p)$ analysis [@Kra01; @Wak17]. Based on this result and the investigations of Refs. [@Gade08; @Gom18] on the ambiguity of different effective nucleon-nucleon (NN) interactions used in the Hartree-Fock calculation, we adopt an uncertainty of 10% for single-particle wave functions. For the $pN$ elementary cross section, we employ the one generated by the $t$-matrix parametrization of Franey and Love [@FL85] with a final-energy prescription, which has been suggested to be the best approximation for the half-off-shell amplitude [@Red70]. Different choices of the on-shell approximation such as initial-energy and average-energy prescriptions give an uncertainty of 2% for $(p,2p)$ and 8% for $(p,pn)$ processes. The large discrepancies in neutron knockout processes are due to the asymmetric shape of the $pn$ scattering angular distribution that makes it more sensitive to the choice of energy prescription. Note that other choices of the NN cross section such as those from the Reid93 potential [@Reid93] in TC [@Gom18], the CD-Bonn potential [@Mac01] in FAGS [@Fer18], or from experimental database fitting [@Arn03] in the eikonal DWIA [@Aum13] give essentially the same NN cross sections up to 350 MeV. The theoretical spectroscopic factor for each state of the bound residual core is the same as in Refs. [@Gom18; @Holl19], which is computed by the standard SM with the WBT interaction [@War92] and includes the c.m. correction [@Die74]. Since many of the considered nuclei are weakly bound, the use of SF calculated from the SM, compared with the IPM limit, provides a more proper description of single-particle-strength fragmentation near the Fermi level. Reduction factors {#sec.redfac} ----------------- [lcddc]{} Reaction & & & &\ $^{10}$C($p$,$pn$)$^{9}$C & 386 & 12.95 & 16.3(22)\[14\] & 1.26(29)\ $^{11}$C($p$,$2p$)$^{10}$B & 325 & 15.68 & 18.2(9)\[10\] & 1.16(19)\ $^{11}$C($p$,$pn$)$^{10}$C & 325 & 14.07 & 17.0(15)\[21\] & 1.21(27)\ $^{12}$C($p$,$2p$)$^{11}$B & 398 & 22.04 & 19.2(18)\[12\] & 0.87(16)\ $^{12}$C($p$,$pn$)$^{11}$C & 398 & 27.43 & 30.0(32)\[27\] & 1.09(23)\ $^{13}$O($p$,$2p$)$^{12}$N & 401 & 5.77 & 5.78(91)\[37\] & 1.00(22)\ $^{14}$O($p$,$2p$)$^{13}$N & 351 & 13.28 & 10.23(80)\[65\] & 0.77(13)\ $^{15}$O($p$,$2p$)$^{14}$N & 310 & 18.07 & 18.92(182)\[120\] & 1.05(19)\ $^{16}$O($p$,$2p$)$^{15}$N & 451 & 27.78 & 26.84(90)\[170\] & 0.97(15)\ $^{17}$O($p$,$2p$)$^{16}$N & 406 & 9.16 & 7.90(26)\[50\] & 0.86(14)\ $^{18}$O($p$,$2p$)$^{17}$N & 368 & 20.01 & 17.80(104)\[113\] & 0.89(15)\ $^{21}$O($p$,$2p$)$^{20}$N & 449 & 5.58 & 5.31(23)\[34\] & 0.95(15)\ $^{21}$N($p$,$2p$)$^{20}$C & 417 & 3.25 & 2.27(34) & 0.70(14)\ $^{21}$N($p$,$pn$)$^{20}$N & 417 & 38.87 & 48.52(404) & 1.25(23)\ $^{22}$O($p$,$2p$)$^{21}$N & 414 & 6.90 & 6.01(41) & 0.87(14)\ $^{22}$O($p$,$pn$)$^{21}$O & 414 & 36.24 & 39.24(234) & 1.08(19)\ $^{23}$O($p$,$2p$)$^{22}$N & 445 & 4.97 & 4.93(96) & 0.99(24)\ $^{23}$O($p$,$pn$)$^{22}$O & 445 & 50.05 & 54.0(108) & 1.08(28)\ \[tab.result\] The results of our DWIA calculations are presented in Table \[tab.result\]. Because of the semi-inclusive nature of the data concerned, the calculation results are the sum of the cross sections corresponding to several bound-state configurations of the residual nucleus. The beam energy in the middle of the target is shown in the second column. The third column indicates the theoretical cross sections $\sigma_\textrm{th}=\sum C^2S \times \sigma_\textrm{sp}$, where $\sigma_\textrm{sp}$ is the single-particle cross section of a specific configuration. The fourth column shows the experimental cross section, with statistical (round brackets) and systematic (square brackets) uncertainties, corresponding to the beam of $^{10-12}$C [@Pan16; @Holl19], $^{13-21}$O [@Atar18] and $^{21}$N,$^{22,23}$O [@Fer18]. The reduction factor $R_s=\sigma_\textrm{exp}/\sigma_\textrm{th}$ is given in the last column. ![ Reduction factors deduced from $(p,pN)$ reactions using DWIA compared with the reduced SFs calculated from CC [@Jen11] (black circles) and SCGF [@Cip15; @Atar18] with different interactions (blue and green triangles). The black (red) texts in the header label the isotope whose proton (neutron) is knocked out. See the text for details. []{data-label="fig.main-theory"}](Main-theory.eps){width="48.00000%"} The reduction factors as a function of the proton-neutron asymmetry $\Delta S$ is shown in Fig. \[fig.main-theory\]. The value calculated from the present DWIA analysis is indicated by red squares with error bars propagated from the experimental uncertainties reported in Refs. [@Pan16; @Atar18; @Fer18; @Holl19] and the theoretical uncertainties estimated in Sec. \[secformalism\], which is about 14%–16%. The total relative uncertainties for the extracted reduction factor are ranging from 15% to 25%, which is similar to those reported in the recent systematic $(p,d)$ analysis [@Xu19]. In general, a trend of the reduction factor about 0.9–1.0 with a very weak asymmetry dependence is observed. Properly taking the uncertainties into account, the reduction factors $R_s=0.87(16)$ for $^{12}$C and $R_s=0.97(15)$ for $^{16}$O are overestimated compared with the $(e,e'p)$ [@Lap93] results of $R_s=0.57(6)$ and $R_s=0.65(5)$, respectively. By performing a linear function fitting, the $\Delta S$ dependence of $R_s$ is obtained as $R_s=0.947(36)-2.6(27)\times 10^{-3}\Delta S$ with a reduced $\chi^2/N$ of 0.74. As discussed in Sec. \[sec.reaction\], the close-to-unity reduction factor does not necessarily mean that the quenching effect observed in $(p,pN)$ reactions is weak but rather indicates a fundamental problem in current reaction models. The reduction factors from DWIA are compared with the reduced SFs, which is the ratio of the SF to the IPM limit, from ab initio self-consistent Green’s function (SCGF) [@Atar18; @Cip15] and coupled-cluster (CC) [@Jen11] models. We note that all of these values are presented as a function of experimental $\Delta S$. The CC calculation [@Jen11] uses the chiral NN interaction at next-to-next-to-next-to-leading order (N3LO) [@Ent03] with the cutoff at $\Lambda_\textit{NN}=500$ MeV while the SCGF of Ref. [@Cip15] uses the same NN interaction in addition to an NNLO three-nucleon force [@Roth12] with $\Lambda_{3N}=400$ MeV. We compare also with the result of the more recent SCGF calculation in Ref. [@Atar18] based on the NNLO-sat [@Eks15] for the NN interaction, which is more optimized for the mass region in that study. The present DWIA calculation shows a reasonable agreement with ab initio results, especially about the weak dependence of the trend on proton-neutron asymmetry. Comparison with $\bm{(e,e'p)}$ results and other reaction models {#sec.reaction} ---------------------------------------------------------------- Figure \[fig.main-reaction\] shows the reduction factors from the present DWIA calculation in comparison with those analyzed by other theoretical reaction models. The slope of the reduction factor obtained by the present analysis is in excellent agreement with the value deduced by TC model with fixed-energy EDAD2 OP [@Gom18]. It is also consistent (within the uncertainty range) with the slopes observed in Refs. [@Atar18; @Fla18]. ![ Same as Fig. \[fig.main-theory\] but compared with other $(p,pN)$ analysis performed with the eikonal DWIA [@Atar18] (green triangles), FAGS [@Fer18] (pink triangles), TC [@Gom18] (blue dashed line) for the GSI data, and similar partial-wave DWIA for the RIKEN/RCNP data [@Kaw18] (black circles). The results from Refs. [@Gom18; @Kaw18] are those performed with Dirac OP. The blue shaded band indicates the trend observed in the analysis of nucleon removal reaction with composite targets [@Gade08; @Tos14]. []{data-label="fig.main-reaction"}](Main-reaction.eps){width="48.00000%"} The values of the reduction factor given by DWIA overestimate those reported by $(e,e'p)$ analysis [@Lap93] and other reaction models such as TC [@Gom18] and the eikonal DWIA [@Atar18]. This means the present DWIA calculation gives a smaller cross section compared with these. However, the better agreement between the $(e,e'p)$ results and those given by other $(p,pN)$ analyses of the same GSI data will be, in fact, caused by the lack of several crucial corrections in the reaction models used in these analyses. We will return to this point below. The overshooting of the $(e,e'p)$ results means that the DWIA cross section is smaller than the observed value. This indicates that some contributions from higher-order processes such as multistep scattering or channel coupling are included in the GSI data, especially in the large recoil momentum region. In fact, a smaller reduction factor around 0.7 is observed from the analysis of $(p,2p)$ data measured at RIKEN/RCNP [@Kaw18] using the same DWIA framework as in this study. An important feature of the measurement at RIKEN/RCNP is a very constrained kinematics corresponding to the quasifree condition. This supports the conclusion that the lack of higher-order effects will be the main reason for the underestimation of the GSI data with the current DWIA calculation. The discrepancies between the current DWIA results and those using other reaction models can be explained by several factors. First, the nonlocality corrections in the single-particle and scattering wave functions are not presented in the TC and eikonal DWIA models. Second, the TC calculation [@Gom18] uses the energy-independent optical potentials evaluated at $E_\textrm{beam}/2$. On the other hand, the large discrepancy with the FAGS method [@Fer18] is due to two reasons pointed out in Ref. [@Gom18], the lack of relativistic kinematics in the FAGS framework as shown in [@Yos18] and different choices of optical potentials used in Ref. [@Fer18]. The latter has been confirmed by recent FAGS calculation with a more proper OP for the energy range considered, where the cross section has been reduced by almost 30% [@Cre19]. We further investigate the possible source of discrepancies between different reaction models used in $(p,pN)$ studies. Impacts of the lack of nonlocality corrections and energy-dependent potentials on the reduction factors are illustrated in Fig. \[fig.deltaNL\]. The relative difference $\Delta R$ is evaluated with respect to the “reference” DWIA result shown in Sec. \[sec.redfac\]. As seen from the lines in Fig. \[fig.deltaNL\], these effects are strongly associated with the separation energy and affect the reduction factor though the cross section in the opposite ways. For the weakly bound nucleus, the single-particle wave function is more extended and since the nonlocality correction only affects the interior of the wave function, its effect gets weaker with decreasing separation energy. On the other hand, since the energy-independent potentials of the outgoing nucleons are evaluated at half the incident beam, the deviation from energy-dependent OP is minimized for the small separation energy, where the kinematics most resembles the quasifree condition. Some deviations from both lines are because of the cross section for each case is the sum of single-particle wave functions with different orbitals. In general, by neglecting both the nonlocality correction and the energy dependence of the OP, DWIA calculations are expected to have a similar magnitude to those from TC [@Yos18]. ![ The relative difference with respect to the reference DWIA results as a function of knocked-out-nucleon separation energy. The red squares and black circles represent the DWIA calculations with energy-dependent potentials and without nonlocality corrections, respectively. []{data-label="fig.deltaNL"}](deltaNL.eps){width="48.00000%"} Finally, the effect of the Møller factor is presented in Fig. \[fig.deltaMol\]. This factor, shown in Eq. (\[eq.mol\]), has a relativistic origin [@Mol45; @Ker59; @Udi95] and therefore is directly related to the incident energy. For the considered energy range of 300–450 MeV, the Møller factor can contribute about 18%–26% to the reduction factor, which also explains the small magnitude of the nonrelativistic FAGS model. We note that, although the TC model does not explicitly include the Møller factor, its consistency with the DWIA model as found in [@Yos18] suggests that the Møller-factor effect has been implicitly included in the relativistic treatment of TC [@Mor15]. ![ The relative difference between the reference DWIA results and those without the Møller factor as a function of beam energy. []{data-label="fig.deltaMol"}](deltaMol.eps){width="48.00000%"} Although the lack of nonlocality correction, energy dependence in OP, and the Møller factor may cancel with the lack of higher-order effects in some models and give a better agreement with $(e,e'p)$ results, they also obscure the true nature of the problem. As we have seen, the effects of these corrections are highly dependent on the separation energy of the single-particle wave function and beam energy, so more proper inclusion of these corrections will be very essential for future knockout studies. Transverse momentum distribution -------------------------------- In inverse kinematics nucleon knockout reactions, a comparison between the measured and calculated momentum distribution of the residual nucleus can reveal a lot of information about the reaction mechanism as well as the validity of the theoretical model. We consider the cylindrical TMD of $^{12}$C($p$,$2p$)$^{11}$B [@Pan16], which is some of the highest resolution data from the R$^3$B/LAND experiments. ![ Cylindrical transverse momentum distribution of the $^{12}$C($p$,$2p$)$^{11}$B reaction. The experimental data are taken from Ref. [@Pan16]. The DWIA calculations with and without the nonlocality corrections are presented as the black solid and red dashed lines, respectively. The blue dotted line represents the result with nonlocality corrections scaled down by a factor of 0.655. []{data-label="fig.12C"}](12C.eps){width="46.00000%"} The possible absence of higher-order effects is pronounced in the comparison between the DWIA result (black solid line) and the experimental data [@Pan16] shown in Fig. \[fig.12C\]. The calculation results are the sum of the theoretical TMD corresponding to the bound states of the $^{11}$B core multiplied by the reduction factor $R_s$ of 0.87. One sees that the calculation result including the nonlocality corrections underestimates the experimental data in the momentum region about 150–300 MeV/$c$. We note that this undershooting of theoretical prediction is also observed in the FAGS calculation for the same data using OP fit to elastic-scattering data at the proper energy range [@Cre19]. Moreover, a similar discrepancy with the data in the high-momentum region found in the Cartesian TMD calculated with TC [@Gom18] implies that some contributions are also not properly accounted for in other reaction models. Because of the success of the DWIA framework for the same system in the quasifree condition [@Wak17], the discrepancies reinforce our claim that they originate from the higher-order effects that take place when the recoil momentum becomes high. The red dashed line represents the DWIA result without nonlocality correction, i.e., neglecting the Perey and Darwin factors. The corresponding $R_s$ is 0.77. Although this line may seem to improve the result, we emphasize that such a prescription is inappropriate and inconsistent with the known properties of $(p,pN)$ scattering as reviewed in Ref. [@Wak17]. The blue dotted line is the DWIA result with nonlocality corrections multiplied by 0.66, which is the ratio of the $R_\textrm{s}$ in Table \[tab.result\] to the one determined by $(e,e’p)$ [@Lap93]. If the $(e,e’p)$ $R_s$ is correct, the blue dotted line should agree with the experimental data. Therefore, it is clear that some additional contributions are necessary to reproduce the data. Summary {#secsum} ======= We have performed an analysis on all the published data to date for the $(p,pN)$ reaction in inverse kinematics by the R$^3$B collaboration by using the standard partial-wave DWIA formalism. Our study focuses on evaluating the source of ambiguity in DWIA calculations of the reduction factor and investigating the discrepancies between various reaction models currently used for inverse kinematics $(p,pN)$ data. Our study suggests a very weak dependence of the reduction factor on the proton-neutron asymmetry $\Delta S$. This result is consistent with previous analyses on $(p,pN)$ using TC, eikonal and partial-wave DWIA [@Gom18; @Atar18; @Kaw18; @Holl19] and ab initio calculations [@Cip15; @Atar18; @Jen11]. More importantly, the present study suggests that the lack of a proper treatment of higher-order effects may considerably affect the cross section at kinematics far from the quasifree condition. That effect, which does not manifest in experiments performed around the recoilless condition, becomes more crucial in the semi-inclusive type of integrated cross-section measurements like those carried out at the GSI R$^3$B/LAND setup. However, it is mostly hindered by the lack of essential corrections such as nonlocality, relativistic, Møller factor, and energy-dependent OP as well as the considerably large uncertainty from various choices of distorting potentials and single-particle wave functions. The proper inclusion of these corrections in the future $(p,pN)$-reactions analyses will be required. Furthermore, unless measurements with restricted kinematics similar to that in Ref. [@Kaw18] are performed, a reaction model that takes into account higher-order processes will be necessary. Recently, a consistent description of the nonlocality in bound and scattering wave functions for $(e,e'p)$ has been done with the nonlocal dispersive optical model approach [@Atk18]. The incorporation of such treatment in the DWIA framework for proton-induced nucleon knockout reactions is in progress and will be reported elsewhere. We thank Mario Gómez-Ramos and Matthias Holl for providing essential details of their calculations. We also thank Tetsuo Noro for valuable discussions on the DWIA calculation. One of the authors (N.T.T.P) would like to thank the Research Center for Nuclear Physics, Osaka University for their hospitality during his stay in which this work was initiated. He also wants to thank the Vietnam MOST for its support through the Physics Development Program Grant No. TLCN.25/18. The computation was carried out with the computer facilities at the Research Center for Nuclear Physics, Osaka University. This work was supported in part by Grants-in-Aid of the Japan Society for the Promotion of Science (Grants No. JP16K05352), and the RCNP Young Foreign Scientist Promotion Program. [00]{} G. Jacob and Th. A. J. Maris, Rev. Mod. Phys. 38, 121 (1966). G. Jacob and Th. A. J. Maris, Rev. Mod. Phys. 45, 6 (1973). P. Kitching, W. J. McDonald, Th. A. J. Maris, and C. A. Z. Vasconcellos, Adv. Nucl. Phys. 15, 43 (1985). T. Wakasa, K. Ogata, and T. Noro, Prog. Part. Nucl. Phys. 96, 32 (2017). N. S. Chant and P. G. Roos, Phys. Rev. C 15, 57 (1977). N. S. Chant and P. G. Roos, Phys. Rev. C 27, 1060 (1983). T. Aumann, C. A. Bertulani, and J. Ryckebusch, Phys. Rev. C 88, 064610 (2013). K. Ogata, K. Yoshida and K. Minomo, Phys. Rev. C 92, 034616 (2015). K. Yoshida, M. Gómez-Ramos, K. Ogata, and A. M. Moro, Phys. Rev. C 97, 024608 (2018). L. Olivier, S. Franchoo, M. Niikura, Z. Vajta, D. Sohler, P. Doornenbal, A. Obertelli, Y. Tsunoda, T. Otsuka, G. Authelet et al., Phys. Rev. Lett. 119, 192501 (2017). Z. Elekes, Á. Kripkó, D. Sohler, K. Sieja, K. Ogata, K. Yoshida, P. Doornenbal, A. Obertelli, G. Authelet, H. Baba et al., Phys. Rev. C 99, 014312 (2019). R. Taniuchi, C. Santamaria, P. Doornenbal, A. Obertelli, K. Yoneda, G. Authelet, H. Baba, D. Calvet, F. Château, A. Corsi et al., Nature 569, 53 (2019). S. Chen, J. Lee, P. Doornenbal, A. Obertelli, C. Barbieri, Y. Chazono, P. Navrátil, K. Ogata, T. Otsuka, F. Raimondi et al., Phys. Rev. Lett. 123, 142501 (2019). S. Kawase, T. Uesaka, T. L. Tang, D. Beaumel, M. Dozono, T. Fukunaga, T. Fujii, N. Fukuda, A. Galindo-Uribarri, S. Hwang et al., Prog. Theor. Exp. Phys. 2018, 021D01 (2018). R. Crespo, A. Deltuva, E. Cravo, M. Rodríguez-Gallardo, and A. C. Fonseca, Phys. Rev. C 77, 024601 (2008). R. Crespo, E. Cravo, and A. Deltuva, Phys. Rev. C 99, 054622 (2019). A. M. Moro, Phys. Rev. C 92, 044605 (2015). L. Lapikás, Nucl. Phys. A 553, 297c (1993). V. R. Pandharipande, I. Sick, and P. K. A. d. Huberts, Rev. Mod. Phys. 69, 981 (1997). W. Dickhoff and C. Barbieri, Prog. Part. Nucl. Phys. 52, 377 (2004). I. Sick, Prog. Part. Nucl. Phys. 59, 447 (2007). C. Barbieri, Phys. Rev. Lett. 103, 202502 (2009). Ø. Jensen, G. Hagen, M. Hjorth-Jensen, B. A. Brown, and A. Gade, Phys. Rev. Lett. 107, 032501 (2011). A. Cipollone, C. Barbieri, and P. Navrátil, Phys. Rev. C 92, 014306 (2015). A. Gade, P. Adrich, D. Bazin, M. D. Bowen, B. A. Brown, C. M. Campbell, J. M. Cook, T. Glasmacher, P. G. Hansen, K. Hosier et al., Phys. Rev. C 77, 044306 (2008). J. A. Tostevin and A. Gade, Phys. Rev. C 90, 057602 (2014). J. Lee, J. A. Tostevin, B. A. Brown, F. Delaunay, W. G. Lynch, M. J. Saelim, and M. B. Tsang, Phys. Rev. C 73, 044608 (2006). F. Flavigny, A. Gillibert, L. Nalpas, A. Obertelli, N. Keeley, C. Barbieri, D. Beaumel, S. Boissinot, G. Burgunder, A. Cipollone et al., Phys. Rev. Lett. 110, 122503 (2013). F. Flavigny, N. Keeley, A. Gillibert, and A. Obertelli, Phys. Rev. C 97, 034601 (2018). B. P. Kay, J. P. Schiffer, and S. J. Freeman, Phys. Rev. Lett. 111, 042502 (2013). Y. P. Xu, D. Y. Pang, X. Y. Yun, C. Wen, C. X. Yuan, and J. L. Lou, Phys. Lett. B 790, 308 (2019). V. Panin, J. T. Taylor, S. Paschalis, F. Wamers, Y. Aksyutina, H. Alvarez-Pol, T. Aumann, C. A. Bertulani, K. Boretzky, C. Caesar et al., Phys. Lett. B 753, 204 (2016). L. Atar, S. Paschalis, C. Barbieri, C. A. Bertulani, P. Díaz Fernández, M. Holl, M. A. Najafi, V. Panin, H. Alvarez-Pol, T. Aumann et al., Phys. Rev. Lett. 120, 052501 (2018). P. Díaz Fernández, H. Alvarez-Pol, R. Crespo, E. Cravo, L. Atar, A. Deltuva, T. Aumann, V. Avdeichikov, S. Beceiro-Novo, D. Bemmerer et al., Phys. Rev. C 97, 024311 (2018). M. Holl, V. Panin, H. Alvarez-Pol, L. Atar, T. Aumann, S. Beceiro-Novo, J. Benlliure, C.A. Bertulani, J.M. Boillos, K. Boretzky et al., Phys. Lett. B 795, 682 (2019). M. Gómez-Ramos and A. M. Moro, Phys. Lett. B 785, 511 (2018). M. A. Franey and W. G. Love, Phys. Rev. C 31, 488 (1985). C. Møller, Kgl. Danske Videnskab. Selsbak, Mat-fys. Medd. 23, 1 (1945). A. K. Kerman, H. McManus, and R. M. Thaler, Ann. Phys. (NY) 8, 551 (1959). F. G. Perey, Direct Interactions and Nuclear Reaction Mechanism (Gordon and Breach Science Publishers, New York, 1963), p. 125. F. Perey and B. Buck, Nucl. Phys. 32, 353 (1962). L. G. Arnold, B. C. Clark, R. L. Mercer, and P. Schwandt, Phys. Rev. C 23, 1949 (1965). S. Hama, B. C. Clark, E. D. Cooper, H. S. Sherif, and R. L. Mercer, Phys. Rev. C 41, 2737 (1990). J. M. Udías, P. Sarriguren, E. Moya de Guerra, E. Garrido, and J. A. Caballero, Phys. Rev. C 51, 3246 (1995). E. D. Cooper, S. Hama, B. C. Clark, and R. L. Mercer, Phys. Rev. C 47, 297 (1993). E. D. Cooper, S. Hama, and B. C. Clark, Phys. Rev. C 80, 034605 (2009). B.A. Brown, Phys. Rev. C 58, 220 (1998). G. J. Kramer, H. P. Blok, and L. Lapikás, Nucl. Phys. A 679, 267 (2001). E. F. Redish, G. J. Stephenson, Jr., and G. M. Lerner, Phys. Rev. C 2, 1665 (1970). V. G. J. Stoks, R. A. M. Klomp, C. P. F. Terheggen, and J. J. de Swart, Phys. Rev. C 49, 2950 (1994). R. Machleidt, Phys. Rev. C 63, 024001 (2001). R. A. Arndt, I. I. Strakovsky, and R. L. Workman, Int. J. Mod. Phys. A 18, 449 (2003). E. K. Warburton and B. A. Brown, Phys. Rev. C 46, 923 (1992). A. E. L. Dieperink and T. de Forest, Jr., Phys. Rev. C 10, 543 (1974). D. R. Entem and R. Machleidt, Phys. Rev. C 68, 041001 (R) (2003). R. Roth, S. Binder, K. Vobig, A. Calci, J. Langhammer, and P. Navrátil, Phys. Rev. Lett. 109, 052501 (2012). A. Ekström, G. R. Jansen, K. A. Wendt, G. Hagen, T. Papenbrock, B. D. Carlsson, C. Forssén, M. Hjorth-Jensen, P. Navrátil, and W. Nazarewicz, Phys. Rev. C 91, 051301 (2015). M. C. Atkinson, H. P. Blok, L. Lapikás, R. J. Charity, and W. H. Dickhoff, Phys. Rev. C 98, 044627 (2018).	{ "pile_set_name": "ArXiv" }
--- abstract: 'To provide better access of the inventory to buyers and better search engine optimization, e-Commerce websites are automatically generating millions of easily searchable browse pages. A browse page consists of a set of slot name/value pairs within a given category, grouping multiple items which share some characteristics. These browse pages require a title describing the content of the page. Since the number of browse pages are huge, manual creation of these titles is infeasible. Previous statistical and neural approaches depend heavily on the availability of large amounts of data in a language. In this research, we apply sequence-to-sequence models to generate titles for high- & low-resourced languages by leveraging transfer learning. We train these models on multi-lingual data, thereby creating one joint model which can generate titles in various different languages. Performance of the title generation system is evaluated on three different languages; English, German, and French, with a particular focus on low-resourced French language.' author: - Prashant Mathur - Nicola Ueffing - \| Gregor Leusch\ MT Science Team\ eBay\ Kasernenstra[ß]{}e 25\ Aachen, Germany bibliography: - 'naaclhlt2016.bib' title: 'Multi-lingual neural title generation for e-Commerce browse pages' --- Introduction {#sec:intro} ============ Related work {#sec:relwrk} ============ Lexicalization {#sec:lexicalization} ============== Sequence-to-Sequence Models {#sec:seq2seq} =========================== Multilingual Generation {#sec:mutlinlg} ======================= Experiments {#sec:exp} =========== Results {#sec:res} ======= Conclusion {#sec:conclusion} ========== Acknowledgments {#acknowledgments .unnumbered} =============== Thanks to Pavel Petrushkov for all the help with the neural MT toolkit.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We demonstrate optical tuning of the scattering length in a Bose-Einstein condensate as predicted by Fedichev [et al.]{} \[Phys. Rev. Lett. [77]{}, 2913 (1996)\]. In our experiment atoms in a $^{87}$Rb condensate are exposed to laser light which is tuned close to the transition frequency to an excited molecular state. By controlling the power and detuning of the laser beam we can change the atomic scattering length over a wide range. In view of laser-driven atomic losses we use Bragg spectroscopy as a fast method to measure the scattering length of the atoms.' address: \| $^{1}$Institut für Experimentalphysik, Universität Innsbruck, Technikerstra[ß]{}e 25, 6020 Innsbruck, Austria\ $^{2}$Institut für Quantenoptik und Quanteninformation, Österreichische Akademie der Wissenschaften, 6020 Innsbruck, Austria author: - 'M. Theis,$^{1}$ G. Thalhammer,$^{1}$ K. Winkler,$^{1}$ M. Hellwig,$^{1}$ G. Ruff,$^{1,\dagger}$ R. Grimm,$^{1,2}$ and J. Hecker Denschlag$^{1}$' title: Tuning the scattering length with an optically induced Feshbach resonance --- The great progress in the field of ultracold quantum gases in recent years can be largely attributed to the existence of magnetically tunable Feshbach resonances [@Tiesinga]. Since their first experimental introduction into the field [@Inouye; @Courteille; @Roberts] they have been widely used to arbitrarily tune the interactions between atoms. A plethora of experiments has been performed ranging for example from ultra-high resolution molecular spectroscopy [@Chin] to the creation of bright matter wave solitons [@Kaykovich] as well as the production of new atomic [@Cornish] and molecular [@Jochim] Bose-Einstein condensates (BEC). In general a Feshbach resonance occurs when a colliding pair of atoms is resonantly coupled to a molecular bound state. A magnetically tunable Feshbach resonance is based on Zeeman shifting a bound molecular state into resonance with the scattering state. Alternative coupling schemes for inducing Feshbach resonances have been proposed but never experimentally applied to control atomic interactions. The use of radiofrequency [@Moerdijk] and static electric fields [@Marinescu] was suggested. Fedichev [et al.]{} [@Fedichev] proposed optical coupling of the scattering state with the molecular state which was theoretically analyzed further in [@Bohn; @Kokoouline]. This scheme, often referred to as “optical Feshbach resonance”, can be controlled via laser detuning and laser power. Inducing Feshbach resonances with optical fields offers experimental advantages compared to magnetic fields. The intensity and detuning of optical fields can be rapidly changed. Furthermore complex spatial intensity distributions can be easily produced which result in corresponding scattering length patterns across the sample. Optical transitions are always available even when no magnetic Feshbach resonances exist. Recently Fatemi [et al.]{} [@Fatemi] observed optical Feshbach resonances in photoassociation spectroscopy. They used photoionization to probe optically induced changes in the scattering wave function. However, the direct influence of the optical Feshbach resonance on the atomic scattering properties was not studied. In this Letter we report a direct measurement of the atomic scattering length $a$ in a BEC of $^{87}$Rb $\| F = 1, m_F = -1 \rangle$ as we cross an optical Feshbach resonance. With moderate laser intensities of about 500 W/cm$^2$ we can change the scattering length over one order of magnitude from 10 a$_0$ to 190 a$_0$ (a$_0$ = 1 Bohr radius). ![Scattering length $a$ (solid line) and inelastic collision rate coeffient K$_\text{inel}$ (broken line) as a function of the laser detuning from the photoassociation resonance. The curves are based on Eqs. (\[eq:scatt\_length\_Julienne\]) and (\[eq:scatt\_length\_Julienne2\]) for typical experimental parameters: $\Gamma_\text{stim}/2\pi= 54\text{ kHz}$, $\Gamma_\text{spon}/2\pi= 20\text{ MHz}$, $k_{i}= 2.47\times 10^{5}\, \text{m}^{-1}$, $a_\text{bg} = 100$ a$_0$ (dotted line). Inset: Scheme for optically coupling the scattering state with an excited molecular state.[]{data-label="theoryplot"}](figure1){width="3in"} To optically modify the scattering length we use laser light tuned close to a photoassociation resonance which couples the continuum state of incoming free atoms to an excited molecular level (see inset in Fig. \[theoryplot\]). This changes the wavefunction and consequently the scattering length of the scattering state. It also leads to atomic loss due to spontaneous decay via the molecular state. The resonant transition rate between the continuum state and the molecular state, which we denote $\Gamma_\text{stim}$, is proportional to the laser intensity. In our experiment $\Gamma_\text{stim}/2\pi$ is on the order of a few 10 kHz. This is three orders of magnitude less than the spontaneous decay rate $\Gamma_\text{spon}$ from the excited molecular state. In [@Bohn] Bohn and Julienne give convenient expressions for the scattering length $a$ and the inelastic collision rate coefficient $K_\text{inel}$ which describes the photoassociation loss. For $\Gamma_\text{stim} \ll \Gamma_\text{spon}$ these expressions can be approximated and, for a condensate [^1], read: $$\begin{aligned} \label{eq:scatt_length_Julienne} a = a_\text{bg} + \frac{1}{2 k_i} \ \frac{ \Gamma_\text{stim} \ \Delta} {\Delta^2 + (\Gamma_\text{spon}/2)^2} \\ \label{eq:scatt_length_Julienne2} K_\text{inel} = \frac{2\pi\hbar}{ m} \ \frac{1}{k_i} \ \frac{\Gamma_\text{stim} \ \Gamma_\text{spon}} {\Delta^2 + (\Gamma_\text{spon}/2)^2}\end{aligned}$$ where $a_\text{bg}$ is the scattering length in the absence of light, $\Delta$ is the detuning from the photoassociation line, $m$ the atomic mass and $\hbar k_i$ the relative momentum of the collision. Fig. \[theoryplot\] shows $a$ and $K_\text{inel}$ as a function of the detuning $\Delta$ for typical experimental parameters. According to Eqs. (\[eq:scatt\_length\_Julienne\]) and (\[eq:scatt\_length\_Julienne2\]) one should in general choose large detuning and strong coupling in order to maximize the change in scattering length while keeping the losses low. Our experiments are carried out with an almost pure $^\text{87}$Rb condensate in the $\| F=1, m_F = -1 \rangle$ spin state with typically $1\times 10^6$ atoms. The setup uses a magnetic transport scheme [@Greiner] to transfer atoms from a magneto-optical trap (MOT) chamber to a glass cell where the BEC is produced by rf-evaporation in a cigar shaped QUIC trap [@QUIC] with trap frequencies $\omega_\text{axial}/2\pi = 15\text{ Hz}$ and $\omega_\text{radial}/2\pi = 150\text{ Hz}$ [^2]. The intensity stabilized photoassociation laser beam ($\approx 40$ mW) is derived from a Ti:Sa laser. It is aligned along the axial direction of the cigar shaped BEC and has a waist radius of 76 $\mu$m. Its linear polarization is perpendicular to the trapping magnetic bias field of 2 Gauss. In our experiments we limit the maximum laser intensities to about 500 W/cm$^2$ because we observe the appearance of a growing component of thermal atoms for higher intensities [^3]. This effect is negligible for laser powers below 500 W/cm$^2$. ![Photoassociation spectrum of the excited molecular state used in the experiment. The two lines belong to the state $\| 0_g^- (\sim S_{1/2}+P_{3/2}), \nu = 1\rangle$ and have rotational quantum numbers $J=0, 2$ respectively. Shown is the remaining atom number after exposing a BEC to a 70 $\mu$s light pulse of 460 W/cm$^2$ intensity. The detuning is given relative to the $J=2$ line. Each data point is an average of three measurements. []{data-label="molecularlines"}](figure2){width="3in"} In order to identify a suitable molecular level with strong coupling to the continuum state we investigated molecular lines in the $1_g$ and $0_g^-$ potentials which connect to the $(S_{1/2} + P_{3/2})$ and $(S_{1/2} + P_{1/2})$ asymptotes. We choose the excited state $\| 0_g^- (\sim S_{1/2} + P_{3/2}), \nu = 1, J = 2 \rangle$ which is located 26.8 cm$^{-1}$ below the $D$2 line [@Fioretti]. Fig. \[molecularlines\] shows the corresponding photoassociation line together with the line for $J = 0$. At a laser intensity of 460 W/cm$^2$ the measured atom losses yield a peak inelastic collision rate $K_\text{inel}=(2 \pm 1) \times 10^{-10} \text{cm$^3$/s}$ which is a factor of 5 weaker than $K_\text{inel}$ in the example of [@Bohn]. Losses due to excitation of the $D$2 line can be neglected. We observe a strong intensity dependent light shift of 215 MHz/(kW cm$^{-2}$) of the photoassociation line which might be mainly explained by coupling to a $d$-wave shape resonance [@Simoni]. ![Two Bragg resonance curves with an optically induced relative shift of 0.75 kHz. The percentage of the diffracted atoms is plotted against the frequency difference of the lattice beams. The two curves correspond to a detuning $\Delta/2\pi = - 47$ MHz (filled circles) and $\Delta/2\pi = + 47$ MHz (open circles) at a photoassociation laser intensity of 460 W/cm$^2$. The lines shown are fits to the data. For better comparison the right curve (open circles) has been scaled by a factor of 1.09 to the same height as the left one.[]{data-label="braggspectroscopy"}](figure3){width="3in"} Measuring the scattering length close to a photoassociation resonance requires a fast experimental method as atom losses restrict the observation time to below 100 $\mu$s in our experiments. Thus the scattering length can neither be extracted from measurements of the collision rate [@Roberts] nor from the mean-field energy in a condensate expansion [@Inouye] both of which require a few ms. Instead we use Bragg spectroscopy [@Stenger] to determine the mean-field energy by imposing on the atoms a moving optical lattice composed of two counter-propagating laser beams with wavenumber $k$ and an adjustable frequency difference $\Delta f$. The Bragg lattice transfers a momentum of $2\hbar k$ to the atoms in a first order diffraction process. This is resonant when energy conservation is fulfilled, which for [noninteracting]{} atoms reads $ h \Delta f_0 = (2\hbar k)^2/2m $. For a condensate, however, the resonance frequency $\Delta f_r$ is shifted by the mean-field energy. In the Thomas-Fermi approximation this yields a value of $$\label{eq:braggspectroscopy} \Delta f_r = \Delta f_0 + \frac{8 \hbar}{7 m} \ n_0 \ a$$ where $ n_0 $ denotes the atomic peak density [@Stenger]. Observing this shift of the Bragg resonance frequency therefore allows to measure the product of density and scattering length. We derive the two Bragg beams from a laser which is 1.4 nm blue detuned relative to the $^{87}$Rb $D$2 line. This determines $\Delta f_0$ to be 15.14 kHz. Two acousto-optical modulators are used to control the frequency difference $\Delta f$ between the two counter-propagating beams. The beams have a diameter of $\approx 900$ $\mu$m and are aligned along the radial trap axis in a horizontal direction. In our measurements we apply a 70 $\mu$s square-pulse of Bragg light to the condensate. After 12 ms of time of flight, when the momentum components of the condensate have spatially separated, we use absorption imaging to measure the portion of condensate atoms that have been diffracted. We always choose the intensity of the lattice such that about 15% to 20% of the atoms are diffracted at resonance. Scanning $\Delta f$ and determining the percentage of diffracted atoms yields curves as shown in Fig. \[braggspectroscopy\] from which we extract the resonance positions. Shining in a photoassociation laser pulse (70 $\mu$s square pulse) at the same time as the Bragg pulse shifts the resonance position. This shift depends on the detuning $\Delta $ from the molecular line (filled and open circles in Fig. \[braggspectroscopy\]). For short illumination times $T$ as in our experiment the shape of the spectra fits well to the Fourier transform of the rectangular light pulse, $\sin^2 \left(\pi (\Delta f-\Delta f_r) T \right)/ (\Delta f-\Delta f_r)^2$, which we use to fit the data (see Fig. 3). Our measurements show that in spite of the Fourier-limited width of the Bragg resonance of 13 kHz (FWHM) we can resolve the peak position to better than $\pm$100 Hz. When we invert the frequency difference of the Bragg laser beams and diffract atoms to a momentum state with $-2\hbar k$ instead of $+2\hbar k$ we notice that the absolute value of the resonance frequency $\|\Delta f_r\|$ changes. This can be explained by an initial condensate momentum of up to 0.05 $\hbar k$ which we find to slowly vary from day to day. This initial momentum is due to residual experimental imperfections like optical dipole forces of a slightly non-centered photoassociation beam. To eliminate this effect we always measure $\Delta f_r$ for $+2\hbar k$ as well as for $-2\hbar k$ and then take the difference. ![Optical Feshbach resonance. In (a) the final atom number is plotted versus the detuning of the photoassociation laser (the dashed line is a Lorentz curve to guide the eye). The data in (b) displays the measured Bragg resonance frequencies. In (c) the values for the scattering length obtained from the data in (a) and (b) are plotted. The continuous line is a fit of Eq. (\[eq:scatt\_length\_Julienne\]) to the data.[]{data-label="scatteringlength"}](figure4){width="3in"} Figure \[scatteringlength\] shows the data we obtain from scanning the photoassociation laser over the optical resonance for a fixed laser intensity of 460 W/cm$^2$. The number of atoms in the condensate at the end of the laser pulse is plotted in Fig. \[scatteringlength\](a) indicating the position of the molecular line. On resonance about 90% of the atoms are lost after the 70 $ \mu$s of interaction time. Fig. \[scatteringlength\](b) shows the resonance frequency $\Delta f_r$ for Bragg diffraction as a function of laser detuning $\Delta$. For large positive (and negative) detuning $\Delta $ the value of $\Delta f_r$ agrees with the 16.6 kHz expected from theory for the background scattering length $a_\text{bg}=100\,\text{a}_0$ [@privTiemann; @privJulienneTiesinga] and a BEC with $\approx 8.2 \times 10^5$ atoms. As we tune across the molecular resonance the measured resonance frequencies exhibit a distorted dispersive shape. Following Eq. (\[eq:braggspectroscopy\]) this is the result of the combination of two effects: First the scattering length $a$ varies with $\Delta$ which alone should result in a dispersive line shape as in Fig. \[theoryplot\]. Second the atomic density $n_0$ decreases due to photoassociation losses which would, if the scattering length was constant, result in a symmetrical dip for $\Delta f_r$. On the right hand side of the resonance these two effects nearly compensate each other whereas on the left hand side the effects add up to produce a prominent dip in $\Delta f_r$. In order to extract the scattering length $a$ from the measured frequencies one can in a first approach replace the dynamically changing density $n_0$ in Eq. (\[eq:braggspectroscopy\]) by a time averaged value $\langle n_0 \rangle_t$. The average $\langle n_0 \rangle_t$ can be derived from the rate equation for the local density [@McKenzie] describing two-atom losses. This yields values for $a$ which differ only marginally from the ones in Fig. \[scatteringlength\](c). The data in Fig. \[scatteringlength\](c) were obtained from a more detailed examination which takes into account the full spatially resolved time evolution of the condensate density [^4]. This includes the dynamical flattening of the condensate density profile caused by the rapid atom loss which is much faster than the trap frequencies [@McKenzie]. Fig. \[scatteringlength\](c) shows that with a laser power of 460 W/cm$^2$ we can tune the scattering length over a range from 10$\text{a}_0$ to 190$\text{a}_0$. A fit of Eq. (\[eq:scatt\_length\_Julienne\]) to these data for $a$ yields a spontaneous decay width $\Gamma_\text{spon}/2\pi $ of 20 MHz and a resonant inelastic collision rate coefficient $K_\text{inel}=1.7 \times 10^{-10} \text{cm}^3/\text{s}$. These values agree with those we obtain from atom loss measurements. Thus our data consistently confirm the intrinsic relation between $a$ and $K_\text{inel}$ as expressed in Eqs. (\[eq:scatt\_length\_Julienne\]) and (\[eq:scatt\_length\_Julienne2\]). The measured width $\Gamma_\text{spon}/2\pi $ of 20 MHz is larger than the expected molecular decay width of 12 MHz (corresponding to two times the atomic width). This might be explained by the line width of the Ti:Sa laser of about 4 MHz and a power broadening of the line due to different light shifts of unresolved molecular hyperfine states [@privJulienneTiesinga; @McKenzie]. ![Dependence of the optically induced mean-field shift on the laser intensity.[]{data-label="shiftvsintensity"}](figure5){width="3in"} Fig. \[shiftvsintensity\] demonstrates the linear dependence of the scattering length $a$ on the photoassociation laser intensity. For these measurements we determine the Bragg resonance frequency for the detunings $\Delta/2\pi \approx \pm 50$MHz at various photoassociation laser intensities. This is slightly complicated by the light shift and broadening of the photoassociation line which lead to an uncertainty in $\Delta / 2\pi$ of $\pm 10$MHz. We keep the final atom number and density fixed by adjusting the pulse duration for each laser intensity. This ensures that only changes in $a$ are reflected in the varying mean-field shift. In Fig. \[shiftvsintensity\] we plot the frequency difference $\Delta f(\text{+50\,MHz})-\Delta f(-\text{50\,MHz)}$ which increases our signal. In conclusion, our experiments demonstrate the tunability of the scattering length in ultracold samples by optically coupling free atoms to a bound molecular state. Because of the exquisite control one has over laser fields we expect optical Feshbach resonances to be valuable when it comes to control atom-atom interactions in demanding applications. The inherent losses suggest the use of high laser intensities at large detuning and a good choice of the molecular state in order to optimize the ratio of change in scattering length and loss rate. Optical Feshbach tuning could be particularly useful to control atomic interactions in optical lattices which are discussed as potential future quantum information processors. We thank Paul Julienne, Eite Tiesinga, John Bohn, Olivier Dulieu, Peter Fedichev, Andrea Micheli and Helmut Ritsch for very helpful discussions. This work was supported by the Austrian Science Fund (FWF) within SFB 15 (project part 17) and by the European Union in the frame of the Cold Molecules TMR Network under contract No. HPRN-CT-2002-00290. $\dagger$ Permanent address: Department of Physics, Bates College, Lewiston, ME 04240. [99]{} E. Tiesinga, B.J. Verhaar, and H.T.C. Stoof, Phys. Rev. A [47]{}, 4114 (1993). S. Inouye [et al.]{}, Nature [392]{}, 151 (1998). Ph. Courteille [et al.]{}, Phys. Rev. Lett. [81]{}, 69 (1998). J.L. Roberts [et al.]{}, Phys. Rev. Lett. [81]{}, 5109 (1998). C. Chin [et al.]{}, Phys. Rev. Lett. [85]{}, 2717 (2000); A. Marte [et al.]{}, Phys. Rev. Lett. [89]{}, 283202 (2002). L. Khaykovich [et al.]{}, Science [296]{}, 1290 (2002); K.E. Strecker [et al.]{}, Nature [417]{}, 150 (2002). S. Cornish [et al.]{}, Phys. Rev. Lett. [85]{}, 1795 (2000); T. Weber [et al.]{}, Science [299]{}, 232 (2003). S. Jochim [et al.]{}, Science [302]{}, 2101 (2003); M. Greiner, C.A. Regal, and D.S. Jin, Nature [426]{}, 537 (2003); M. Zwierlein [et al.]{}, Phys. Rev. Lett. [91]{}, 250401 (2003); T. Bourdel [et al.]{}, cond-mat/0403091. A.J. Moerdijk, B.J. Verhaar, and T.M. Nagtegaal, Phys. Rev. A [53]{}, 4343 (1996). M. Marinescu and L. You, Phys. Rev. Lett. [81]{}, 4596 (1998). P. Fedichev, Yu Kagan, G.V. Shlyapnikov, and J.T.M. Walraven, Phys. Rev. Lett [77]{}, 2913 (1996). J. Bohn and P.S. Julienne, Phys. Rev. A [56]{}, 1486 (1997). V. Kokoouline, J. Vala, and R. Kosloff, J. Chem. Phys. [114]{}, 3046 (2001). F.K. Fatemi, K.M. Jones, and P.D. Lett, Phys. Rev. Lett. [85]{}, 4462 (2000). M. Greiner, I. Bloch, T.W. Hänsch, and T. Esslinger, Phys. Rev. A [63]{}, R031401 2001. T. Esslinger, I. Bloch, and T.W. Hänsch, Phys. Rev. A 58, R2664 (1998). A. Fioretti [et al.]{}, Eur. Phys. J. D [15]{}, 189 (2001). A. Simoni, P.S. Julienne, E. Tiesinga, and C.J. Williams, Phys. Rev. A [66]{}, 063406 (2002). J. Stenger [et al.]{}, Phys. Rev. Lett. [82]{}, 4569 (1999). Eberhard Tiemann, private communication. Paul Julienne and Eite Tiesinga, private communication. C. McKenzie [et al.]{}, Phys. Rev. Lett. [88]{}, 120403 (2002). [^1]: For condensed atoms the collision rate coefficient is only half of the coefficient for thermal atoms as all atoms share the same quantum state. [^2]: Details of our setup will be described elsewhere. [^3]: We attribute this heating to the appearance of an uncontrolled corrugation in the photoassociation laser beam which transfers momentum components to the condensate atoms. The corrugation could be a result of interfering backreflections from the glass cell windows. [^4]: This calculation is a simulation of the Rabi flopping between two levels corresponding to the condensate component at rest and the component with momentum $2\hbar k$. The changing density due to the loss is included by introducing a time dependent detuning.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We study the statistics of charge transport in a mesoscopic three-terminal device with one superconducting terminal and two normal-metal terminals. We calculate the full distribution of transmitted charges into the two symmetrically biased normal terminals. In a wide parameter range, we find large positive crosscorrelations between the currents in the two normal arms. We also determine the third cumulant that provides additional information on the statistics not contained in the current noise.' author: - 'J. Börlin' - 'W. Belzig' - 'C. Bruder' date: '30 April 2002; published in Phys. Rev. Lett. [88]{}, 197001 (2002)' title: Full Counting Statistics of a Superconducting Beam Splitter --- The number of charges transfered in a transport process fluctuates due to quantum-mechanical uncertainty and statistics. Therefore, the outcome of a current measurement accumulated over some time period $t_0$ is in general described by a probability $P(N)$, where $N$ is the total number of charges transfered. $P(N)$ is called the full counting statistics (FCS) of the transport process [@levitov:96]. The first two moments of the FCS are related to the average current and the current noise and are accessible to present experimental techniques. Higher-order correlations are likely to be measured in the future. Several schemes to measure either higher correlators or the full distribution have been proposed recently [@levitov:96; @belzig:01-2; @levitov:01; @kindermann:01; @pumps; @beenakker]. The current noise, i. e., the second moment of the FCS, is of particular interest. It can be used as a diagnostic tool to probe the nature and the quantum statistics of the charge carriers [@blanter] and the existence of entanglement [@loss]. For superconductor(S)-normal metal(N) heterostructures, a doubling of the shot noise in comparison to the normal case was predicted [@2e] and measured in diffusive heterostructures [@2e-measure]. Recent calculations taking into account the proximity effect in such structures [@belzig:01-1] are in good agreement with experimental results [@kozhevnikov]. Multi-terminal SN structures have been suggested to produce entangled electron pairs [@recher; @lesmarblat]. So far, crosscorrelations, i. e., current correlations involving different terminals, were measured only in normal single-channel heterostructures [@hantwissexp]. These have confirmed the prediction [@buettiker:91] that current crosscorrelations in a fermionic system are always negative. To our knowledge, there is no measurement of crosscorrelations in a system with superconducting contacts up to now. Theoretically, positive crosscorrelation with a single-channel beam splitter for Andreev pairs injected from a superconductor have been predicted [@martin]. In a setup in which crosscorrelations between a normal lead and a tunneling probe are considered, the sign of the correlations was found to depend crucially on the sample geometry [@gramesbacher]. A numerical study found positive crosscorrelations in a three-terminal device with a few channels with ferromagnetic contacts [@taddei2]. In this Letter we find the full counting statistics of a many-channel beam splitter that divides a supercurrent in two normal quasiparticle currents. We calculate the distribution of the transmitted charges taking the proximity effect into account. For comparison we also calculate the FCS for the case in which the superconducting terminal is replaced by a normal one. ![Three-terminal beam-splitter. One superconducting or normal terminal (S or N) and two normal terminals ($N_1$ and $N_2$) are tunnel-coupled by conductances $g$, $g_1$ and $g_2$ to a common central node . A current is passed from S/N into the two normal leads that are kept at the same voltage. Ideal passive charge counters are indicated by the counting fields $\chi$, $\chi_1$ and $\chi_2$.[]{data-label="fig:system"}](boerfig1){width="4.5cm"} The setup of our three-terminal device with one superconducting and two normal-metal terminals is shown in Fig. \[fig:system\]. All three terminals are connected by tunnel junctions to a small normal-metal island. We assume the island to be large enough, that we can neglect charging effects, and small enough that we can treat the Green’s functions of the island as constant. Thus, we are restricted to energies below the Thouless energy of the island. The system is then appropriately described by the circuit theory of mesoscopic transport [@yuli:9499-1], within which the counting statistics is easily accessible [@yuli:99-2; @belzig:01-1; @belzig:01-2]. The circuit elements that are important for our case are normal, resp. superconducting terminals and tunnel connectors. The terminals are described by quasiclassical Green’s functions, which depend on the type of the terminal (N or S), chemical potential, temperature $T$ and a counting field $\chi$. We assume zero temperature and a symmetric bias at the two normal terminals. Since we are interested in correlations between currents in different terminals, we introduce different counting fields. The voltage is chosen such that $eV\ll\Delta$. Charge transport (at $T=0$) occurs then only in the interval $\|E\|\leq eV$ and we need to consider only this energy interval below. The Green’s functions of the two normal terminals are then given by $$\label{eq:normal-greensfunc} \check G_{1,2} = e^{i\chi_{1,2}\check \tau_K/2} \check G_{N} e^{-i\chi_{1,2}\check \tau_K/2}\,,$$ where $\check G_N$ is the same for both normal terminals. At zero temperature $\check G_N=\hat \sigma_3 \bar \tau_3 + (\bar \tau_1+i\bar \tau_2)\hat 1$ for $\|E\|\leq eV$ and $\check G_N=\hat \sigma_3 \bar \tau_3 + \mbox{sgn}(E)\hat \sigma_3(\bar \tau_1+i\bar \tau_2)$ for $\|E\|>eV$. Here $\hat \sigma_i (\bar \tau_i)$ denote Pauli matrices in Nambu(Keldysh)-space. The counting rotation matrix is $\check \tau_K=\hat \sigma_3\bar \tau_1$. The superconducting terminal in equilibrium is characterized by $\check G_S=\hat\sigma_1\bar 1$ and a counting field $\chi$, that enters as in (\[eq:normal-greensfunc\]). If one node is connected to $M$ terminals by means of tunnel connectors, one can find a general form of the FCS, i.e. the probability $P(N_1,\ldots,N_M)$ that $N_{1(2,\ldots,M)}$ charges are counted in terminal $1(2,\ldots,M)$. The unknown Green’s function of the central node is denoted by $\check G_c$. The matrix currents into the central node are given by $%\begin{equation} \check I_k = \frac{g_k}{2} \left[ \check G_c,\check G_k \right] %\end{equation} $, where the index $k=1,\ldots,M$ labels the terminals and $g_k$ is the conductance of the respective junction. The Green’s function of the central node is determined by matrix current conservation on the central node, reading $ \sum_{k=1}^M \check I_k = \frac12 \left[ \sum_{k=1}^M g_k \check G_k , \check G_c \right]=0$. Employing the normalization condition $\check G_c^2=1$, the solution is $$\check G_c = \frac{\sum_{k=1}^M g_k \check G_k}{ \sqrt{\sum_{k,m=1}^M g_kg_m \left\{\check G_k,\check G_m\right\}/2}}\,.$$ To find the cumulant-generating function (CGF) $S$ of $P(N_1,\ldots,N_M)$ we integrate the equations $(-it_0/e) \partial S(\chi_1,\ldots,\chi_M)/\partial \chi_k = \int dE \mbox{Tr}\check\tau_K\check I_{k}/8e$ [@yuli:01-1]. We obtain $$\label{eq:cgf-general} S(\chi_1,\ldots,\chi_M) =- \frac{t_0}{e}\int \frac{dE}{2} \mbox{Tr} \sqrt{\sum_{k,m=1}^M \frac{g_kg_m}{2} \left\{\check G_k,\check G_m\right\}}\,.$$ This is the general result for an M-terminal geometry in which all terminals are tunnel-coupled to a common node. We now evaluate Eq. (\[eq:cgf-general\]) for our three terminal setup. Introducing $p_i=2 g g_i /(g^2+(g_1+g_2)^2)$ we find $$\label{eq:cgf} S(\chi_1,\chi_2,\chi) = -\frac{Vt_0\sqrt{g^2+(g_1+g_2)^2} }{\sqrt 2 e} \sqrt{1+ \sqrt{1+ \left(p_1 e^{i(\chi_1-\chi)}+p_2e^{i(\chi_2-\chi)}\right)^2-(p_1+p_2)^2}}\,.$$ This result for the cumulant-generating function incorporates all statistical transport properties for our present setup. The inner argument contains counting factors for the different possible processes. A term $\exp(i(\chi_k+\chi_l-2\chi)-1)$ corresponds to an event in which two charges leave the superconducting terminal and one charge is counted in terminal $k$ and one charge in terminal $l$. The prefactors are related to the corresponding probabilities. For instance, $p_1$ is proportional to the probability of a coherent tunneling event of an electron from the superconductor into terminal 1. A coherent pair-tunneling process is therefore weighted with $p_1^2$. This is accompanied by counting factors which describe either the tunneling of two electrons into terminal 1(2) \[counting factor $\exp(i2(\chi_{1(2)}-\chi))-1$\] or tunneling into different terminals \[counting factor $\exp(i(\chi_1+\chi_2-2\chi))-1$\]. The double square-root function shows that these different processes are non-separable. It is interesting to compare Eq. (\[eq:cgf\]) with the case in which the superconductor is replaced by a normal metal. The resulting CGF is $$\begin{array}[t]{l} S^N(\chi_1,\chi_2,\chi)=-\frac{Vt_0}{2e}(g+g_1+g_2)\times\\\\ \quad\sqrt{1+ p^N_1 \left(e^{i(\chi_1-\chi)}-1\right) + p^N_2 \left(e^{i(\chi_2-\chi)}-1\right)}\,, \end{array} \label{eq:cgf-normal}$$ where $p^N_{1(2)}=4 g g_{1(2)}/(g+g_1+g_2)^2$. Thus, one of the square roots in Eq. (\[eq:cgf\]) can be attributed to the multiple tunnel-junction geometry, which is already present in the normal configuration. The second square root in the CGF for the superconducting case must then be due to the proximity effect. We now evaluate some average transport properties of the S$\mid$NN-system and compare them to the N$\mid$NN-case. The currents into the different terminals are obtained from derivatives of the CGF: $I_k=(-ie/t_0) \partial S/\partial\chi_k\|_{\chi_1=\chi_2=\chi=0}$. The trans-conductances $G_k=I_k/V$ into terminal $k$ $(=1,2)$ are then given by $$\begin{aligned} G_k^S = \frac{g^2g_k(g_1+g_2)}{ \left(g^2+\left(g_1+g_2\right)^2\right)^{3/2}}&,& G_k^N = \frac{gg_k}{g+g_1+g_2}\,.\end{aligned}$$ The superscript $S(N)$ denotes the S$\mid$NN(N$\mid$NN)-case. Noise and crosscorrelations are obtained from second derivatives of the CGF, i. e., $P_{kl}^I=(2e^2/t_0) \partial^2 S(\chi_1,\chi_2,\chi)/\partial \chi_k\partial \chi_l\|_{\chi_1=\chi_2=\chi=0}$. We define Fano factors $F_{kl}=P^I_{kl}/2eI$, and we denote the Fano factor of the total current with $F=F_{11}+F_{22}+2F_{12}$. We also calculate the third cumulant of the total charge transfer (normalized to the Poisson value) $C_3=(ie/It_0)\partial^3 S(0,0,\chi)/\partial\chi^3\|_{\chi=0}$. The results in the superconducting case are $$\begin{aligned} \label{eq:fss} F_{12}^S = \frac{g_1g_2}{(g_1+g_2)^2}(1-5x^2)&,&F^S = 2- 5x^2\,, \\\nonumber C^S_3= 4-30 x^2+63 x^4 &,& x=\frac{g (g_1+g_2)}{g^2 +(g_1+g_2)^2}\,.\end{aligned}$$ In the N$\mid$NN case, on the other hand, we find $$\begin{aligned} \label{eq:fnn} F_{12}^N = -\frac{g_1g_2}{(g_1+g_2)^2}x_N&,& F^N = 1-2x_N\,, \\\nonumber C_3^N=1-6x_N+3x_N^2 &,& x_N=\frac{g(g_1+g_2)}{(g+g_1+g_2)^2}\,.\end{aligned}$$ ![Conductance, Fano factors, crosscorrelations and third cumulant of the beam splitter. The thick lines correspond to the S$\mid$NN-case and the thin lines to the N$\mid$NN-case. The conductance (upper-left panel) in the superconducting case shows a maximum around $g=g_1+g_2$. In the normal state, the conductance varies between $g$ and $g_1+g_2$. In both cases, the current noise (upper-right panel) shows a suppression around $g=g_1+g_2$ as compared to the limiting values of 2 in the S$\mid$NN-case and 1 in the N$\mid$NN-case. Large positive crosscorrelations occur in the superconducting case (lower-left panel), whereas they are always negative in the normal case. Around $g=g_1+g_2$, the superconducting crosscorrelations become negative. Note, that what is plotted here is $f_{12}^{S/N}= F_{12}^{S/N}(g_1+g_2)^2/g_1g_2$. The third cumulants (lower-right panel) are always positive. Around $g=g_1+g_2$ they are strongly suppressed. In the S$\mid$NN-case, $C_3^S$ has a double-minimum here, as shown in the blow-up.[]{data-label="fig:noise1"}](boerfig2){width="7cm"} All other Fano factors can be deduced from $F_{12}$ and $F$ using the relations $\sum_k F_{kl}=0$ and $F_{kl}=F_{lk}$. The transport properties are summarized in Fig. \[fig:noise1\]. In the figure the crosscorrelations are plotted as $f_{12}=F_{12} (g_1+g_2)^2/g_1g_2$. Most remarkably, the crosscorrelations $F_{12}^S$ are positive if $x$ is small, whereas $F_{12}^N$ is always negative in the normal state. Here, the Fano factor $F^S$($F^N$) is close to $2$($1$). Going to the regime $g \approx (g_1+g_2)$ suppresses the Fano factor $F^{S(N)}$ below 2(1) and leads to negative crosscorrelations $F_{12}^S$. In the limiting case $g = (g_1+g_2)$ the Fano factors are $F^{S}=3/4$ and $F^N=1/2$, and the crosscorrelations are $F_{12}^S=F_{12}^N = -g_1g_2/4(g_1+g_2)^2$. The third cumulant is always positive, but shows a strong suppression around the resonant conductance ratio $g = (g_1+g_2)$. In the limit of small $x$ ($x_N$) the third cumulant is $4$ ($1$), corresponding to the effective charge squared transfered in a tunneling process [@levitov:01]. However, the variation with $g/(g_1+g_2)$ in the S$\mid$NN-case is more pronounced than in the N$\mid$NN-case. As an interesting side remark we point out that $F_{12}^S=0$ and $F^S=1$ for $x^2=1/5$. This looks like a signature of uncorrelated charge transfer in units of $e$. However, the third cumulant $C_3^S=13/25$ differs from the corresponding value for uncorrelated $1e$-charge transfer, viz., $C_3=1$. Thus, higher correlations show that the charge transfer is still correlated. We briefly discuss the influence of an asymmetry $g_1\neq g_2$ of the beam splitter. The crosscorrelations are reduced, both in the S$\mid$NN and in the N$\mid$NN case. However, the positive crosscorrelations in the superconducting state persist for all values of the asymmetry. Cumulants of the total charge transfer like the conductance, $F^{S,N}$ and $C_3^{S,N}$ are independent of this asymmetry. Using the CGF from Eq. (\[eq:cgf\]), we can identify the physical processes leading to our previous results. We have seen from (\[eq:fss\]) that positive crosscorrelations are found if $g/(g_1+g_2)$ is not close to 1. Then, $p_{1,2}\ll 1$ and we can expand Eq. (\[eq:cgf\]) in $p_{1,2}$. Dropping the trivial dependence on $\chi$, the CGF can be written as $$\begin{aligned} \label{eq:cgf-poisson} S(\chi_1,\chi_2) = -\frac{t_0V}{e}\frac{g^2}{(g^2+(g_1+g_2)^2)^{3/2}}\times\qquad\qquad \\\nonumber \qquad\left( g_1^2 e^{i2\chi_1}+ g_2^2 e^{i2\chi_2}+2 g_1g_2 e^{i(\chi_1+\chi_2)} \right) \; .\end{aligned}$$ The CGF is composed of three different terms, corresponding to a charge transfer of $2e$ either into terminal 1 or terminal 2 (the first two terms in the bracket) or separate charge transfer into terminals 1 and 2. According to the general principles of statistics, sums of CGFs of independent statistical processes are additive. Therefore, the CGF (\[eq:cgf-poisson\]) is a sum of CGFs of independent Poissonian processes. Crosscorrelations are obtained from derivatives with respect to $\chi_1$ and $\chi_2$. Thus, the first two terms in (\[eq:cgf-poisson\]) corresponding to two-particle tunneling either into terminal 1 or 2 do not contribute. It is only the last term which yields crosscorrelations, and those are positive. Poissonian statistics are the statistics of uncorrelated events, which in our case means all tunneling events are independent. Thus, a two-particle tunneling event into one of the normal terminals is not correlated with other tunneling events and does not contribute to crosscorrelations, but only to the autocorrelations. The two-particle tunneling into different terminals, however, is automatically positively crosscorrelated. The crosscorrelations are therefore positive. The total probability distribution $P(N_1,N_2)$ corresponding to (\[eq:cgf-poisson\]) can be found. It vanishes for odd values of $(N_1+N_2)$ and for even values it is $$\label{eq:distribution-poisson} P(N_1,N_2) = \frac{e^{-\frac{\bar N}{2}} \left(\frac{\bar N}{2}\right)^{\frac{N_1+N_2}{2}}}{ \left(\frac{N_1+N_2}{2}\right)!} \binom{N_1+N_2}{N_1} T_1^{N_1}T_2^{N_2}\,.$$ Here we have defined the average number of transfered electrons $\bar N = t_0 G^S V/e$ and the probabilities $T_{1(2)}=g_{1(2)}/(g_1+g_2)$ that one electron leaves the island into terminal $1(2)$. If one would not distinguish electrons in terminals 1 and 2, the charge counting distribution can be obtained from (\[eq:cgf-poisson\]) by setting $\chi_1=\chi_2=\chi$ and performing the integration. This leads to $P^S_{\textrm{tot}}(N)=\exp(-\bar N/2) (\bar N/2)^{N/2}/(N/2)!$, which corresponds to a Poisson distribution of an uncorrelated transfer of electron pairs. The full distribution (\[eq:distribution-poisson\]) is given by $P^S_{\textrm{tot}}(N_1+N_2)$, multiplied with a partitioning factor, which corresponds to the number of ways how $N_1+N_2$ identical electrons can be distributed among the terminals 1 and 2, with respective probabilities $T_1$ and $T_2$. Note, that $T_1+T_2=1$, since the electrons have no other possibility to leave the island. In contrast to that, we obtain in the normal case for $t^N_{1,2}\ll 1$ the probability distribution: $$\label{eq:normal-poisson} P^N(N_1,N_2) = e^{-\bar N_1} \frac{\bar N_1^{N_1}}{N_1!} e^{-\bar N_2}\frac{\bar N_2^{N_2}}{N_2!}\,.$$ Here we have abbreviated the average number transfered into terminal $i$ by $\bar N_i$. Thus, the distribution in the normal case is the product of two Poisson distributions of charge transfers into the two terminals. In the superconducting case such a factorisation is not possible. In conclusion, we have studied the full counting statistics of a three-terminal device with one superconducting and two normal leads. The system is biased such that a supercurrent is passed from the superconductor into the two normal leads, with no net current between the normal leads. Thus, the device acts as a sort of beam splitter. We have calculated the full distribution of transmitted charges using the extended Keldysh-Green’s function method fully accounting for the proximity effect. Our main finding are large positive crosscorrelations of the currents in the two normal terminals in a wide parameter range. These should be easily accessible experimentally. These positive correlations originate from independent Poisson processes of coherent tunneling of charges into the different terminals. These dominate the crosscorrelations, since two-particle tunneling into the same lead does not contribute to the crosscorrelations. We have also calculated the third cumulant which provides additional information on the current statistics not contained in the current noise. We would like to thank G. Burkard, D. Loss, and Yu. V. Nazarov for discussions. During the preparation of this manuscript, we became aware of similar work by P. Samuelsson and M. Büttiker [@samuelsson]. Our work was supported by the Swiss NSF and the NCCR Nanoscience. [99]{} L. S. Levitov, H. W. Lee, and G. B. Lesovik, J. Math. Phys. [37]{}, 4845 (1996). W. Belzig and Yu. V. Nazarov, Phys. Rev. Lett. [87]{}, 197006 (2001). L. S. Levitov and M. Reznikov, cond-mat/0111057. Yu. V. Nazarov and M. Kindermann, cond-mat/0107133. C. W. J. Beenakker and H. Schomerus, Phys. Rev. Lett. [86]{}, 700 (2001). A. Andreev and A. Kamenev, Phys. Rev. Lett. [85]{}, 1294 (2000); Yu. Makhlin and A. D. Mirlin, ibid. [87]{}, 276803 (2001); L. S. Levitov, cond-mat/0103617. Ya. M. Blanter and M. Büttiker, Phys. Rep. [336]{}, 1 (2000). D. Loss and E. V. Sukhorukov, Phys. Rev. Lett. [84]{}, 1035 (2000); G. Burkard, D. Loss, and E. V. Sukhorukov, Phys. Rev. B [61]{}, R16303 (2000). V. A. Khlus, Sov. Phys. JETP [66]{}, 1243 (1987); B. A. Muzykantskii and D. E. Khmelnitskii, Phys. Rev. B [50]{}, 3982 (1994); M. J. M. de Jong and C. W. J. Beenakker, ibid. [49]{}, 16070 (1994); K. E. Nagaev and M. Büttiker, ibid. [63]{}, 081301(R) (2001). X. Jehl et al., Nature [405]{}, 50 (2000). W. Belzig and Yu. V. Nazarov, Phys. Rev. Lett. [87]{}, 067006 (2001). A. A. Kozhevnikov, R. J. Schoelkopf, and D. E. Prober, Phys. Rev. Lett [84]{}, 3398 (2000). P. Recher, E. V. Sukhorukov, and D. Loss, Phys. Rev. B [63]{}, 165314 (2001). G. B. Lesovik, T. Martin, and G. Blatter, Eur. Phys. J. B [24]{}, 287 (2001). M. Henny et al., Science [284]{}, 296 (1999); W. D. Oliver et al., Science [284]{}, 299 (1999); S. Oberholzer et al., Physica (Amsterdam) E [6]{}, 314 (2000). M. Büttiker, Phys. Rev. Lett. [65]{}, 2901 (1990); Physica B [175]{}, 199 (1991); Phys. Rev. B [46]{}, 12485 (1992). T. Martin, Phys. Lett. A [220]{}, 137 (1996); M. P. Anantram and S. Datta, Phys. Rev. B [53]{}, 16390 (1996); J. Torres and T. Martin, Eur. Phys. J. B [12]{}, 319 (1999). T. Gramespacher and M. Büttiker, Phys. Rev. B [61]{}, 8125 (2000). F. Taddei and R. Fazio, Phys. Rev. B [65]{}, 134522 (2002). Yu. V. Nazarov, Phys. Rev. Lett. [73]{}, 134 (1994); Yu. V. Nazarov, Superlattices Microst. [ 25]{}, 1221 (1999). Yu. V. Nazarov, Ann. Phys. (Leipzig) [8]{}, SI-193 (1999). Yu. V. Nazarov and D. Bagrets, Phys. Rev. Lett. [88]{}, 196801 (2002). P. Samuelsson and M. Büttiker, cond-mat/0203188.	{ "pile_set_name": "ArXiv" }
=1 Introduction ============ The story of beta integrals (the integrals of hypergeometric type admitting exact evaluation) starts from Euler’s proof of the following formula [@aar] $$\int_0^1t^{\alpha-1}(1-t)^{\beta-1}{\rm d}t=\frac{\Gamma(\alpha)\Gamma(\beta)} {\Gamma(\alpha+\beta)}, \qquad \operatorname{Re}(\alpha), \operatorname{Re}(\beta)>0, \label{Bfunction}$$ where $\Gamma(x)$ is the Euler gamma function. Note that the Gaussian integral $\int_{-\infty}^\infty {\rm e}^{-x^2}{\rm d}x=\sqrt{\pi}$ emerges in a special degeneration limit of this exact relation. Over the years identity has found many generalizations. In particular, the $q$-hypergeometric line of developments brought to light the following Askey–Wilson $q$-beta integral $$\frac{(q;q)_\infty}{4\pi {\rm i}}\int_{{\mathbb T}} \frac{\big(z^2;q\big)_\infty\big(z^{-2};q\big)_\infty} {\prod\limits_{j=1}^4(t_jz;q)_\infty\big(t_jz^{-1};q\big)_\infty} \frac{{\rm d}z}{z} =\frac{(t_1t_2t_3t_4;q)_\infty}{\prod\limits_{1\le j<k\le 4}(t_jt_k;q)_\infty}, \qquad \|q\|, \|t_j\|<1, \label{AWint}$$ serving as a measure for the most general classical orthogonal polynomials [@aw]. Here $\mathbb{T}$ denotes the unit circle of counterclockwise orientation and $(z;q)_\infty:=\prod\limits_{n=0}^\infty\big(1-zq^n\big)$. Jumping over the Rahman $q$-beta integral [@Rahman] extending , we come to the elliptic beta integral evaluation formula [@spi:umn] – currently the top identity of the type of interest, $$\frac{(p;p)_\infty(q;q)_\infty}{4\pi {\rm i}} \int_\mathbb{T}\frac{\prod\limits_{j=1}^6 \Gamma\big(t_jz^{{\pm 1}};p,q\big)}{\Gamma\big(z^{\pm 2};p,q\big)}\frac{{\rm d}z}{z} =\prod_{1\leq j<k\leq6}\Gamma(t_jt_k;p,q), \label{ellbeta}$$ where $\|p\|, \|q\|, \|t_j\|<1$, $\prod\limits_{j=1}^6t_j=pq$, and $$\Gamma(z;p,q):= \prod_{j,k=0}^\infty\frac{1-z^{-1}p^{j+1}q^{k+1}}{1-zp^{j}q^{k}}$$ is the elliptic gamma function. Here we apply the standard compact notation $$\Gamma\big(tz^{\pm1};p,q\big):=\Gamma(tz;p,q)\Gamma\big(tz^{-1};p,q\big).$$ The form of this identity is somewhat universal – 14 generalized gamma functions in the integral definition and 15 gamma functions in its exact evaluation expression, a pattern that will be seen several times below in other instances. As formally shown in [@stok] (see also [@DS]), in the limit $\|p\|, \|q\| \to 1$ relation reduces to the following hyperbolic beta integral evaluation formula, which is a hyperbolic analogue of the Rahman integral identity [@Rahman], $$\begin{gathered} \int_{-{\rm i}\infty}^{{\rm i}\infty} \frac{\prod\limits_{k=1}^6\gamma^{(2)}(g_k\pm z;\mathbf{\omega}) } {\gamma^{(2)}(\pm 2 z;\mathbf{\omega}) } \frac{{\rm d}z}{2{\rm i}\sqrt{\omega_1\omega_2}} = \prod_{1\leq j<k\leq 6}\gamma^{(2)}(g_j+g_k;\mathbf{\omega}), \label{hyper}\end{gathered}$$ and the following balancing condition holds true $$\label{balcon} \sum_{k=1}^6 g_k=Q,\qquad Q:=\omega_1+\omega_2.$$ Since this identity plays the key role in the following considerations, we shall describe its ingredients in full detail. First, we explain the compact notation $\gamma^{(2)}(g\pm u;\mathbf{\omega}) :=\gamma^{(2)}(g+u;\mathbf{\omega})\gamma^{(2)}(g-u;\mathbf{\omega})$, where $$\gamma^{(2)}(u;\mathbf{\omega})= \gamma^{(2)}(u;\omega_1,\omega_2):={\rm e}^{-\frac{\pi{\rm i}}{2} B_{2,2}(u;\mathbf{\omega}) } \gamma(u;\mathbf{\omega}), \label{HGF}$$ with the second order multiple Bernoulli polynomial $$B_{2,2}(u;\mathbf{\omega})=\frac{1}{\omega_1\omega_2} \left(\left(u-\frac{\omega_1+\omega_2}{2}\right)^2-\frac{\omega_1^2+\omega_2^2}{12}\right),$$ and $$\gamma(u;\mathbf{\omega}):= \frac{\big(\tilde q {\rm e}^{2\pi {\rm i} \frac{u}{\omega_1}};\tilde q\big)_\infty} {\big({\rm e}^{2\pi {\rm i} \frac{u}{\omega_2}};q\big)_\infty} =\exp\left(-\int_{{\mathbb R}+{\rm i}0}\frac{{\rm e}^{ux}} {(1-{\rm e}^{\omega_1 x})(1-{\rm e}^{\omega_2 x})}\frac{{\rm d}x}{x}\right). \label{int_rep}$$ The latter function is known as the Faddeev modular quantum dilogarithm [@Fad95; @Fad94] or the hyperbolic gamma function [@Ruij]. The relation between its integral and product forms is described explicitly in [@KLS], where the inverse of was called the double sine function (see also Appendix A in [@spi:conm] for a description of different notations used in the literature). If $\operatorname{Re}(\omega_1), \operatorname{Re}(\omega_2)>0$, then the integral in converges for $0<\operatorname{Re}(u)< \operatorname{Re}(\omega_1+\omega_2)$. For $\operatorname{Re}(\omega_1), \operatorname{Re}(\omega_2)<0$ it is well defined in the strip $\operatorname{Re}(\omega_1+\omega_2)<\operatorname{Re}(u)< 0$. For $\operatorname{Re}(\omega_2)\leq 0$ and $\operatorname{Re}(\omega_1)>0$ convergence takes place for $\operatorname{Re}(\omega_2)<\operatorname{Re}(u)< \operatorname{Re}(\omega_1)$ (by symmetry this is also true for $\operatorname{Re}(\omega_1)\leq 0$ and $\operatorname{Re}(\omega_2)>0$, if $\operatorname{Re}(\omega_1)<\operatorname{Re}(u)< \operatorname{Re}(\omega_2)$). The infinite product representation is well defined and allows analytical continuation in $u$ to the whole complex plane, provided $\|q\|<1$, where $$q= {\rm e}^{2\pi {\rm i}\frac{\omega_1}{\omega_2}},\qquad \tilde q= {\rm e}^{-2\pi {\rm i}\frac{\omega_2}{\omega_1}}, \qquad \text{for} \quad \operatorname{Im}(\omega_1/\omega_2)>0, \label{q}$$ or $q= {\rm e}^{2\pi {\rm i}\frac{\omega_2}{\omega_1}}$, $\tilde q= {\rm e}^{-2\pi {\rm i}\frac{\omega_1}{\omega_2}}$, if $\operatorname{Im}(\omega_2/\omega_1)>0$. Note that the integral representation in is manifestly symmetric in $\omega_1$ and $\omega_2$ and, moreover, it shows that this function still remains analytical for $\omega_1/\omega_2\in{\mathbb R}/\{0\}$ (i.e., when $\|q\|=1$) in appropriate domains of $u$. In the following we stick to the parametrisation . Now it is necessary to explain admissible choices of the integration contour in . It is not difficult to see that true poles of function $\gamma^{(2)}(u;\mathbf{\omega})$ are located at the following points $$u_p\in \{ -n\omega_1 -m\omega_2\},\qquad n,m \in {\mathbb Z}_{\geq0}.$$ Therefore poles of the integrand function in form two separate arrays going to infinity in different directions $$z_{\rm poles}\in \{ g_k+n\omega_1+ m\omega_2\}\cup \{ -g_k-n\omega_1- m\omega_2,\},\qquad n,m \in {\mathbb Z}_{\geq0}, \qquad k=1,\ldots, 6.$$ The contour of integration in should separate these two sets of points. It remains to explain conditions of the convergence of the integral in . For that one should use the following asymptotic formulas [@KLS]: $$\begin{gathered} \label{asy1} A\colon \quad \lim_{z\to \infty}{\rm e}^{{\pi{\rm i}\over 2}B_{2,2}(z;\omega_1,\omega_2)}\gamma^{(2)}(z;\omega_1,\omega_2)=1, \qquad {\rm for}\quad \operatorname{arg}\omega_1<\operatorname{arg} z<\operatorname{arg}\omega_2+\pi, \\ \label{asy2} B\colon \quad \lim_{z\to \infty}{\rm e}^{-{\pi{\rm i}\over 2}B_{2,2}(z;\omega_1,\omega_2)}\gamma^{(2)}(z;\omega_1,\omega_2)=1, \qquad {\rm for}\quad \operatorname{arg}\omega_1-\pi<\operatorname{arg} z<\operatorname{arg}\omega_2.\end{gathered}$$ Applying these formulas to the integrand in when the integration variable goes to infinity within the indicated cones, one finds the asymptotics $$A \to {\rm e}^{6\pi{\rm i}\frac{z(\omega_1+\omega_2)}{\omega_1\omega_2}}, \qquad B \to {\rm e}^{-6\pi{\rm i}\frac{z(\omega_1+\omega_2)}{\omega_1\omega_2}}. \label{AB}$$ So, the contour of integration should be chosen in such a way that both these factors are vanishing exponentially fast. It is standard to assume that $\operatorname{Re}(\omega_1), \operatorname{Re}(\omega_2)>0$ in which case it is sufficient to take $z\to +{\rm i}\infty$ in the region $A$ and $z\to -{\rm i}\infty$ in the region $B$. Assuming that $\operatorname{Re}(g_k )> 0$, the contour of integration can be taken as the imaginary axis. After rotating the integration contour by passing to the integration variable $x= z/\sqrt{\omega_1\omega_2}$ the integral converges for $x\to \pm {\rm i}\infty$, if $\operatorname{Re}(\sqrt{\omega_1/\omega_2})>0$, and for $\operatorname{Re}(g_k/\sqrt{\omega_1\omega_2})>0$ the imaginary axis of $x$ can be taken as the integration contour. For completeness we indicate also the way how formula is reduced to . Namely, one should parametrise [@stok] $$t_j={\rm e}^{-2\pi v g_j}, \qquad z={\rm e}^{-2\pi v u},\qquad p={\rm e}^{-2\pi v\omega_1}, \qquad q={\rm e}^{-2\pi v\omega_2}$$ and take the limit $v\to 0^+$ using the limiting relation $$\label{parlim2} \Gamma\big({\rm e}^{-2\pi vu};{\rm e}^{-2\pi v\omega_1},{\rm e}^{-2\pi v\omega_2}\big) \underset{v\to 0^+}{=} {\rm e}^{-\pi\frac{2u-\omega_1-\omega_2}{12v\omega_1\omega_2}}\gamma^{(2)}(u;\omega_1,\omega_2).$$ As shown in [@rai:limits] this transition from the elliptic gamma function to the hyperbolic one is uniform on compacta. Therefore the degeneration procedure from to is actually fully legitimate. We squeezed the history of beta integrals to a few examples and wish to state that it is far from complete, i.e., its ending is not seen yet, even at the univariate integrals level. To justify this claim we shall present two more beta integrals which extend the picture in the directions not expected even after discovery of the elliptic beta integral . The first new case. Recently we have extended identity to a beta integral associated with the general lens space [@SS23; @SS24]. Corresponding formula has the following form $$\begin{gathered} \sum_{m\in {\mathbb Z}_c+\nu}\int_{-{\rm i}\infty}^{{\rm i}\infty} \frac{\prod\limits_{j=1}^6\Gamma_M(g_j\pm z,n_j\pm m)} {\Gamma_M(\pm 2z,\pm 2m)} \frac{{\rm d}z}{2{\rm i}c\sqrt{\omega_1\omega_2}} = \prod_{1\leq \ell<j\leq 6} \Gamma_M(g_\ell+g_j,n_\ell+n_j), \label{integral}\end{gathered}$$ where ${\mathbb Z}_c = \{0, 1,\ldots,c -1\}$, $n_j\in {\mathbb Z} +\nu$, $\nu = 0,\frac{1}{2}$. The continuous variables $\omega_{1}$, $\omega_{2}$, $g_j\in {\mathbb C}$, $\operatorname{Re}(\omega_{1}), \operatorname{Re}(\omega_{2}),\operatorname{Re}(g_j) > 0$, and discrete ones $n_j$ satisfy the balancing condition $$\sum_{j=1}^6g_j = \omega_1 + \omega_2,\qquad \sum_{j=1}^6 n_j=-d-1. $$ Here $\Gamma_M(\mu\pm z, n\pm m):= \Gamma_M(\mu+z,n+m)\Gamma(\mu-z,n-m)$ and the rarefied hyperbolic gamma function $\Gamma_M(\mu,m)$ has the form $$\label{gammakp} \Gamma_M(\mu,m):=Z(m){\rm e}^{-\frac{\pi {\rm i}}{2c}B_{2,2}(\mu;\omega_1,\omega_2)}\gamma_M(\mu,m).$$ The $\gamma_M(\mu,m)$-function was introduced by Dimofte [@Dimofte] as the modular quantum dilogarithm associated with the general lens space $$\gamma_M(\mu,m)=\gamma_M(\mu,m;\omega_1,\omega_2):= \frac{\big(\tilde q {\rm e}^{2\pi {\rm i} \tilde u(\mu, m)};\tilde q\big)_\infty} {\big({\rm e}^{2\pi {\rm i} u(\mu,m)};q\big)_\infty}, \qquad \|q\|<1,$$ where $q:={\rm e}^{2\pi {\rm i} \tau}$, $\tilde q:={\rm e}^{2\pi {\rm i}\tilde \tau}$, $$\tau:=\frac{\omega_1-d\omega_2}{c\omega_2},\qquad \tilde \tau:=\frac{a\tau+b}{c\tau+d}=\frac{a\omega_1-\omega_2}{c\omega_1}, \qquad M=\left( \begin{matrix} a & b \\ c & d \end{matrix} \right)\in {\rm SL}(2,{\mathbb Z}),$$ and $$u(\mu,m):=\frac{\mu+m\omega_2}{c\omega_2}, \qquad \tilde u(\mu,m):=\frac{\mu+am\omega_1}{c\omega_1} =m\tilde \tau+\frac{u(\mu,m)}{c\tau+d}.$$ The normalizing factors of were suggested in [@SS23; @SS24]. In particular, $Z(m)$ was chosen in the form $$Z(m)=\frac{{\rm e}^{-\frac{\pi {\rm i}}{4} (1-\frac{a+d+3}{3c})}}{\varepsilon(a,b,c,d)} {\rm e}^{\pi {\rm i} \frac{(1+b)c+a}{2c}m(m+d+1)}, $$ where $\varepsilon(a,b,c,d)$ is a 24-th root of unity emerging in the general modular transformation law for the Dedekind $\eta$-function $$\eta\left(\frac{a\tau+b}{c\tau+d}\right)=\varepsilon(a,b,c,d)\sqrt{-{\rm i}(c\tau+d)}\eta(\tau),\qquad \eta(\tau)={\rm e}^{\frac{\pi {\rm i}\tau}{12}}\big({\rm e}^{2\pi {\rm i}\tau};{\rm e}^{2\pi {\rm i}\tau}\big)_\infty. \label{eta}$$ As promised, relation contains 14 generalized gamma functions on the left-hand side and 15 such gamma functions on the right-hand side. As the second new formula, which is the main goal of this work, we describe a special degeneration limit $\omega_1+\omega_2\to 0$ of the ordinary hyperbolic beta integral . In a further step, we describe also similar reduction of symmetry transformations of the most general hyperbolic analogue of the Euler–Gauss hypergeometric function. Gamma function over the complex numbers ======================================= Let us take $\alpha, \alpha'\in\mathbb{C}$ such that $\alpha-\alpha'=n_\alpha \in\mathbb{Z}$ and for $z\in\mathbb{C}$ denote $$[z]^\alpha:= z^\alpha \bar z^{\alpha'}=\|z\|^{2\alpha'} z^{n_\alpha},\qquad \int_{\mathbb{C}} {\rm d}^2z:=\int_{\mathbb{R}^2}{\rm d}(\operatorname{Re} z)\, {\rm d}(\operatorname{Im} z),$$ where $\bar z$ is a complex conjugate of $z$. Then one has the following complex beta integral evaluation formula [@GGV] $$\label{CB} \int_{\mathbb{C}}[w-z_1]^{\alpha-1} [z_2-w]^{\beta-1} \frac{{\rm d}^2w}{\pi} =\frac{\Gamma(\alpha)\Gamma(\beta)}{\Gamma(\alpha+\beta)} \frac{\Gamma(1-\alpha'-\beta')}{\Gamma(1-\alpha')\Gamma(1-\beta')}[z_2-z_1]^{\alpha+\beta-1},$$ which is a clear analogue of relation . This formula suggests the definition of a gamma function over the field of complex numbers as a particular ratio of Euler’s gamma functions $${\bf \Gamma}(x,n)={\bf \Gamma}(\alpha\|\alpha'):=\frac{\Gamma(\alpha)}{\Gamma(1-\alpha')} =\frac{\Gamma\big(\frac{n+{\rm i}x}{2}\big)}{\Gamma\big(1+\frac{n-{\rm i}x}{2}\big)}, \qquad \alpha=\frac{n+{\rm i}x}{2},\qquad \alpha'=\frac{-n+{\rm i} x}{2}, \label{Cgamma}$$ where $x\in {\mathbb C}$ and $n\in {\mathbb Z}$. From the reflection relation for the Euler gamma function $\Gamma(x)\Gamma(1-x)=\pi/\sin\pi x$, it follows that $${\bf \Gamma}(\alpha\|\alpha') =(-1)^{\alpha-\alpha'}{\bf \Gamma}(\alpha'\|\alpha), \qquad {\bf \Gamma}(x,-n)=(-1)^n{\bf \Gamma}(x,n), $$ and $${\bf \Gamma}(\alpha\|\alpha'){\bf \Gamma}(1-\alpha\|1-\alpha') =(-1)^{\alpha-\alpha'}, \qquad {\bf \Gamma}(x,n){\bf \Gamma}(-x-2{\rm i},n)=1. \label{reflCgamma}$$ The functional equation takes the form $${\bf \Gamma}(\alpha+1\|\alpha'+1) =-\alpha\alpha' {\bf \Gamma}(\alpha\|\alpha'),\qquad {\bf \Gamma}(x-2{\rm i},n)=\frac{n^2+x^2}{4}{\bf \Gamma}(x,n).$$ Now one can rewrite the right-hand side of in the following forms $$\frac{{\bf\Gamma}(\alpha\|\alpha'){\bf\Gamma}(\beta\|\beta')} {{\bf\Gamma}(\alpha+\beta\|\alpha'+\beta')} [z_2-z_1]^{\alpha+\beta-1}= \frac{{\bf\Gamma}(\alpha,\beta,\gamma)}{[z_1-z_2]^{\gamma}}, \qquad {\bf\Gamma}(\alpha_1,\ldots,\alpha_k):=\prod_{j=1}^k{\bf\Gamma}(\alpha_j\|\alpha_j'),$$ where $\alpha+\beta+\gamma= \alpha'+\beta'+\gamma'=1$. After making the inversion transformations $w\to w^{-1}$, $z_1\to z_1^{-1}$, $z_2 \to z_2^{-1}$ and the shifts $w\to w-z_3$, $z_1\to z_1-z_3$, $z_2\to z_2-z_3$, relation takes the form of a star-triangle relation: $$\begin{gathered} \int_{\mathbb{C}}[z_1-w]^{\alpha-1} [z_2-w]^{\beta-1} [z_3-w]^{\gamma-1}\frac{{\rm d}^2w}{\pi} \nonumber\\ \qquad{} =\frac{ {\bf\Gamma}(\alpha,\beta,\gamma) } {[z_3-z_2]^{\alpha}[z_1-z_3]^{\beta}[z_2-z_1]^{\gamma}}, \qquad \alpha+\beta+\gamma=1. \label{STR}\end{gathered}$$ Multidimensional analogues of complex integrals and were considered by Dotsenko and Fateev within the context of $2d$ conformal field theory [@DF]. An independent study of the complex Selberg integral was performed in [@Aomoto]. Such integrals naturally emerge also in the theory of non-compact ${\rm SL}(2, {\mathbb C})$ spin chains [@DMV2017; @DMV2018]. The well known trigonometric $q$-gamma function [@aar] is defined as $$\Gamma_q(x):=\frac{(q;q)_\infty}{(q^x;q)_\infty}(1-q)^{1-x},\qquad \|q\|<1,\qquad x\in\mathbb{C}. $$ For fixed $x$, in the limit $q\to 1^-$ one obtains the Euler gamma function $$\label{gamlim} \lim_{q\to 1^-} \Gamma_q(x)= \Gamma(x).$$ As shown in [@Tom; @rai:limits], this convergence to gamma function is uniform on compacta, which allows degeneration of the $q$-beta integral with a compact measure support to the de Branges–Wilson integral [@aar] with an infinite Mellin–Barnes type integration contour. We would like to consider now a similar limit for the hyperbolic gamma function $$\gamma(u;\omega_1,\omega_2)={\big({\rm e}^{2\pi {\rm i}\frac{u}{\omega_1}}{\rm e}^{-2\pi {\rm i} {\omega_2\over \omega_1}};{\rm e}^{-2\pi {\rm i} {\omega_2\over \omega_1}}\big)_{\infty}\over \big({\rm e}^{2\pi {\rm i} \frac{u}{\omega_2}};{\rm e}^{2\pi {\rm i} {\omega_1\over \omega_2}}\big)_{\infty}}. \label{hypergamma}$$ In the context of $2d$ quantum Liouville theory it is customary to use notations [@DF; @ILT; @RW] $$b:=\sqrt{\omega_1\over \omega_2}, \qquad q={\rm e}^{2\pi {\rm i} b^2},\qquad \tilde q={\rm e}^{-2\pi {\rm i} b^{-2}}.$$ The central charge $c$ of this theory, the key characteristic of $2d$ conformal field theory [@DF], has the form $c=1+6\big(b+b^{-1}\big)^2$, i.e., $b$ is the variable parametrising $c$. Let us consider the cases when simultaneously $q\to 1$ and $\tilde q\to 1$, so that in there starts to emerge a ratio of the Euler gamma functions. Clearly this is possible, if $b^2\to n$, $b^{-2}\to m$, $n, m\in {\mathbb Z}$. Evidently the only admissible choices are $b=\pm{\rm i}$, when $c=1$ (this case can be considered as a $p\to\infty$ limit of the minimal models [@RW]), and $b=\pm1$, when $c=25$. Consider the first of these possibilities (the case $b=\pm 1$ will be considered in Section \[newlimit\]). Namely, let us take small $\delta>0$ and set $$\label{om1om22} b=\sqrt{\omega_1\over \omega_2}={\rm i}+\delta, \qquad \delta\to 0^+.$$ Obviously one has now $$\label{om1om2} \sqrt{\omega_2\over \omega_1}=-{\rm i}+\delta+O\big(\delta^2\big),\qquad {\omega_1\over \omega_2}=-1+2{\rm i}\delta+\delta^2,\qquad {\omega_2\over \omega_1}=-1-2{\rm i}\delta+O\big(\delta^2\big),$$ as well as $Q=\omega_1+\omega_2=2\delta\sqrt{\omega_1\omega_2}+O\big(\delta^2\big).$ In addition to this choice, we parametrise the argument $u$ in as follows $$u={\rm i}\sqrt{\omega_1\omega_2}(n+x\delta), \qquad n\in {\mathbb Z}, \qquad x\in {\mathbb C},$$ and consider the limit $\delta\to 0^+$. Let us investigate behavior of each of the infinite products in . In the denominator we have $$\big({\rm e}^{2\pi {\rm i} \frac{u}{\omega_2}};{\rm e}^{2\pi {\rm i} {\omega_1\over \omega_2}}\big)_{\infty} =\big({\rm e}^{-2\pi\delta(n+{\rm i}x+\delta x)};q\big)_\infty =\frac{(q;q)_\infty(1-q)^{1-\frac{n+{\rm i}x}{2}+O(\log q)}} {\Gamma_q\big(\frac{n+{\rm i}x}{2}+O(\log q)\big)},$$ where $q={\rm e}^{-4\pi\delta(1-{\rm i}\delta/2)}$. Analogously, for the numerator we obtain $$\big({\rm e}^{2\pi {\rm i}\frac{u}{\omega_1}}{\rm e}^{-2\pi {\rm i} {\omega_2\over \omega_1}};{\rm e}^{-2\pi {\rm i} {\omega_2\over \omega_1}}\big)_{\infty} = \big(\tilde q^{1+\frac{n-{\rm i}x}{2}+O(\log \tilde q)};\tilde q\big)_\infty =\frac{(\tilde q;\tilde q)_\infty(1-\tilde q)^{\frac{-n+{\rm i}x}{2}+O(\log \tilde q)}} {\Gamma_{\tilde q}\big(1+\frac{n-{\rm i}x}{2}+O(\log \tilde q)\big)},$$ where $$\tilde q={\rm e}^{-4\pi\delta\frac{1-{\rm i}\delta/2}{(1-{\rm i}\delta)^2}}.$$ As a result we obtain, $$\gamma(u;\omega_1,\omega_2)=\frac{\Gamma_q\big( \frac{n+{\rm i}x}{2}+O(\log q) \big)} {\Gamma_{\tilde q}\left( 1+\frac{n-{\rm i}x}{2}+O(\log \tilde q)\right)} \frac{(\tilde q;\tilde q)_\infty}{(q;q)_\infty} \frac{(1-\tilde q)^{\frac{-n+{\rm i}x}{2}+O(\log \tilde q)}} {(1-q)^{1-\frac{n+{\rm i}x}{2}+O(\log q)}}.$$ Now we apply a slightly stronger limiting relation than , namely for $q\to 1$ we take $\Gamma_q(x+O(\log q))\to \Gamma(x)$. As follows from the considerations of [@Tom] (see there Appendix B) and [@rai:limits] this limit is uniform over the compact domains excluding poles similar to for real $q\to 1^-$ and complex $x$ away from the poles. The arguments of [@rai:limits] show that this property is preserved even for complex $q$, provided it approaches 1 with the angle away from $\pm \pi/2$, which is satisfied in our case, since this angle is proportional to $\delta$. This uniformity will be very useful for consideration of such a limit for hyperbolic integrals. Using the modular transformation rule $\eta(-1/\tau)=\sqrt{-{\rm i}\tau} \eta(\tau)$ for the Dedekind eta-function , we find $$\frac{(\tilde q;\tilde q)_\infty}{(q;q)_\infty} ={\rm e}^{{\pi{\rm i} \over 12}\big({\omega_2\over \omega_1}+{\omega_1\over \omega_2}\big)} \left(-{\rm i} {\omega_1\over \omega_2}\right)^{1\over 2} \underset{\delta\to 0^+}{=}{\rm e}^{ \pi{\rm i} \over 12}. \label{ratio}$$ Note that for $\delta\to 0^+$ one has $\tilde q= q+O\big(\delta^2\big)$, but formal substitution of this relation to $(\tilde q;\tilde q)_\infty$ in and termwise cancellation of the individual multipliers in the ratio of infinite products of interest yields 1, instead of the nontrivial phase factor ${\rm e}^{ \pi{\rm i} \over 12}$. Finally, we come to the leading asymptotics $$\gamma(u;\omega_1,\omega_2)\underset{\delta\to 0^+}{=} {\rm e}^{\pi {\rm i}\over 12} (4\pi\delta)^{{\rm i}x-1}{\bf \Gamma}(x,n),$$ where ${\bf \Gamma}(x,n)$ is the complex gamma function defined in . This result was presented first in [@BMS] without any derivation details and rigorous justifications. Recalling definition (\[HGF\]), we obtain $$\label{gam2lim2} \gamma^{(2)}({\rm i}\sqrt{\omega_1\omega_2}(n+x\delta);\omega_1,\omega_2)\underset{\delta\to 0^+}{=} {\rm e}^{\frac{\pi {\rm i}}{2}n^2} (4\pi\delta)^{{\rm i}x-1}{\bf \Gamma}(x,n), \qquad \sqrt{\omega_1\over \omega_2}={\rm i}+\delta,$$ where $n\in {\mathbb Z}$, $x\in {\mathbb C}$. Thus in this limit the function $\gamma^{(2)}(u;\mathbf{\omega})$ starts to blow up around a special discrete set of points of the argument $u$ passing through the whole complex plane along a particular line. Note that the choice $b= -{\rm i}$ is equivalent to : the ansatz $b:=-{\rm i}+\delta$, $\delta<0$, yields for $\delta\to 0^-$ the same limit as in and leads to results identical to the ones described below for . General complex beta integral ($\boldsymbol{\omega_1=-\omega_2}$) {#genbet} ================================================================= We are going to apply the limit considered in the previous section to an integral of the form $$\label{intdel} \int_{-{\rm i}\infty}^{{\rm i}\infty}\Delta(z){{\rm d}z\over {\rm i}\sqrt{\omega_1\omega_2}} =\int_{-{\rm i}\infty}^{{\rm i}\infty}\Delta(\sqrt{\omega_1\omega_2}x){{\rm d}x\over {\rm i}}, \qquad x=\frac{z}{\sqrt{\omega_1\omega_2}},$$ where $\Delta(z)$ is a product of $\gamma^{(2)}(u;\omega_1,\omega_2)$ functions. Here we assume that these integrals converge when the integration contour is taken as the imaginary axis for both integration variables $z$ and $x$, i.e., the integration contour can be rotated by the angle $\arg \sqrt{\omega_1\omega_2}$. The function $\gamma^{(2)}(u;\omega)$ is uniform, i.e., we can scale all variables and it does not change the form of this function, $$\gamma^{(2)}(\lambda u;\lambda\omega_1,\lambda\omega_2)=\gamma^{(2)}(u;\omega_1,\omega_2), \qquad \lambda\neq 0,$$ Therefore we can fix the product $\omega_1\omega_2$ to be any nonzero number. E.g., in the quantum Liouville theory the standard normalization is $\omega_1\omega_2=1$, or $\omega_1=b$, $\omega_2=b^{-1}$ [@ILT; @RW; @DS], which corresponds to the choice $\lambda=1/\sqrt{\omega_1\omega_2}$. First, we rewrite this integral as an infinite sum $$\begin{gathered} \int_{-{\rm i}\infty}^{{\rm i}\infty}\Delta\big(\sqrt{\omega_1\omega_2} x\big){{\rm d}x\over {\rm i}}= \sum_{N\in {\mathbb Z}} \int_{{\rm i}(N-1/2)}^{{\rm i}(N+1/2)}\Delta\big(\sqrt{\omega_1\omega_2} x\big){{\rm d}x\over {\rm i}}\\ \hphantom{\int_{-{\rm i}\infty}^{{\rm i}\infty}\Delta\big(\sqrt{\omega_1\omega_2} x\big){{\rm d}x\over {\rm i}}}{} = \sum_{N\in {\mathbb Z}} \int_{N-1/2}^{N+1/2}\Delta\big( {\rm i}\sqrt{\omega_1\omega_2} x\big){\rm d}x =\sum_{N\in {\mathbb Z}} \int_{-1/2}^{1/2} \Delta\big({\rm i}\sqrt{\omega_1\omega_2}(N+x)\big){\rm d}x. \end{gathered}$$ In the last two steps we changed the variable $x\to {\rm i} x$ with the subsequent shift $x\to x+N$. Now we parametrise $x=y\delta$, $\delta>0,$ and take the limit $\delta\to 0^+$. Then we have $$\lim_{\delta\to 0}\sum_{N\in {\mathbb Z}} \int_{-1/2}^{1/2} \Delta\big({\rm i}\sqrt{\omega_1\omega_2}(N+x)\big){\rm d}x= \lim_{\delta\to 0} \sum_{N\in {\mathbb Z}} \int_{-1/2\delta}^{1/2\delta} \delta \Delta\big({\rm i}\sqrt{\omega_1\omega_2}(N+y\delta)\big){\rm d}y.$$ The sum over $N$ is infinite and for $\delta\to 0^+$ the integration contour becomes the noncompact real axis $(-\infty,\infty)$. Therefore, in order to interchange the $\lim\limits_{\delta\to 0}$ sign with the summation and integration, we need the uniform convergence of the limit , which is true and thus justifies our formal manipulations. Finally, we obtain $$\sum_{N\in {\mathbb Z}} \int_{-\infty}^{\infty} \Big[ \lim_{\delta\to 0}\delta\Delta\big({\rm i}\sqrt{\omega_1\omega_2}(N+y\delta)\big)\Big] {\rm d}y. \label{integral_result}$$ So, the behaviour of the initial integral (\[intdel\]) for $\delta\to 0$ is determined by the asymptotics of integrands in , provided they are well defined. Now we apply this reasoning to the beta integral . Besides taking the integration variable in the above mentioned form $$\label{zkom} z= {\rm i}\sqrt{\omega_1\omega_2}(N+\delta y), \qquad y\in {\mathbb C},\qquad N\in {\mathbb Z}+\nu,\qquad \nu=0, \frac{1}{2},$$ where $\delta$ is taken to $0^+$, we scale also the parameters $g_k$ according to the same rule, $$\label{gkom} g_k={\rm i}\sqrt{\omega_1\omega_2}(N_k+\delta a_k),\qquad a_k\in {\mathbb C}, \qquad N_k\in {\mathbb Z}+\nu,\qquad \nu=0, \frac{1}{2},$$ where $a_k$ and $N_k$ satisfy the constraints $$\sum_{k=1}^6 a_k=-2{\rm i},\qquad \sum_{k=1}^6 N_k=0,$$ following from the balancing condition . Note the appearance of a new discrete parameter $\nu=0, \frac{1}{2}$ in formulas and . It emerges from the fact that only the sums $N+N_k$ or the differences $N-N_k$ should be integers in the arguments of the hyperbolic gamma functions, as required in the limit . Considerations given in the derivation of formula remain valid after the replacement of $N$ by $N+\nu$, $\nu=0, \frac{1}{2}$, i.e., we can replace summations over $N\in {\mathbb Z}$ by the sums over $N\in {\mathbb Z}+\nu$. In terms of the integration variable $x=z/\sqrt{\omega_1\omega_2}$ the original integrand asymptotics for $x\to \pm {\rm i}\infty$ takes the form ${\rm e}^{- 12\pi \delta \|x\|}$, i.e., for finite $\delta$ the integral does converge. However, for $\delta\to 0^+$ is starts to diverge and we need to estimate the rate of this divergence. Inserting parametrisations (\[zkom\]) and (\[gkom\]) in (\[hyper\]), and recalling the asymptotics (\[gam2lim2\]), we find the limiting relations $$\begin{gathered} \prod_{k=1}^6\gamma^{(2)}(g_k\pm z;\mathbf{\omega})\to \frac{(-1)^{2\nu}}{(4\pi\delta)^{8}} \prod_{k=1}^6{\bf \Gamma}(a_k+y,N_k+N){\bf \Gamma}(a_k-y,N_k-N),\nonumber\\ \prod_{1\leq j < k\leq 6} \gamma^{(2)}(g_j+g_k;\mathbf{\omega})\to \frac{(-1)^{2\nu}}{(4\pi\delta)^{5}}\prod_{1\leq j < k \leq 6} {\bf \Gamma}( a_j+a_k,N_j+N_k),\nonumber\\ \gamma^{(2)}(\pm 2z;\mathbf{\omega}) \to \frac{(-1)^{2\nu}}{(4\pi\delta)^{2}} {\Gamma(N+ {\rm i}y)\over \Gamma(1+N- {\rm i}y)} {\Gamma(-N- {\rm i}y)\over \Gamma(1-N+ {\rm i}y)}={(4\pi\delta)^{-2}\over y^2+N^2}. \label{extrasign}\end{gathered}$$ Collecting all the multipliers and cancelling the diverging factor $(4\pi\delta)^{-5}$ on both sides of the equality , we obtain our key complex beta integral: $$\frac{1}{8\pi}\sum_{N\in {\mathbb Z}+\nu}\int_{-\infty}^{\infty}\big(y^2+N^2\big) \prod_{k=1}^6{\bf \Gamma}(a_k\pm y,N_k\pm N)dy=\prod_{1\leq j< k \leq 6}{\bf \Gamma}( a_j+a_k,N_j+N_k), \label{keybeta}$$ where $\sum\limits_{k=1}^6 a_k=-2{\rm i}$, $\sum\limits_{k=1}^6 N_k=0$, and $${\bf \Gamma}(x_1\pm x_2 ,n_1\pm n_2):={\bf \Gamma}(x_1+ x_2 ,n_1+ n_2){\bf \Gamma}(x_1- x_2 ,n_1- n_2) .$$ Here we have the variables $N_k$, $N \in {\mathbb Z}+\nu$, $\nu=0, \frac{1}{2}$, so that their sums or differences take integer values. Again, this formula contains 14 generalized gamma functions on the left-hand side and 15 of them on the right-hand side. It can be considered as a complex analogue of the plain hypergeometric Rahman beta integral [@Rahman], since it also contains five free continuous (complex) parameters (five discrete parameters are associated with them in accordance to the ${\rm SL}(2,{\mathbb C})$ principal series representations parametrization). Formal poles of the integrands in are located at the points $$y^{(1)}_p\in\{{\rm i}(N+N_k)-a_k+2{\rm i}\ell_1\}, \qquad y^{(2)}_p\in \{-{\rm i}(N_k-N)+a_k-2{\rm i}\ell_2\}, \qquad \ell_1, \ell_2\in {\mathbb Z}_{\geq 0},$$ and the corresponding formal zeros are $$\begin{gathered} y^{(1)}_z\in \{-2{\rm i}-{\rm i}(N_k+N)-a_k-2{\rm i}\ell_3\},\\ y^{(2)}_z \in \{2{\rm i}+{\rm i}(N_k-N)+a_k+2{\rm i}\ell_4\}, \qquad \ell_3, \ell_4\in {\mathbb Z}_{\geq 0}.\end{gathered}$$ The sets $y^{(1)}_p$ and $y^{(2)}_z$ (or $y^{(2)}_p$ and $y^{(1)}_z$) may overlap only if simultaneously $\operatorname{Re}(a_k)=0$ and $\operatorname{Im}(a_k)\in {\mathbb Z}$. Let us demand that $\operatorname{Im}(a_k)\notin {\mathbb Z}/\{0\}$ and discuss the special case of real $a_k$ separately. Analyzing the overlap of $y^{(1)}_p$ with $y^{(1)}_z$ jointly with the overlap of $y^{(2)}_p$ with $y^{(2)}_z$, leading to cancellations of poles and zeros, we come to the conclusion that for any $N$ true poles on the integrand are located at $$y_{\rm poles} \in \{ {\rm i}\|N+N_k\|-a_k+2{\rm i}\ell_1, -{\rm i}\|N-N_k\|+a_k-2{\rm i}\ell_2\}, \qquad \ell_1, \ell_2\in {\mathbb Z}_{\geq 0}.$$ Therefore for the choice $\operatorname{Im}(a_k)<0$ (which follows from the conditions $\operatorname{Re}\big(g_k/\sqrt{\omega_1\omega_2}\big)>0$ and $\delta\to 0^+$) the real axis separates sequences of poles going to infinity upwards from the ones falling down and the derived formula is true under these conditions. As to the case $a_k\in {\mathbb R}$, we can perform analytical continuation. Namely, we deform the contour of integration slightly below the real axis in such a way that for $\operatorname{Im}(a_k)=0$ no poles emerge on the integration contour and the formula remains true in this case as well. For $\nu=0$ the functions of the type standing in the left-hand side of appeared for the first time in Naimark’s investigation of the representation theory of the Lorentz group ${\rm SL}(2,{\mathbb C})$ [@Naimark]. In particular, for $a_k\in {\mathbb R}$ one deals with the unitary principal series representation of this group. In the modern time, continuation of the investigation of such functions has been launched by Ismagilov [@Ismag2], who constructed $6j$-symbols for the ${\rm SL}(2, {\mathbb C})$ group (for a verification of his result reached via a different approach, see [@DS2017]). As shown in [@DMV2017] (see there Appendix B), a particular subcase of relation (see details below) corresponds to the Mellin–Barnes representation of the complex beta integral , which is equivalent to identity initially considered in [@GGV]. The first understanding that such mathematical structures emerge as a special limit of $q$-hypergeometric functions defined with the help of Faddeev’s modular quantum dilogarithm, or the hyperbolic gamma function, was reached in [@BMS]. Thus, joint efforts of the works [@Ismag2] and [@BMS; @DMV2017] have shown that the representation theory of Faddeev’s modular double [@fad:mod] comprises the representation theory of the ${\rm SL}(2,{\mathbb C})$ group. A rigorous consideration of the Hilbert space aspects of this class of special functions of hypergeometric type is given in [@MN; @Neretin2019]. Another recent related study can be found in [@Mimachi]. The class of functions emerging for $\nu=1/2$ is a new one and its group-theoretical understanding is still missing. For the first time existence of such a nontrivial additional discrete parameter was noticed in [@spi:rare] in the investigation of elliptic hypergeometric functions related to the lens space (such functions were considered also in [@KY]), where the choice $\nu=1/2$ resulted in the discovery of a novel family of trigonometric $q$-hypergeometric integrals. A similar situation holds true for the rarefied hyperbolic functions described above [@SS23; @SS24]. Existence of the discrete variable $\nu=1/2$ in the Mellin–Barnes type representation of complex hypergeometric integrals was noticed first in [@DMV2018]. As shown in [@spi:conm], the original hyperbolic beta integral evaluation formula can be represented in the star-triangle form useful for solving the Yang–Baxter equation [@CS]. Therefore its limiting relation we have derived also can be written in this attractive form which is useful for solvable models in statistical mechanics. A special case of identity corresponding to $\nu=0$ and a reduced number of discrete parameters $N_k$ appeared first in [@Kels2014] exactly in the form of the star-triangle relation.[^1] Beta integrals play a key role in the construction of symmetry transformations for higher order hypergeometric functions of the corresponding type. At the top elliptic level such consequences of identity were considered in [@spi:theta] and in a more general setting in [@spi:rare; @spi:rareYBE]. Let us derive symmetry transformations for the top complex hypergeometric function generalizing the Euler–Gauss $_2F_1$-function by reducing such transformations for an elliptic hypergeometric function. Transformation rule I ===================== Consider the $V$-function, an elliptic analogue of the Euler–Gauss hypergeometric function [@spi:essays], $$V(t_1,\ldots,t_8;p,q)=\frac{(p;p)_\infty(q;q)_\infty}{4\pi {\rm i}}\int_{\mathbb{T}} {\prod\limits_{a=1}^8\Gamma\big(t_az^{\pm1}; p,q\big)\over \Gamma\big(z^{\pm 2}; p,q\big)}{{\rm d}z\over z},$$ where the parameters satisfy constraints $\|t_a\|<1$ and the balancing condition $ \prod\limits_{a=1}^8t_a=p^2q^2$. This function has the $W(E_7)$ Weyl group symmetry transformations, whose key generating relation has been established in [@spi:theta]: $$\begin{gathered} \label{vt1} V(t_1,\ldots,t_8;p,q)=\prod_{1\leq j< k\leq 4}\Gamma(t_jt_k;p,q) \prod_{5\leq j< k\leq 8}\Gamma(t_jt_k;p,q) V(s_1,\ldots,s_8;p,q),\end{gathered}$$ where $$s_j=\rho^{-1}t_j,\qquad s_{j+4}=\rho t_{j+4},\qquad j=1,2,3,4, \qquad \rho=\sqrt{t_1t_2t_3t_4\over pq}=\sqrt{pq\over t_5t_6t_7t_8}.$$ Consider the function $I_h(\underline{g})$ defined by the integral $$I_h(\underline{g})=\int_{-{\rm i}\infty}^{{\rm i}\infty}{\prod\limits_{j=1}^8\gamma^{(2)}(g_j\pm z;\omega_1,\omega_2)\over \gamma^{(2)}(\pm 2z;\omega_1,\omega_2)}\frac{{\rm d}z}{2{\rm i}\sqrt{\omega_1\omega_2}},$$ with $g_j$ satisfying the conditions $\operatorname{Re}(g_j)>0$ and $$\label{mu8} \sum_{j=1}^8 g_j=2Q,\qquad Q:=\omega_1+\omega_2.$$ This is the most general hyperbolic analogue of the Euler–Gauss hypergeometric $_2F_1$-function satisfying a second order difference equation. It represents a one-parameter extension of the function built in [@Ruij]. Applying the hyperbolic degeneration limit to the transformation rule , one comes to the following relation [@BRS] $$\label{ide1} I_h(\underline{g})=\prod_{1\leq j< k \leq 4}\gamma^{(2)}(g_j+g_k;\omega_1,\omega_2) \prod_{5\leq j< k \leq 8}\gamma^{(2)}(g_j+g_k;\omega_1,\omega_2) I_h(\underline{\lambda}) ,$$ where $$\lambda_j=g_j+\xi, \qquad \lambda_{j+4}=g_{j+4}-\xi, \qquad j=1,2,3,4,\qquad \xi={1\over 2}\left(\omega_1+\omega_2-\sum_{j=1}^4 g_j\right).$$ As a next step, let us take the parametrisations , and of variables in (\[ide1\]) and consider the limit $\delta\to 0^+$. The balancing condition (\[mu8\]) passes to the following constraints $$\label{balanca} \sum_{k=1}^8 a_k=-4{\rm i},\qquad a_k\in {\mathbb C},\qquad \sum_{k=1}^8 N_k=0, \qquad N_k\in {\mathbb Z}+\nu, \qquad \nu=0,\frac{1}{2}.$$ This leads to the parametrisation $${\xi\over {\rm i}\sqrt{\omega_1\omega_2}}=-{L\over 2}+\delta\left(-{\rm i}-{X\over 2}\right),\qquad X:=\sum_{j=1}^4 a_j, \qquad L:=\sum_{j=1}^4 N_j.$$ Using considerations of the previous section, now it is straightforward to see that the asymptotic formula (\[gam2lim2\]) allows the $\delta\to 0^+$ reduction of identity (\[ide1\]) to the following symmetry transformation relation $$\begin{gathered} \nonumber \sum_{N\in {\mathbb Z}+\nu}\int_{-\infty}^{\infty}\left(y^2+N^2\right) \prod_{k=1}^8{\bf \Gamma}(a_k\pm y,N_k\pm N){\rm d}y \\ \nonumber \qquad{} =(-1)^L\prod_{1\leq j< k \leq 4}{\bf \Gamma}(a_j+a_k,N_j+N_k)\prod_{5\leq j< k \leq 8}{\bf \Gamma}( a_j+a_k,N_j+N_k) \\ \nonumber \qquad\quad{} \times \sum_{N\in {\mathbb Z}+\mu}\int_{-\infty}^{\infty}\big(y^2+N^2\big) \prod_{k=1}^4{\bf \Gamma}\big(a_k\pm y- \tfrac{1}{2}X-{\rm i},N_k\pm N-\tfrac{1}{2}L\big) \\ {} \qquad\quad{}\times \prod_{k=5}^8{\bf \Gamma}\big(a_k\pm y+\tfrac{1}{2}X+{\rm i},N_k\pm N+\tfrac{1}{2}L\big){\rm d}y, \label{complextrafoI}\end{gathered}$$ where the balancing condition holds true. Here the simultaneous choice of the integration contours as the real axis is valid under the constraints $\operatorname{Im}(a_k)<0$ for all $k$ together with $\operatorname{Im}\big(a_k-\frac12 X\big)<1$, $k=1,2,3,4,$ and $\operatorname{Im}\big(a_k+\frac12 X\big)<-1$, $k=5,6,7,8$. In this relation we have two discrete parameters $\nu, \mu=0,\frac{1}{2}$. If the integer $L$ is even, then one has $\mu=\nu$. If $L$ is an odd integer, then $\mu\neq \nu$. This is completely similar to the $W(E_7)$ group generating transformation for the rarefied elliptic hypergeometric function derived in [@spi:rare] (see also [@spi:rareYBE]). Analytical continuation of the functions standing in to other domains of parameters can be reached by proper deformations of the integration contours. We note that the symmetry transformation is a general complex analogue of the four term Bailey transformation for non-terminating hypergeometric $_9F_8$-series. Transformation rule II ====================== The second type of identities follows from after a group action composition, $$V(t_1,\ldots,t_8;p,q)=\prod_{j, k=1}^4\Gamma(t_jt_{k+4};p,q) V\Big(\tfrac{T^{1/2}}{t_1},\ldots,\tfrac{T^{1/2}}{t_4},\tfrac{U^{1/2}}{t_5},\ldots,\tfrac{U^{1/2}}{t_8};p,q\Big), \label{vt2}$$ where $T=t_1t_2t_3t_4$ and $U=t_5t_6t_7t_8$. The hyperbolic degeneration limit for integrals described in the previous sections reduces to the following identity for $I_h(\underline{g})$ function $$\begin{gathered} \label{ide2} I_h(\underline{g})=\prod_{j,k=1}^4\gamma^{(2)}(g_j+g_{k+4};\mathbb{\omega}) I_h(G-g_1,\ldots,G-g_4,Q-G-g_5,\ldots,Q-G-g_8),\end{gathered}$$ where $G:={1\over 2}\sum\limits_{j=1}^4 g_j$ and $Q=\omega_1+\omega_2$. Let us apply a further degeneration limit to the complex hypergeometric integrals. For $a_k$ and $N_k$ satisfying the balancing condition (\[balanca\]), we denote $$Y_1=\sum_{j=1}^4 a_j,\qquad L_1=\sum_{j=1}^4 N_j,\qquad Y_2=\sum_{j=5}^8 a_j,\qquad L_2=\sum_{j=5}^8 N_j,$$ so that $Y_1+Y_2=-4{\rm i}$, $L_1+L_2=0$. Now the asymptotic relation and the arguments given in previous sections reduce to $$\begin{gathered} \sum_{N\in {\mathbb Z}+\nu}\int_{-\infty}^{\infty}\big(y^2+N^2\big) \prod_{k=1}^8{\bf \Gamma}(a_k\pm y,N_k\pm N){\rm d}y= (-1)^{L_1}\prod_{j,k=1}^4{\bf \Gamma}(a_j+a_{k+4},N_j+N_{k+4}) \nonumber\\ \qquad{} \times \sum_{N\in {\mathbb Z}+\mu}\int_{-\infty}^{\infty}\big(y^2+N^2\big) \prod_{k=1}^4{\bf \Gamma}\big(\tfrac{1}{2}Y_1-a_k\pm y,\tfrac{1}{2}L_1-N_k\pm N\big)\nonumber \\ \qquad{} {} \times \prod_{k=5}^8{\bf \Gamma}\big(\tfrac{1}{2}Y_2-a_k\pm y, \tfrac{1}{2}L_2-N_k\pm N\big){\rm d}y. \label{cide2}\end{gathered}$$ Here we have $\mu=\nu$ for even integers $L_1$ (and, so, even $L_2$ as well), whereas $\mu\neq \nu$ for odd $L_1$. The contours of integration can be taken as the real axis, provided imaginary parts of the continuous parameters entering arguments of the complex gamma functions in are negative. Transformation rule III ======================= The third form of the symmetry transformation for the $V$-function follows from equating right-hand side expressions in and , $$\begin{gathered} V(t_1,\ldots,t_8;p,q)=\prod_{1\leq j< k\leq 8}\Gamma(t_jt_{k};p,q) V\left(\frac{\sqrt{pq}}{t_1},\ldots,\frac{\sqrt{pq}}{t_8};p,q\right).\end{gathered}$$ The hyperbolic degeneration limit (\[parlim2\]) brings the following relation for the $I_h(\underline{g})$ function $$\label{reflt} I_h(\underline{g})=\prod_{1\leq j< k\leq 8}\gamma^{(2)}(g_j+g_k;\omega_1,\omega_2) I_h\left(\underline{\lambda}\right), \qquad \lambda_j =\frac{\omega_1+\omega_2}{2}-g_j.$$ Now we use the parametrisation , and , and consider the $\delta\to 0^+$ limit . Then we have again the balancing condition (\[balanca\]) and the relation $${\omega_1+\omega_2\over 2}-g_k ={\rm i}\sqrt{\omega_1\omega_2}(-N_k+\delta(-a_k-{\rm i}))+O\big(\delta^2\big),\qquad k=1,\ldots,8.$$ As a result of the same steps as in previous cases, we obtain the formula $$\begin{gathered} \sum_{N\in {\mathbb Z}+\nu}\int_{-\infty}^{\infty}\big(y^2+N^2\big) \prod_{k=1}^8{\bf \Gamma}(a_k\pm y,N_k\pm N){\rm d}y =\prod_{1\leq j< k\leq 8}{\bf \Gamma}( a_j+a_k,N_j+N_k) \nonumber\\ \qquad{} \times \sum_{N\in {\mathbb Z}+\nu}\int_{-\infty}^{\infty}\big(y^2+N^2\big) \prod_{k=1}^8{\bf \Gamma}(-{\rm i}-a_k\pm y,-N_k\pm N){\rm d}y,\end{gathered}$$ where $-1<\operatorname{Im}(a_k)<0$. Limiting case of the beta integral I ==================================== The beta integral (\[hyper\]) can serve as a source of many other calculable integrals with a smaller number of the hyperbolic gamma functions in the kernel. To derive them one should take to infinity some of the parameters $g_k$ in a smart way and use the asymptotic behaviour (\[asy1\]) and (\[asy2\]). Here we will consider a couple of examples. Let us set in (\[hyper\]) $$g_j=f_j+{\rm i}\xi,\qquad g_{j+3}=h_j-{\rm i}\xi,\qquad j=1,2,3,\qquad \sum_{j=1}^3 (f_j+h_j)=Q,$$ and also shift the integration variable $z\to z-{\rm i}\xi$. Now we take the limit $\xi\to -\infty$ using the asymptotics (\[asy1\]) and (\[asy2\]) and obtain $$\label{hyper2} \int_{-{\rm i}\infty}^{{\rm i}\infty}\prod_{j=1}^3\gamma^{(2)}(f_j+z;\mathbf{\omega}) \gamma^{(2)}(h_j-z;\mathbf{\omega}){dz\over {\rm i}\sqrt{\omega_1\omega_2}} =\prod_{j, k =1}^3 \gamma^{(2)}(f_j+h_k;\mathbf{\omega}).$$ Let us apply now the parametrisation and set in (\[hyper2\]) $$f_j={\rm i}\sqrt{\omega_1\omega_2}(N_j+\delta s_j), \qquad h_j={\rm i}\sqrt{\omega_1\omega_2}(M_j+\delta t_j),\qquad j=1,2,3,$$ and $z={\rm i}\sqrt{\omega_1\omega_2}(N+\delta y)$, where $N, N_j, M_j \in {\mathbb Z}+\nu$, $\nu=0,\frac{1}{2}$. As a result, the balancing condition takes the form $$\sum_{j=1}^3( N_j+M_j)=0,\qquad \sum_{j=1}^3( s_j+t_j)=-2{\rm i} .$$ In the limit $\delta\to 0^+$, using the arguments of previous sections and formula (\[gam2lim2\]), one can see that relation (\[hyper2\]) reduces to $$\begin{gathered} \frac{{\rm i}^{2\nu}}{4\pi} \sum_{N\in {\mathbb Z}+\nu} \int_{-{\rm i}\infty}^{{\rm i}\infty}\prod_{j=1}^3{\bf \Gamma}(s_j+y,N+N_j) {\bf \Gamma}(t_j-y,N-M_j){\rm d}y\nonumber\\ \qquad{} =\prod_{j,k=1}^3{\bf \Gamma}(s_j+t_k,N_j+M_k). \label{limtrafoI} \end{gathered}$$ For $\nu=0$ this identity was obtained earlier in [@BMS] for special values of $N_j$, $M_j$ and in [@DMV2017] it was derived for general $N_j$, $M_j$. As shown in the latter paper, for $\nu=0$ this is nothing else than the Mellin–Barnes representation of the general complex star-triangle relation . The case $\nu=1/2$ of formula defines a new complex beta integral. Limiting case of the beta integral II ===================================== Consider now the following limit in the identity (\[hyper\]) $$g_5\to {\rm i}\infty,\qquad g_6=Q-g_5-\sum_{j=1}^4g_j\to -{\rm i}\infty.$$ Using the asymptotics (\[asy1\]) and (\[asy2\]), we come to the hyperbolic analogue of the Askey–Wilson $q$-beta integral established by Ruijsenaars [@Ruij] $$\label{hyper4} \int_{-{\rm i}\infty}^{{\rm i}\infty}{\prod\limits_{k=1}^4\gamma^{(2)}(g_k\pm z;\mathbf{\omega})\over \gamma^{(2)}(\pm 2z;\mathbf{\omega})}{{\rm d}z\over 2{\rm i}\sqrt{\omega_1\omega_2}}= {\prod\limits_{1\leq j < k \leq 4} \gamma^{(2)}(g_j+g_k;\mathbf{\omega})\over \gamma^{(2)}\Big(\sum\limits_{k=1}^ 4 g_k;\mathbf{\omega}\Big)}.$$ Setting $${g_k\over {\rm i}\sqrt{\omega_1\omega_2}}=N_k+\delta a_k,\qquad k=1,2,3,4, \qquad \frac{z}{{\rm i}\sqrt{\omega_1\omega_2}}=N+\delta y,$$ and using the same limit $\delta\to 0^+$ as before (\[gam2lim2\]), one can see that (\[hyper4\]) reduces to $$\begin{gathered} \nonumber {1\over 8\pi}\sum_{N\in {\mathbb Z}+\nu}\int_{-\infty}^{\infty}\big(y^2+N^2\big)\prod_{k=1}^4{\bf \Gamma}(a_k\pm y,N_k\pm N) \\ \qquad{} =(-1)^{2\nu}{\prod\limits_{1\leq j < k \leq 4} {\bf \Gamma}(a_j+a_k,N_j+N_k)\over {\bf \Gamma}\Big(\sum\limits_{k=1}^ 4 a_k,\sum\limits_{k=1}^ 4 N_k\Big)}. \label{hyper44}\end{gathered}$$ This is a general complex analogue of the de Branges–Wilson integral [@aar]. For $\nu=0$ (i.e., for integer values of $N$) and all $N_k=0$ this relation was obtained in [@Neretin2018] (with the contradiction that instead of our $1/8\pi$ factor on the left-hand side there stands $1/4\pi^2$).[^2] Limiting case of the transformation rules I =========================================== Obviously, the same limiting transitions can be performed also for the transformation rules (\[ide1\]), (\[ide2\]), and (\[reflt\]). In this way we will obtain a number of new relations between integrals with a smaller number of hyperbolic gamma functions. Let us apply this procedure to the last rule (\[reflt\]). It is straightforward to do the same with relations (\[ide1\]) and (\[ide2\]), but we skip them for brevity. So, consider the limit $$g_7\to {\rm i}\infty,\qquad g_8=2Q-\sum_{j=1}^6g_j-g_7\to -{\rm i}\infty$$ in (\[reflt\]). As a result we obtain $$\begin{gathered} \nonumber \int_{-{\rm i}\infty}^{{\rm i}\infty}{\prod\limits_{j=1}^6\gamma^{(2)}(g_j\pm z;\mathbf{\omega})\over \gamma^{(2)}(\pm 2z;\mathbf{\omega})}{\rm d}z \\ \qquad{} = {1\over \gamma^{(2)}(G-Q;\mathbf{\omega})} \prod_{1\leq j< k \leq 6}\gamma^{(2)}(g_j+g_k;\mathbf{\omega}) \int_{-{\rm i}\infty}^{{\rm i}\infty}{\prod\limits_{j=1}^6\gamma^{(2)}(\tilde{g_j}\pm z;\mathbf{\omega})\over \gamma^{(2)}(\pm 2z;\mathbf{\omega})}{\rm d}z, \label{gmro} \end{gathered}$$ where $$\tilde{g_j}=\tfrac{1}{2}Q-g_j, \qquad j=1,\ldots,6,\qquad G=\sum_{j=1}^6g_j, \qquad Q=\omega_1+\omega_2.$$ Taking the parametrisation of variables , , and , we also have $$\frac{G-Q}{{\rm i}\sqrt{\omega_1\omega_2}} =\sum_{k=1}^6 N_k+\delta\left(\sum_{k=1}^6a_k+2{\rm i}\right)+O\big(\delta^2\big).$$ Now, in the limit $\delta\to 0^+$ we use formula (\[gam2lim2\]), and then relation (\[gmro\]) reduces to[^3] $$\begin{gathered} \nonumber \sum_{N\in {\mathbb Z}+\nu} \int_{-\infty}^{\infty}\big(y^2+N^2\big)\prod_{j=1}^6{\bf \Gamma}(a_j\pm y,N_j\pm N){\rm d}y \\ \qquad{} =(-1)^{2\nu}{\prod\limits_{1\leq j< k \leq 6}{\bf \Gamma}(a_j+a_k, N_j+N_k) \over \Gamma\Big(\sum\limits_{j=1}^6 a_j+2{\rm i},\sum\limits_{j=1}^6 N_j\Big)}\nonumber \\ \qquad\quad{} \times \sum_{N\in {\mathbb Z}+\nu} \int_{-\infty}^{\infty}\big(y^2+N^2\big) \prod_{j=1}^6{\bf \Gamma}(-{\rm i}-a_j\pm y,-N_j\pm N){\rm d}y, \label{degtrafo1}\end{gathered}$$ where we assume that the contours of integration are either real axes for $-1<\textup{Im}(a_k)<0$ or their proper deformations allowing analytical continuations of the functions on both sides in the variables $a_k$. Limiting case of transformation rules II ======================================== For deriving another symmetry transformation, we replace in (\[reflt\]) $$g_j \to g_j+{\rm i}\xi, \qquad g_{j+4}=f_j-{\rm i}\xi, \qquad j=1,\ldots, 4, \qquad z\to z-{\rm i}\xi.$$ The balancing condition takes the form $\sum\limits_{j=1}^4 (f_j+g_j)=2Q$. After taking the limit $\xi\to -\infty$, we come to the identity $$\begin{gathered} \nonumber \int_{-{\rm i}\infty}^{{\rm i}\infty}\prod_{j=1}^4\gamma^{(2)}(f_j+z;\mathbf{\omega}) \gamma^{(2)}(g_j-z;\mathbf{\omega}) {\rm d}z =\prod_{j,k=1}^4 \gamma^{(2)}(g_j+f_k;\mathbf{\omega}) \\ \qquad{} \times \int_{-{\rm i}\infty}^{{\rm i}\infty}\prod_{j=1}^4\gamma^{(2)}\big(\tfrac{1}{2}Q-f_j+z;\mathbf{\omega}\big) \gamma^{(2)}(\tfrac{1}{2}Q-g_j-z;\mathbf{\omega}) {\rm d}z. \label{infy} \end{gathered}$$ Now we take the parametrisation , jointly with $$\frac{g_j}{{\rm i}\sqrt{\omega_1\omega_2}}=N_j+\delta s_j,\qquad \frac{f_j}{{\rm i}\sqrt{\omega_1\omega_2}}=M_j+\delta t_j,\qquad j=1,2,3,4. \label{newpar}$$ The balancing condition takes the form $$\sum_{j=1}^4( N_j+M_j)=0,\qquad \sum_{j=1}^4(s_j+t_j)=-4{\rm i}.$$ In the limit $\delta\to 0^+$, in the same way as in many cases before, equation (\[infy\]) yields the following identity $$\begin{gathered} \sum_{N\in {\mathbb Z}+\nu}\int_{-{\rm i}\infty}^{{\rm i}\infty}\prod_{k=1}^4{\bf \Gamma}(s_k+y,N_k+N) {\bf \Gamma}(t_k-y,M_k-N) {\rm d}y \nonumber\\ \qquad{}=(-1)^{\sum\limits_{k=1}^4N_k}\prod_{j,k=1}^4 {\bf \Gamma}(s_j+t_k,N_j+M_k) \nonumber\\ \qquad\quad{} \times \sum_{N\in {\mathbb Z}+\nu}\int_{-{\rm i}\infty}^{{\rm i}\infty}\prod_{k=1}^4{\bf \Gamma}(-{\rm i}-t_k+y,N-M_k) {\bf \Gamma}(-{\rm i}-s_k-y,-N-N_k) {\rm d}y. \end{gathered}$$ After resolving the balancing condition in favor of $f_4$, $$f_4=2Q-g_4-\sum_{k=1}^3(f_k+g_k),$$ and taking the limit $g_4\to {\rm i}\infty$ in (\[infy\]), we obtain $$\begin{gathered} \int_{-{\rm i}\infty}^{{\rm i}\infty}{\rm e}^{{\pi{\rm i} z\over \omega_1\omega_2}\big(Q-\sum\limits_{k=1}^3(f_k+g_k)\big)}\prod_{k=1}^3\gamma^{(2)}(g_k+z;\mathbf{\omega}) \gamma^{(2)}(f_k-z;\mathbf{\omega}){\rm d}z \nonumber\\ \qquad {} ={\rm e}^{{\pi {\rm i}\over 2\omega_1\omega_2}\big(Q\sum\limits_{k=1}^3(f_k-g_k)+2\sum\limits_{1\leq j< k\leq 3}(g_jg_k-f_jf_k)\big)} {\prod\limits_{j,k=1}^3 \gamma^{(2)}(g_j+f_k;\mathbf{\omega})\over \gamma^{(2)}\Big(\sum\limits_{k=1}^3(g_k+f_k)-Q;\mathbf{\omega}\Big)} \\ \qquad\quad {} \times \int_{-{\rm i}\infty}^{{\rm i}\infty}{\rm e}^{{\pi {\rm i}z\over \omega_1\omega_2}\big({-}2Q+\sum\limits_{k=1}^3(g_k+f_k)\big)}\prod_{k=1}^3\gamma^{(2)}\big(\tfrac{1}{2}Q-f_k+z;\mathbf{\omega}\big) \gamma^{(2)}\big(\tfrac{1}{2}Q-g_k-z;\mathbf{\omega}\big) {\rm d}z.\nonumber $$ Taking the same parametrization , and , the limit $\delta\to 0^+$ yields the identity[^4] $$\begin{gathered} \sum_{N\in \mathbb{Z}+\nu} (-1)^{N-\nu} \int_{-\infty}^{\infty}\prod_{k=1}^3{\bf \Gamma}(y+s_k,N+N_k) {\bf \Gamma}(-y+t_k,M_k-N) {\rm d}y\nonumber\\ \qquad{} ={(-1)^{\sum\limits_{k=1}^3(N_k+M_k)}\prod\limits_{k,j=1}^3 {\bf \Gamma}(s_k+t_j,N_k+M_j)\over {\bf \Gamma}\Big(\sum\limits_{k=1}^3(s_k+t_k)+2{\rm i},\sum\limits_{k=1}^3(N_k+M_k)\Big)} \nonumber\\ \qquad\quad{} \times\sum_{N\in \mathbb{Z}+\nu}\int_{-\infty}^{\infty}(-1)^{N-\nu} \prod_{k=1}^3{\bf \Gamma}(y-{\rm i}-t_k, N-M_k) {\bf \Gamma}(-y-{\rm i}-s_k,-N-N_k) {\rm d}y.\end{gathered}$$ It is easy to see that for $\nu=1/2$ this formula is identical with the $\nu=0$ case after replacing $N_k$ by $N_k-1$, i.e., the parameter $\nu$ becomes redundant. As usual, the contours of integration separate sequences of poles of the integrands going upwards from the ones falling down. A new degeneration of the hyperbolic gamma function\ ($\boldsymbol{\omega_1=\omega_2}$) {#newlimit} ==================================================== Now we consider the limit $\omega_2\to\omega_1$, or $b\to 1$ corresponding to the central charge $c=25$ (the case $b\to -1$ is equivalent to it and we skip it). In this case, for generic values of $u$ the hyperbolic gamma function (or ) remains a well defined meromorphic function of $u$. However, for special choices of $u$ we have a divergence which we describe below. Namely, we take small $\delta>0$ and set $$b=\sqrt{\omega_1\over \omega_2}=1+{\rm i}\delta.$$ As a consequence, $Q=\omega_1+\omega_2=2\sqrt{\omega_1\omega_2}+O\big(\delta^2\big)$ and $$\sqrt{\omega_2\over \omega_1}=1-{\rm i}\delta+O\big(\delta^2\big),\qquad {\omega_1\over \omega_2}=1+2{\rm i}\delta-\delta^2,\qquad {\omega_2\over \omega_1}=1-2{\rm i}\delta+O\big(\delta^2\big).$$ Also we parametrise the argument $u$ in as $$u=\sqrt{\omega_1\omega_2}(n+y\delta), \qquad n\in {\mathbb Z}, \qquad y\in {\mathbb C},$$ and consider the limit $\delta\to 0^+$. For infinite products entering we have $$\big({\rm e}^{2\pi {\rm i} \frac{u}{\omega_2}};{\rm e}^{2\pi {\rm i} {\omega_1\over \omega_2}}\big)_{\infty} =\big(q^{\frac{n-{\rm i}y}{2}+O(\log q)};q\big)_\infty =\frac{(q;q)_\infty(1-q)^{1+\frac{-n+{\rm i}y}{2}+O(\log q)}} {\Gamma_q\big(\frac{n-{\rm i}y}{2}+O(\log q)\big)},$$ where $q={\rm e}^{-4\pi\delta(1+{\rm i}\delta/2)}$. Analogously, $$\big({\rm e}^{2\pi {\rm i}\frac{u}{\omega_1}}{\rm e}^{-2\pi {\rm i} {\omega_2\over \omega_1}};{\rm e}^{-2\pi {\rm i} {\omega_2\over \omega_1}}\big)_{\infty} = \big(\tilde q^{1-\frac{n+{\rm i}y}{2}+O(\log \tilde q)};\tilde q\big)_\infty =\frac{(\tilde q;\tilde q)_\infty(1-\tilde q)^{\frac{n+{\rm i}y}{2}+O(\log \tilde q)}} {\Gamma_{\tilde q}\big(1-\frac{n+{\rm i}y}{2}+O(\log \tilde q)\big)},$$ where $$\tilde q={\rm e}^{-4\pi\delta\frac{1+{\rm i}\delta/2}{(1+{\rm i}\delta)^2}}.$$ As a result we obtain, $$\gamma(u;\omega_1,\omega_2)=\frac{\Gamma_q\big( \frac{n-{\rm i}y}{2}+O(\log q) \big)} {\Gamma_{\tilde q}\big( 1-\frac{n+{\rm i}y}{2}+O(\log \tilde q)\big)} \frac{(\tilde q;\tilde q)_\infty}{(q;q)_\infty} \frac{(1-\tilde q)^{\frac{n+{\rm i}y}{2}+O(\log \tilde q)}} {(1-q)^{1+\frac{-n+{\rm i}y}{2}+O(\log q)}}.$$ Using the modular transformation rule for the Dedekind eta-function, we find $$\frac{(\tilde q;\tilde q)_\infty}{(q;q)_\infty} ={\rm e}^{{\pi{\rm i} \over 12}\big({\omega_2\over \omega_1}+{\omega_1\over \omega_2}\big)} \left(-{\rm i} {\omega_1\over \omega_2}\right)^{1\over 2} \underset{\delta\to 0^+}{=}{\rm e}^{ -\frac{\pi{\rm i}}{12}}. $$ Finally, applying the strong limit $\Gamma_q(x+O(\log q))\to \Gamma(x)$ for $q\to 1$, which is uniform on compacta [@Tom; @rai:limits], and combining all factors together, we come to the leading asymptotics $$\gamma(u;\omega_1,\omega_2)\underset{\delta\to 0^+}{=} {\rm e}^{-\frac{\pi {\rm i}}{12}} (4\pi\delta)^{n-1} \left(1-\frac{n+{\rm i}y}{2}\right)_{n-1}, \qquad (a)_n:=\frac{\Gamma(a+n)}{\Gamma(a)},$$ where $(a)_0=1$ and $$(a)_n= \begin{cases} a(a+1)\cdots(a+n-1), & \text{for} \ n>0,\quad \\ \dfrac{1}{(a-1)(a-2)\cdots(a+n)}, &\text{for} \ n<0, \end{cases}$$ is the standard Pochhammer symbol. For function this yields $$\gamma^{(2)}(u;\omega_1,\omega_2)\underset{\delta\to 0^+}{=} {\rm e}^{-\frac{\pi {\rm i}}{2}(n-1)^2} (4\pi\delta)^{n-1}\left(1-\frac{n+{\rm i}y}{2}\right)_{n-1}.$$ Since here $n$ is an arbitrary integer, we can shift $n\to n+1$ in this formula, which noticeably simplifies its form $$\gamma^{(2)}\big(\sqrt{\omega_1\omega_2}(n+1+y\delta);\mathbf{\omega}\big) \underset{\delta\to 0^+}{=} {\rm e}^{-\frac{\pi {\rm i}}{2}n^2} (4\pi\delta)^n\left(\frac{1-n-{\rm i}y}{2}\right)_n, \qquad \sqrt{\omega_1\over \omega_2}=1+{\rm i}\delta,$$ where $n\in {\mathbb Z}$, $y\in {\mathbb C}$. This is a new degeneration limit for the hyperbolic gamma function or the Faddeev modular quantum dilogarithm. In this case the function $\gamma^{(2)}(u;\mathbf{\omega})$ starts either to vanish or to blow up around a special discrete set of points of the argument $u$ passing through the whole complex plane along a particular line. This is partially similar to the picture taking place for $\omega_1+\omega_2\to 0$, but the direction of the corresponding line on the complex plane is different. We postpone consideration of the consequences of such a degeneration for hyperbolic beta integrals and its possible applications to a later work. Conclusion ========== In the present paper we have performed a rigorous and complete analysis of the degeneration of hyperbolic integrals to the complex hypergeometric functions in the Mellin–Barnes representation, which was noticed for the first time in [@BMS]. The limit $\omega_1+\omega_2\to 0$ corresponds to $b\to \pm {\rm i}$ in the context of $2d$ Liouville quantum field theory (which is related to the $p\to\infty$ limit of minimal models leading to the central charge value $c=1$ [@RW]). Additionally, we have discovered a new nontrivial degeneration of the hyperbolic gamma function, or Faddeev’s modular quantum dilogarithm, in the limit $\omega_1\to\omega_2$, which corresponds to $b\to\pm1$ for the Liouville theory leading to the central charge $c=25$. It would be interesting to investigate applications of our formulas in the context of fusion matrices of the corresponding two-dimensional quantum field theories. Another interesting application should emerge within the theory of Painlevé transcendents. Namely, as follows from the analysis of $c=1$ conformal blocks in [@ILT], the degeneration of the function standing in the left-hand side of the identity to the one in should define the fusion matrix of interest with a direct relation to a particular tau-function of the Painlevé-VI function. Similar question can be raised for the $c=25$ conformal blocks and their manifestations for the Painlevé equations. The third application in the quantum field theory is expected to emerge in the topological field theory. Namely, one can extend considerations of [@KLV] and realize the corresponding Pachner moves using the unusual integral identities derived by us. From the point of view of representation theory of the complex group ${\rm SL}(2,{\mathbb C})$, it is necessary to clarity the group-theoretical meaning of the parameter value $\nu=1/2$, showing very interesting phenomena [@spi:rare]. In this context, one of the unsolved problems is the inversion of the Mellin–Barnes form of our functions – infinite bilateral sums of integrals standing in identities , , etc. – to the integrals over complex planes of the type and . For $\nu=0$ some examples of such conversions are given in [@DMV2017; @DS2017; @Ismag2]. However, for $\nu=1/2$ it is not clear how it can be performed yet. There are also applications to integrable systems. Namely, since hyperbolic beta integrals serve as measures of the orthogonality relations for wave functions of the Ruijsenaars type many body systems [@Ruij], one can consider what happens with them in the taken limits. In particular, this requires construction of proper generalizations of the hypergeometric equation to a finite-difference equation for the function standing in the left-hand side of equality . As another manifestation in integrable systems, it is straightforward to build the corresponding solution of the Yang–Baxter equation just by appropriate degeneration of the results of [@CS] (or, in a more general setting, the results of [@spi:rareYBE]). Obviously, the degenerations similar to the one we considered in the present paper for the most general univariate hyperbolic beta integral and symmetry transformations for the corresponding analogue of the Euler–Gauss hypergeometric function can be applied to the rarefied hyperbolic integrals associated with the general lens space [@SS23; @SS24]. In particular, the first task would be to consider the most general way of approaching the unit circle simultaneously by $q$ and $\tilde q$ in the rarefied $q$-beta integral . Finally, we have presented the degeneration hierarchy only for the simplest identities for elliptic hypergeometric integrals [@spi:umn; @spi:essays] and one can extend our approach to all relations between multidimensional integrals of such type established to the present moment. Acknowledgements {#acknowledgements .unnumbered} ---------------- This paper is based on the talk given by V.S. at the conference “Elliptic Integrable Systems, Special Functions and Quantum Field Theory”, June 16–20, 2019, Nordita, Stockholm. The key results of this work were obtained within the research program of project no. 19-11-00131 supported by the Russian Science Foundation. We thank T.H. Koornwinder and E.M. Rains for explanations on the uniformness of the limit for $q$-gamma function following from their works [@Tom] and [@rai:limits]. [99]{} Andrews G.E., Askey R., Roy R., Special functions, Encyclopedia of Mathematics and its Applications, Vol. 71, [Cambridge University Press](https://doi.org/10.1017/CBO9781107325937), Cambridge, 1999. Aomoto K., On the complex [S]{}elberg integral, [Quart. J. Math. Oxford](https://doi.org/10.1093/qmath/38.4.385) 38 (1987), 385–399. Askey R., Wilson J., Some basic hypergeometric orthogonal polynomials that generalize [J]{}acobi polynomials, [Mem. Amer. Math. Soc.](https://doi.org/10.1090/memo/0319) 54 (1985), iv+55 pages. Bazhanov V.V., Mangazeev V.V., Sergeev S.M., Exact solution of the [F]{}addeev–[V]{}olkov model, [Phys. Lett. A](https://doi.org/10.1016/j.physleta.2007.10.053) 372 (2008), 1547–1550, [arXiv:0706.3077](https://arxiv.org/abs/0706.3077). Chicherin D., Spiridonov V.P., The hyperbolic modular double and the [Y]{}ang–[B]{}axter equation, in Representation Theory, Special Functions and [P]{}ainlevé equations – [RIMS]{} 2015, Adv. Stud. Pure Math., Vol. 76, [Math. Soc. Japan](https://doi.org/10.2969/aspm/07610095), Tokyo, 2018, 95–123, [arXiv:1511.00131](https://arxiv.org/abs/1511.00131). Derkachov S.É., Manashov A.N., On complex gamma-function integrals, [SIGMA](https://doi.org/10.3842/SIGMA.2020.003) 16 (2020), 003, 20 pages, [arXiv:1908.01530](https://arxiv.org/abs/1908.01530). Derkachov S.É., Manashov A.N., Valinevich P.A., Gustafson integrals for [${\rm SL}(2,{\mathbb C})$]{} spin magnet, [J. Phys. A: Math. Theor.](https://doi.org/10.1088/1751-8121/aa7480) 50 (2017), 294007, 12 pages, [arXiv:1612.00727](https://arxiv.org/abs/1612.00727). Derkachov S.É., Manashov A.N., Valinevich P.A., [${\rm SL}(2,{\mathbb C})$]{} [G]{}ustafson integrals, [SIGMA](https://doi.org/10.3842/SIGMA.2018.030) 14 (2018), 030, 16 pages, [arXiv:1711.07822](https://arxiv.org/abs/1711.07822). Derkachov S.É., Spiridonov V.P., The $6j$-symbols for the ${\rm SL}(2,{\mathbb C})$ group, [Theoret. and Math. Phys.](https://doi.org/10.1134/S0040577919010033) 198 (2019), 29–47, [arXiv:1711.07073](https://arxiv.org/abs/1711.07073). Dimofte T., Complex [C]{}hern–[S]{}imons theory at level [$k$]{} via the 3d-3d correspondence, [Comm. Math. Phys.](https://doi.org/10.1007/s00220-015-2401-1) 339 (2015), 619–662, [arXiv:1409.0857](https://arxiv.org/abs/1409.0857). Dotsenko V.S., Fateev V.A., Four-point correlation functions and the operator algebra in [$2$]{}[D]{} conformal invariant theories with central charge [$C\leq 1$]{}, [Nuclear Phys. B](https://doi.org/10.1016/S0550-3213(85)80004-3) 251 (1985), 691–734. Faddeev L.D., Discrete [H]{}eisenberg–[W]{}eyl group and modular group, [Lett. Math. Phys.](https://doi.org/10.1007/BF01872779) 34 (1995), 249–254, [arXiv:hep-th/9504111](https://arxiv.org/abs/hep-th/9504111). Faddeev L.D., Currentlike variables in massive and massless integrable models, in Quantum Groups and their Applications in Physics ([V]{}arenna, 1994), Proc. Internat. School Phys. Enrico Fermi, Vol. 127, [IOS](https://doi.org/10.3254/978-1-61499-213-4-117), Amsterdam, 1996, 117–135, [arXiv:hep-th/9408041](https://arxiv.org/abs/hep-th/9408041). Faddeev L.D., Modular double of a quantum group, in Conférence [M]{}oshé [F]{}lato 1999, [V]{}ol. [I]{} ([D]{}ijon), Math. Phys. Stud., Vol. 21, Kluwer Acad. Publ., Dordrecht, 2000, 149–156, [arXiv:math.QA/9912078](https://arxiv.org/abs/math.QA/9912078). Gel’fand I.M., Graev M.I., Vilenkin N.J., Generalized functions, Vol. 5, Integral geometry and representation theory, Fizmatgiz, Moscow, 1962. Iorgov N., Lisovyy O., Tykhyy Yu., Painlevé [VI]{} connection problem and monodromy of [$c=1$]{} conformal blocks, [J. High Energy Phys.](https://doi.org/10.1007/JHEP12(2013)029) 2013 (2013), no. 12, 029, 27 pages, [arXiv:1308.4092](https://arxiv.org/abs/1308.4092). Ismagilov R.S., Racah operators for principal series of representations of the group ${\rm SL}(2,{\mathbb C})$, [Sb. Math.](https://doi.org/10.1070/SM2007v198n03ABEH003840) 198 (2007), 369–381. Kashaev R., Luo F., Vartanov G., A [TQFT]{} of [T]{}uraev–[V]{}iro type on shaped triangulations, [Ann. Henri Poincaré](https://doi.org/10.1007/s00023-015-0427-8) 17 (2016), 1109–1143, [arXiv:1210.8393](https://arxiv.org/abs/1210.8393). Kels A.P., A new solution of the star-triangle relation, [J. Phys. A: Math. Theor.](https://doi.org/10.1088/1751-8113/47/5/055203) 47 (2014), 055203, 7 pages, [arXiv:1302.3025](https://arxiv.org/abs/1302.3025). Kels A.P., Yamazaki M., Elliptic hypergeometric sum/integral transformations and supersymmetric lens index, [SIGMA](https://doi.org/10.3842/SIGMA.2018.013) 14 (2018), 013, 29 pages, [arXiv:1704.03159](https://arxiv.org/abs/1704.03159). Kharchev S., Lebedev D., Semenov-Tian-Shansky M., Unitary representations of [$U_q({\mathfrak{gl}}(2,{\mathbb R}))$]{}, the modular double and the multiparticle [$q$]{}-deformed [T]{}oda chains, [Comm. Math. Phys.](https://doi.org/10.1007/s002200100592) 225 (2002), 573–609, [arXiv:hep-th/0102180](https://arxiv.org/abs/hep-th/0102180). Koornwinder T.H., Jacobi functions as limit cases of [$q$]{}-ultraspherical polynomials, [J. Math. Anal. Appl.](https://doi.org/10.1016/0022-247X(90)90026-C) 148 (1990), 44–54. Mimachi K., Complex hypergeometric integrals, in Representation Theory, Special Functions and [P]{}ainlevé equations – [RIMS]{} 2015, Adv. Stud. Pure Math., Vol. 76, [Math. Soc. Japan](https://doi.org/10.2969/aspm/07610469), Tokyo, 2018, 469–485. Molchanov V.F., Neretin Yu.A., A pair of commuting hypergeometric operators on the complex plane and bispectrality, J. Spectr. Theory, [t]{}o appear, [arXiv:1812.06766](https://arxiv.org/abs/1812.06766). Naimark M.A., Decomposition of a tensor product of irreducible representations of the proper Lorentz group into irreducible representations, [Amer. Math. Soc.](https://doi.org/10.1090/trans2/036/04), Providence, RI, 1964, 101–229. Neretin Yu.A., An analog of the [D]{}ougall formula and of the de [B]{}ranges–[W]{}ilson integral, [Ramanujan J.](https://doi.org/10.1007/s11139-019-00218-0), [t]{}o appear, [arXiv:1812.07341](https://arxiv.org/abs/1812.07341). Neretin Yu.A., Barnes–[I]{}smagilov integrals and hypergeometric functions of the complex field, [SIGMA](https://doi.org/10.3842/SIGMA.2020.072) 16 (2020), 072, 20 pages, [arXiv:1910.10686](https://arxiv.org/abs/1910.10686). Rahman M., An integral representation of a [$_{10}\varphi_9$]{} and continuous bi-orthogonal [$_{10}\varphi_9$]{} rational functions, [Canad. J. Math.](https://doi.org/10.4153/CJM-1986-030-6) 38 (1986), 605–618. Rains E.M., Limits of elliptic hypergeometric integrals, [Ramanujan J.](https://doi.org/10.1007/s11139-007-9055-3) 18 (2009), 257–306, [arXiv:math.CA/0607093](https://arxiv.org/abs/math.CA/0607093). Ruijsenaars S.N.M., A generalized hypergeometric function satisfying four analytic difference equations of [A]{}skey–[W]{}ilson type, [Comm. Math. Phys.](https://doi.org/10.1007/s002200050840) 206 (1999), 639–690. Runkel I., Watts G.M.T., A nonrational CFT with $c = 1$ as a limit of minimal models, [J. High Energy Phys.](https://doi.org/10.1088/1126-6708/2001/09/006) 2001 (2001), no. 9, 006, 41 pages, [arXiv:hep-th/0107118](https://arxiv.org/abs/hep-th/0107118). Sarkissian G.A., Spiridonov V.P., The modular group and a hyperbolic beta integral, [Russian Math. Surveys](https://doi.org/10.4213/rm9951) 75 (2020), 187–188. Sarkissian G.A., Spiridonov V.P., General modular quantum dilogarithm and beta integrals, [Proc. Steklov Inst. Math.](https://doi.org/10.1134/S0081543820030190) 309 (2020), 251–270, [arXiv:1910.11747](https://arxiv.org/abs/1910.11747). Spiridonov V.P., On the elliptic beta function, [Russian Math. Surveys](https://doi.org/10.1070/rm2001v056n01ABEH000374) 56 (2001), 185–186. Spiridonov V.P., Theta hypergeometric integrals, [St. Petersburg Math. J.](https://doi.org/10.1090/S1061-0022-04-00839-8) 15 (2004), 929–967, [arXiv:math.CA/0303205](https://arxiv.org/abs/math.CA/0303205). Spiridonov V.P., Essays on the theory of elliptic hypergeometric functions, [Russian Math. Surveys](https://doi.org/10.1070/RM2008v063n03ABEH004533) 63 (2008), 405–472, [arXiv:0805.3135](https://arxiv.org/abs/0805.3135). Spiridonov V.P., Elliptic beta integrals and solvable models of statistical mechanics, in Algebraic Aspects of [D]{}arboux Transformations, Quantum Integrable Systems and Supersymmetric Quantum Mechanics, Contemp. Math., Vol. 563, [Amer. Math. Soc.](https://doi.org/10.1090/conm/563/11170), Providence, RI, 2012, 181–211, [arXiv:1011.3798](https://arxiv.org/abs/1011.3798). Spiridonov V.P., Rarefied elliptic hypergeometric functions, [Adv. Math.](https://doi.org/10.1016/j.aim.2018.04.014) 331 (2018), 830–873, [arXiv:1609.00715](https://arxiv.org/abs/1609.00715). Spiridonov V.P., The rarefied elliptic [B]{}ailey lemma and the [Y]{}ang–[B]{}axter equation, [J. Phys. A: Math. Theor.](https://doi.org/10.1088/1751-8121/ab3358) 52 (2019), 355201, 15 pages, [arXiv:1904.12046](https://arxiv.org/abs/1904.12046). Stokman J.V., Hyperbolic beta integrals, [Adv. Math.](https://doi.org/10.1016/j.aim.2003.12.003) 190 (2004), 119–160, [arXiv:math.QA/0303178](https://arxiv.org/abs/math.QA/0303178). van de Bult F.J., Rains E.M., Stokman J.V., Properties of generalized univariate hypergeometric functions, [Comm. Math. Phys.](https://doi.org/10.1007/s00220-007-0289-0) 275 (2007), 37–95, [arXiv:math.CA/0607250](https://arxiv.org/abs/math.CA/0607250). van Diejen J.F., Spiridonov V.P., Unit circle elliptic beta integrals, [Ramanujan J.](https://doi.org/10.1007/s11139-005-4846-x) 10 (2005), 187–204, [arXiv:math.CA/0309279](https://arxiv.org/abs/math.CA/0309279). [^1]: After presenting relation at the Nordita conference in June 2019, there appeared the work by Derkachov and Manashov [@DM2019] where it was independently derived (as well as its substantially more complicated multidimensional version) by a completely different method. Namely, our identity corresponds to the choice $N=2$ in formulas (3.7) and (3.8) in [@DM2019] (formula (3.8) was obtained from (3.7) after applying the reflection formula to the $\bf \Gamma$-functions depending on $x_6$ with a small typo on the right-hand side, where $2N+3$ should be replaced by $2N+1$). The difference in the sign factors on the right-hand sides emerges from different representations of the product of complex gamma functions in the kernel denominator . Since the star-triangle relation form of was considered in detail in [@DM2019], we skip its discussion here. [^2]: Exactly the same relation was obtained also independently in [@DM2019] as formula (2.3b) for $N=2$. [^3]: This formula corresponds to the choice $n=m=1$ in formula (6.7) in [@DM2019]. [^4]: This identity was also obtained in [@DM2019], see there formula (6.6) for $n=m=1$.	{ "pile_set_name": "ArXiv" }
--- abstract: 'In sneutrino hybrid inflation the superpartner of one of the right-handed neutrinos involved in the seesaw mechanism plays the role of the inflaton field. It obtains its large mass after the “waterfall” phase transition which ends hybrid inflation. After this phase transition the oscillations of the sneutrino inflaton field may dominate the universe and efficiently produce the baryon asymmetry of the universe via nonthermal leptogenesis. We investigate the conditions under which inflation, with primordial perturbations in accordance with the latest WMAP results, as well as successful nonthermal leptogenesis can be realized simultaneously within the sneutrino hybrid inflation scenario. We point out which requirements successful inflation and leptogenesis impose on the seesaw parameters, i.e. on the Yukawa couplings and the mass of the right-handed (s)neutrino, and derive the predictions for the CMB observables in terms of the right-handed (s)neutrino mass and the other relevant model parameters.' --- 0.05cm [Sneutrino Hybrid Inflation and\ Nonthermal Leptogenesis]{} Stefan Antusch, Jochen P. Baumann, Valerie F. Domcke and\ Philipp M. Kostka\ [[ Max-Planck-Institut für Physik (Werner-Heisenberg-Institut),\ Föhringer Ring 6, 80805 München, Germany]{} ]{} 0.75cm Introduction ============ The paradigm of cosmic inflation [@Guth:1980zm; @Liddle:2000cg; @Bailin:2004zd] (for recent reviews see e.g. [@Baumann:2009ds; @Mazumdar:2010sa]) has proven very successful in resolving the flatness and horizon problems of the early universe and in explaining the absence of relics from early phase transitions. However, the connection to particle physics is still unclear. One possibility to establish such a connection is provided by sneutrino hybrid inflation [@Antusch:2004hd], where the superpartner of one of the right-handed neutrinos involved in the seesaw mechanism [@seesaw] plays the role of the inflaton field. In sneutrino hybrid inflation, a large vacuum energy density is present which drives inflation and the sneutrino direction in field space has an almost flat potential suitable for slow-roll inflation. The right-handed (s)neutrinos obtain their large masses after the “waterfall” phase transition which ends hybrid inflation. Inflation in this scenario is closely linked to the physics generating the small neutrino masses via the seesaw mechanism. Another attractive connection between the seesaw mechanism and early universe cosmology is the possibility of generating the observed baryon asymmetry of the universe via the out-of-equilibrium decays of the right-handed (s)neutrinos in leptogenesis [@Fukugita:1986hr] (for recent reviews see [@Chen:2007fv; @Davidson:2008bu]). For calculating the produced baryon asymmetry, the knowledge of the phase of (p)reheating after inflation is in general mandatory, since it may lead to the nonthermal production of right-handed (s)neutrinos and since it determines the reheat temperature, which in turn governs the possibility of thermal (s)neutrino production. In most inflation models the nonthermal (s)neutrino production must arise from the decays of the inflaton field. In sneutrino hybrid inflation, on the other hand, the inflaton itself is a right-handed sneutrino, which means that this intermediate step is skipped and the sneutrino inflaton field after inflation may directly dominate the universe and, when it decays, most efficiently produce the baryon asymmetry and reheat the universe. In previous works, leptogenesis after sneutrino inflation has been studied in the context of chaotic sneutrino inflation [@Murayama:1992ua; @Ellis:2003sq] which however requires a quite heavy sneutrino with a mass of about $10^{13}$ GeV and correspondingly very small Yukawa couplings in order to realize a low reheat temperature as suggested by gravitino and similar constraints in supersymmetric cosmology. Furthermore, chaotic sneutrino inflation with a quadratic potential for the inflaton gives rise to a comparatively large tensor-to-scalar ratio of $r \sim 0.16$. On the other hand, sneutrino hybrid inflation [@Antusch:2004hd], as typical for hybrid-type inflation scenarios [@Linde:1990gz; @Linde:1991km; @Copeland:1994vg; @Linde:1997sj], predicts a much smaller ratio $r \lesssim 0.01$ and is thus clearly distinguishable from chaotic sneutrino inflation by future observations (e.g. by the Planck satellite). Recently, it has been shown that sneutrino hybrid inflation [@Antusch:2004hd] belongs to a wider class of hybrid-like inflation models, dubbed “tribrid inflation” in [@Antusch:2009vg], which are very suitable for being embedded into supergravity (SUGRA) theories with the SUGRA $\eta$-problem solved by either a shift symmetry [^1] [@Antusch:2009ef] or a Heisenberg symmetry [@Antusch:2008pn] in the Kähler potential. While the sneutrino was a gauge singlet in [@Antusch:2004hd], it has been demonstrated in [@Antusch:2010va] that it may be embedded into a Grand Unified Theory (GUT) representation, e.g. into a 16-plet of SO(10), establishing a possible link between sneutrino hybrid inflation and left-right symmetric GUTs. Nonthermal leptogenesis after sneutrino hybrid inflation, on the other hand, was only briefly discussed in [@Antusch:2004hd] for an example set of model parameters. In this paper, we therefore investigate in detail the conditions under which inflation, with primordial perturbations in accordance with the latest WMAP results, as well as successful nonthermal leptogenesis can be realized simultaneously within the sneutrino hybrid inflation scenario. We point out which requirements successful inflation and leptogenesis impose on the seesaw parameters, i.e. on the Yukawa couplings and the mass of the right-handed (s)neutrino, and derive the predictions for the CMB observables in terms of the right-handed (s)neutrino mass and the other relevant model parameters. Our results are meant as a guideline for the construction of explicit particle physics models incorporating sneutrino hybrid inflation and baryogenesis via nonthermal leptogenesis. The paper is organized as follows: In section \[framework\] we introduce the sneutrino hybrid inflation scenario in a simple setup. Section \[inflation\] is dedicated to the inflationary phase and the predictions for the CMB observables. In section \[leptogenesis\] we discusses the reheating of the universe after inflation and the production of the baryon asymmetry of the universe via nonthermal leptogenesis. We conclude in section \[infl\_lepto\] by combining our results from inflation and leptogenesis to highlight the preferred ranges of the model parameters. Framework ========= We will discuss sneutrino hybrid inflation and subsequent baryogenesis via nonthermal leptogenesis in an extension of the minimal supersymmetric standard model (MSSM) with conserved R-parity, where three additional right-handed (s)neutrinos acquire large masses after the “waterfall” phase transition ending inflation. The superpotential defining our framework is given by $$\label{eq_W} W = W_{\text{MSSM}} + (y_{\nu})_{ij} \,\hat{N}^i\, \hat{h}_{ a}\, \epsilon^{a b} \hat{L}^j_b + \frac{\lambda_{ii}}{M_P}\,(\hat{N}^i)^2 \hat{H}^2 + \kappa\, \hat{S}\left(\hat{H}^2 - M^2\right) + \ldots\,,$$ where the $\hat{N}^i$ (the index $i=1, 2, 3$ denotes the different generations) are gauge singlet superfields describing the heavy right-handed (s)neutrinos and where the reduced Planck scale is given by $M_P\simeq2.4\cdot 10^{18}\,\text{GeV}$. The canonically normalized imaginary parts[^2] $N^i$ of the respective scalar components are inflaton candidates as will be described below. $\hat{L}$ and $\hat{h}$ are SU(2)$_L$-doublet superfields which contain the standard model leptons and up-type Higgs, respectively. The Yukawa coupling term of $\hat{N}$ with the Higgs and lepton doublet, i.e. the second term in Eq. , allows to identify $\hat{N}$ with the right-handed neutrino superfield. $\hat{H}$ and $\hat{S}$ are two additional gauge singlet superfields. Here the canonically normalized real part $H$ of the scalar component of $\hat{H}$ is the so-called “waterfall” field responsible for ending inflation. The F-term of $\hat{S}$, the so-called “driving superfield”, provides the large vacuum energy density that drives inflation. The scalar component of $\hat{S}$ is fixed at zero during inflation by SUGRA corrections (cf. section \[sugra\]) and does not affect the inflationary dynamics. Furthermore, we assume $\lambda_{ii}$ and $\kappa$ to be real coupling parameters for simplicity. The form of the superpotential Eq. is motivated as follows: The latter two terms generate the scalar potential suitable for inflation. In the false vacuum with large values of $N^i$ and $H$ stabilized at zero, the large vacuum energy $V_0=\kappa^2 M^4$ drives the quasi-exponential growth of the scale factor in inflation. Once the slow-rolling fields $N^i$ fall below a critical value, the negative contribution to the squared mass of $H$ from the term $\kappa\, \hat{S}\,(\hat{H}^2 - M^2)$ starts dominating over the positive contribution from the terms $ \frac{\lambda_{ii}}{M_P}\,(\hat{N}^i)^2 \hat{H}^2$. Therefore, $H$ becomes tachyonic which triggers the “waterfall” ending inflation as $H$ acquires a non-zero vacuum expectation value (vev). After inflation, close to the global minimum of the potential where $N^i \approx 0$ and $H \approx \sqrt{2} \,M$ and where the large vacuum energy contribution vanishes, the fields $N^i$ and $H$ perform damped oscillations accounting for a matter dominated universe. The field which decays last and finally dominates the universe is generically the right-handed sneutrino with the smallest mass and smallest Yukawa couplings, since it only decays via the second term in Eq. proportional to $(y_{\nu})_{ij}$. This decay reheats the universe which thus enters its radiation dominated epoch. For illustration, we have plotted a typical scalar potential resulting from the scenario described above in Fig. \[hybridpotential\]. For further details see e.g. Ref. [@Antusch:2009ef]. ![\[hybridpotential\] Typical F-term scalar potential resulting from the model of Eq. for the scalar components of $\hat{S},\, \hat{L},\, \hat{h}$ set to zero. For the plot, we have used example parameters $\kappa=\lambda=1$ and $M=M_P$.](hybrid) The first three terms in Eq. describe the MSSM with masses for the additional right-handed neutrinos generated after inflation. In particular the term $\frac{\lambda_{ii}}{M_P}(\hat{N}^i)^2 \hat{H}^2$ generates mass terms for the heavy (s)neutrinos as $H$ acquires its non-zero vev. The vev of $H$ in the true minimum is governed by the fourth term in Eq. . In a realistic scenario, we would expect inflation to end by a phase transition, i.e. the $H$ field to be a non-singlet under some symmetry group [^3]. A realization within the context of GUTs can be found in Ref. [@Antusch:2010va], and the idea of relating the “waterfall” of hybrid inflation to the breaking of a family symmetry was discussed in Ref. [@Antusch:2008gw]. For simplicity, we keep $\hat{H}$ a gauge singlet here. Finally, the ellipsis represent possible higher dimensional operators. We note that a superpotential of the form given in Eq. can be realized in an explicit model with discrete symmetries, as has been shown in [@Antusch:2009ef; @Antusch:2010va]. These discrete symmetries distinguish between the gauge singlet fields of Eq. . The parameters which appear in Eq. , and which govern our framework, can be understood as follows: - The phase transition scale $M$ is the vev of the scalar component of the $\hat{H}$ superfield after the phase transition ending inflation and is the mass scale relevant for inflation. - The parameter $\lambda_{11}$ determines the seesaw scale which corresponds to the mass of the lightest right-handed neutrino $m_{N^1} = 2 \frac{\lambda_{11}}{M_P} M^2$ in the true vacuum of the theory. - The vacuum energy parameter $\kappa$ fixes the the vacuum energy density with regard to the phase transition scale. - The effective Yukawa coupling $\tilde{y}_1 \equiv \sqrt{(y_{\nu} y_{\nu}^{\dagger})_{11}}$ is linked to the light neutrino masses. We will be working in a SUGRA framework with SUGRA corrections stabilizing the scalar components of $\hat{S}$, $\hat{L}$ and $\hat{h}$ during inflation and with a symmetry in the Kähler potential that guarantees tree-level flat $N^i$ directions. This solves the $\eta$-problem of SUGRA inflation. The details of such a SUGRA framework are discussed in section \[sugra\]. However, to illustrate the underlying physics more clearly we first focus on a global supersymmetry (SUSY) model and take the features mentioned above for granted. In order to produce the CP-violation necessary for leptogenesis we work with three (s)neutrino generations. Assuming that the right-handed neutrinos are strongly hierarchical, i.e. one of them is significantly lighter than the other two, the scalar components of the latter superfields can be stabilized at their minima before the final 60 e-folds of inflation begin. Thus the time evolution of the lightest sneutrino controls the relevant slow-roll dynamics and it can therefore be identified as the inflaton. On the other hand, the outcome of leptogenesis is governed by the sneutrino with the smallest decay rate. This implies a comparatively small mass and small Yukawa couplings. In the following, we shall concentrate on the case where the lightest sneutrino drives both inflation and leptogenesis. Hence, the three generation model can be simplified to an effective one generation model in the right-handed neutrino sector, with the only remnant of the other two generations being a non-vanishing CP-asymmetry necessary for leptogenesis. We can thus concentrate on $i = 1$ in Eqs. and we denote the relevant inflaton direction by $N \equiv N^1$ and the respective coupling constant by $\lambda \equiv \lambda_{11}$. Inflation \[inflation\] ======================= Based on the framework described in the previous section, we now have a closer look at the inflationary dynamics in our model. Furthermore, we derive restrictions on the model parameters from the requirement of successful inflation and the latest observational data. We start with a short introduction to slow-roll inflation and then discuss a realization of the model of section \[framework\] in a globally supersymmetric context. We then refine this discussion by including SUGRA effects and close the section by listing the inflationary predictions from our model and comparing them to the latest observational data. Short Overview \[overview\] --------------------------- A common way to realize inflation is the so-called slow-roll paradigm, where a classical scalar field with a strong dominance of its potential energy over its kinetic energy $V \gg \mathcal{L}_{\text{kin}}$ drives the accelerated expansion of the universe. At the same time, the quantum fluctuations of the inflaton field can account for the metric perturbations which give rise to the small scale CMB anisotropies. Inflation ends when the slow-roll conditions are violated, i.e. when the slow-roll parameters parameterizing the scalar potential and its derivatives $$\label{slow_roll_pars} \epsilon = \frac{M_P^2}{2}\, \left( \frac{V'}{V} \right)^2, \qquad \eta = M_P^2 \,\frac{V''}{V} \,, \qquad \xi^2 = M_P^4 \,\frac{V' V'''}{V^2} \,,$$ become of order one. Here, a prime denotes derivative w.r.t. the inflaton field $N$. In the slow-roll approximation, when $\epsilon \ll 1$, $\left\| \eta \right\| \ll 1$ and $\xi^2\ll1$, the equation of motion of a homogeneous (classical) scalar field $$\label{eom_scalar} \ddot{N} + 3\,\mathcal{H}\, \dot{N} + V' = 0\,,$$ simplifies to $$\label{eomI} 3 \,\mathcal{H} \,\dot{N} = -V'\,.$$ Here $\cal H$ denotes the Hubble expansion parameter. Models of inflation typically predict the power spectra of the gauge invariant scalar and tensor perturbations at the time when the relevant fluctuations exited the horizon, roughly $\mathcal{N}_{\text{e}}\simeq 50 - 70$ e-folds before the end of inflation. The amplitude of the scalar perturbations $\Delta_s^2$, the scalar spectral index $n_s$, the running of the scalar spectral index $\alpha_s$, the tensor-to-scalar ratio $r$ and the tensor spectral index $n_t$ can be estimated in terms of the potential and the slow-roll parameters [@Baumann:2009ds] as $$\label{eq_infl_pred} \begin{split} \Delta_s^2 &\simeq \frac{1}{M_P^6}\,\frac{1}{12 \pi^2} \,\frac{V^3}{(V')^2}\,,\\ n_s &\simeq 1 - 6\, \epsilon + 2 \,\eta\,, \\ \alpha_s &\simeq 16 \,\epsilon\, \eta - 24 \,\epsilon^2 - 2\, \xi^2\,,\\ r &\simeq 16 \,\epsilon \,,\\ n_t &\simeq - 2\, \epsilon\,, \end{split}$$ where these expressions have to be evaluated at the field value $N = N(\mathcal{N}_{\text{e}})$ when the relevant scales leave the horizon. This value can be computed from Eq. . In order to test our model against observations, we compare the predictions from Eqs. to the experimental data obtained from the 7 year WMAP survey combined with measurements of the baryon acoustic oscillations (BAO) [@Percival] and measurements of the present value of the Hubble parameter $H_0$ [@Riess] using the six parameter $\Lambda$CDM fit [@wmap], which are given by $$\label{infl_obs} \begin{split} &0.951 < n_s < 0.975 \qquad \text{(68 $\%$ CL)}\,,\\ &0.939 < n_s < 0.987 \qquad \text{(95 $\%$ CL)} \,,\\ &\Delta_s^2 = (2.441^{+0.088}_{-0.092}) \cdot 10^{-9}\,. \end{split}$$ Realization in Global Supersymmetry \[susy\] -------------------------------------------- In the model described in section \[framework\], we assumed a symmetry in the Kähler potential guaranteeing a flat $N$ direction (imaginary direction of the scalar component of $\hat{N}^1$) at tree-level (see also section \[sugra\]). Thus loop corrections must be taken into account and these can indeed generate a small slope as required for slow-roll inflation. According to [@coleman_weinberg_potential], the Coleman-Weinberg one-loop effective potential is given by $$\label{vloop} V_{\text{loop}} = \frac{1}{64 \pi^2} \, \text{STr} \left[ {\cal M}^4 \left( \ln \frac{{\cal M}^2}{Q^2} - \frac{3}{2}\right) \right] \,,$$ with ${\cal M}$ denoting the mass matrix of the theory and $Q$ a renormalization scale. The $N$-dependent bosonic and fermionic mass terms generating a slope for the inflaton via the loop potential can be calculated from the scalar F-term potential $$\label{global_scalar_pot} V_F = \sum_i \left\| \; \frac{\delta W(\hat{\Phi})}{\delta \hat{\Phi}^i} \, \bigg\|_{\hat{\Phi} \rightarrow \Phi} \; \right\|^2 \,,$$ and the fermionic mass matrix $$({\cal M}_F)_{ij} = \frac{\delta^2 W(\hat{\Phi})}{\delta \hat{\Phi}^i \delta \hat{\Phi}^j} \, \bigg\|_{\hat{\Phi} \rightarrow \Phi}\,.$$ $\hat{\Phi}^i$ denote the superfields of the theory, $\Phi^i$ the respective scalar components. The relevant (i.e. $N$-dependent) contributions to the loop potential are give by the $\hat{H}, \, \hat{L}^j$ and $\hat{h}$ mass terms [^4] $$\label{eq_global_mass} \begin{split} (m_{h_{ a}}^{(S)})^2 &= (m_{h_{ a}}^{(P)})^2 = (m_{h_{ a}}^{(F)})^2 = \frac{1}{2} N^2 \sum_j \|(y_{\nu})_{1j}\|^2\,, \\ (m_{L^j_a}^{(S)})^2 &= (m_{L^j_a}^{(P)})^2 = (m_{L^j_a}^{(F)})^2 = \frac{1}{2} N^2 \|(y_{\nu})_{1j}\|^2\,, \\ (m_H^{(S)})^2 &= 2 \,\kappa^2 M^2\left(x - 1\right) \,, \\ (m_H^{(P)})^2 &= 2 \,\kappa^2 M^2\left(x + 1\right) \,, \\ (m_H^{(F)})^2 &= 2 \,\kappa^2 M^2 x \,, \end{split}$$ with $$\begin{aligned} x \equiv \frac{N^4 \lambda^2}{2 \,\kappa^2 M^2 M_P^2 }\,.\end{aligned}$$ Note that the $\hat{L}^j$ and $\hat{h}$ terms in the supertrace vanish since the degeneracy in the respective fermionic and bosonic masses leads to a cancellation of these contributions. Embedding this model in SUGRA provides the necessary stabilization of the scalar components of the $\hat{L}^j$ and $\hat{h}$ superfields during inflation and removes this degeneracy. However in the parameter range of interest, the contribution of the $\hat{L}^j$ and $\hat{h}$ terms to the loop potential turn out to be negligible (see also section \[sugra\]). In the following, we fix the renormalization scale to $Q = \sqrt{2} \,\kappa \,M$, which is the order of magnitude of the SUSY breaking scale. In our model, inflation ends when the $H$-field destabilizes at the critical value $N^c$ characterized by $m_H^{(S)} = 0$ and thus $x = 1$, $$(N^c)^2 = \sqrt{2} \,\frac{\kappa }{\lambda}\, M \, M_P \, .$$ We can now determine the observables describing the CMB fluctuations given by Eqs. , thus obtaining expressions depending on the phase transition scale $M$, the seesaw scale $m_N \sim \lambda$ and the vacuum energy parameter $\kappa$. With $x$ as defined above, a Taylor expansion in $1/x$ (with $1/x < 1$ because $N > N^c$ during inflation) yields $$V_{\text{loop}} \simeq \frac{\kappa^4 M^4}{8\, \pi^2} \, \ln x \, .$$ Inserting this into the equation of motion Eq. , with ${\cal H}$ approximately constant, gives the value for $N$ at ${\cal N}$ e-folds before the end of inflation: $$\label{eq_nc_global} N^2({\cal N}) = (N^c)^2 + \frac{{\cal N} \kappa^2}{\pi^2}\, M_P^2\, .$$ With this, the inflationary predictions of Eqs. are given by $$\label{pred_global} \begin{split} \Delta_s^2 &\simeq \frac{\pi^2 M^4 N^2}{3\, \kappa^2 M_P^6} \,, \\ n_s &\simeq 1 - \left(1 + \frac{3\kappa^2 }{4 \,\pi^2} \right) \frac{\kappa^2 M_P^2}{\pi^2 N^2}\,, \\ \alpha_s &\simeq - \left(1 + \frac{\kappa^2}{\pi^2} + \frac{ 2 \,\kappa^4}{8\, \pi^4} \right) \frac{\kappa^4 M_P^4}{\pi^4 N^4}\,, \\ r &\simeq 2\, \frac{\kappa^4 M_P^2}{\pi^4 N^2}\,, \\ n_t &\simeq - \frac{\kappa^4 M_P^2}{4\, \pi^4 N^2} \,. \\ \end{split}$$ Embedding in Supergravity {#sugra} ------------------------- We next consider a possible embedding of our model in SUGRA. We focus on a Kähler potential with the $\eta$-problem [@Copeland:1994vg; @Dine:1995uk] resolved by a shift symmetry [@Antusch:2009ef] in the inflaton direction $$\label{eq_K} K = \|\hat{S}\|^2 + \|\hat{H}\|^2 + \|\hat{h}\|^2 + \sum_i \frac{1}{2}\left(\hat{N}^i+(\hat{N}^i)^{\dagger}\right)^2 + \sum_j \|\hat{L}^j\|^2 + \frac{\kappa_{SH}}{M_P^2}\|\hat{S}\|^2\|\hat{H}\|^2 + \ldots\,.$$ The Kähler potential can be seen as a general expansion in the superfields of the theory with the additional feature of a shift symmetry which guarantees tree-level flat directions for the imaginary parts of the scalar components of $\hat{N}^i$ and thus possible inflaton directions. The $N$-dependent mass terms generating a slope for the inflaton via the loop potential can be calculated from the scalar F-term potential and the fermionic mass matrix as before by $$\label{sugrapotential} \begin{split} V_F &= e^K \left[ K^{i\overline{j}}D_i W D_{\overline{j}} W^* - 3\|W\|^2 \right] \; \Big\|_{\hat{\Phi} \rightarrow \Phi} \,, \\ ({\cal M}_F)_{ij} &= e^{K/2}(W_{ij} + K_{ij}W + K_iW_j + K_j W_i + K_i K_j W - K^{k\overline{l}}K_{ij\overline{l}} {\cal D}_k W) \; \Big\|_{\hat{\Phi} \rightarrow \Phi} \,. \end{split}$$ The relevant contributions to the loop potential are given by the $\hat{H}, \, \hat{L}^j$ and $\hat{h}$ mass terms which now obtain SUGRA corrections: $$\begin{split} (m_H^{(S)})^2 &= 2\, \kappa^2 M^2\left[ x - 1 +\left(\frac{M}{M_P}\right)^2 \left(1 - \kappa_{SH}\right)/2\right] \,,\\ (m_H^{(P)})^2 &= 2 \,\kappa^2 M^2\left[ x + 1 + \left(\frac{M}{M_P}\right)^2 \left(1 - \kappa_{SH}\right)/2\right] \,,\\ (m_H^{(F)})^2 &= 2 \,\kappa^2 M^2 x \,,\\ (m_{h_{ a}}^{(S)})^2 &= (m_{h_{ a}}^{(P)})^2 = \kappa^2\frac{M^4}{M_P^2} + \frac{1}{2}\, N^2 \sum_j \|(y_{\nu})_{1j}\|^2\,,\\ (m_{h_{ a}}^{(F)})^2 &= \frac{1}{2}\, N^2 \sum_j \|(y_{\nu})_{1j}\|^2\,, \\ (m_{L^j_a}^{(S)})^2 &= (m_{L^j_a}^{(P)})^2 = \kappa^2\frac{ M^4}{M_P^2} + \frac{1}{2}\, N^2 \|(y_{\nu})_{1j}\|^2\,, \\ (m_{L^j_a}^{(F)})^2 &= \frac{1}{2}\, N^2 \|(y_{\nu})_{1j}\|^2 \,. \end{split}$$ Comparing these expressions to the mass terms calculated in section \[susy\] we note some important points. A second mass scale, the scale of the SUGRA mass splitting $\kappa M^2 / M_P$, has appeared. However this scale is much smaller than the SUSY mass splitting scale $ \sqrt{2}\, \kappa \,M$ and thus we shall keep the latter scale as the renormalization scale. The additional mass splitting implies that the $\hat{h}$ and $\hat{L}^j$ contributions no longer cancel. However since the mass splitting is small compared to the SUSY mass splitting appearing in the $\hat{H}$ mass terms and since the remaining parts of the $\hat{h}$ and $\hat{L}^j$ mass terms are proportional to $\|(y_{\nu})_{1j}\|$ these additional contributions to the loop potential are negligible for $\tilde{y}_1 < 10^{-2}$. We will see later that this easily holds in our model. Furthermore, a new parameter has appeared in the loop potential: - The SUGRA correction parameter $\kappa_{SH}$ controls the SUGRA corrections to the loop potential. $\kappa_{SH}= 1$ recovers the phenomenology of global SUSY. Predictions \[predictions\] --------------------------- As in the globally supersymmetric case, predictions for observables describing the CMB spectrum can now be obtained by solving Eq. and evaluating Eqs. at the time when the CMB fluctuations exited the horizon. In the SUGRA scenario, this was done numerically for ${\cal N}_e = 60$. Fixing the phase transition scale $M$ by the experimental value for the amplitude of the CMB fluctuations $\Delta_s^2$, the behavior of the spectral index $n_s$, its running $\alpha_s$ and the tensor-to-scalar ratio $r$ is shown in Fig. \[fig\_pred\]. Interpreting the results visualized in Fig. \[fig\_pred\] and enforcing the experimental bounds of Eqs. implies restrictions on the model parameters. - The phase transition scale $M$ is fixed to $M \simeq 0.0032 \, M_P \simeq 8 \cdot 10^{15}$ GeV with a slight deviation in the region of large SUGRA corrections. This is consistent with the global SUSY calculation (from Eqs. and ) which gives $M^4 \simeq 3 \Delta_s^2/{\mathcal{N}_{\text{e}}}$ for $N(\mathcal{N}_{\text{e}}) \gg N^c$. - The width of the band in Fig. \[fig\_pred\] is given by the variation of the vacuum energy parameter $\kappa$. A priori we would expect $\kappa$ to be an ${\cal O}(1)$ parameter, thus we shall assume $0.5 < \kappa < 2$. In Fig. \[fig\_pred\], larger values of $\kappa$ are associated with SUGRA corrections becoming relevant at smaller values of $m_N$. In particular the tensor-to-scalar ratio $r$ is quite sensitive to $\kappa$ with $\kappa = 2$ leading to comparatively larger $r \sim {\cal O}(10^{-2})$. - Fig. \[fig\_pred\] also demonstrates the effect of the SUGRA correction parameter $\kappa_{SH}$. The respective quantities are marked in black for $\kappa_{SH} -1 = 0$ which corresponds to the globally supersymmetric limit and in blue (red) for $\kappa_{SH} -1 = -1 \, (+1)$ which corresponds to turning on the SUGRA corrections in the $\hat{H}$ mass terms with positive (negative) sign. In the considered SUGRA context the value of $\kappa_{SH}$ is a priori undetermined. Thus we would in general not expect to find global SUSY restored, which would correspond to $\kappa_{SH}$ exactly equal to one. - The second parameter controlling the effect of the SUGRA corrections is the seesaw scale $m_N = 2 \,\lambda\, M^2/M_P$. Fig. \[fig\_pred\] shows that these corrections are suppressed for small $m_N$, i.e. the observables are independent of $\kappa_{SH} $ for small values of the seesaw scale and the model predicts (for $M$ fixed by the experimental value of $\Delta_s^2$) $$0.98 < n_s < 1 \,,\qquad 3 \cdot 10^{-4} < \alpha_s < 0 \,, \qquad r < 0.013\,.$$ Note that in this case Eqs. hold. For very small values of the seesaw scale $m_N$, the spectral index $n_s$ approaches 1, which is not preferred by the latest WMAP data. On the other hand, all solutions with $\kappa_{SH} \neq 1$ leave the experimentally preferred region for the spectral index at large values of $m_N$. In combination, we find the preferred regions $2 \cdot 10^{10}\, \text{GeV} \lesssim m_N \lesssim 7 \cdot 10^{12}\, \text{GeV}$ for $\kappa_{SH}-1= + 1$ and $2 \cdot 10^{10}\, \text{GeV} \lesssim m_N \lesssim 2 \cdot 10^{12}\, \text{GeV}$ for $\kappa_{SH} -1= -1$, respectively. [^5] Motivated by the above results, we take the phase transition scale to be fixed at $M \simeq 8 \cdot 10^{15} \, \text{GeV}$ and concentrate on the parameter ranges $$\label{eq_infl_preferred} 2 \cdot 10^{10}\, \text{GeV} < m_N < 7 \cdot 10^{12} \, \text{GeV} \,, \qquad 0.5 < \kappa < 2 \,, \qquad \|\kappa_{SH} - 1\| > 0.1 \,,$$ in the further discussion. This yields $$-0.0004 \lesssim \alpha_s \lesssim 0.0002\, , \qquad r \lesssim 0.015 \,,$$ for the running of the spectral index and the tensor-to-scalar ratio. Reheating and Leptogenesis {#leptogenesis} ========================== After the end of the inflationary epoch the homogeneous classical fields and their quantum fluctuations evolve according to their respective equations of motion. The universe enters a matter dominated regime until the decay of heavy particles and the thermalization of the light particles result in the total energy density being dominated by radiation (Fig. \[fig\_numerics\]). This out-of-equilibrium decay of heavy particles can furthermore produce the necessary lepton asymmetry. In the following we will study these processes for the classical fields (reheating) and briefly comment on possible effects originating from their fluctuations (preheating). To this end we will start with the equations of motion for classical scalar fields, justify a simplification to Boltzmann equations and finally derive analytical expressions for the generated baryon asymmetry and the reheat temperature. We finish by commenting on preheating via parametric resonance and tachyonic preheating in this context. Classical Field Dynamics after Inflation {#Classical Field Dynamics after Inflation} ---------------------------------------- The equations of motion for the scalar fields can be obtained by adding a phenomenological decay term [@Kofman:1997yn] to Eq. thus giving $$\ddot{\phi} + 3\, {\cal H} \,\dot{\phi} + V'(\phi) + \Gamma \dot{\phi} = 0 \quad \text{with} \quad \phi =\{N,\, H\}\, .$$ Adding a Boltzmann equation for the quickly thermalizing[^6] ultra-relativistic particles and the Friedmann equation, we arrive at a closed set of differential equations: $$\begin{aligned} \label{n} \ddot{N} + 3\, {\cal H} \,\dot{N} + \frac{\partial V}{\partial N} + \Gamma_N \dot{N} &= 0\,, \\ \label{eq h} \ddot{H} + 3\, {\cal H}\, \dot{H} + \frac{\partial V}{\partial H} + \Gamma_H \dot{H} &= 0\, ,\\ \label{rrad} \dot{\rho_{R}} + 4\, {\cal H}\, \rho_{R} - \Gamma_N \rho_N - \Gamma_H \rho_H &= 0\,,\\ \label{friedmann} \frac{1}{3\,M_P^2}\left(\rho_N + \rho_H + \rho_R\right) &= {\cal H}^2 \,,\end{aligned}$$ with $\rho_R$, $\rho_N$ and $\rho_H$ denoting the energy densities of the ultra-relativistic particles, the $N$-field and the $H$-field respectively with $\rho_{N} + \rho_H = \dot{N}^2/2 + \dot{H}^2/2 + V(N,H)$. Having solved Eq. - , the lepton number density $n_L$ can be calculated from the Boltzmann equation $$\label{n asy} \dot{n_L} + 3\, {\cal H}\, n_L = \epsilon_{1} \Gamma_N \frac{\rho_N}{m_{N_1}} + \epsilon_{3} \Gamma_H \frac{\rho_H}{m_{N_3}} \,,$$ with the CP-violation per (s)neutrino decay $\epsilon$ for a hierarchical spectrum of right-handed neutrinos bounded by [@Covi:1996wh; @Davidson:2002qv; @Hamaguchi:2001gw] $$\label{epsilon} \epsilon_{i} < \frac{3}{8 \pi} \frac{\sqrt{\Delta m^2_{\text{atm}}} m_{N_i}}{ \langle v \rangle ^2} \,.$$ Here $\langle v \rangle$ denotes the vacuum expectation value of the up-type Higgs. The lepton asymmetry is typically normalized to the entropy density $s = 2 \pi^2 g_* T^3 / 45$ [@Bailin:2004zd] with the effective number of degrees of freedom $g_* = 915/4$ for the MSSM particles. The asymmetry $n_L / s$ is transferred to the baryon sector via sphaleron processes $n_B~=~\frac{C}{C - 1} n_L$ where $C$ is a number ${\cal O}(1)$ depending on the field content of the model and the temperature $T_{sph}$ when the sphalerons leave equilibrium. In the MSSM $C \sim 1/3$ [@Davidson:2008bu]. The quantity measured today is $\eta \equiv \frac{n_B}{n_{\gamma}}$ which can be calculated from the results above using the current conversion factor $s = 7.04 \, n_{\gamma}$ [@Bailin:2004zd]. The second important physical quantity in the theory of reheating is the temperature of the universe when the universe becomes radiation dominated ($\Gamma_N \approx {\cal H}$), the so-called reheat temperature. It can be calculated from the results above using $T^4 = \rho_R \cdot 30 / (g_* \pi^2)$. Eqs. - assume that both the $N$-particles and the $H$-particles decay into ultra-relativistic particles with the respective decay rates $\Gamma_N$ and $\Gamma_H$. We will now describe a possibility how to evaluate these quantities in our framework. At the beginning of the reheating phase ${\cal H} \gg \Gamma_N, \Gamma_H$ holds which implies that the decaying particles are damped predominantly by Hubble expansion, not by decays, and the produced ultra-relativistic particles are strongly diluted. The decays become significant for $t \sim {\cal H}^{-1} \sim \text{min} \{ \Gamma_N^{-1}, \Gamma_H^{-1} \}$. At this stage it is safe to assume $N \ll 1$. In this limit the respective decay rates derived from Eq. are $$\begin{aligned} \Gamma_N &\simeq \lim_{N \rightarrow 0} \frac{(y_{\nu} y_{\nu}^{\dagger})_{11}}{4 \pi} m_N = \frac{(y_{\nu} y_{\nu}^{\dagger})_{11}}{2 \pi} \frac{\lambda}{M_P} M^2 \,, \\ \Gamma_H &\simeq \lim_{N \rightarrow 0} \min\{ \Gamma_{H \rightarrow N_3 N_3}, \Gamma_{N_3 \rightarrow hL} \} = \frac{2}{\pi} \frac{\lambda_{33}}{M_P} M^2 \min \{ (y_{\nu} y_{\nu}^{\dagger})_{33}, \frac{\lambda_{33} \kappa M}{16 M_P} \} \,, \end{aligned}$$ with the sneutrino $N$ decaying directly into lepton and Higgsino or slepton and Higgs and the $H$ particles decaying predominantly into the heaviest fermionic neutrino (assuming this is not strongly suppressed by kinematics) [^7], which then in turn decays into lepton and Higgs or slepton and Higgsino. Note that the Boltzmann equations and imply a splitting of the total matter energy density $\rho_M$ into $\rho_N$ and $\rho_H$, which is not straightforward if the respective degrees of freedom are highly coupled. However, since $\Gamma_N \ll \Gamma_H$ in our setting [^8] in the preferred region of parameter space (see section \[infl\_lepto\]), any radiation energy density produced by $H$-decays will be strongly diluted during the following matter dominated phase governed by oscillations of the sneutrino. With $\rho_M \approx \rho_N$ shortly after the end of inflation due to the strong damping of the $H$-field (see below) we can thus substitute and by $$\begin{aligned} \label{rrad2} &\dot{\rho_{R}} + 4 \,{\cal H}\, \rho_{R} - \Gamma_N \rho_M \simeq 0 \,, \\ \label{n asy 2} &\dot{n_L} + 3\, {\cal H}\, n_L \simeq \epsilon_{1} \Gamma_N \frac{\rho_M}{m_{N_1}} \,, \end{aligned}$$ without introducing a significant error for the finally produced radiation density. The effect of this approximation on the Hubble expansion rate is negligible since the Friedmann equation is predominantly governed by $\rho_M$ for $t < \Gamma_N^{-1}$. Solving Eqs. , , , and numerically, we obtain the time evolution of scalar fields, the energy densities, the scale factor and the lepton asymmetry, respectively. The former two are displayed in Fig. \[fig\_numerics\]. The regime of reheating is characterized by oscillating scalar fields and can be divided into distinct phases: After the end of inflation both $N$ and $H$ fall to their true minimum and begin to oscillate. After only a few oscillations the classical field $H$ settles at its minimum and the dynamics of the system is governed by the oscillation of the $N$ field. The further evolution of the $N$ oscillations is governed by Hubble damping. As long as ${\cal H} \gg \Gamma_N$ the universe is governed by (damped) oscillating scalar fields which can be interpreted as (decaying) heavy particles. This implies a matter dominated universe out of thermal equilibrium. Ultra-relativistic particles are produced through the decays of the heavy particles, however they are diluted by the expansion of the universe. As soon as ${\cal H} \approx \Gamma_N$ the radiation energy density becomes dominant and the light particles begin to thermalize. This marks the end of reheating and determines the reheat temperature and the asymmetry $n_L/s$. Simplified Treatment with Boltzmann Equations {#Simplified Treatment with Boltzmann Equations} --------------------------------------------- Since the set of equations , , and is quite involved, a common attempt in the literature (e.g. [@Kolb:1990vq]) is to simplify these equations by time-averaging the equations of motion of the scalar fields. The result is a set of Boltzmann equations for the matter energy density $\rho_M = \rho_N + \rho_H$ and the radiation energy density $\rho_R$ completed by the Friedmann equation $$\begin{aligned} \label{rhon} \dot{\rho_{M}} + 3\, {\cal H}\, \rho_{M} &= -\Gamma_N \rho_{M}, \\ \dot{\rho_R} + 4\, {\cal H}\, \rho_R &= \Gamma_N \rho_{M}, \\ \label{H} \frac{1}{3\, M^2_P}\left(\rho_M + \rho_R\right) &= {\cal H}^2 .\end{aligned}$$ The lepton asymmetry is determined by Eq. . The big advantage is that these equations have approximate analytical solutions. However their derivation (see e.g. [@Kolb:1990vq]) implies an important assumption concerning the scalar potential $V(N,H)$. In order to rewrite the time-averaged kinetic energy density in terms of the total energy density by exploiting the Virial theorem we must assume that we can write the scalar potential as $V(N,H) = V_N(N) + V_H(H)$ with $V_N \sim N^r$ and $V_H \sim H^r$. Eqs. - are obtained with $r = 2$. Numerical simulations of the full system , , , and show that this assumption is not justified in the early oscillation phase in the model described by Eq. since the large oscillations of the $N$-field result in a highly coupled system with higher orders terms in the scalar potential playing a non-negligible role. However they do also show that for $t \approx \Gamma^{-1} \approx {\cal H}^{-1}$ the simpler system of differential equations - does give a good approximation. This is the point of time relevant for the predictions of the reheating phase. Having seen that the results of the numerical solutions to the full field equations for $t \approx \Gamma^{-1}$ can be approximated reasonably well by the simpler set of Boltzmann differential equations, we can now find approximate analytical solutions to the latter and use these expressions to find estimates for the reheat temperature and the produced baryon asymmetry $$\begin{aligned} \label{Trh} T_{RH} &\approx \left(\frac{9}{4 \pi^{4} g_}\right)^{1/4} \sqrt{(y_{\nu} y_{\nu}^{\dagger})_{11} m_N M_P}\,, \\ \label{nB} \frac{n_B}{n_{\gamma}} (t_0) &\approx 3.45 \,\frac{C}{C-1} \, g_^{-1/4} \epsilon\, \sqrt{\frac{(y_{\nu} y_{\nu}^{\dagger})_{11}}{m_N/M_P}}\,.\end{aligned}$$ Combining and reproduces the familiar relation $n_B/s \sim \epsilon\ T_{RH}/ m_N$ (see e.g. [@Bailin:2004zd; @Kolb:1990vq]). These results must be compared with existing bounds on the reheating process. The WMAP 7 year data combined with measurements of the baryon acoustic oscillations and todays Hubble parameter imply $\frac{n_B}{n_{\gamma}} = (6.19 \pm 0.15) \cdot 10^{-10}$ [@Komatsu:2010fb], thus yielding $T_{RH} > 1.4 \cdot 10^6$ GeV. Additionally, the reheat temperature is bounded from above by the so-called gravitino problem [@Khlopov:1984pf; @Ellis:1984eq; @Ellis:1984er; @Moroi:1993mb; @Kawasaki:2004yh]. A high reheat temperature would result in an overproduction of gravitinos. If these are stable, then the fact that their energy density can not be larger than the present total energy density of the universe leads to a bound on the reheat temperature in terms of the gravitino mass $m_{3/2}$. On the other hand, if gravitinos are not stable, they can either decay before or during and after the Big Bang Nucleosynthesis (BBN). In the former case (i.e. heavy gravitinos), with R-parity conserved the gravitinos will decay into the lightest supersymmetric particle (LSP) and their production is thus constrained by the dark matter abundance. This yields a fairly model independent bound of $T_{RH} < 2 \cdot 10^{10}$ GeV for an LSP mass of about 100 GeV to 150 GeV. In the latter case (i.e. light gravitinos), the decay of the gravitinos would alter the outcome of BBN and create a conflict between BBN predictions and observations. This yields even stronger, however model dependent, constraints on the reheat temperature. Combining these arguments yields a constraint on the reheat temperature of typically $T_{RH} < 10^7 - 10^{10} \, \text{GeV}$, depending mainly on the model under consideration and on the value of $m_{3/2}$. The resulting preferred region in ($m_N, \, \tilde{y}_1$)-parameter space is depicted in blue in Fig. \[fig\_infl\_lepto\]. Remarks on Preheating {#Remarks on Preheating} --------------------- Note that throughout this chapter we have focussed on the evolution of the homogeneous fields $N$ and $H$. It has been pointed out that under certain circumstances this might not be sufficient, since modes with $k \neq 0$ of all fields in the model can be strongly excited at the end of inflation and before the beginning of reheating in a process referred to as preheating. There are two types of preheating worth mentioning in the context of hybrid inflation, namely preheating via parametric resonance [@Kofman:1997yn; @BasteroGil:1999fz; @GarciaBellido:2000dc; @Kofman:1994rk] and tachyonic preheating [@GarciaBellido:2001cb; @Felder:2000hj]. In the former case, the coupling of fermions and bosons to the oscillating inflaton field results in oscillating mass terms for these particles. Solving the respective equations of motion (roughly the equation of an harmonic oscillator with an oscillating mass as described, e.g., by the Mathieu equation) can yield explosive particle production. However, in the region of parameter space of interest to us, any heavy particles that are produced by this mechanism will decay back into heavy (s)neutrinos or into radiation. The radiation produced directly or indirectly through this process at the beginning of the reheating phase will however be strongly diluted during the ongoing matter dominated phase and thus be insignificant for the outcome of the reheating phase. Thus in our model, parametric resonance will not affect the results discussed above, mainly due to the structure of the mass spectrum and the very small effective Yukawa coupling $\tilde{y}_1$. Tachyonic preheating occurs when the squared mass of the $H$ field becomes negative, triggering the waterfall ending inflation. Modes of the $H$ field with $k < \|m^{(S)}_H\|$ grow exponentially [^9], causing particle production of bosonic and fermionic fields coupled to the waterfall field [@GarciaBellido:2001cb] and creating an inhomogeneous field $H(x,t)$ which can cause the formation of topological defects when the waterfall occurs [@Felder:2000hj]. It was stated in [@Felder:2000hj] that the production of fermions and bosons coupling to the waterfall field with a coupling strength $g$ is suppressed by $\rho_{B,F}/\rho_V \sim 10^{-3} \, g$ with $\rho_V$ denoting the energy density during inflation. Thus in the parameter region of interest, this is negligible in our model. On the other hand, the production of topological defects could indeed dominate the evolution of the universe in an early stage. However, since we have not observed any topological defects yet, a mechanism to prevent or dilute these objects (e.g. a preferred waterfall direction or a slight shift of the potential energy of the discrete vacua) is typically implemented. We will assume that the higher dimensional operators denoted by dots in Eq. provide such a solution so that at some time after the waterfall, the universe is dominated by the lightest right-handed sneutrino. The evolution from this point on is correctly described by the classical theory of reheating, as discussed above. Other possible scenarios in which the evolution of the universe may not be dominated by the homogeneous component of the inflaton field remain to be explored in this context. Summary and Conclusions: Combining Inflation and Leptogenesis \[infl\_lepto\] ============================================================================= In sections \[inflation\] and \[leptogenesis\] we have investigated the conditions under which inflation, with primordial perturbations in accordance with the latest WMAP results, as well as successful leptogenesis can be realized simultaneously in simple models of sneutrino hybrid inflation as outlined in section \[framework\]. The combined results are summarized in Fig. \[fig\_infl\_lepto\]. The dynamics of inflation is governed by the scale $M$ of the phase transition ending hybrid inflation, the mass of the lightest right-handed (s)neutrino $m_N$, the vacuum energy parameter $\kappa$ (= waterfall field self coupling) and the parameter $\kappa_{SH}$ controlling the SUGRA corrections. In principle, terms depending on the neutrino Yukawa coupling matrix could contribute, too. However, in our case the comparatively small first generation Yukawa couplings make these contributions negligible. With $M$ fixed by the amplitude of the scalar CMB fluctuations and $\kappa \sim {\cal O} (1)$, the spectrum of the CMB fluctuations is primarily dependent on the lightest right-handed (s)neutrino mass $m_N$. For large values of $m_N$, SUGRA corrections controlled by $\kappa_{SH}$ become important, with the sign of these contributions depending on the sign of $\kappa_{SH} - 1$. For the spectral index, its running and the tensor-to-scalar ratio the predictions are shown in Fig. \[fig\_pred\]. Recent WMAP observations constrain the preferred region for the spectral index $n_s$, thus imposing a constraint on the preferred region for $m_N$. For $\kappa = 1$ and $\kappa_{SH} -1 = \pm 1$ this is marked in grey in Fig. \[fig\_infl\_lepto\]. On the other hand, the decisive quantities of reheating and leptogenesis, namely the reheat temperature $T_{RH}$ and the baryon asymmetry $n_B/n_{\gamma}$ depend on the effective first generation neutrino Yukawa coupling $\tilde{y}_1$, the CP asymmetry $\epsilon_1$ and the mass of the lightest right-handed (s)neutrino $m_N$ (see Eqs. and ). The latter parameter is thus the link between inflation and leptogenesis. The preferred region of parameter space resulting from bounds on these quantities is marked in blue in Fig. \[fig\_infl\_lepto\]. It is bounded from below by the experimental value of the baryon asymmetry measured by WMAP and by an upper bound on the CP-violation per (s)neutrino decay Eq. . From above it is bounded by constraints imposed on the reheat temperature from the gravitino problem. Since these are model dependent, we have plotted the regions satisfying $T_{RH} < 10^{10}, \, 10^9, \, 10^8, \, 10^7$ GeV in different shadings. Note that a higher reheat temperature at a fixed value for $m_N$ automatically corresponds to a smaller value of $\epsilon_1$ in order to match the measured baryon asymmetry. The resulting preferred region in parameter space implies an effective first generation Yukawa coupling $\tilde{y}_1 = {\cal O}(10^{-9} - 10^{-4})$. The upper part of this range is of the same order as the first family quark and charged lepton Yukawa couplings, which in the MSSM with moderate $\tan \beta$ are of the order $10^{-4}$ to $10^{-6}$. Throughout this paper, we have assumed nonthermal leptogenesis and hierarchical masses of left-handed as well as right-handed neutrinos. Assuming that the light neutrinos obtain masses via a type I seesaw mechanism[^10], both assumptions depend on the value of the effective light neutrino mass parameter (also dubbed washout parameter) $\widetilde{m}_1 \equiv \tilde{y}_1^2 \langle v \rangle^2 / m_N$. More explicitly, one can easily see from Eq. that $(T_{RH}/m_N)^2 \approx 4.0 \cdot 10^2 \, \widetilde{m}_1/\text{eV}$. Lines of constant $\widetilde{m}_1$ are marked in red in Fig. \[fig\_infl\_lepto\], corresponding to a fixed relation between $T_{RH}$ and $m_N$. Simultaneously, they give the order of magnitude for the mass of the left-handed neutrino $m_{\nu_1} \sim \widetilde{m}_1$. In the preferred region of parameter space, we find $\widetilde{m}_1 < 3.4 \cdot 10^{-5}$ eV, thus implying nonthermal leptogenesis with $T_{RH} \ll m_N$ and $m_{\nu_1} \ll \sqrt{\Delta m^2_{\text{atm, sol}}}$. Finally, we want to comment on possible extensions of this scenario and the significance of cosmological observations in the near future. In Fig. \[fig\_infl\_lepto\] we have set $\kappa = 1$. Allowing for $0.5 < \kappa < 2$ gives qualitatively the same picture (see section \[inflation\], in particular Fig. \[fig\_pred\]) with a somewhat shifted grey region. For example, for $\kappa = 2$ the grey region is extended to the left to $m_N = 2 \cdot 10^{10}$ GeV whereas for $\kappa = 0.5$ it is extended to the right to $m_N = 7 \cdot 10^{12}$ GeV for $\kappa_{SH} -1= +1$.[^11] Another interesting possibility would arise if the experimentally preferred region for the spectral index was raised, favoring a spectral index closer to 1. This would lower the preferred range for the lightest (s)neutrino mass $m_N$ significantly and thus open up the region of thermal leptogenesis and allow for $m_{\nu_1} \sim {\cal O}(\sqrt{\Delta m^2_{\text{atm, sol}}})$. The forthcoming results of the Planck satellite will make the requirements for $m_N$ more accurate. In summary, we have pointed out that successful sneutrino hybrid inflation and leptogenesis can be achieved in this framework, and that combining both imposes requirements on the parameters of the underlying particle physics model. We obtain a mass for the lightest right-handed (s)neutrino of $m_N = {\cal O}(10^{10} - 10^{13})$ GeV, an effective first generation neutrino Yukawa coupling $\tilde{y}_1 = {\cal O}(10^{-9} - 10^{-4})$ and a very light left-handed neutrino with $m_{\nu_1} < {\cal O} (10^{-4})$ eV. Furthermore, we find that leptogenesis occurs via nonthermal leptogenesis (with $T_{RH}/m_N < 0.1$ for $\kappa = 1$). Concerning the spectrum of the CMB fluctuations, we predict a running of the spectral index of $-0.0004 < \alpha_s < 0.0002$ and a tensor-to-scalar ratio of $r \lesssim 0.015$. Our results provide a guideline for the construction of explicit particle physics models incorporating sneutrino hybrid inflation and subsequent baryogenesis via nonthermal leptogenesis. Acknowledgments {#acknowledgments .unnumbered} =============== We would like to thank Alejandro Ibarra and Koushik Dutta for discussions. We acknowledge partial support by the DFG cluster of excellence “Origin and Structure of the Universe”. [999]{} A. H. Guth, Phys. Rev. D23 (1981), 347–356; A. D. Linde, Phys. Lett. B108 (1982), 389–393; A. Albrecht and P. J. Steinhardt, Phys. Rev. Lett. 48 (1982), 1220–1223; for a review containing an extensive list of references, see e.g.: D. H. Lyth and A. Riotto, Phys. Rept. [314]{} (1999) 1. For textbook reviews on inflation see: A. R. Liddle and D. H. Lyth, “Cosmological inflation and large-scale structure,” [Cambridge, UK: Univ. Pr. (2000) 400 p]{}; A. D. Linde, “Particle Physics and Inflationary Cosmology,” \[arXiv:hep-th/0503203\]; V. Mukhanov, “Physical Foundations of Cosmology,” [Cambridge, UK: Univ. Pr. (2005) 421 p]{} D. Bailin and A. Love, [Bristol, UK: IOP (2004) 313 p]{}. D. Baumann, arXiv:0907.5424 \[hep-th\]. A. Mazumdar and J. Rocher, arXiv:1001.0993 \[hep-ph\]. M. Fukugita and T. Yanagida, Phys. Lett. B [174]{}, 45 (1986). P. Minkowski, Phys. Lett. B [67]{} (1977) 421; M. Gell-Mann, P. Ramond and R. Slansky in Sanibel Talk, CALT-68-709, Feb 1979, and in [Supergravity]{} (North Holland, Amsterdam 1979); T. Yanagida in [Proc. of the Workshop on Unified Theory and Baryon Number of the Universe]{}, KEK, Japan, 1979; S.L.Glashow, Cargese Lectures (1979); R. N. Mohapatra and G. Senjanovic, Phys. Rev. Lett. [44]{} (1980) 912; J. Schechter and J. W. Valle, Phys. Rev. D [25]{} (1982) 774. M. C. Chen, arXiv:hep-ph/0703087. S. Davidson, E. Nardi and Y. Nir, Phys. Rept. [466]{}, 105 (2008) \[arXiv:0802.2962 \[hep-ph\]\]. H. Murayama, H. Suzuki, T. Yanagida and J. Yokoyama, Phys. Rev. Lett. [70]{} (1993) 1912. J. R. Ellis, M. Raidal and T. Yanagida, Phys. Lett. B [581]{}, 9 (2004) \[arXiv:hep-ph/0303242\]. S. Antusch, M. Bastero-Gil, S. F. King and Q. Shafi, Phys. Rev. D [71]{}, 083519 (2005) \[arXiv:hep-ph/0411298\]. A. D. Linde, Phys. Lett. B [249]{}, 18 (1990). A. D. Linde, Phys. Lett. B [259]{}, 38 (1991). E. J. Copeland, A. R. Liddle, D. H. Lyth, E. D. Stewart and D. Wands, Phys. Rev. D [49]{}, 6410 (1994) \[arXiv:astro-ph/9401011\]. A. D. Linde and A. Riotto, Phys. Rev. D [56]{}, 1841 (1997) \[arXiv:hep-ph/9703209\]. S. Antusch, K. Dutta and P. M. Kostka, AIP Conf. Proc. [1200]{}, 1007 (2010) \[arXiv:0908.1694 \[hep-ph\]\]. M. Kawasaki, M. Yamaguchi and T. Yanagida, Phys. Rev. Lett. [85]{}, 3572 (2000) \[arXiv:hep-ph/0004243\]. M. Yamaguchi and J. Yokoyama, Phys. Rev. D [63]{}, 043506 (2001) \[arXiv:hep-ph/0007021\]. M. Kawasaki, M. Yamaguchi and T. Yanagida, Phys. Rev. D [63]{}, 103514 (2001) \[arXiv:hep-ph/0011104\]. S. Antusch, K. Dutta and P. M. Kostka, Phys. Lett. B [677]{}, 221 (2009) \[arXiv:0902.2934 \[hep-ph\]\]. S. Antusch, M. Bastero-Gil, K. Dutta, S. F. King and P. M. Kostka, JCAP [0901]{}, 040 (2009) \[arXiv:0808.2425 \[hep-ph\]\]. S. Antusch, M. Bastero-Gil, J. P. Baumann, K. Dutta, S. F. King and P. M. Kostka, arXiv:1003.3233 \[hep-ph\]. S. Antusch, S. F. King, M. Malinsky, L. Velasco-Sevilla and I. Zavala, Phys. Lett. B [666]{}, 176 (2008) \[arXiv:0805.0325 \[hep-ph\]\]. W. J. Percival [et al.]{}, Mon. Not. Roy. Astron. Soc. [401]{}, 2148 (2010) \[arXiv:0907.1660 \[astro-ph.CO\]\]. A. G. Riess [et al.]{}, Astrophys. J. [699]{}, 539 (2009) \[arXiv:0905.0695 \[astro-ph.CO\]\]. WMAP cosmological parameters. Model: lcdm+sz+lens. Data: wmap7+bao+h0\ http://lambda.gsfc.nasa.gov/product/map/current/params/ lcdm$\_$sz$\_$lens$\_$wmap7$\_$bao$\_$h0.cfm E. J. Weinberg, arXiv:hep-th/0507214. G. Gamberini, G. Ridolfi and F. Zwirner, Nucl. Phys. B [331]{} (1990) 331. S. Weinberg, Phys. Rev. D [7]{}, 2887 (1973). M. Dine, L. Randall and S. D. Thomas, Phys. Rev. Lett. [75]{}, 398 (1995) \[arXiv:hep-ph/9503303\]. L. Kofman, A. D. Linde and A. A. Starobinsky, Phys. Rev. D [56]{}, 3258 (1997) \[arXiv:hep-ph/9704452\]. R. Allahverdi and A. Mazumdar, JCAP [0610]{}, 008 (2006) \[arXiv:hep-ph/0512227\]. R. Allahverdi and A. Mazumdar, Phys. Rev. D [76]{}, 103526 (2007) \[arXiv:hep-ph/0603244\]. L. Covi, E. Roulet and F. Vissani, Phys. Lett. B [384]{} (1996) 169 \[arXiv:hep-ph/9605319\]. S. Davidson and A. Ibarra, Phys. Lett. B [535]{}, 25 (2002) \[arXiv:hep-ph/0202239\]. K. Hamaguchi, H. Murayama and T. Yanagida, Phys. Rev. D [65]{}, 043512 (2002) \[arXiv:hep-ph/0109030\]. E. W. Kolb and M. S. Turner, Front. Phys. [69]{}, 1 (1990). E. Komatsu [et al.]{}, arXiv:1001.4538 \[astro-ph.CO\]. M. Y. Khlopov and A. D. Linde, Phys. Lett. B [138]{}, 265 (1984). J. R. Ellis, J. E. Kim and D. V. Nanopoulos, Phys. Lett. B [145]{}, 181 (1984). J. R. Ellis, D. V. Nanopoulos and S. Sarkar, Nucl. Phys. B [259]{}, 175 (1985). T. Moroi, H. Murayama and M. Yamaguchi, Phys. Lett. B [303]{}, 289 (1993). M. Kawasaki, K. Kohri and T. Moroi, Phys. Lett. B [625]{}, 7 (2005) \[arXiv:astro-ph/0402490\]. M. Bastero-Gil, S. F. King and J. Sanderson, Phys. Rev. D [60]{}, 103517 (1999) \[arXiv:hep-ph/9904315\]. J. Garcia-Bellido, S. Mollerach and E. Roulet, JHEP [0002]{}, 034 (2000) \[arXiv:hep-ph/0002076\]. L. Kofman, A. D. Linde and A. A. Starobinsky, Phys. Rev. Lett. [73]{}, 3195 (1994) \[arXiv:hep-th/9405187\]. J. Garcia-Bellido and E. Ruiz Morales, Phys. Lett. B [536]{}, 193 (2002) \[arXiv:hep-ph/0109230\]. G. N. Felder, J. Garcia-Bellido, P. B. Greene, L. Kofman, A. D. Linde and I. Tkachev, Phys. Rev. Lett. [87]{}, 011601 (2001) \[arXiv:hep-ph/0012142\]. J. McDonald, Phys. Rev. D [66]{}, 043525 (2002) \[arXiv:hep-ph/0105235\]. [^1]: In the context of chaotic inflation, shift symmetry has been used e.g. in [@Kawasaki:2000yn; @Yamaguchi:2000vm; @Kawasaki:2000ws]. [^2]: At this point we will assume that inflation proceeds along the imaginary direction of the complex scalar sneutrino field. We will see later that this can be obtained by a shift symmetry in the Kähler potential, which protects this direction against the SUGRA $\eta$-problem as was demonstrated in [@Antusch:2009ef]. [^3]: In this case, the terms $\hat{H}^2$ might be replaced by a different combination of fields, i.e. by $\hat{H}_1\hat{H}_2$, in order to form a singlet under the symmetry group. The symmetry could, e.g., be a unifying gauge symmetry or a family symmetry. [^4]: Here the index (S) (for scalar) denotes mass terms of the real parts of the complex spin-0 components of the superfields whereas the index (P) (for pseudoscalar) marks the mass terms corresponding to the purely imaginary parts. The index (F) marks the mass terms of the fermionic components of the superfields. [^5]: Equivalently, we obtain $5 \cdot 10^{-4} < \lambda < 0.13 (0.043)$ for $\kappa_{SH}-1 = + 1$ and $\kappa_{SH} -1 = -1$, respectively. [^6]: In principle, thermalization in the MSSM could be delayed if MSSM flat directions obtain large vevs (see e.g. [@Allahverdi:2005mz; @Allahverdi:2007zz]). However in our scenario, this is not the case since only the $N$ direction is protected against large SUGRA corrections. [^7]: Note that $\lim_{N \rightarrow 0} \Gamma_{H \rightarrow N_i N_i} \simeq \frac{\lambda_{ii}^2 \kappa M^3}{8 \pi M_P^2} (1 - 2\frac{\lambda_{ii}^2 M^2}{\kappa^2 M_P^2}) (1 - 4 \frac{\lambda_{ii}^2 M^2}{\kappa^2 M_P^2})^{1/4}$. The expressions in brackets emphasize that the decay is kinematically possible if $m_H^{(S)} > 2 m_{N_i}^{(F)}$. However in the parameter range of interest, $\lambda_{ii}/M_P \ll \kappa/M$ holds, as can be seen from Eqs. . Thus we have $\Gamma_{H \rightarrow N_i N_i} \simeq \frac{\lambda_{ii}^2 \kappa M^3}{8 \pi M_P^2}$. [^8]: The assumption $\Gamma_{N_1} < \Gamma_{N_3}$ (see section \[framework\]) implies $\lambda_{33} (y_{\nu} y_{\nu}^{\dagger})_{33} > \lambda_{11} (y_{\nu} y_{\nu}^{\dagger})_{11}$. The assumption of hierarchical heavy neutrinos implies $\lambda_{33} \gg \lambda_{11}$. [^9]: It was pointed out in [@McDonald:2001iv] that in some hybrid inflation models a fragmentation of the inflaton condensate can occur, causing the evolution of the universe to be dominated by these ’lumps’ instead of by the homogeneous component of the inflaton field. However this ’lump’ formation requires a flatter than $\phi^2$ potential (with $\phi = \{N, \ H\}$), which does not appear in our model as can easily be checked from Eq. . [^10]: This implies a mass matrix for the left-handed neutrinos $(m_{\nu})_{ij} = -(y_{\nu}^T M^{-1} y_{\nu})_{ij} \langle v \rangle^2/2$. [^11]: For $\kappa_{SH} = -1$ the respective region is extended to $m_N = 2 \cdot 10^{12}$ GeV.	{ "pile_set_name": "ArXiv" }
--- abstract: 'In this demonstration, we will present the world’s first molecular multiple-input multiple-output (MIMO) communication link to deliver two data streams in a spatial domain. We show that chemical signals such as concentration gradients could be used in MIMO fashion to transfer sequential data. Until now it was unclear whether MIMO techniques, which are used extensively in modern radio communication, could be applied to molecular communication. In the demonstration, using our devised MIMO apparatus and carefully designed detection algorithm, we will show that we can achieve about 1.7 times higher data rate than single input single output (SISO) molecular communication systems.' author: - '\' bibliography: - 'IEEEabrv.bib' - 'Ref.bib' title: Molecular MIMO Communication Link --- Background ========== Molecular communication is a biologically inspired form of communication, where chemical signals are used to transfer information [@Farsard_arXiv14]. Molecular communication could be used in places where radio based communication fails or is inefficient: for example, city infrastructure monitoring in smart cities at macroscale [@Wassell10], and body area nanonetworks for health monitoring and targeted drug delivery at microscale [@ata12CM]. Most previous work on molecular communication has focused on microscale systems and nanonetworks such as diffusion-based intra- and inter-cell communications [@lla12WC]. Most of these works have been theoretical, and only recently the have been experimental implementations of molecular communication systems [@Farsad13], where reliable communication was achieved. There have also been a number of attempts at mimicking pheromone-based communication [@col09]. Although these systems were not designed for transferring sequential data. More recently, it was demonstrated that the nonlinearity in [@Chae_JSAC14] could be modelled as noise. In our prior work, the world’s first macro scale molecular communication link [@Farsad13] was demonstrated at IEEE INFOCOM 2014 [@infocom_demo]. In this demonstration we will show the world’s first molecular multiple-input multiple-output (MIMO) communication link, where the transmitter and the receiver are equipped with multiple sprays and sensors to further increase the data rate. MIMO is a technique, which is used in modern radio communication to increase transmission data rate. The feasibility of using MIMO in molecular communication, however, has not been demonstrated in the past. In our MIMO design, we implement our own signal detection algorithms that are different from classical RF MIMO communication. The algorithms will be described in more detail in an extended draft. Test-bed and Demonstration Output ================================= Hardware Layout --------------- Our system consists of a molecular MIMO transmitter and receiver as shown in Fig. \[Fig:test\_bed\_mimo\]. The propagation channel in between is several meters of free-space. The transmission consists of: 1) a simple user interface for text entry, 2) a microcontroller for executing transmitter algorithms, 3) two reservoir for chemicals, and iv) two chemical release mechanism (i.e. two sprays). At the receiver, the hardware consists of: 1) two chemical sensors for MIMO operation, 2) two microcontrollers that demodulate and decode the signal, and 3) a computer for displaying and visualizing results. \[MIMO test bed\] Although any sequential data could be transported by our setup, for this demonstration we will consider short string of text data. Text-based information is very important to sensor networks and command-based communication systems. The information delivery rate for the prior platform we demonstrated at IEEE INFOCOM 2014 was low because of inexpensive components and use of single type of chemical [@infocom_demo]. This year, we utilize spatial domain to increase the transmission data rate. Unlike prior work in RF communication, non-coherent detection is required since the coherent time of the molecular channel is zero. Health and Safety ----------------- We have the same demonstration conditions as described in [@infocom_demo]. As part of the demonstration low volumes of alcohol is diffused in open air. There will not be any chemical risks since the alcohol used will be safe for human consumption and of a small quantity and concentration. Moreover, we will perform the demonstration behind a transparent shielded screen to respect religious sensitivities and avoid any unwanted alcohol odours in the conference venue. Application =========== The main goal of the demonstration is to show that messages can be continuously and reliably carried via molecular MIMO. Fig. \[Fig:sampleTxRx\] shows the sample text entered at the transmitter and Fig. \[Fig:receiver\_screen\] illustrates the receiver screen. As can be seen from the figure, we decode three alphabets at each receiver sensor. Table I compares the transmission time and the data rate of SISO and MIMO systems, from which we observe that the MIMO system show 1.7 times higher data rate than the SISO system. The data rate enhancement is not exactly two times even though we use two sprays and two sensors. This is because of the need for interference compensation and the system overhead due to start and end of communication indicators. ![Sample text message entered at the transmitter.[]{data-label="Fig:sampleTxRx"}](text_entry.png){width="3.1in"} Most mathematical models developed for molecular communication have relied on Fick’s diffusion equation and Monte-Carlo simulation [@Yilmaz_simul]. Moreover, most prior work has assumed perfect transmission, propagation, and reception [@Chae_JSAC13]. These assumptions, however, do not hold in practice, and more realistic models based on experimental data are necessary [@Chae_JSAC14]. Thus, we believe that more accurate MIMO channel models can also be derived through using experimental data obtained from our MIMO platform. In terms of industrial interest, our platform could be extended towards structural high speed health monitoring (smart cities) applications, and for transmitting commands to robots in subterranean areas [@Wassell10]. Conclusion ========== In this demonstration, we present the first macroscale molecular MIMO communication system that could reliably transmit short text messages. Our goal is to show that molecular communication can be used as an alternative to radio communication in challenging environments. To improve the low transmission rate of molecular communication, we implement novel molecular MIMO detection algorithms. The main challenge in our design was implementing a signal separation algorithm for the molecular MIMO channel, since MIMO detection algorithms for classical RF communications could not be directly applied. We hope to motivate researches, and fill a gap between theory and practice of molecular communication.\ Acknowledgement =============== This research is funded by the MSIP (Ministry of Science, ICT & Future Planning), under the ÔÔIT Consilience Creative ProgramÕÕ (NIPA-2014-H0201-14-1002) supervised by the NIPA (National IT Industry Promotion Agency) and by the Basic Science Research Program (2014R1A1A1002186) funded by the MSIP, through the National Research Foundation of Korea and by the ICT R& D programme of MSIP/IITP. The authors would like to thank C. Kim for his help in implementing the hardware. ![On the receiver screens the decoded characters are seen.[]{data-label="Fig:receiver_screen"}](receiver_screen2.pdf){width="1.0\columnwidth"} Type Transmission time (s) Data rate (bps) ------ ----------------------- ----------------- SISO 108 0.28 MIMO 63 0.48 : Experiment results of the macro-scale SISO and MIMO molecular communication testbed.[]{data-label="tab_comparison_testbed"}	{ "pile_set_name": "ArXiv" }
--- author: - title: RRAM based neuromorphic algorithms --- Introduction ============ Human brain performs massively parallel and low power operations. It can outperform present age microprocessors on many tasks involving pattern recognition and input classification. The underlying neurons are heavily inter-connected; on average each neuron is connected to 10,000 (or up to 100,000) other neurons [@book_koch; @AIP; @Laughlin]. Despite the complexity of the human brain, research to understand the human brain is on-going, with the hope of emulating it in terms of its functionalities. In recent years, RRAM devices have emerged as a major memory component in mimicking the functionality of synapses in the human brain[@Indiveri_2016]. This is mainly because RRAM can be used both as a memory element and computation unit. As mentioned in [@Cao_2015], there are two ways of looking at RRAM based neuromorphic algorithms. From the deep learning perspective, one is to design algorithms for inference only, i.e to map the pre-trained deep learning models which fulfil certain hardware constraints onto the RRAM based neuromorphic hardware without any further training. While another way is to perform on-chip training on the RRAM based neuromorphic hardware, which will require additional interface circuitry for specific algorithms. Inference alone requires the conversion of existing pre-trained deep learning algorithms in high precision digital domain to the binary event-based (or spiking) domain so as to be able to be mapped onto RRAM based neuromorphic hardware. Whereas, on-chip training may be implemented at the RRAM synapse in the neuromorphic hardware by emulating local spike timing based algorithms such as spike timing dependent plasticity or its variants. These two methods belong to a new computational paradigm known as spiking deep neural network (SDNN). Other than the aforementioned learning algorithms that can be implemented on RRAM based neuromorphic hardware, low precision convolutional neural networks (CNN) such as the binarized neural network [@BNN], binaryNet [@BinaryNet], XNOR-NET [@XNORNet] and DoReFa-NET [@DoReFaNet] can be mapped onto a chip containing RRAM based synaptic crossbar array [@Yu_Hao_RRAM]. In such an approach, the computations performed in the CNN maybe converted to bitwise operations, such as bitwise convolution, batch normalization and pooling etc., as shown in figure 1 of [@Yu_Hao_RRAM]. Contrary to other paradigms, mapping is much easier with such an approach as it does not involve spiking neurons. Irrespective of the mapping algorithms implemented on the RRAM based neuromorphic hardware, we should expect a drop in accuracy due to hardware noise, especially the noise inherent in RRAM synapses (Set or reset variability [@Stefano_set_reset], Random Telegraph Noise (RTN) [@Stefano_RTN] etc.). One plausible approach to mitigate the drop in accuracy is to account for the noise itself during training, which may help to alleviate the accuracy loss to some extent. RRAM Synapse {#sec:1.3.1} ------------ ![RRAM synapse between an axon and a neuron.[]{data-label="RRAM_synapse"}](RRAM_synapse.pdf) RRAM is a two terminal non-volatile device with a conducting dielectric layer sandwiched between two electrodes as shown in fig. \[RRAM\_synapse\]. Electrically induced resistive switching effects shown in metal-insulator-metal systems are the basis for RRAM [@RRAM_nano]. Manipulation of oxygen vacancies in the conductance layer using positive and negative voltages helps in controlling current flow in RRAM. The state of RRAM reflects the current passed through it in the history, making it useful for modelling the synaptic weights of neurological synapses and implementing neural network architectures. In the case of spiking neural network (SNN), the tunable resistive state of RRAM synapses is analogous to the synaptic plasticity in brain. The electrical connection between a presynaptic neuron and a postsynaptic neuron (as shown in fig. \[RRAM\_synapse\]) changes, strengthening or weakening the synaptic impulses thus making it a case for brain-like pattern recognition. Crossbar array of RRAM Synapses {#sec:1.3.2} ------------------------------- ![Crossbar array of synapses in a neuromorphic core.[]{data-label="Crossbar"}](Crossbar_array.pdf) A crossbar architecture is shown in Fig. \[Crossbar\]. Input axons are the input connections from the output neurons in the previous convolution layer which was mapped onto another neuromorphic core. Output neurons are spiking neurons. Spiking neurons recieve input current from many other spiking neurons (input axons as per the figure) and fire a spike when the integrated current input reaches the neuron threshold. These building blocks like axons, neurons and synapses together can perform mathematical operations. Matrix dot vector multiplcations can be performed efficiently with these crossbar structure [@MVM]. Each column in a crossbar produces the sum of product of input from axons and the weights stored at each RRAM synapses. Computation in a crossbar array of RRAM synapses {#sec:1.3.3} ------------------------------------------------ ![The weights and input activations used in the crossbar architecture of a neuromorphic core. Weights and inputs marked in different color is corresponding to its sign as marked with the similar color in fig. \[Crossbar\]. []{data-label="Crossbar_weights"}](Crossbar_weights.pdf) ![Illustration of computation in a crossbar array of synapses in a neuromorphic core.[]{data-label="Crossbar"}](Crossbar_computation.pdf) The convolution operation in a convolutional neural network is implemented with the help of crossbar array of synapses in a neuromorphic core. Suppose a 3x3 convolution filter kernel after training a convolutional neural network is saved as weights, W as shown in fig. \[Crossbar\_weights\]. Out of these weights W2, W4, W6 and W8 are negative weights and rest of the weights are positive. Thus for the implementation of convolution operation this weight kernel W has to be separated into positive weights (marked in blue) and negative weights (marked in pink). Similarly for the input, X, it is divided into two matrices one is positive (marked in green) and another is negated input, -X (marked in orange). Once the weight matrix and input matrix is ready, the weight matrix can be written into the RRAM synapses – positive weights occupy the top of the crossbar column while negative weights occupy the bottom part, while the inputs are also respectively fed into axons – positive inputs are given to the top part of the axons and negative inputs to the bottom part of the axons. The convolution operation in a crossbar array between the inputs and the weights kernel is explicitly illustrated in fig. \[Crossbar\] as mentioned in [@Alom_2016]. One of the disadvantage of such implementation is the utilization of double the amount of input axons (2 x filter size) needed as well as double the number of RRAM synapses. Half of the RRAM synapses has to be written with low conductance state. This work is also extended to make the architecture extremely parallel by stretching the separated weight matrices as in the toeplitz matrix [@Alom_2017]. But, the same disadvantages of poor utilization of axons and synapses as mentioned above will remain. A slightly different approach of implementation is utilized in IBM’s truenorth chip [@IBM_Esser]. They have only ternary weights (-1, 0, +1) and uses two crossbar synapses in a column as a single synapse to implement ternary weights. This will also end up using double the amount of physical synapses on neuromorphic chip compared to actual number of synapses in a weight kernel. Hence, truenorth also has a disadvantage of poor utilization of axons and synapses. Truenorth’s actual physical core size, meaning number of axons X number of neurons, is 256 X 256, but literally their core size is only 128 X 256 to implement ternary weights. Deep Neural Network (DNN) architecture {#sec:1.3.4} -------------------------------------- ![Block diagram of a conventional deep neural network architecture (the convolutional neural network). It comprises of mainly three blocks: the convolutional block, the fully connected block and the softmax layer.[]{data-label="fig:dnn_block"}](DNN_blocks.pdf) Deep learning has made much progress in recent years so much so that it has even outperformed humans in certain tasks, for instance, beating the current GO world champion [@Alphago]. DNN or deep CNN has achieved state-of-the-art accuracy in many image classification or patten recognition tasks such as handwritten digit recognition [@MNIST], and several other datasets such as CIFAR [@CIFAR], and ImageNet [@IMAGENET]. However, these networks typically need large amount of labeled training data; ImageNet has over 1 million labeled images for training. A conventional CNN is shown in figure \[fig:dnn\_block\]. It comprises of mainly three blocks: the first block is made up of convolution layers, the second of fully connected layers and the third is the softmax layer. The convolution block contains convolution layers that perform the convolution operation on intermediate output activations. The convolution block also contains other layers that perform batch normalization or pooling. The fully connected block contains several layers of fully connected neural network. These two blocks are mainly for feature extraction. The final layer is a fully connected classifier which gives an output based on the softmax function. A typical learning algorithm used in a CNN is backpropagation of errors with stochastic gradient descent. The network parameters such as weights and biases are adjusted during training so as to predict the object label of an input image during testing. Spiking Neural Network (SNN) architecture {#sec:1.3.5} ----------------------------------------- ![Block diagram of a neuron in a SNN: implemented in blocks as shown, namely, the synapse and the neuron.[]{data-label="fig:snn_block"}](SNN_figure.pdf) Spiking neural network (SNN) is considered as the third generation of neural networks [@Maass]. SNN is inspired by biological neural networks while the DNN less so; hence the DNN is also commonly referred to as artificial neural networks (ANN). DNN does not have any biological roots apart from the hierarchical structure it possess [@Neocognitron]. SNN is event based: neural activations are communicated through spikes. Spiking neurons integrate incoming input spikes and emit a spike which is a threshold crossing event, as and when new information needs to be processed or communicated. These spikes are communicated through synapses which are associated with a weight quantity. A neuron in a SNN and its hardware implementation is shown in figure \[fig:snn\_block\]. The above figure \[fig:snn\_block\] (a) shows a single neuron (as part of a SNN) with its input and output mechanisms. The synapse is the connection between the axon of a pre-synaptic neuron and the dendrite of a post-synaptic neuron. A neuron integrates the incoming spikes received through its dendrites and may then emit a spike in the event of threshold crossing through its axon to its post-synaptic neurons. Figure \[fig:snn\_block\] (b) shows a block diagram representation of the biological model as in figure \[fig:snn\_block\] (a). The synapse is a storage element with input spikes and output current. Neuron computation is done using an integrator and a comparator. The integrator accumulates the input currents in terms of potential difference, which emulates the membrane potential in biological neurons. The comparator then checks if the membrane potential crosses the voltage threshold; a spike is emitted if crossed and the membrane potential is then reset to its baseline value. Conversion of DNN to the spike-based domain: Spiking Deep Neural Network (SDNN) {#sec:1.3.6} ------------------------------------------------------------------------------- In a conventional CPU or GPU, it requires more time and energy to run a SDNN, whereas the power consumption and computational latency in neuromorphic analog or digital dedicated hardwares [@Neurogrid; @Spinnaker; @IBM-truenorth] are orders of magnitude less. The substantial computational cost incurred during training and inference in a deep network for real world practical applications has created [@Peter_2015] : - a need for specialized hardware acceleration. - a new computational paradigm. One emerging approach is to convert the pre-trained DNN into SNN (while retaining its parameters) so that it can be mapped directly onto a neuromorphic hardware with little performance loss. The spike based computation in the SNN consumes much less power compared to the high precision digital computation in the DNN. DNN has better classification accuracy compared to SNN. Hence, mapping a deep CNN to a SDNN potentially allows us to achieve better accuracy with high energy efficiency. $$\begin{split} & Spiking \; Neural \; Network => Low \; Energy \\ & Deep \; Neural \; Network => Better \; Accuracy \\ & SDNN => Better \; Accuracy \; Low \; Energy \; (BALE)\; Neural\; Network \end{split}$$ While it is difficult to achieve in a mapped SDNN the same level of accuracy as the DNN, research is ongoing to develop better mapping techniques. ### DNN to SNN conversion: SDNN background {#sec:1.3.6.1} DNN to SNN conversion techniques were developed in the ongoing research to map a trained neural network in conventional frame-based vision system representation to an event-based one [@Perez_2013]. Neurons in the frame-based CNN were converted to event-based neurons with leak, membrane potential reset and refractory periods. One of the first research paper on CNN to SNN conversion is [@Cao_2015]. The conventional CNN is first converted into a tailored CNN which fulfils the requirements of the SNN. This tailored CNN is then trained. Finally, this tailored CNN is converted into a spiking CNN, while retaining the trained weights. The requirements imposed by the SNN on the tailored CNN are - using RELU [@Alex_2012] as activation functions, - removing biases from convolution and fully connected layers and - using spatial linear subsampling in place of maxpooling. [@Peter_2015] extended the work of [@Cao_2015] by adding weight normalization techniques to improve the conversion accuracy. The approximation errors in SNNs due to either excessive or too little spikes are avoided by rescaling of weights. Model based and data based weight normalization techniques were proposed; data based normalization gives no loss in conversion accuracy for classification of MNIST dataset. The integrate and fire (IF) neuron model was extensively used in SDNN until [@Huns_2015] demonstrated that a CNN can also be mapped onto a SDNN made up of leaky integrate and fire (LIF) neurons which are more biological plausible. This is achieved by using a modified LIF neuron known as the softened LIF neuron and by training the network with noise so as to improve network robustness against the variability inherent in spikes. The hardware constrained neuromorphic algorithm is implemented in [@IBM_Esser] on the IBM Truenorth neuromorphic chip. The hardware constraints are namely, low precision weights and restricted connectivity among spiking neurons. Adapting SNN is introduced in [@Bohte_2016], which is based on adaptive spiking neurons. Asynchronous pulsed sigma-delta coding scheme is used by these spiking neurons to efficiently encode information in spike trains, while homeostatically optimizing the firing rate. This method uses an order of magnitude less spikes compared to other SDNN approaches; the RELU neurons in an ANN could be directly mapped to adaptive spiking neurons during conversion. ### General steps for conversion {#sec:1.3.6.2} The conversion of a pre-trained DNN to the event-based domain is for inference purposes. The principle of the conversion technique as mentioned in [@Cao_2015] is that the time averaged firing rate of a spiking neuron must be correlated with the activation value of the corresponding neuron in the ANN. The generic steps involved for network conversion is as mentioned below: - Choose a CNN to train. - Use ReLU for activation functions in the CNN. - Fix the bias to zero throughout training using stochastic gradient descent. - Save all the weights after training. - Replace neurons in the CNN with integrate and fire neurons without refractory period. - Map the saved weights to the SNN. - Convert the input image to poisson spike trains with firing rates proportional to each pixel intensity value. ### Factors affecting conversion accuracy {#sec:1.3.6.3} The issues affecting conversion accuracy as mentioned in [@Cao_2015] are: in the CNN the weights and biases can be negative. Since input integration is a weighted sum of inputs and the bias, the output can be negative. If the sigmoid function is used for activation it may also be negative. It is difficult to represent negative activations in the CNN on a SNN. It is also difficult to represent biases in the SNN. Two layer neural network is needed to implement spatial maxpooling in the SNN. CNN to SNN mapping requires the input image to be converted to poisson spike trains with firing rates proportional to the pixel intensity value. As a result, the loss of accuracy during conversion can happen due to the factors [@Peter_2015]: Input spikes are not enough to result in threshold crossing, hence no output spike is emitted when activation values in the CNN are below threshold. If the spiking neuron receives too many input spikes in a single timestep or if some of its synaptic weights are higher than threshold, then the spiking neuron should emit more than one spike per timestep, which it cannot, and hence introducing error in the process. Due to the non-uniformity of the spike trains or the stochastic nature of the spiking input, a specific feature set could be over- or under- activated by incoming spikes. An analysis of conversion and its theory is proposed in [@Bodo_2016]. One on one mapping of the spiking neuron and the activation function of the CNN, reveals that during threshold crossing, the membrane potential reached maybe of any value above threshold. This error would accumulate over time. ### Solution to the issues affecting conversion accuracy {#sec:1.3.6.4} The solution to the above-mentioned issues are the following (listed as above): 1. as mentioned in [@Cao_2015], are to remove biases from convolution layers, use ReLU as activation function and use spatial linear subsampling instead of maxpooling. 2. as mentioned in [@Peter_2015] use weight normalization. 3. as mentioned in [@Bodo_2016] use reset by subtraction instead of reset to zero for spiking neurons. Instead of removing biases from convolutional layers, a constant input current can be applied to emulate the biases. Also apply normalization techniques. 4. as mentioned in [@Bohte_2016], to reduce the variability of input spikes, the multi-bit values of the input maybe fed directly into the first hidden layer and spikes are then output henceforth. 5. as mentioned in [@Pool], pooling layers can be avoided in deep neural network. Hence, even though there are techniques to convert pooling layers in SDNN, we can remove these layers from the DNN for simplicity sake. Spike based algorithms {#sec:1.3.4} ---------------------- In the past decade, spike timing dependent plasticity (STDP) has been a popular unsupervised learning method due to its biological plausibility . STDP mechanism depends on the timing difference between the pre-synaptic and post-synaptic spikes to adjust the synaptic weight. In the simple, doublet STDP [@Shubha; @D-STDP_Roshan; @D-STDP_Roshan_TNNLS], when a post-synaptic spike happens after a pre-synaptic spike has arrived (pre-post event), then the weight of the synapse increases i.e. synaptic potentiation takes place; whereas, if a post-synaptic spike happens before a pre-synaptic spike (post-pre event), then the weight of the synapse decreases, i.e. depotentiation takes place. Similar to the doublet STDP, there is another variant of STDP called the triplet STDP [@Pfister_2006; @Mostafa; @T-STDP_Roshan; @T-STDP_Roshan_ISCAS], whereby, three spike events are considered (pre-post-pre, post-pre-post etc.). There are CMOS devices such as the floating gate MOSFET or nano-technology devices such as the memristors, Resistive Random Access Memories (RRAM), Phase Change Memories (PCM) and Spin-Transfer Torque Magnetic Random Access Memories (STT-MRAMs) used for the implementation of artificial synapses. One of the challenge is to integrate these nano-technology devices with CMOS. The characteristics of high synaptic density on neuromorphic hardware has to be compromised. Understanding the device physics becomes the key for the implementation of artificial synapses, especially while using any of the technologies such as floating gate MOSFET, memristors or the more recent spin devices to implement plasticity rules such as STDP. Conclusion {#sec:1.4} ---------- For future work, it maybe worth investigating conversion of DNN using different encoding schemes such as temporal coding or latency coding instead of just rate coding. This would reduce the number of spikes required to represent an input and result in more efficient computing. In the long run however, hardware compatible SNN algorithms should be developed that enable on-chip learning and inference for various applications. This will eliminate the need for conversion of DNN to SNN; the challenge would be how one may improve the accuracy of such SNN algorithms. Here, we have given an overview of the current state-of-art neuromorphic algorithms on RRAM based neuromorphic devices. While research is on-going to develop SNN for on-chip learning, the current reliable approach for a real-world application is to do inferencing on-chip based on a converted DNN that is pre-trained off-chip. During the mapping of a DNN to neuromorphic hardware, hardware constraints such as number of neurons and synapses, core size, fan in-fan out degrees, routing, spike traffic congestion etc. have to be taken into consideration. Should any of the above constraints not be met, the DNN architecture will have to be modified accordingly, so as to fit into a specific neuromorphic hardware. Given its small form factor and energy efficiency, neuromorphic hardware is well suited for edge computing applications in the fields of robotics, surveillance, unmanned aerial vehicles etc. Funding {#funding .unnumbered} ======= This research is supported by Programmatic grant no. A1687b0033 from the Singapore governments Research, Innovation and Enterprise 2020 plan (Advanced Manufacturing and Engineering domain). [99.]{} Koch C. Biophysics of Computation: Information Processing in Single Neurons. Oxford Univ. Press; New York: 1999 Geoffrey W. Burr, Robert M. Shelby, Abu Sebastian, Sangbum Kim, Seyoung Kim, Severin Sidler, Kumar Virwani, Masatoshi Ishii, Pritish Narayanan, Alessandro Fumarola, Lucas L. Sanches, Irem Boybat, Manuel Le Gallo, Kibong Moon, Jiyoo Woo, Hyunsang Hwang and Yusuf Leblebici (2017) Neuromorphic computing using non-volatile memory, Advances in Physics: X, 2:1, 89-124, DOI:10.1080/23746149.2016.1259585 Laughlin SB, Sejnowski TJ. Communication in Neuronal Networks. Science (New York, NY). 2003;301(5641):1870-1874. DOI:10.1126/science.1089662. G. Indiveri, E. Linn, and S. Ambrogio, “ ReRAM-based neuromorphic computing," in Resistive Switching: From Fundamentals of Nanoionic Redox Processes to Memristive Device Applications. Weinheim, Germany : Wiley-VCH Verlag GmbH & Co. KGaA, 2016, pp. 715-735. Itay Hubara and Matthieu Courbariaux and Daniel Soudry and Ran El-Yaniv and Yoshua Bengio, Binarized Neural Networks, NIPS, 2016. Matthieu Courbariaux and Yoshua Bengio, BinaryNet: Training Deep Neural Networks with Weights and Activations Constrained to +1 or -1, CoRR, abs/1602.02830, 2016. M Rastegari, V Ordonez, J Redmon, A Farhadi, Xnor-net: Imagenet classification using binary convolutional neural networks, European Conference on Computer Vision, 525-542, 2016. S Zhou, Y Wu, Z Ni, X Zhou, H Wen, Y Zou, DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients, arXiv preprint arXiv:1606.06160, 2016. Leibin Ni, Zichuan Liu, Hao Yu and Rajiv V. Joshi, An Energy-Efficient Digital ReRAM-Crossbar-Based CNN With Bitwise Parallelism, IEEE Journal on Exploratory Solid-State Computational Devices and Circuits, 2017. Stefano Ambrogio, Simone Balatti, Antonio Cubeta, Alessandro Calderoni, Nirmal Ramaswamy and Daniele Ielmini, Statistical Fluctuations in HfOx Resistive-Switching Memory: Part I - Set/Reset Variability, IEEE Transactions on Electron Devices, vol. 61, no. 8, Aug 2014. Stefano Ambrogio, Simone Balatti, Antonio Cubeta, Alessandro Calderoni, Nirmal Ramaswamy and Daniele Ielmini, Statistical Fluctuations in HfOx Resistive-Switching Memory: Part II - Random Telegraph Noise, IEEE Transactions on Electron Devices, vol. 61, no. 8, Aug 2014. R. Waser and M. Aono, “Nanoionics-based resistive switching memories," Nature Materials, vol. 6, no. 11, pp. 833-840, 2007. M. Hu et al., “Dot-product engine for neuromorphic computing: Programming 1T1M crossbar to accelerate matrix-vector multiplication," 53nd ACM/EDAC/IEEE Design Automation Conference (DAC), pp. 1-6, Austin, TX, 2016. Chris Yakopcic, Md Zahangir Alom and Tarek M. Taha, “Memristor Crossbar Deep Network Implementation Based on a Convolutional Neural Network," International Joint Conference on Neural Networks (IJCNN), 2016, pp. 963-970. Chris Yakopcic, Md Zahangir Alom and Tarek M Taha, “Extremely parallel memristor crossbar architecture for convolutional neural network implementation," International Joint Conference on Neural Networks (IJCNN), 2017, pp. 1696-1703. Steven K. Esser, Paul A. Merolla, John V. Arthur, Andrew S. Cassidy, Rathinakumar Appuswamy, Alexander Andreopoulos, David J. Berg, Jeffrey L. McKinstry, Timothy Melano, Davis R. Barch, Carmelo di Nolfo, Pallab Datta, Arnon Amir, Brian Taba, Myron D. Flickner and Dharmendra S. Modha, “Convolutional networks for fast, energy-efficient neuromorphic computing," Proceedings of the National Academy of Sciences, Oct 2016, 113 (41) 11441-11446. https://deepmind.com/research/alphago/ LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86, 2278-2324. Alex Krizhevsky, Learning Multiple Layers of Features from Tiny Images, 2009. J. Deng and W. Dong and R. Socher and L. Li and Kai Li and Li Fei-Fei, ImageNet: A large-scale hierarchical image database, IEEE Conference on Computer Vision and Pattern Recognition, 248-255, 10.1109/CVPR.2009.5206848, June, 2009. Maass, Wolfgang (1997). “Networks of spiking neurons: The third generation of neural network models”. Neural Networks. 10 (9): 1659-1671. doi:10.1016/S0893-6080(97)00011-7. ISSN 0893-6080. K. Fukushima, “ Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position, " Biological Cybernetics, vol. 36, no. 4, pp. 193-202, 1980. P. U. Diehl, D. Neil, J. Binas, M. Cook, S. Liu and M. Pfeiffer, “Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing,” 2015 International Joint Conference on Neural Networks (IJCNN), Killarney, 2015, pp. 1-8. B. V. Benjamin et al., “Neurogrid: A Mixed-Analog-Digital Multichip System for Large-Scale Neural Simulations,” in Proceedings of the IEEE, vol. 102, no. 5, pp. 699-716, May 2014. Plana, L. A., Clark, D., Davidson, S., Furber, S., Garside, J., Painkras, E., Pepper, J., Temple, S., and Bainbridge, J. 2011. SpiNNaker: Design and implementation of a GALS multicore system-on-chip. ACM J. Emerg. Technol. Comput. Syst. 7, 4, Article 17 (December 2011), 18 pages. F. Akopyan et al., “TrueNorth: Design and Tool Flow of a 65 mW 1 Million Neuron Programmable Neurosynaptic Chip,” in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 34, no. 10, pp. 1537-1557, Oct. 2015. J. A. Pérez-Carrasco and B. Zhao and C. Serrano and B. Acha and T. Serrano-Gotarredona and S. Chen and B. Linares-Barranco, IEEE Transactions on Pattern Analysis and Machine Intelligence, Mapping from Frame-Driven to Frame-Free Event-Driven Vision Systems by Low-Rate Rate Coding and Coincidence Processing–Application to Feedforward ConvNets. Nov 2013, 35(11), 2706-2719. Cao, Y. and Chen, Y. and Khosla, D. Spiking Deep Convolutional Neural Networks for Energy-Efficient Object Recognition, International Journal of Computer Vision (2015) 113: 54. https://doi.org/10.1007/s11263-014-0788-3. Krizhevsky, A., Sutskever, I, and Hinton, G., ImageNet classification with deep convolutional neural networks. In: Advances in Neural Information Processing Systems 25 (pp. 1106-1114), 2012. Hunsberger, E. and Eliasmith, C., Spiking Deep Networks with LIF Neurons. arXiv:1510.08829 \[cs\], pages 1-9. 2015 Zambrano, D. and Bohte, S. M., Fast and efficient asynchronous neural computation with adapting spiking neural networks. arXiv preprint arXiv:1609.02053, 2016. Bodo Rueckauer, Iulia-Alexandra Lungu, Yuhuang Hu, and Michael Pfeiffer, Theory and Tools for the Conversion of Analog to Spiking Convolutional Neural Networks. Workshop “Computing with Spikes", 29th Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain. Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox and Martin Riedmiller, “Striving for simplicity: The all convolutional net," accepted as a workshop contribution at ICLR 2015. G.-Q. Bi and M.-M. Poo, Synaptic modifications in cultured hippocampal neurons: Dependence on spike timing, synaptic strength, and postsynaptic cell type, J. Neurosci., vol. 18, no. 24, pp. 10464-10472, Dec. 1998. L. I. Zhang, H. W. Tao, C. E. Holt, W. A. Harris, and M.-M. Poo, “A critical window for cooperation and competition among developing retinotectal synapses," Nature Neurosci., vol. 395, pp. 37-44, Sep. 1998. L. F. Abbott and S. B. Nelson, Synaptic plasticity: Taming the beast, Nature Neurosci., vol. 3, pp. 1178-1183, Nov. 2000. S. Ramakrishnan, P. Hasler, and C. Gordon. Floating gate synapses with spike-time-dependent plasticity. IEEE Transactions on Biomedical Circuits and Systems, 5(3):244-252, Jun. 2011. Roshan Gopalakrishnan and Arindam Basu, Robust Doublet STDP in a Floating-Gate Synapse, IJCNN, July 6-11, 2014, Beijing, China. Roshan Gopalakrishnan and Arindam Basu, On the Non-STDP Behavior and Its Remedy in a Floating-Gate Synapse, in IEEE Transactions on Neural Networks and Learning Systems, vol. 26, no. 10, pp. 2596-2601, Oct. 2015. J.-P. Pfister and W. Gerstner, Triplets of spikes in a model of spike timing-dependent plasticity, J. Neurosci., vol. 26, no. 38, pp. 9673-9682, Sep. 2006. M. R. Azghadi, S. Al-Sarawi, D. Abbott, and N. Iannella. A neuromorphic vlsi design for spike timing and rate based synaptic plasticity. Neural Networks, 45:70-82, 2013. Roshan Gopalakrishnan and Arindam Basu, Triplet Spike Time-Dependent Plasticity in a Floating-Gate Synapse, IEEE TNNLS, vol. 28, no. 4, April 2017. Roshan Gopalakrishnan and Arindam Basu, Triplet spike time dependent plasticity in a floating-gate synapse, IEEE International Symposium on Circuits and Systems (ISCAS), pp. 710-713, 2015.	{ "pile_set_name": "ArXiv" }
Resolutions of $p$-Modular TQFT’s and\ Representations of Symmetric Groups[^1] {#resolutions-of-p-modular-tqfts-and-representations-of-symmetric-groups .unnumbered} ======================================= [Thomas Kerler]{}\ September 2001 [Contents]{} [1. Introduction and Survey of Results]{} [\[S1\]]{} In this article we bring together several areas of representation theory in a series of interrelated results. The first is the rather established theory of $p$-modular representations of the symmetric groups, followed by the representation theory of groups of Lie type over the finite field ${{\mathbb F}}_p$, and, finally, the area of topological quantum field theories (TQFT’s) over ${{\mathbb F}}_p$. The latter have been our original motivation since they appear as constant order reduction in the Reshetikhin Turaev Theories. In fact, as we shall outline in more detail at the end of in this section, the identities we will find in this article will, for example, imply relations between the Lescop invariant and the Reshetikhin Turaev invariant for 3-manifolds for $p=5$. Our results in each area concern resolutions and expansions of $p$-modular representations and invariants into their respective characteristic zero counterparts, and, thus, naturally build on each other. Let us state the results in order, beginning with the case of the symmetric groups. #### The Symmetric Groups: The representation theory of the symmetric group $S_n$ in $n$ letters over ${\mathbb Q}$ (or ${\mathbb Z}$) is well known. The theory is semisimple and the simple representations are isomorphic to the Specht modules ${\cal S}^{\tau}$, where $\tau$ is a Young diagram with $n$ boxes. They have a natural basis given by Young tablaux, and the $S_n$-action preserves the free ${{\mathbb Z}}$-modules (or lattices) ${\cal S}^{\tau}_{{{\mathbb Z}}}$ generated by these basis vectors. Passing to ${\cal S}^{\tau}_p=\raise .3ex\hbox{${\cal S}^{\tau}_{{\mathbb Z}}$}\! \big/\! \raise -.3ex\hbox{$p{\cal S}^{\tau}_{{\mathbb Z}}$}$ we thus obtain representations of the same rank over the field ${{\mathbb F}}_p$ for any given prime $p\geq 3$. The ${\cal S}^{\tau}_p$, however, are not longer irreducible, but they have a unique simple quotient ${\cal D}^{\tau}_p$ obtained from canonical inner forms on the Specht modules, see [@Jam78]. The representations ${\cal S}^{\tau}_{{\mathbb Z}}$ and ${\cal D}^{\tau}_p$ define characters $\chi^{\tau}$ and $\varphi_p^{\tau}$ on $S_n$ with values in ${{\mathbb Z}}$ and ${{\mathbb F}}_p$, respectively. We also denote by $\chi^{\tau}_p$ the $p$-reduction of $\chi^{\tau}$, which may, alternatively, be thought of as the character asscociated to ${\cal S}^{\tau}_p$. The relationship between the “ordinary” representations ${\cal S}^{\tau}_{p\, \mbox{\em\scriptsize or}\,{{\mathbb Z}}}$ and “ordinary” characters $\chi^{\tau}_{p\, \mbox{\em\scriptsize or}\,{{\mathbb Z}}}$ and their $p$-modular counterparts ${\cal D}^{\tau}_p$ and $\varphi_p^{\tau}$ is, even after decades of research, still intensely investigated with many open questions remaining. While several algorithms exist for expressing the $\chi^{\tau}_p$ in terms of the $\varphi_p^{\tau}$, fewer results exist for the converse relations, and even fewer results relate the modules themselves. The exact modular structure of the ${\cal S}^{\tau}_p$ in terms of a modular ordering of the ${\cal D}^{\tau}_p$-components has only very recently been uncovered by Kleshchev and Sheth in [@KS99] for the special case of Young diagrams $\tau$ with two rows. The first result of this article may be thought of as the inverse relation of the result in [@KS99], and it implies a similarly inverse relation for the characters. \[thm-SnReso\] Let $p\geq 3$ be a prime and $\tau=[a,b]$ be the two row Young diagram with row lengths $a$ and $b$ with $0\leq a-b\leq p-2$. Consider the sequence of Young diagrams $\tau_j$ given by $\tau_{2i}=[a+ip,b-ip]$ and $\tau_{2i-1}=[b+ip-1, a-ip+1]$, whenever defined. (That is, $\tau_0=\tau=[a,b]$, $\tau_1=[b+p-1,a-p+1]$, $\tau_2=[a+p,b-p]$, $\tau_3=[b+2p-1,a-2p+1]$, etc.) 1. There is a resolution of the modular, simple representation ${\cal D}^{\tau}_p$ of $S_n$ in terms ordinary representations given by an exact sequence of $S_n$-equivariant maps as follows. \[eq-thm-Cseq\] …\^[\_2]{}\_p\^[\_1]{}\_p\^[\_0]{}\_p\^\_p0. 2. We have the following expansion of $p$-modular characters in terms of ordinary characters. \[eq-charactersX\] \^\_p=\_[i0]{} (-1)\^[i]{} \^[\_[i]{}]{}\_p, It is an intriguing fact that the maps in the sequence (\[eq-thm-Cseq\]) are powers of generators of ${\mathfrak s}{\mathfrak l}_2$ acting dually on $({{\mathbb Z}}^2)^{\otimes n}$, which, as an $S_n$-module, contains the permutation modules $M^{\tau}$. The precise action is constructed in Corollary \[cor-Cseq\] of Section 4, where we prove that it yields a well defined complex. The proof of exactness of this complex uses the results of [@KS99] and occupies most of Section 5. A generalization of Theorem \[thm-SnReso\] to $n$-row diagrams using a dual ${\mathfrak s}{\mathfrak l}_n$ is likely to yield more insights the structure of $p$-modular Specht modules for general Young diagrams. The character identity (\[eq-charactersX\]) is stated as an immediate consequence in Corollary \[cor-characters\] of Section 6. #### The Symplectic Groups: The symmetric groups typically appear as or within Weyl groups of groups of Lie type. In this article we are particularly interested in the symplectic groups ${\rm Sp}(2g,{\mathbb M})$ where ${\mathbb M}$ may be ${\mathbb R}$, ${{\mathbb Z}}$ or ${\mathbb F}_p$. There are obvious generalizations of our results to most other groups of Lie type, which we leave to the reader. Let $H=H_1(\Sigma_g,{{\mathbb Z}})$ be the fundamental representation of ${\rm Sp}(2g,{{\mathbb Z}})$ with symplectic basis $\langle a_1,\ldots,a_g,b_1,\ldots,b_g\rangle$, and denote by $V(\varpi_k)\subset {\mbox{\raisebox{.4ex}{$\bigwedge^{\!k}$}}\mkern-1mu} H$ the subrepresentation generated by the highest weight vector $a_1\wedge\ldots \wedge a_k$. Over ${\mathbb R}$ this is the irreducible representation of fundamental heighest weight $\varpi_j=\epsilon_1+\ldots+\epsilon_j$. Denote by $V_p(\varpi_j)$ the respective $p$-modular reduction, and by $\dov V_p(\varpi_j)$ the unique irreducible subquotient over ${\mathbb F}_p$ generated by the same highest weight vector. Let $g-p+2\leq l\leq g$ and $\hat l=2(g-p+1)-l$. From Theorem \[thm-SnReso\] now we derive resolutions of Sp-representations that are given by exact sequences as follows. \[eq-SpReso\] … V\_p(\_[l-4p]{}) V\_p(\_[[l]{}-2p]{}) V\_p(\_[l-2p]{}) V\_p(\_[l]{}) V\_p(\_l) V\_p(\_l) 0 . Evidently, (\[eq-SpReso\]) implies similar expansions of ${\rm Sp}$-characters, and such expansions also exist for other groups of Lie type. The generalization more intersting to us, which in fact implies (\[eq-SpReso\]), are resolutions of topological quantum field theories (TQFT’s). In Lemmas \[lm-upsilambda\] and \[lm-induce\] of Section 2 we establish the relation between the ${\rm Sp}(2g,{{\mathbb Z}})$ weight spaces of ${\mbox{\raisebox{.4ex}{$\bigwedge^{\!}$}}\mkern-1mu}H$ with the $S_k$-module $({{\mathbb Z}}^2)^{\otimes k}$. In Lemma \[lm-Spe-action\] we also give the explicit action of the Serre generators of ${\mathfrak s}{\mathfrak p}_{2g}$ in the language of the permutation modules. The relation between the weight spaces of $V(\varpi_k)$ and the Specht modules and the respective explicit actions of the ${\mathfrak s}{\mathfrak p}_{2g}$-generators on the Young diagrams is derived in Section 3. #### Homological TQFT’s: Recall that a TQFT is a functor ${\cal V}:\Cob\to {\mathbb M}{\rm -mod}$ from a category of 2+1-dimensional cobordisms into a category of free modules over a commutative ring ${\mathbb M}\,$. Specifically, $\cal V$ assigns to a surface $\Sigma$ a free ${\mathbb M}\,$-module ${\cal V}(\Sigma)$ and to a cobordism between two surfaces an ${\mathbb M}\,$-linear map between the respective ${\mathbb M}\,$-modules. In [@FroNic92] Frohman and Nicas construct a TQFT ${\cal V}^{FN}$ over ${\mathbb M}={{\mathbb Z}}$, where the free ${{\mathbb Z}}$-module for a surface is given by the (co)homology of its Jacobian, specifically, ${\cal V}^{FN}(\Sigma)=H^({\rm Hom}(\pi_1(\Sigma),U(1)),{{\mathbb Z}}) ={\mbox{\raisebox{.4ex}{$\bigwedge^{\!}$}}\mkern-1mu}H_1(\Sigma,{{\mathbb Z}})\,$. The basics of this construction are reviewed in the beginning of Section 2. Further, in Section 3, we will recall the decomposition of this TQFT into its irreducible components ${\cal V}^{(j)}_{{\mathbb Z}}$, with $j=1,2,\ldots\,$, of ${\cal V}^{FN}$. They are again TQFT’s over ${{\mathbb Z}}$ and are obtained in [@Ker01] from a dual Lefschetz ${\mathfrak s}{\mathfrak l}_2$-action. Composing ${\cal V}^{(j)}_Z$ with the canonical (rank-preserving) functor $\,{\mathbb Z}{\rm -mod}\,\to\!\!\!\!\to\,{\mathbb F}_p{\rm -mod}\,$ for primes $p\geq 3$ we obtain a family of TQFT’s ${\cal V}^{(j)}_p$ over ${\mathbb M}={\mathbb F}_p$. As before, the $p$-modular TQFT’s are generally highly reducible. However, they have a unique irreducible TQFT subquotient $\dov {\cal V}^{(j)}_p$. \[thm-TQFTreso\] For any prime $p\geq 3$ and integer $k$ with $0<k<p$ we have an exact sequence of natural transformations of TQFT’s \[eq-TQFTreso\] …[V]{}\^[((i+1)p+k\_[i+1]{})]{}\_p [V]{}\^[(ip+k\_i)]{}\_p… [V]{}\^[(2p-k)]{}\_p [V]{}\^[(k)]{}\_p\^[(k)]{}\_p0, where we set $k_i=k$ for even $i$ and $k_i=p-k$ for odd $i$. The sequence is constructed from its symmetric group summands in Corollary \[cor-seqTQFT\] of Section 4. Exactness follows in Section 5 from the respective results for Specht modules. In order to see why (\[eq-SpReso\]) is indeed a special case of Theorem \[thm-TQFTreso\] observe that the mapping class group $\Gamma_g$ is identical with the group of invertible cobordisms in $\Cob$ from a surface $\Sigma_g$ to itself. Hence, any TQFT $\cal V$ entails for every $g$ a representation of $\Gamma_g$ on ${\cal V}(\Sigma)$, which, in the case of ${\cal V}^{FN}$, factors through the symplectic quotient $\Gamma_g\to\!\!\!\!\to {\rm Sp}(2g,{{\mathbb Z}})$. The sequence in (\[eq-SpReso\]) is now obtained from Theorem \[thm-TQFTreso\] by evaluation on a particular surface and using the identifications of ${\rm Sp}(2g,{{\mathbb Z}})$-modules, given by ${\cal V}^{(j)}_{{\mathbb Z}}(\Sigma_g)\cong V(\varpi_{g-j+1})$ and $\dov {\cal V}^{(j)}_p(\Sigma_g)\cong \dov V_p(\varpi_{g-j+1})\,$. In the TQFT framework characters are endowed with an interpretion as topological invariants of infinite cyclic covers of closed 3-manifolds. More precisely, let $\cal V$ be any TQFT, $M$ a closed 3-manifold with $b_1(M)\geq 1$, and $\varphi:H_1(M)\to\!\!\!\to {{\mathbb Z}}$ an epimorphism. We define an invariant of a pair $(M,\varphi)$ as follows. Pick any two-sided surface $\Sigma\subset M$ which is dual to $\varphi\,$, and define $C_{\Sigma}=\overline{M-\Sigma}$ seen as a cobordism from $\Sigma$ to itself. The value ${\cal V}(M,\varphi)=trace({\cal V}(C_{\Sigma}))$ is now independent of the choice of $\Sigma$. As an important example, Frohman and Nicas extracted in [@FroNic92] the Alexander polynomial $\Delta_{\varphi}(M)$ as a Lefschetz trace from ${\cal V}^{FN}$. Translated into the decomposition of [@Ker01] this result says that the invariant ${\cal V}^{(j)}_{{\mathbb Z}}(M,\varphi)$ is the difference of two successive coefficients of the Alexander polynomial. Combining this observation with Theorem \[thm-TQFTreso\] we derive in the end of Section 6 the following relation between these invariants over ${{\mathbb Z}}$ and the respective invariants $\dov {\cal V}^{(j)}_p(M,\varphi)$ obtained from the [irreducible]{} TQFT’s over ${\mathbb F}_p$. \[thm-pAlexander\] Let $\Delta_{\varphi}(M)\in{{\mathbb Z}}[x,x^{-1}]$ be the Alexander Polynomial of a closed, compact, oriented 3-manifold $M$ with respect to an epimorphism $\varphi:H_1(M)\to\mkern-15mu\to{{\mathbb Z}}$. Let $\dov\Delta_{\varphi, p}^{\pm}(M)$ be the reduction of $\Delta_{\varphi}(M)$ obtained by substituting $x=\pm \zeta_p$, with $\zeta_p$ a $p$-th root of unity, and taking the ${{\mathbb Z}}$ coefficients modulo $p$. Then $$\dov\Delta_{\varphi, p}^{\mp}(M)\;\;=\;\;\sum_{k=1}^{p-1}(\pm1)^{k-1}[k]_{\zeta_p} \dov {\cal V}^{(k)}_p(M,\varphi)\;\;\;\;\in\;\;{{\mathbb F}}_p[\zeta_p]\;.$$ #### Johnson-Morita-Extensions and TQFT-Overview: Note, that all TQFT’s constructed up to this point have representations of $\Gamma_g$ that factor through ${\rm Sp}(2g,{{\mathbb Z}})$, that is, they contain the Torelli group ${\cal I}_g$ in their kernels. In Section 7 of this article we will also consider extensions of these TQFT’s with slightly smaller kernels, at least at the level of representations of $\Gamma_g$. In [@Mor93] Morita constructs a homomorphism $\tilde k: \Gamma_g\to {\rm Sp}(2g,{{\mathbb Z}})\ltimes \frac 12 U$ which has a smaller kernel ${\cal K}_g\subset{\cal I}_g$, and whose image ${\bf Q}_g\cong\Gamma_g/{\cal K}_g$ has finite index in ${\rm Sp}(2g,{{\mathbb Z}})\ltimes \frac 12 U$. It extends the Johnson homomorphism ${\cal T}_g\to\!\!\!\to U$ previously constructed in [@Joh80], where $U$ denote the free abelian group $U\cong\,\raise .4ex\hbox{${\mbox{\raisebox{.4ex}{$\bigwedge^{\!3}$}}\mkern-1mu} H$}\!\Big / \!\raise -.4ex\hbox{$\omega\wedge H$} \,\cong \,V(\varpi_3)$. As before we denote $H=H_1(\Sigma_g)$, considered as an Sp-module, and we let $\omega\in{\mbox{\raisebox{.4ex}{$\bigwedge^{\!2}$}}\mkern-1mu}H$ be the standard Sp-invaraiant symplectic form. Non-trivial representations of ${\bf Q}_g$ are readily obtained in Proposition \[propos-Umods\] of Section 7 as the extension of a pair of representations $V(\varpi_l)$ and $V(\varpi_{l+3})$ using the (up to scale unique) Sp-equivariant map $U\,\to\,{\rm Hom}(V(\varpi_l), V(\varpi_{l+3}))$. We prove that these extension also exist for the irreducible $p$-modular representations $\dov V_p(\varpi_{l})$. In TQFT language this translates to the following result. \[thm-JMext\] For $0<k<p-3$ there are indecomposable representations $\dov {\cal U}^{(k)}_p(\Sigma_g)$ of the mapping class group $\Gamma_g$ that factor through ${\bf Q}_g=\Gamma_g/{\cal K}_g$ but represent the Torelli group ${\cal I}_g$ non-trivially. There is a short, non-split exact sequence of $\Gamma_g$-equivariant maps as follows: \[eq-JMext\] 0 \^[(k+3)]{}\_p(\_g) \^[(k)]{}\_p(\_g) \^[(k)]{}\_p(\_g)0. We know from Theorem \[thm-KerFib\] below that $\dov {\cal U}^{(1)}_5$ does on fact extend to a TQFT. However, the question whether or how the $\dov {\cal U}^{(k)}_p$ and ${\cal U}^{(k)}_{{{\mathbb Z}}}$ descend from TQFT’s for general $p$ and $k$ is still open, and will be discussed in future work. We summarize the different TQFT’s of this paper, the constructions connecting them, and the included modules of the symmetric groups $S_n$ in the following table: --------------------------------- -------------------------------- ------------------------------------------------------------------- ------------------ -------------------------------------------------- ${\cal V}^{FN}_{{\mathbb Z}}$ Jacobian TQFT [@FroNic92] $\hookleftarrow$ Permutation Modules Fully reducible over ${{\mathbb Z}}$ $SL(2,{{\mathbb R}})$-Lefschetz Decomposition ${\cal V}^{(k)}_{{\mathbb Z}}$ Lefschetz Component, $\hookleftarrow$ Specht Modules ${\cal S}_{{{\mathbb Z}}}^{\tau}$ irreducible over ${{\mathbb Z}}$ $p$-Reduction ${\cal V}^{(k)}_p$ Reducible over ${{\mathbb F}}_p$ $\hookleftarrow$ $p$-Specht Modules ${\cal S}_p^{\tau}$ with inner form. Null space quotient $\dov {\cal V}^{(k)}_p$ Irreducible over ${{\mathbb F}}_p$. $\hookleftarrow$ Simple $S_n$-modules ${\cal D}_p^{\tau}$ Not ${{\mathbb Z}}$-liftable, but resolutions in ${\cal V}^{(j)}_p$’s Johnson-Morita Extension $\dov {\cal U}^{(k)}_p$ Indecomposable with two factors $\dov {\cal V}^{(k)}_p$ and $\dov {\cal V}^{(k+3)}_p$ --------------------------------- -------------------------------- ------------------------------------------------------------------- ------------------ -------------------------------------------------- #### Relations to the Reshetikhin Turaev Theory: The original motivation of this article comes from the study of the TQFT’s ${\cal V}_{\zeta_p}^{RT}$ constructed by Reshetikhin and Turaev in [@ResTur91] from the quantum group $U_{\zeta_p}({\mathfrak s}{\mathfrak o}_3)$ for a $p$-th root of unity $\zeta_p$. Thus, in order to put the TQFT’s of this article into their broader context and illustrate their relevance let us briefly sketch the pertinent relations and results from other work in quantum topology. It is shown by Gilmer in [@Gil01] that, for a restricted set of cobordisms, ${\cal V}_{\zeta_p}^{RT}$ can be written as a TQFT over the ring of cyclotomic integers ${{\mathbb Z}}[\zeta_p]$. For the Reshetikhin Turaev invariants of closed 3-manifolds this integrality property was previously proved in [@Mur95; @MasRob97]. Applying the ring reduction ${{\mathbb Z}}[\zeta_p]\to\!\!\!\to {\mathbb F}_p$, with $\zeta_p\mapsto 1$, we obtain a TQFT ${\cal V}_p^{I}$ over the finite field ${\mathbb F}_p$ from ${\cal V}_{\zeta_p}^{RT}$ for a given $p$. It is now natural to ask whether or not there exists a relationship between the ${\cal V}_p^{I}$ and the $\dov{\cal V}^{(j)}_p\,$, since they are both TQFT’s over ${\mathbb F}_p$ and are conjectured to share a list of other features. Identifications between the TQFT’s ${\cal V}_p^{I}$ obtained from quantm groups on the one hand and the ones obtained from the homological theory ${\cal V}_p^{FN}$ (and their Johnson Morita extensions) on the other hand will entail many insights into the relation between classical and quantum invariants as well as establish natural bases for the quantum theory in the language of the tensor calculus of the symplectic groups. Moreover, we expect this to lead to a deeper understanding of the geometric interpretations of the higher order terms in the cyclotomic integer expansions of the Reshetikhin Turaev theories, which include the Ohtsuki and, particularly, the Casson-Walker-Lescop invariants. Since the ${\cal V}_{\zeta_p}^{RT}$ are “unitarizable” theories and the dimensions of the vector spaces do not match any combination of ordinary Sp-representations we must expect the simple quotients $\dov{\cal V}^{(j)}_p$ to enter the relations rather than the reducible TQFT’s ${\cal V}_p^{(j)}$. This is precisely the point where the main results of this article are crucially needed. The resolutions of $p$-modular TQFT’s provide the missing link between the classical invariants defined over ${{\mathbb Z}}$ and the quantum invariants defined over ${\mathbb F}_p$. Thus far we are able to understand these relations and their applications in a rather detailed manner for $p=5$. We state next the main result from [@KerFib]. \[thm-KerFib\] We have the following isormophism of $\Gamma_g$-representations over ${{\mathbb F}}_5$: \[eq-TQFTID\] [V]{}\^[I]{}\_[5]{}(\_g)\^[(1)]{}\_5(\_g). In particular, for pairs $(M,\varphi)$ as in Theorem \[thm-pAlexander\] we have \[eq-Lescop\] [V]{}\^[I]{}\_[5]{}(M,)=\_5\^[(1)]{}(M,)+ \_5\^[(4)]{}(M,)=-2\_[Lescop]{}(M)5 The first equality in (\[eq-Lescop\]) follows directly from (\[eq-TQFTID\]). The second equality is now a consequence of the identification with the Alexander polynomial at a root of unity from Theorem \[thm-pAlexander\]. It is used in [@KerKyoto] to identify ${\cal V}^I_5(M,\varphi)$ with the ${\mathbb F}_5$-reduction of the Lescop invariant $\lambda_{Lescop}(M)\,$. This identity is special to $p=5$ and does not hold for $p>5$. Observe also that (\[eq-Lescop\]) implies that the invariant ${\cal V}^I_5(M,\varphi)$ is really independent of $\varphi$. This is not surprising since it is also equal to the lowest oder non-vanishing coefficient of the cyclotomic integer expansion of ${\cal V}^{RT}_{\zeta_5}(M)$. Nevertheless, this raises an interesting question, namely, which linear or polynomial combinations of the $\dov {\cal V}_p^{(j)}(M,\varphi)$ are independent of the choice of $\varphi$ and, thus, yield true invariants of closed 3-manifolds. #### Dimensions, Combinatorics, and Asymptotics: An important special case for any character formula is the implied relation for the dimensions of modules, that is, the characters evaluated at the unit element. The dimensions of the irreducible modules over ${\mathbb F}_5$ in this article are Fibonacci numbers, while the dimensions of the corresponding modules over ${{\mathbb Z}}$ are Catalan numbers. As a result we obtain in (\[eq-fibcatideven\]) and (\[eq-fibcatidodd\]) of Section 6 identities that express the Fibonacci numbers as 5-periodic, alternating sums in these Catalan numbers. Despite the simplicity of these relations they appear to be unknown in the available literature. We also determine in Section 6 the asymptotics of the dimensions of the TQFT modules for $g\to\infty$, which is summarized in the following lemma. \[lm-asym\] For a fixed prime $p\geq 3$ and fixed $j$, the dimensions of the vector spaces grow for large $g$ as follows: \[eq-PFasy\] [dim]{}([V]{}\^[I]{}\_p(\_g))\~[[F]{}-8mu[F]{}]{}\_p\^g (\^[(j)]{}\_p(\_g))\~[[f-4.4mu f]{}]{}\_p\^g where $\displaystyle \\| {{\sf F}\mkern -8mu{\sf F}}_p\\|=\frac p{4\sin^2(\frac \pi p)}\qquad{\rm and }\qquad \\| {{\sf f\mkern -4.4mu f}}_p\\|=4\cos^2(\frac \pi{2p})\;. $ Moreover, there are polynomials $R_j(f)\in{{\mathbb Z}}[f]$ of degree $\,{\rm deg}(R_j)=\frac{p-3}2\,$ with \[eq-FfPoly\] [[F]{}-8mu[F]{}]{}\_p=R\_p([[f-4.4mu f]{}]{}\_p). The polynomial dependence between the asymptotic behaviors from (\[eq-FfPoly\]) suggests a similar “polynomial” dependence between the TQFT’s. If the products of the dimensions are replaced by tensor products of the respective TQFT’s, this suggests that $\sim R_p(\\|{{\sf f\mkern -4.4mu f}}_p\\|)^g$ describes the asymptotics of parts of an $\frac {p-3}2$-fold tensor product of the $\dov {\cal V}^{(j)}_p$. We are thus led to the following conjecture on the constant order structure of the Reshetikhin Turaev theory, which can be verified for the genus=1 mapping class group (${\rm SL}(2,{{\mathbb Z}})$) representations. \[conj-sums\] The irreducible components of ${\cal V}^I_p$ can be found as irreducible components of $\displaystyle \bigotimes^{\frac {p-3}2}\dov {\cal V}^{FN}_p$. #### Acknowledgements: I am much indebted to Alexander Kleshchev for pointing out his results in [@KS99], and to Gordon James and Alain Reuter for checking the arguments in Section 5. I thank Ronald Solomon, Stephen Rallis, Adalbert Kerber, and Charles Frohman for comments and interest. Also, I have benefited much from discussions with Pat Gilmer on related integral TQFT’s and their applications. Finally, I would like to express my thanks to Tomotada Ohtsuki and Hitoshi Murakami for hospitality at the R.I.M.S., Kyoto, where final parts of this paper were completed, and for the opportunity to present and discuss the results. [2. Frohman-Nicas TQFT’s over ${{\mathbb Z}}$]{} [\[S2\]]{} In [@FroNic92] Frohman and Nicas construct a TQFT that allows the interpretation of the Alexander Polynomial as a weighted Lefschetz trace. The vector spaces are the cohomology rings of the Jacobians of the surfaces. In [@Ker01] we extract a natural Lefschetz $SL(2,{{\mathbb R}})$-action on these spaces, with respect to which this TQFT is equivariant. Let us describe in this sections the basic ingredients of the Frohman Nicas theory, reviewing also the conventions of [@Ker01]. We denote by $\Cob^{22fr}$ the category of [evenly framed]{} cobordisms between connected standard surfaces $\Sigma_g$. These are essentially the cobordisms obtained by an even number of surgeries, see Lemma 10, [@Ker01]. The Frohman Nicas TQFT is then given as a functor $$\label{eq-FNfunct} {\cal V}^{FN}\;: \;\;\Cob^{22fr}\;\longrightarrow\;SL(2,{{\mathbb R}})-mod_{{{\mathbb R}}}\;,$$ for which $${\cal V}^{FN}(\Sigma_g)\;\;=\;\;{\mbox{\raisebox{.4ex}{$\bigwedge^{\!}$}}\mkern-1mu} H_1(\Sigma_g)\;\;$$ for the surface $\Sigma_g$ of genus $g$. For each surface we pick a symplectic basis $\{a_1,\ldots, a_g, b_1,\ldots,b_g\}$ for $H_1(\Sigma_g)$ and introduce the inner form $\lz \_,\_\rz$, for which this basis orthonormal. We denote $i_:H_1(\Sigma_g)\to H_1(\Sigma_{g+1})$ the inclusion map compatible with the choice of bases and the construction of $\Sigma_{g+1}$ from $\Sigma_g$ by handle addition. The bases naturally provide lattice bases for ${\mbox{\raisebox{.4ex}{$\bigwedge^{\!}$}}\mkern-1mu} H_1(\Sigma_g,{{\mathbb Z}})$, given by monomials in the $a_i$ and $b_i$. Moreover, the inner form extends to to the exterior product to make the monomial basis an orthonormal one. We also denote by $\omega_g=\sum_{i=1}^ga_i\wedge b_i\in {\mbox{\raisebox{.4ex}{$\bigwedge^{\!2}$}}\mkern-1mu} H_1(\Sigma_g,{{\mathbb Z}})$ the symplectic form. The $\sl (2,{{\mathbb R}})$-action on ${\mbox{\raisebox{.4ex}{$\bigwedge^{\!}$}}\mkern-1mu} H_1(\Sigma_g)$ is given in terms of the standard generators $E$, $F$, and $H$ as [\[eq-defsl2act\]]{} H=(j-g) jH\_1(\_g), E=\_g, F = E\^\. From the explicit actions of the generators in [@Ker01] we see that the lattices ${\mbox{\raisebox{.4ex}{$\bigwedge^{\!}$}}\mkern-1mu} H_1(\Sigma_g, {{\mathbb Z}})$ are preserved by the maps ${\cal V}^{FN}(M)$ for cobordisms $M$ in $\Cob^{22fr}$. Moreover, also the standard generators of $\sl (2,{{\mathbb R}})$ preserve the lattices. Hence also the universal enveloping algebra over ${{\mathbb Z}}$ generated by the operators in (\[eq-defsl2act\]), which we shall (abusively) denote by $\sl (2,{{\mathbb Z}})$. We denote the respective functor into ${{\mathbb Z}}$-modules by: $$\label{eq-FNSL2funct} {\cal V}_{{{\mathbb Z}}}\;: \;\;\Cob^{22fr}\;\longrightarrow\;\sl(2,{{\mathbb Z}})-mod_{{{\mathbb Z}}}\;,$$ More specifically, the extension of the mapping class group in $\Cob^{22fr}$ factors under ${\cal V}_{{{\mathbb Z}}}$ through a split ${{\mathbb F}}_2={{\mathbb Z}}/2{{\mathbb Z}}$ extension of ${\rm Sp}(2g,{{\mathbb Z}})$, acting in the natural way on ${\mbox{\raisebox{.4ex}{$\bigwedge^{\!}$}}\mkern-1mu} H_1(\Sigma_g, {{\mathbb Z}})$. We also denote by $\sp(2g,{{\mathbb Z}})$ the algebra over ${{\mathbb Z}}$ generated by the standard ${\rm Sp}(2g,{{\mathbb R}})$ Lie algebra generators. It is not hard to show that in this representation it coincides with the group algebra ${{\mathbb Z}}[{\rm Sp}(2g,{{\mathbb Z}})]$. Besides the mapping class groups the other generators of $\Cob^{22fr}$ are the handle attachment cobordisms $H^+_g:\Sigma_g\to\Sigma_{g+1}$ and $H^-_g:\Sigma_{g+1}\to\Sigma_g$ with actions given by $$\label{eq-handle} {\cal V}_{{{\mathbb Z}}}(H^+_g)\;:\;\;\alpha\,\mapsto\,i_(\alpha)\wedge a_{g+1} \qquad\qquad\mbox{and}\qquad {\cal V}_{{{\mathbb Z}}}(H^-_g)={\cal V}_{{{\mathbb Z}}}(H^+_g)^\;,$$ where we use the notation $i_$ also for the inclusion $H_1(\Sigma_g)\into H_1(\Sigma_{g+1})$ extended to the exterior powers. The structure of ${\cal V}_{{{\mathbb Z}}}$ and the dual $\sl (2,{{\mathbb Z}})$-action can be better understood if we decompose the lattices ${\cal V}_{{{\mathbb Z}}}(\Sigma_g)$ according to $\sp(2g,{{\mathbb Z}})$-weights. We write an $\sp (2g)$-weight in the standard basis $\lambda=\sum_{i=1}^g\lambda_i\epsilon_i$ as given in [@GooWal98]. Specifically, we have $ha_i=\lz \epsilon_i, h\rz a_i$ and $hb_i=-\lz \epsilon_i, h\rz b_i$ for $h\in {\mathfrak h}$, the diagonal matrices in $\sp(2g,{{\mathbb Z}})$. We denote by $\nabla_g=\{\lambda\in {\mathfrak h}^:\;\lambda_i\in\{-1,0,1\}\}$ the set of possible weights of vectors in ${\mbox{\raisebox{.4ex}{$\bigwedge^{\!}$}}\mkern-1mu} H_1(\Sigma_g)$. This yields a decomposition into weight spaces denoted as follows. $${\cal V}_{{{\mathbb Z}}}(\Sigma_g)\;\;=\;\;\bigoplus_{\lambda\in\nabla_g} {\cal W}_{{{\mathbb Z}}}(\lambda,g)$$ For a given weight $\lambda\in \nabla_g$ let $N({\lambda})=\{ i:\,\lambda_i=0\}\subset\{1,\ldots,g\}$ and $n(\lambda)=\|N({\lambda})\|$. A special vector $w(\lambda)\in {\cal W}_{{{\mathbb Z}}}(\lambda,g)$ is given by $$w(\lambda)=w_1(\lambda_1)\wedge\ldots\wedge w_g(\lambda_g) \;\in{\mbox{\raisebox{.4ex}{$\bigwedge^{\!g-n(\lambda)}$}}\mkern-1mu} H_1(\Sigma_g) \qquad\quad\mbox{where}\quad \left\{ \begin{array}{rcl}w_i(1)&=&a_i \\ w_i(-1)&=&b_i\\ w_i(0)&=&1 \end{array}\right. \;\;.$$ Let $e_{\pm}$ be generators of a 2-dimensional lattice $\lz e_-, e_+\rz_{{{\mathbb Z}}}$, and $L_{{{\mathbb Z}}}^n=\lz e_-, e_+\rz_{{{\mathbb Z}}}^{\otimes n}$ the lattice of rank $2^n$. Given $N(\lambda)=\{j_1,\ldots, j_{n(\lambda)}\}\,$ with $j_1<\ldots < j_{n(\lambda)}$ we thus define maps $$\begin{aligned} \label{eq-upsilon} \Upsilon_{\lambda}\,&: &\;\;\;\quad L_{{{\mathbb Z}}}^{n(\lambda)}\, \;\;\stackrel{\cong}{-\!\!\!-\!\!\!-\!\!\!\longrightarrow} \;\;\, {\cal W}_{{{\mathbb Z}}}(\lambda,g)\\ \;&:&\;\;e_{\epsilon_1}\otimes\ldots\otimes e_{\epsilon_{n}} \;\mapsto\;w(\lambda)\wedge o_{j_1}(\epsilon_1)\wedge\ldots\wedge o_{j_{n}}(\epsilon_{n}) \nonumber \\ &&\mbox{where}\quad n=n(\lambda)\quad\mbox{and} \quad\quad\left\{\begin{array}{l} o_j(+)=a_j\wedge b_j\\ o_j(-)= 1\nonumber \end{array} \right.\;\;\;. \end{aligned}$$ The lattices $L_{{{\mathbb Z}}}^n$ also have a natural inner product for which the monomials $e_{\epsilon_1}\otimes\ldots\otimes e_{\epsilon_n}$ form an orthonormal basis, and carry a natural $\sl (2,{{\mathbb Z}})$-action given by $Ee_-=e_+$, $Fe_+=e_-$, $He_{\pm}=\pm e_\pm$, and $Ee_+=Fe_-=0$, for which also $E^=F$ and $H^=H$. \[lm-upsilambda\] The $\Upsilon_{\lambda}$ are $\sl (2,{{\mathbb Z}})$-equivariant isomorphisms of lattices with inner forms. [Proof:]{} They are obviously isomorphisms of lattices as they map orthonormal bases to each other. It is also easy to see that the $H$-weight for a monomial is $-n(\lambda)+2\sum_{i=1}^{n(\lambda)}\epsilon_i$ on both sides of (\[eq-upsilon\]). Now, $E$ is multiplication with $\omega_g=\sum_ja_j\wedge b_j$. Clearly, $w(\lambda)\wedge a_j\wedge b_j=0$ if $j\not\in N(\lambda)$ so we multiply only with $\omega_{\lambda}=\sum_{i=1}^{n(\lambda)}a_{j_i}\wedge b_{j_i}= \sum_{i=1}^{n(\lambda)}o_{j_i}(+)$. The $E$-equivariance then follows from $o_j(+)\wedge o_j(+)=0$ and $o_j(-)\wedge o_j(+)=o_j(+)$. $F$-equivariance follows from $F^=E$. Although not of immediate necessity for the main result let us record here also how the actions of morphisms ${\cal V}_{{{\mathbb Z}}}(M)$ on the lattices $L_{{{\mathbb Z}}}^n$ look like. We give them in terms of generators of $\sp(2g,{{\mathbb Z}})$ and the handle attaching maps. We introduce $\sl (2,{{\mathbb Z}})$-equivariant morphisms $$coev_k\;=\;{{\mathbb I}}_{k-1}\otimes coev\otimes {{\mathbb I}}_{n-k}\;\;:\;\; \qquad L_{{{\mathbb Z}}}^n\;\;\longrightarrow \;\; L_{{{\mathbb Z}}}^{n+2} \qquad\qquad\qquad\qquad$$ $$\mbox{where} \qquad coev\,:\; 1\,\mapsto\, (e_-\otimes e_+-e_+\otimes e_-) \qquad\qquad \mbox{and}\quad {{\mathbb I}}_n\,=\,id_{L^n_{{{\mathbb Z}}}}\;.$$ $$\mbox{and}\qquad ev_k= - (coev_k)^:\;\;L_{{{\mathbb Z}}}^{n+2}\longrightarrow L_{{{\mathbb Z}}}^n \qquad\qquad\qquad$$ The obey relation $$ev_{k\pm 1}\circ coev_k\;=\;{{\mathbb I}}\qquad\qquad\mbox{and}\qquad\qquad ev_{k}\circ coev_k\;=\; -2\cdot{{\mathbb I}}\;.$$ For a map $\phi:{\cal W}_{{{\mathbb Z}}}(\lambda,g)\to {\cal W}_{{{\mathbb Z}}}(\lambda',g')$ we denote $\phi^{\Upsilon}=\Upsilon_{\lambda'}^{-1}\circ \phi\circ \Upsilon_{\lambda}: L_{{{\mathbb Z}}}^{n(\lambda)}\to L_{{{\mathbb Z}}}^{n(\lambda')}\,$. Moreover, as in [@GooWal98], we denote the standard standard generators $e_{\alpha_i}$ and $f_{\alpha_j}$ of $\sp (2g,{{\mathbb Z}})$ for simple roots $\alpha_i=\epsilon_i-\epsilon_{i+1}$ for $i<g$ and $\alpha_g=2\epsilon_g$. We have $e_{\alpha_i}a_{i+1}=a_i$, $e_{\alpha_i}b_i=-b_{i+1}$, and $e_{\alpha_i}v=0$ for all other basis vectors $v$ if $i<g$. Also $e_{\alpha_g}b_g=e_g$ and $e_{\alpha_g}v=0$ for all others. This further determines the action of the other generators with $f_{\alpha_i}=e_{\alpha_i}^$. Denote now by $e_{\alpha,\lambda}$ the restriction $e_{\alpha}:{\cal W}_{{{\mathbb Z}}}(\lambda,g)\to {\cal W}_{{{\mathbb Z}}}(\lambda+\alpha,g)$, where we put $e_{\alpha,\lambda}=0$ of $\lambda+\alpha\not \in \nabla_g$. Let us also denote the restriction of the handle attaching map $H_{\lambda}={\cal V}_{{{\mathbb Z}}}(H^+_g): {\cal W}_{{{\mathbb Z}}}(\lambda,g)\to {\cal W}_{{{\mathbb Z}}}(\lambda+\epsilon_{g+1},g+1)$. The following is the result of a straight forward calculation. \[lm-Spe-action\] For $\lambda\in\nabla_g$ and $n=n(\lambda)$ we have, when $i<g$, $$e_{\alpha_i,\lambda}^{\Upsilon}\;\;=\;\; \left\{ \begin{array}{cll} \quad{{\mathbb I}}_n\quad & \mbox{for\ \ \ } (\lambda_i,\lambda_{i+1})=(0,1)&\\ -{{\mathbb I}}_n& \mbox{for\ \ \ } (\lambda_i,\lambda_{i+1})=(-1,0)&\\ coev_k& \mbox{for\ \ \ } (\lambda_i,\lambda_{i+1})=(-1,1)& \mbox{with\ \ \ } j_{k-1}<i<j_k\\ -ev_k& \mbox{for\ \ \ } (\lambda_i,\lambda_{i+1})=(0,0)& \mbox{with\ \ \ } i=j_k,\;i+1=j_{k+1}\\ 0& \mbox{elsewise}&\\ \end{array}\right. \;\;,$$ $$e_{\alpha_g,\lambda}^{\Upsilon}\;\;=\;\; \left\{ \begin{array}{cll} \quad{{\mathbb I}}_n\quad & \mbox{for\ \ \ } \lambda_g\,=\,-1&\\ 0& \mbox{elsewise}&\\ \end{array}\right.\;\;, \qquad\mbox{and}\qquad H^{\Upsilon}_{\lambda}={{\mathbb I}}_n \qquad.\qquad\qquad\qquad$$ Another prominent, and for our purposes more important, action on the lattices ${\cal V}_{{{\mathbb Z}}}(\Sigma_g)$ is that of subgroups of the Weyl group ${\mathfrak W}_g\cong ({{\mathbb F}}_2)^g\rtimes S_g$ of ${\rm Sp}(2g,{{\mathbb Z}})$, where $S_g$ denotes the symmetric group in $g$ letters. The $j$-th ${{\mathbb F}}_2$-generator of ${\mathfrak W}_g$ acts on weights by changing the sign of $\lambda_j$. It is realized as a subgroup $\widehat{\mathfrak W}_g\subset {\rm Sp}(2g,{{\mathbb Z}})$ by an extension $1\to {\mathbb F}_2^g\to\widehat{\mathfrak W}_g\to {\mathfrak W}_g \to 1$, with $\widehat{\mathfrak W}_g\cong ({{\mathbb F}}_4)^g\rtimes S_g$. The ${{\mathbb F}}_4$-generators are given by the “$S$-matrices” ${\sf S}_j\in {\rm Sp}(2g,{{\mathbb Z}})$, see [@Ker01], defined by ${\sf S}_ja_j=-b_j$, ${\sf S}_jb_j=a_j$ and ${\sf S}_ja_i=a_i$ and ${\sf S}_jb_i=b_i$ for $i\neq j$. We specify two relevant representations of $\widehat{\mathfrak W}_k$: 1. $L^k_{{{\mathbb Z}}}\cong ({{\mathbb Z}}^2)^{\otimes k}$: This action factors through $\widehat{\mathfrak W}_k\to S_k$ the symmetric group, which acts canonically on the lattice by permutation of factors. 2. $M^k_{{{\mathbb Z}}}\cong ({{\mathbb Z}}^2)^{\otimes k}$: Here the $j$-th ${{\mathbb F}}_4$-factor is represented by the matrix ${\sf S}_j=\scriptscriptstyle \left[\matrix{0 & 1 \cr -1 & 0}\right ]$ acting on the $j$-th factor of the tensor product. The action of ${\sf S}_k$ on $M^k_{{{\mathbb Z}}}$ is the canonical representation multiplied by the alternating representation, i.e., $\sigma(v_1\otimes\ldots \otimes v_k)=sign(\sigma) v_{\sigma^{-1}(1)}\otimes\ldots \otimes v_{\sigma^{-1}(k)}$. We have natural subgroups $\widehat{\mathfrak W}_{g-n}\times \widehat{\mathfrak W}_n\subset \widehat{\mathfrak W}_g$, for which the right coset $$C^g_n\;=\; \raise3pt\hbox{$\widehat{\mathfrak W}_g$}\Big / \raise-3pt\hbox{$\widehat{\mathfrak W}_{g-n}\times \widehat{\mathfrak W}_n$} \;=\; \raise3pt\hbox{$S_g$}\Big / \raise-3pt\hbox{$S_{g-n}\times S_n$}\;\;,$$ is identified with the set of subsets $A\subset \{1,\ldots, g\}$ of size $\|A\|=n$. We denote $$\begin{aligned} \label{eq-Wnotation} {\cal W}_{{{\mathbb Z}}}(n,g)&\,=\,&\bigoplus_{A\in C^g_n} {\cal W}_{{{\mathbb Z}}}(A,g) \,=\,\bigoplus_{\lambda\in\nabla_g: n(\lambda)=n} {\cal W}_{{{\mathbb Z}}}(\lambda,g) \;, \qquad\qquad\qquad \\ \ \ &\mbox{where}&\qquad {\cal W}_{{{\mathbb Z}}}(A,g)\,=\!\!\bigoplus_{\lambda\in\nabla_g:N(\lambda)=A} {\cal W}_{{{\mathbb Z}}}(\lambda,g) \;. \end{aligned}$$ Clearly, the summands of ${\cal V}_{{{\mathbb Z}}}=\bigoplus_n{\cal W}_{{{\mathbb Z}}}(n,g)$ are invariant under the $\widehat{\mathfrak W}_g$-action for each $n$. These subrepresentations are identified next as induced representations. \[lm-induce\] For every $n$ with $0\leq n\leq g$ there is a natural isomorphism of $\widehat{\mathfrak W}_g$-modules $$\Upsilon\;:\;\; Ind^{\,\,\widehat{\mathfrak W}_g}_{\widehat{\mathfrak W}_{g-n}\times \widehat{\mathfrak W}_n} \Bigl(M^{g-n}_{{{\mathbb Z}}}\otimes L^n_{{{\mathbb Z}}}\Bigr) \;\stackrel{\cong}{\longrightarrow}\;\;{\cal W}_{{{\mathbb Z}}}(n,g)\;.$$ This map is an $\sl(2,{{\mathbb Z}})$-equivariant isometry. [Proof:]{} For $N_n=\{g-n+1,\ldots,g\}$ the $\widehat{\mathfrak W}_{g-n}\times \widehat{\mathfrak W}_n$-module is readily identified via the isomorphism $$M^{g-n}_{{{\mathbb Z}}}\otimes L^n_{{{\mathbb Z}}}\;\stackrel{\cong}{\longrightarrow} \;{\cal W}_{{{\mathbb Z}}}(N_n,g)\;:\;\; e_{\lambda_1}\otimes\ldots\otimes e_{\lambda_{n-g}}\otimes l \;\mapsto\;\Upsilon_{[\lambda_1,\ldots,\lambda_{n-g},0,\ldots,0]}(l)\;.$$ with the submodule ${\cal W}_{{{\mathbb Z}}}(N_n,g)\subset{\mbox{\raisebox{.4ex}{$\bigwedge^{\!}$}}\mkern-1mu} H_1(\Sigma_g)$, where the action is defined by restricting the action of $\widehat{\mathfrak W}_g$. We next define a natural section \[eq-section\] : C\^g\_n S\_g\_g: A\_A as follows. Let $A\subseteq\{1,\ldots,g\}$ with $n=\|A\|$. There is a unique permutation $\pi_A\in S_g$ such that $A=\pi_A(\{g-n+1,\ldots,g\})$, $\pi_A(1)<\pi_A(2)<\ldots<\pi_A(g-n)$, and $\pi_A(g-n+1)<\pi_A(g-n+2)<\ldots<\pi_A(g)$. Clearly, we have $$\pi_A:{\cal W}_{{{\mathbb Z}}}(N_n,g)\;\stackrel{\cong}{\longrightarrow}\; {\cal W}_{{{\mathbb Z}}}(A,g)\;.$$ The induced representation by definition the space of all maps $f: \widehat{\mathfrak W}_g \to M^{g-n}_{{{\mathbb Z}}}\otimes L^n_{{{\mathbb Z}}}$ such that $f(\sigma \eta)=\eta^{-1}.(f(\sigma))$ for $\eta\in \widehat{\mathfrak W}_{g-n}\times \widehat{\mathfrak W}_n$, equipped with the left regular $\widehat{\mathfrak W}_g$-action $(\sigma f)(\sigma')=f(\sigma^{-1}\sigma')$. Now, every $\sigma\in \widehat{\mathfrak W}_g$ has a unique decomposition $\sigma=\pi_{\sigma(N_n)}\eta_\sigma$, with $\eta_\sigma\in \widehat{\mathfrak W}_{g-n}\times \widehat{\mathfrak W}_n$. Thus we may identify the induced representation with the space of maps $\overline f: C_n^g\to M^{g-n}_{{{\mathbb Z}}}\otimes L^n_{{{\mathbb Z}}}$, setting $\overline f (A)=f(\pi_A)$ and hence $f(\sigma)=\eta_\sigma^{-1}.\overline f(\sigma(N_n))$. The isomorphism is now given by \[eq-isomind\] IndMap(C\_n\^g, M\^[g-n]{}\_[[[Z]{}]{}]{}L\^n\_[[[Z]{}]{}]{})\_[[[Z]{}]{}]{}(n,g): f (f)= \_[AC\^g\_n]{}\_A(f(A)). An inverse is obtained by mapping $v\in {\cal W}_{{{\mathbb Z}}}(A,g)$ to $\pi^{-1}_A(v)\otimes \delta_{\{A\}}$, where $\delta_{\{A\}}(B)=1$ for $A=B$ and 0 elsewise. In order to show that it is equivariant let $\overline f$ be an arbitrary map $C^g_n\to {\cal W}_{{{\mathbb Z}}}(N_n,g)$, $\sigma\in \widehat{\mathfrak W}_g$, and $ \overline f'=\sigma(\overline f)$. Let $A\in C^g_n$ and define $\eta_{A,\sigma}\in {\mathfrak W}_{g-n}\times \widehat{\mathfrak W}_n$ by $\sigma^{-1}\pi_A=\pi_{\sigma^{-1}(A)}\eta_{A,\sigma}^{-1}$. Hence we have $\overline f'(A)=f'(\pi_A)=(\sigma f)(\pi_A)=f(\sigma^{-1}\pi_A) = f(\pi_{\sigma^{-1}(A)}\eta_{A,\sigma}^{-1})=\eta_{A,\sigma}f(\pi_{\sigma^{-1}(A)})$. Thus $$\begin{array}{lllll} \Upsilon(\overline f') &=&\displaystyle \bigoplus_{A\in C^g_n}\pi_A(\overline f'(A))&=&\displaystyle \bigoplus_{A\in C^g_n}\pi_A\eta_{A,\sigma}\overline f(\sigma^{-1}A)=\\ &=&\displaystyle \bigoplus_{A\in C^g_n} \sigma\pi_{\sigma^{-1}(A)}\overline f(\sigma^{-1}A) &=& \displaystyle \bigoplus_{B\in C^g_n} \sigma\pi_{B}\overline f(B)=\\ &=& \displaystyle \sigma(\bigoplus_{B\in C^g_n}\pi_B(\overline f(B)))&=& \displaystyle \sigma (\Upsilon(\overline f))\;, \end{array}$$ which is what we needed to show. Isometry of $\Upsilon$ is with respect to the natural inner product on $Map(C_n^g,\, M^{g-n}_{{{\mathbb Z}}}\otimes L^n_{{{\mathbb Z}}})$ given by \[eq-indinner\] f, h=\_[AC\^g\_n]{} f (A), g (A)given the inner form on $M^{g-n}_{{{\mathbb Z}}}\otimes L^n_{{{\mathbb Z}}}$. Also, as $L^n_{{{\mathbb Z}}}$ is an $\sl(2,{{\mathbb Z}})$-module also $Map(C_n^g,\, M^{g-n}_{{{\mathbb Z}}}\otimes L^n_{{{\mathbb Z}}})$ is. Both, equivariance and isometry, follows immediately from the form of the isomorphism in (\[eq-isomind\]) and the fact that the $\pi_A$ are isometric equivariant maps. The relation of the isomorphisms in Lemma \[lm-induce\] to the ones in Lemma \[lm-upsilambda\] is given by the restrictions of $\Upsilon^{-1}$ to the weight spaces \[eq-relupsis\] \_[N()]{}\^[-1]{}=\_[\^]{}\_\^[-1]{} :(,g)(\^,g). (N\_n,g) Here $\lambda^{\pi}$ denotes the the “sign-content” of a weight $\lambda\in\nabla_g$ defined by $$\lambda^{\pi}:=\pi_{N(\lambda)}^{-1}\lambda= \;\sum_{j=1}^{g-n(\lambda)} \lambda_{\pi_{N(\lambda)}(j)}\epsilon_j \;\;=\;\; \pm\epsilon_1+ \dots \pm\epsilon_{g-n(\lambda)}\;.$$ [3. Lefschetz Decompositions and Specht Modules]{} \[S3\] As in [@Ker01] we consider the decomposition of the Frohman Nicas TQFT according to $SL(2,{{\mathbb R}})$-representations: $${\cal V}_{{{\mathbb Z}}}^{FN}\;=\;\;\bigoplus_{j\geq 1} V_{j}\otimes {\cal V}^{(j)}\;,$$ where $V_{j}$ is the $j$-dimensional irreducible representation of $SL(2,{{\mathbb R}})$. Note, that we the convention we use here for the superscript in ${\cal V}^{(j)}$ is shifted by one from the one used in [@Ker01]. For weights we follow the notations of [@GooWal98]. The sublattices of the irreducible TQFT components can be defined as the $SL(2,{{\mathbb R}})$ lowest weight spaces $${\cal V}^{(j)}_{{{\mathbb Z}}}(\Sigma)\;\;=\;\;\{v\in {\cal V}_{{{\mathbb Z}}}(\Sigma)\,: \; Fv=0\;\;\;\mbox{and}\;\;\;Hv=-(j-1)v\}\;\; =\;\;{\mbox{\raisebox{.4ex}{$\bigwedge^{\!g-j+1}$}}\mkern-1mu}H_1(\Sigma,{{\mathbb Z}})\cap ker(F)\;\;.$$ The representation of ${\rm Sp}(2g,{{\mathbb Z}})$ on ${\cal V}^{(j)}_{{{\mathbb Z}}}(\Sigma)$ is irreducible of fundamental heighest weight $\varpi_{g-j+1}=\epsilon_1+\ldots +\epsilon_{g-j+1}\,$ with heighets weight vector $$w(\varpi_{g-j+1})=a_1\wedge\ldots\wedge a_{g-j+1}=\Upsilon_{\varpi_{g-j+1}} (e_-^{\otimes j+1})\;.$$ The possible weights in this representation are given by $$\nabla_{g}^{(j)}\;\;=\;\;\{\lambda\in\nabla_g\,:\;n(\lambda)\geq j-1\;\mbox{and}\;n(\lambda)\equiv j -1\mod 2\,\}\;.$$ We obtain an analogous weight space decomposition \[eq-weightcomp\] [V]{}\_[[[Z]{}]{}]{}\^[(j)]{}(\_g)=\_[\^[(j)]{}\_g]{} [W]{}\^[(j)]{}\_[[[Z]{}]{}]{}(,g) \_[\^[(j)]{}\_g]{} L\^[n()]{}\_[[[Z]{}]{}]{}ker(H+j-1). \[lm-spechtiso\] The spaces ${\cal W}^{(j)}_{{{\mathbb Z}}}(\lambda,g)$ are as $S_{n(\lambda)}$-modules isomorphic to the standard irreducible Specht modules ${\cal S}^{[a,b]}$ for the two-row Young-diagram $$[a,b]\;=\;\Bigr [ \frac {n(\lambda)+j-1}2, \frac {n(\lambda)-j+1}2\Bigr]\;.$$ [Proof:]{} Although this appears to be standard we shall provide a proof to fix conventions. We largely follow here the definitions and notations of [@Jam78]. First we note that $\,ker(H+j-1)$ is naturally isomorphic to the permutation module $M^{[a,b]}$, where $n(\lambda)=a+b$ and $j=a-b+1$. The isomorphism $M^{[a,b]}\stackrel{\cong}{\longrightarrow} ker(H+j-1)\cap L^{n(\lambda)}_{{{\mathbb Z}}}\,$ maps a tabloid $\{t\}=\!\!\!\!\raise-8pt\hbox{ \bstid i_1 & i_2 & \ldots & i_a \cr j_1& \ldots & j_b \cr \estid}\,\, $ to the basis vector $e_{\{t\}}\,=\,e_{\epsilon_1}\otimes e_{\epsilon_2}\otimes\ldots\otimes e_{\epsilon_N}$ with $\epsilon_k=+$ if $k\in\{j_1,\ldots,j_b\}$ and $\epsilon_k=-$ if $k\in\{i_1,\ldots,i_a\}$. It is obvious that the $e_{\{t\}}$ indeed span ${\rm ker}(H+j-1)$. Consider a tableau $t= {\!\!\raise-8pt\hbox{\bsyg i_1 \cr j_1 \cr \esyg}}\! {\,\, \cdots \!\!\!\raise-8pt\hbox{\bsyg i_b & i_{b+1} \cr j_b \cr \esyg}}\! {\,\,\raise5pt\hbox{$\cdots$}\!\!\! \raise2pt\hbox{\bsyg i_a\cr \esyg}}\, $ of shape $[a,b]$. Let $C_t$ be the column stabilizer group and \[eq-defkappa\] \_t=\_[C\_t]{}sign()=\_[k=1]{}\^b(1-(i\_k,j\_k)) the signed column sum. We set $e_t\;=\;\kappa_t e_{\{t\}}\;$. The Specht module ${\cal S}^{[a,b]}$ is the space generated by these $e_t$ and we want to show that it coincider with $ker(F)$. The easy part is to show ${\cal S}^{[a,b]}\subset ker(F)$. Using that $F$ commutes with $\kappa_t$ we compute $Fe_t=F\kappa_t e_{\{t\}}=\kappa_t Fe_{\{t\}}= \sum_{k=1}^b\kappa_t e_{\{t_k\}}$, where $\{t_k\}$ is the tabloid of shape $[a+1,b-1]$, in which we have removed $j_k$ from the bottom row and added to the top row. As a result $(1-(i_k,j_k))\{t_k\}=0$, hence $\kappa_t\{t_k\}=0$ so that $Fe_t=0$. In order to prove $ker(F)\subset {\cal S}^{[a,b]}$ we proceed by induction. $ker(F)$ on $M^{[a,b]}$ is given by $ker(F^2)$ on $M^{[a-1,b]}$. We have a map $$ker(F^2)\cap M^{[a-1,b]}\;\longrightarrow\; ker(F)\cap M^{[a,b]}\;\;:\;\;\; x\;\mapsto\;e_-\otimes x\,-\,e_+\otimes Fx\;\;.$$ It is easy to see that this map is an isomorphism. We use first that $Fx\in ker(F)\cap M^{[a,b-1]}$ and hence by induction $Fx=\sum_{t\in T} b_te_t$. Here $T$ is the set of tableaux of the form $t= {\!\!\raise-8pt\hbox{\bsyg i_1 \cr j_1 \cr \esyg}}\! {\,\, \cdots \!\!\!\raise-8pt\hbox{\bsyg i_{b-1} & i_{b} \cr j_{b-1} \cr \esyg}}\! {\,\,\raise5pt\hbox{$\cdots$}\!\!\! \raise2pt\hbox{\bsyg i_a\cr \esyg}}\, $ with numbers from $\{2,3,\ldots, a+b\}$. Now as $b-1<a$ we have that $i_a$ is not permuted by $C_t$. Thus if we denote $E_j={{\mathbb I}}^{\otimes j-1}\otimes E\otimes {{\mathbb I}}^{\otimes N-j}$ we have that $E_{i_a}$ commutes with $\kappa_t$. Let $ h_t=\kappa_tE_{i_a}e_{\{t\}} $ Now $Fh_t=FE_{i_a}\kappa_t e_{\{t\}}= [F,E_{i_a}]\kappa_t e_{\{t\}}+E_{i_a}F\kappa_t e_{\{t\}} = -H_{i_a}\kappa_t e_{\{t\}}$ since $e_t\in ker(F)$ as shown above. Now, $H_{i_a}$ commutes with $\kappa_t$ and $H_{i_a}e_{\{t\}}=-e_{\{t\}}$ by construction. Hence $ Fh_t=e_t $ In other words ${\cal S}^{[a,b-1]}\subset im(F)$. Consider now $y=x-\sum_{t\in T}b_th_t$. We thus have $y\in ker(F)\cap M^{[a-1,b]}$ so that by induction $y=\sum_{s\in S}c_s e_s$, where $S$ denotes the tableau of the form $ s=\!\!\! {\!\!\raise-8pt\hbox{\bsyg k_1 \cr l_1 \cr \esyg}}\! {\,\, \cdots \!\!\!\raise-8pt\hbox{\bsyg k_b & k_{b+1}\cr l_b \cr \esyg}}\! {\,\,\raise5pt\hbox{$\cdots$}\!\!\! \raise2pt\hbox{\bsyg k_{a-1}\cr \esyg}}\, $ with all numbers in $\{2,3,\ldots, a+b\}$. Inserting everything we find $$z\;=\;e_-\otimes x\,-\,e_+\otimes Fx\;\;=\;\; \sum_{s\in S}c_s e_-\otimes e_s\;\;+\;\; \sum_{t\in T} b_t(e_-\otimes h_t-e_+\otimes e_t)\;.$$ Now, it is not hard to see that $e_{\hat s}= (e_-\otimes e_s)$ with $ \hat s=\!\!\! {\!\!\raise-8pt\hbox{\bsyg k_1 \cr l_1 \cr \esyg}}\! {\,\, \cdots \!\!\!\raise-8pt\hbox{\bsyg k_b & k_{b+1}\cr l_b \cr \esyg}}\! {\,\,\raise5pt\hbox{$\cdots$}\!\!\! \raise2pt\hbox{\bsyg k_{a-1} & 1\cr \esyg}}\, $ given that $a-1\geq b$ so that 1 is not permuted by $C_{\hat s}$. Moreover, $(e_-\otimes h_t-e_+\otimes e_t)=(1-(1,i_a))(e_-\otimes h_t) (1-(1,i_a))\kappa_tE_{i_a}e_{\{t\}}=\kappa_{\hat t}e_{\{\hat t\}}=e_{\hat t}$, where $ \hat t=\!\! {\!\!\raise-8pt\hbox{\bsyg 1 & i_1 \cr i_a & j_1 \cr \esyg}}\! {\,\, \cdots \!\!\!\raise-8pt\hbox{\bsyg i_{b-1} & i_b\cr j_{b-1} \cr \esyg}}\! {\,\,\raise5pt\hbox{$\cdots$}\!\!\! \raise2pt\hbox{\bsyg i_{a-1}\cr \esyg}} $ Thus $z\;=\;\sum_{\hat s}c_se_{\hat s}\,+\,\sum_{\hat t}b_te_{\hat t}$ so that $z\in {\cal S}^{[a,b]}$. Combining Lemma \[lm-induce\] with Lemma \[lm-spechtiso\] and using notation ${\cal W}^{(j)}_{{{\mathbb Z}}}(A,g)$ and ${\cal W}^{(j)}_{{{\mathbb Z}}}(n,g)$ with $n\geq j-1$ and $ n \equiv j-1 \mod 2$ analogous to (\[eq-Wnotation\]) so that ${\cal V}_{{{\mathbb Z}}}^{(j)}(\Sigma_g)=\bigoplus_n {\cal W}^{(j)}_{{{\mathbb Z}}}(n,g)$ we find the following structure. \[cor-UpsSpecWei\] For every $1\leq j-1\leq n\leq g$ with $n \equiv j-1 \mod 2$ there is an isomorphism of $\widehat{\mathfrak W}_g$-modules $$\Upsilon^{(j)}\;\;:\;\;\; Ind^{\,\,\widehat{\mathfrak W}_g}_{\widehat{\mathfrak W}_{g-n}\times \widehat{\mathfrak W}_n} \Bigl(M^{g-n}_{{{\mathbb Z}}}\otimes {\cal S}^{[\frac{n+j-1}2,\frac{n-j+1}2 ]}\Bigr) \;\;\;\stackrel{\cong}{\longrightarrow}\;\;\;\;{\cal W}_{{{\mathbb Z}}}^{(j)}(n,g)\;.$$ Let us also describe the $\sp (2g,{{\mathbb Z}})$-generators on the vectors $e_t$ spanning the Specht modules ${\cal W}^{(j)}(\lambda,g)$. To this end it is convenient to use tableaux in which entries are takes from the set $N(\lambda)$ rather than $\{1,\ldots, n(\lambda)\}$, related to the standard ones by application of $\pi_{N(\lambda)}$. We denote the set of these tableaux by $T^{(j)}(\lambda)$. \[lm-spechtTQFTgen\] The $\sp (2g,{{\mathbb Z}})$-generators act on the tableau vectors of Specht modules in the ${\cal V}^{(j)}_{{{\mathbb Z}}}$-TQFT as follows. For the $e_{\alpha_i,\lambda}$ with $1\leq i\leq g-1$ we have: 1. If $(\lambda_i,\lambda_{i+1})=(0,1)$ then $e_{\alpha_i,\lambda}e_t=e_s$ where $s\in T^{(j)}(\lambda+\alpha_i)$ is obtained from $t\in T^{(j)}(\lambda)$ by replacing the label $i\in N(\lambda)$ in $t$ by the label ${i+1}\in N(\lambda+\alpha_i)$. 2. If $(\lambda_i,\lambda_{i+1})=(-1,0)$ then $e_{\alpha_i,\lambda}e_t=-e_s$ where $s\in T^{(j)}(\lambda+\alpha_i)$ is obtained from $t\in T^{(j)}(\lambda)$ by replacing the label $i+1\in N(\lambda)$ in $t$ by the label $i\in N(\lambda+\alpha_i)$. 3. If $(\lambda_i,\lambda_{i+1})=(-1,1)$ we have $\{i,i+1\}= N(\lambda+\alpha_i)- N(\lambda)$ and $e_{\alpha_i,\lambda}e_t=e_s$ where $s\in T^{(j)}(\lambda+\alpha_i)$ by adding a column -7pt to $t\in T^{(j)}(\lambda)$ 4. If $(\lambda_i,\lambda_{i+1})=(0,0)$ so that $\{i,i+1\}= N(\lambda)- N(\lambda+\alpha_i)$ then $e_{\alpha_i,\lambda}e_t$, with $t\in T^{(j)}(\lambda)$, is 1. 0 if the labels $i$ and $i+1$ occur in columns of height 1. 2. $2 e_s$ if $t$ is given by adding the column -7pt to $s$. 3. $e_s$ if $i$ and $i+1$ occur in different columns of height 2. Here $s$ is obtained from $t$ by replacing the double column -7pt by -7pt 4. $e_s$ if $i$ is in column of height 2 and and $i+1$ in columns of height 1, where $s$ is obtained from $t$ by deleting the column -7pt and replacing -3pt by -3pt 5. $e_s$ if $i+1$ is in column of height 2 and and $i$ in columns of height 1, where $s$ is obtained from $t$ by deleting the column -7pt and replacing -3pt by -3pt All other cases of positions of $i$ and $i+1$ follow from the symmetry properties of the vectors $e_t$ under permutations of columns or within columns. Since $N(\lambda+\alpha_g)=N(\lambda)$ if $\lambda+\alpha_g, \lambda\in \nabla_g$ we have that $e_{\alpha_g, \lambda}$ acts as identity also on the vectors $e_t$. Similarly, $H_{\lambda}$ acts as identity. [4. ${{\mathbb F}}_p={{\mathbb Z}}/p{{\mathbb Z}}$-Reductions and the Sequences ${\cal C}_{p,k}$]{} \[S4\] For the remainder of this article let $p$ be an odd prime number. Since the TQFT’s ${\cal V}^{(j)}_{{{\mathbb Z}}}$ are defined over free ${{\mathbb Z}}$-modules (lattices) we naturally obtain TQFT’s ${\cal V}^{(j)}_p$ over the number field ${{\mathbb F}}_p$ by setting $${\cal V}^{(j)}_p(\Sigma_g)\;=\; \raise5pt\hbox{${\cal V}^{(j)}_{{{\mathbb Z}}}(\Sigma_g)$}\Big / \raise-5pt\hbox{$p {\cal V}^{(j)}_{{{\mathbb Z}}}(\Sigma_g)$}$$ Now, each ${\cal V}^{(j)}_{{{\mathbb Z}}}(\Sigma_g)$ inherits a non-degenerate inner product as a sublattice of ${\mbox{\raisebox{.4ex}{$\bigwedge^{\!}$}}\mkern-1mu}H_1(\Sigma_g,{{\mathbb Z}})$. This, however, will in general degenerate if we consider the the $p$-reduction $\lz\,,\,\rz_p:\Bigr({\cal V}^{(j)}_p(\Sigma_g)\Bigl)^{\otimes 2}\to {{\mathbb F}}_p$. We denote the corresponding null space as follows. $$\nil {\cal V}^{(j)}_p(\Sigma_g)\;\;=\;\;\bigl\{v\in {\cal V}^{(j)}_p(\Sigma_g)\,:\;\lz v,w\rz_p=0\;\; \forall\,w\in\,{\cal V}^{(j)}_p(\Sigma_g) \bigr\}$$ The elements are represented by vectors $v\in {\cal V}^{(j)}_{{{\mathbb Z}}}(\Sigma_g)$ for which $\lz v,w\rz\in p{{\mathbb Z}}$ for all $w\in {\cal V}^{(j)}_{{{\mathbb Z}}}(\Sigma_g)$ although $v\not \in p{\cal V}^{(j)}_{{{\mathbb Z}}}(\Sigma_g)$. There are well defined TQFT’s $\nil {\cal V}^{(j)}_p$ and $\dov {\cal V}^{(j)}_p$ which assign to a surface $\Sigma_g$ the ${{\mathbb F}}_p$-vectors spaces $$\qquad\qquad \nil {\cal V}^{(j)}_p(\Sigma_g)\qquad\qquad\mbox{and}\qquad\qquad \dov {\cal V}^{(j)}_p(\Sigma_g)\;=\; \raise5pt\hbox{${\cal V}^{(j)}_p(\Sigma_g)$}\Big / \raise-5pt\hbox{$\nil {\cal V}^{(j)}_p(\Sigma_g)$}$$ [Proof:]{} We note that since for a cobordism $M$ the map ${\cal V}_{{{\mathbb Z}}}(M)$ commutes with $E$ we have that ${\cal V}_{{{\mathbb Z}}}(M)^$ commutes with $F=E^$ and hence also maps the ${\cal V}_{{{\mathbb Z}}}^{(j)}(\Sigma_g)$ to themselves. Thus if $v_i\in{\cal V}^{(j)}_{{{\mathbb Z}}}(\Sigma_{i})$ for $i=1,2$, $M$ is a cobordism from $\Sigma_{g_1}$ to $\Sigma_{g_2}$, and $v_1$ represents a vector in $\nil {\cal V}^{(j)}_p(\Sigma_{1})$, then $\lz v_2,{\cal V}^{(j)}_{{{\mathbb Z}}}(M)v_1\rz= \lz {\cal V}^{(j)}_{{{\mathbb Z}}}(M)^v_2,v_1\rz\in p{{\mathbb Z}}$ as ${\cal V}^{(j)}_{{{\mathbb Z}}}(M)^v_2\in {\cal V}^{(j)}_{{{\mathbb Z}}}(\Sigma_{1})$. We extend the previous notations to the weight spaces $\nil{\cal W}^{(j)}_p(\lambda,g)$, $\dov {\cal W}^{(j)}_p(\lambda,g)$, $\nil{\cal W}^{(j)}_p(n,g)$, etc.. Since the weight spaces are all orthogonal to each other these subspaces can be defined also as the null spaces from the respective restriction of the inner forms. Now, also the Specht modules ${\cal S}^{\tau}$ for a diagram $\tau=[a,b]$ inherit an inner form from the permutation module $M^{\tau}$, which is via the isometry $\Upsilon$ compatible with the one on the weight spaces. As in the standard literature, e.g., [@Jam78], we set $${\cal D}^{\tau}_p=\raise4pt\hbox{${\cal S}^{\tau}_p$}\Big / \raise-4pt\hbox{$\nil {\cal S}^{\tau}_p$}\qquad\quad\mbox{where}\quad \nil {\cal S}^{\tau}_p={\cal S}^{\tau}_p\cap {{\cal S}^{\tau}_p}^{\perp}$$ and ${\cal S}^{\tau}_p$ is the $p$-reduction of ${\cal S}^{\tau}$. They are related to irreducible TQFT’s as follows. Let $p\geq 3$ be a prime. The TQFT’s $\dov {\cal V}^{(j)}_p$ are [irreducible]{} over ${{\mathbb F}}_p$ and the weight spaces are identified as $\widehat{\mathfrak W}_g$-modules by equivariant isomorphisms: \[eq-weightiso-p\] \^[(j)]{}\_p: Ind\^[\_g]{}\_[\_[g-n]{}\_n]{} (M\^[g-n]{}\_p\^[\[2,2 \]]{}\_p) \_p\^[(j)]{}(n,g). [Proof:]{} The isomorphism in (\[eq-weightiso-p\]) follows from the definitions and properties of $\Upsilon^{(j)}$. We first show that the spaces $\dov {\cal W}_p^{(j)}(n,g)$ are irreducible with respect to the semidirect product ${\mathfrak X}_g$ of the Cartan algebra ${{\mathbb Z}}[{\mathfrak h}_g]\subset \sl(2g,{{\mathbb Z}})$ and the algebra of the Weyl group ${{\mathbb Z}}[\widehat{\mathfrak W}_g]\subset {{\mathbb Z}}[{\rm Sp}(2g,{{\mathbb Z}})]$. For any vector $v=\sum_{\lambda} v_{\lambda}\in \dov {\cal W}_p^{(j)}(n,g)$ the action of ${\mathfrak h}_g$ shows that each weight component $v_{\lambda}\in \dov {\cal W}_p^{(j)}(\lambda,g)$ has to lie in ${\mathfrak X}_gv$. Each $\dov {\cal W}_p^{(j)}(n,g)$ is a module of a symmetric group $S_n\subset \widehat{\mathfrak W}_g\subset{\mathfrak X}_g$ which is equivalent to ${\cal D}^{[\frac{n+j-1}2,\frac{n-j+1}2 ]}_p$. It now follows from Theorem 4.9 in [@Jam78] that these representations are irreducible. In fact, as $p\geq 3$ any two-row diagram is $p$-regular so that these representations are never zero, see Theorem 11.1 in [@Jam78]. In particular, if $v_{\lambda}\neq 0$ then ${{\mathbb Z}}[S_n]v_{\lambda}$ is the entire module. Hence, for $v\neq 0$ we must have $\dov{\cal W}_p^{(j)}(\lambda,g)\subset {\mathfrak X}_gv$ for at least one $\lambda\in\nabla_g$ with $n(\lambda)=n$. Since $\widehat{\mathfrak W}_g$ acts transitively on all of such weights and provides isomorphisms between the weight spaces we thus have $\dov {\cal W}_p^{(j)}(n,g)= {\mathfrak X}_gv$, which implies irreducibility since $v$ was arbitrary. A submodule of $\dov {\cal V}_p^{(j)}(\Sigma_g)$ must therefore be a direct sum of the $\dov {\cal W}_p^{(j)}(n,g)$. Each of these contains a special vector $w_n^g=\Upsilon_{\varpi_{g-j+1}}(e_{t(n,j)})$ with $t(n,j)= {\!\!\raise-8pt\hbox{\bsyg 1 & 3 \cr 2 & 4 \cr \esyg}}\! {\,\, \cdots \!\!\!\raise-8pt\hbox{\bsyg m-1 & m+1 \cr m \cr \esyg}}\! {\,\,\raise5pt\hbox{$\cdots$}\!\!\! \raise2pt\hbox{\bsyg n\cr \esyg}}\, $ where $m=n-j+1$. Using $\kappa_{t(n,j)}^2=2^m\kappa_{t(n,j)}$ for the antisymmetrizer (\[eq-defkappa\]), we find $\lz w^g_n, w^g_n\rz_p=\lz e_{t(n,j)}, e_{t(n,j)}\rz_p= 2^m\lz e_{\{t(n,j)\}},\kappa_t e_{\{t(n,j)\}}\rz_p=2^m\not\equiv 0\mod p$ for $p\geq 3$ so that all of these vectors are non-zero in $\dov {\cal W}_p^{(j)}(n,g)$. It can be computed from the rules [3.]{} and [4.(b)]{} in Lemma \[lm-spechtTQFTgen\] that $e_{\alpha_{g-n-1}}{\sf S}_{g-n-1}w^g_n=w^g_{n-2}$ and $e_{\alpha_{g-n+1}}w^g_n=2w^g_{n+2}$, where ${\sf S}_l\in{\mathfrak W}_g$ maps $a_j$ to $-b_j$ so that with $2\not\equiv 0 \mod p$ we have non-trivial maps between all of the irreducible ${\mathfrak X}_g$-components. Consequently, the $\dov {\cal V}_p^{(j)}(\Sigma_g)$ are irreducible as $\sp(2g,{{\mathbb Z}})$-representations and any sub-TQFT must assign either this space or 0 to a surface $\Sigma_g$. As in [@Ker01] we easily check that the handle attachment maps are non-trivial between these spaces so that $\dov {\cal V}_p^{(j)}$ does in fact contain no proper sub-TQFT. We next construct a sequence of maps between the $p$-reductions of the Specht modules, using the $\sl(2,{{\mathbb Z}})$-actions. As before we fix $n\in{{\mathbb N}}$ and denote the Specht module $${\cal S}^{\{c\}}_{{{\mathbb Z}}}\;\;=\;\;L_{{{\mathbb Z}}}^n\cap\ker(F)\cap\ker(H+c-1)\;\;\cong\;\;{\cal S}^{\tau}\;,$$ where $c=a-b+1$ with tableau $\tau=[a,b]=[\frac {n+c-1}2, \frac {n-c+1}2]$. \[lm-Emaps\] Let $p\geq 3$, $c\not\equiv 0\mod p$, and $c_0\in\{1,\ldots,p-1\}$ such that $c\equiv c_0\mod p$. Then $$E^{c_0}(S^{\{c\}}_{{{\mathbb Z}}})\;\subset\;S^{\{c-2c_0\}}_{{{\mathbb Z}}}\,+\,pL^n_{{{\mathbb Z}}}\;\;.$$ Moreover, $$E^{c_0}(S^{\{c\}}_{{{\mathbb Z}}})\,\not\subset\,pL^n_{{{\mathbb Z}}}\qquad\quad \mbox{and}\qquad\quad E^{p}(L^n_{{{\mathbb Z}}})\subset pL^n_{{{\mathbb Z}}}\;.$$ [Proof:]{} Now, with notation as in the proof of Lemma \[eq-weightiso-p\], $e_{t(n,c-1)}\in S^{\{c\}}_{{{\mathbb Z}}}$ is a cyclic vector so it suffices to show that $E^{c_0}(e_{t(n,c-1)})\in (S^{\{c-2c_0\}}_{{{\mathbb Z}}}\,+\,pL^n_{{{\mathbb Z}}})-pL^n_{{{\mathbb Z}}}$. Furthermore, $e_{t(n,c-1)}=e_{t(m,0)}\otimes e_-^{\otimes c-1}$ and $Ee_{t(m,0)}=0$ so that we really need to show that $E^{c_0}(e_-^{\otimes c-1})\in (ker(F)\,+\,pL^{c-1}_{{{\mathbb Z}}})-pL^{c-1}_{{{\mathbb Z}}}$, where $c=c_0+kp$. We do this by induction in $c_0$. For $c_0=1$ we have $c-1=kp$ and, using $\sl_2$-relations, $FE(e_-^{\otimes kp})= -He_-^{\otimes kp}+EFe_-^{\otimes kp}=kp e_-^{\otimes kp}$. Hence, if we set $w=k e_-^{\otimes kp-1}\otimes e_+$ we find $Fw=e_-^{\otimes kp}$ so that $E(e_-^{\otimes kp})-kpw\in ker(F)$, meaning $E(e_-^{\otimes kp})\in ker(F)+pL^{c-1}_{{{\mathbb Z}}}\,$. Next assume the assertion is true for $c_0$ so that $E^{c_0}(e_-^{\otimes c-1})=y\,+\,pz\,$ with $Fy=0$ and $-Hy=(c-2c_0-1)y$. . We have by binomial formula and $Ee_-=e_+$ and $E^2e_-=0$ that $E^{c_0+1}(e_-^{\otimes c})=E^{c_0+1}(e_-^{\otimes c-1}\otimes e_-)= (E^{c_0+1}e_-^{\otimes c-1})\otimes e_-\,+\,(c_0+1) E^{c_0}e_-^{\otimes c-1})\otimes e_+=(Ey)\otimes e_-\,+\,(c_0+1)y\otimes e_+\,+\,pz'\,$. Thus we need to show $t=(Ey)\otimes e_-\,+\,(c_0+1)y\otimes e_+\in ker(F)\cap pL^{c}_{{{\mathbb Z}}}$. We compute $Ft=(FEy)\otimes e_-\,+\,(c_0+1)y\otimes e_-= (-Hy)\otimes e_-\,+\,(c_0+1)y\otimes e_-=(c-c_0)y\otimes e_-=kp y\otimes e_-$. Also $F(y\otimes e_+)=y\otimes e_-$ so that $t-kp y\otimes e+-\in ker(F)$ and hence $t\in ker(F)\cap pL^{c}_{{{\mathbb Z}}}$. Finally, it is not hard to see that $\lz e_+^{\otimes c_0}\otimes e_-^{\otimes c-c_0-1}, E^{c_0}(e_-^{\otimes c-1})\rz = c_0 !\not\equiv 0\mod p$ if $c_0<p$ so that $E^{c_0}(e_-^{\otimes c-1})\not\in pL^{c-1}_{{{\mathbb Z}}}$ and hence $E^{c_0}e_{t(n,c-1)}\not\in pL^{n}_{{{\mathbb Z}}}$. Also, the binomial formula yields $E^p=\sum_j{p\choose j}E^{p-j}\otimes E^j$ so that we can conclude $E^p=0$ by a similar induction argument. In particular Lemma \[lm-Emaps\] implies that we have well defined, non-zero maps $E^{c_0}\,:\,S^{\{c\}}_p\,\longrightarrow\, S^{\{c-2c_0\}}_p$ on the respective $p$-reductions $S^{\{c\}}_p=S^{\{c\}}_{{{\mathbb Z}}}/pS^{\{c\}}_{{{\mathbb Z}}}= S^{\{c\}}_{{{\mathbb Z}}}/pL^n_{{{\mathbb Z}}}$ using that $pS^{\{c\}}_{{{\mathbb Z}}}=S^{\{c\}}_{{{\mathbb Z}}}\cap pL^n_{{{\mathbb Z}}}$. \[cor-Cseq\] For $p$ and $n$ as above and $k=1,\ldots,p-1$ with $k\equiv n+1\mod 2$ there is a sequence ${\cal C}_{p,k}$ of Specht modules over ${{\mathbb F}}_p$ as follows: \[eq-Cseq\] 0[S]{}\^[{n+1-2l}]{}\_p… [S]{}\^[{2p+k}]{}\_p [S]{}\^[{2p-k}]{}\_p [S]{}\^[{k}]{}\_p[D]{}\^[{k}]{}\_p0. All maps (except the first and last one) are non-zero, and any two consecutive maps compose to zero. More precisely, we have that the $i$-th component of this sequence is $\,{\cal C}^{(i)}_{p,k}={\cal S}^{\{ip+k_i\}}_p$, where $k_i=k$ if $i$ is even and $k_i=p-k$ if $i$ is odd. The maps are $E^{k_i}:{\cal C}^{(i)}_{p,k}\to {\cal C}^{(i-1)}_{p,k}$ so that two consecutive maps compose as $E^{k_i}E^{k_{i+1}}=E^p=0$. We have that ${\cal C}^{(0)}_{p,k}={\cal S}^{\{k\}}_p\to {\cal C}^{(-1)}_{p,k}={\cal D}^{\{k\}}_p$ is the (non-zero) quotient map. Now, it is clear that $im(E)\subset ker(F)^{\perp}\subset \nil {\cal S}^{\{k\}}_p$ given that $E^=F$. Hence also the composite ${\cal C}^{(1)}_{p,k}\to{\cal C}^{(0)}_{p,k}\to {\cal C}^{(-1)}_{p,k}$ is zero. In order to characterize the last index write $\frac{n+1+k} 2= ph+q$ with $h\in{{\mathbb Z}}$ and $q=0,\ldots\, p-1$. We have \[eq-lastindex\] l={ [ll]{} q-k&qk\ q &q<k\ .. The maps in Corollary \[cor-Cseq\] thus extend to a sequence of the $p$-reductions of the induced representations from Corollary \[cor-UpsSpecWei\] as well as the weight spaces ${\cal W}_p^{(c)}(n,g)$. The respective maps on the vector spaces ${\cal V}_p^{(c)}(\Sigma_g)$ are, by equivariance of the $\Upsilon^{(j)}$, given by the restriction and $p$-reductions of the maps $E^{c_0}$ on ${\mbox{\raisebox{.4ex}{$\bigwedge^{\!}$}}\mkern-1mu}H_1(\Sigma_g)$. In particular, these maps commute with the TQFT-images of the cobordisms by equivariance of ${\cal V}_{{{\mathbb Z}}}$. \[cor-seqTQFT\] For $p$, $n$ and $k$ as above there is a sequence of natural transformations of TQFT’s \[eq-seqTQFT\] …[V]{}\^[((i+1)p+k\_[i+1]{})]{}\_p [V]{}\^[(ip+k\_i)]{}\_p… [V]{}\^[(2p-k)]{}\_p [V]{}\^[(k)]{}\_p\^[(k)]{}\_p0 with $k_i=k, p-k$ as above and any two consecutive transformations compose to zero. [5. Exactness of ${\cal C}_{p,k}$]{} \[S5\] Exactness of the sequence ${\cal C}_{p,k}$ defined in Corollary \[cor-Cseq\] follows from the modular structure of the involved Specht modules. The irreducible factors of ${\cal S}^{\tau}_p$ for two row diagrams $\tau$ are determined in Theorem 24.15 of [@Jam78]. For our proof we will, however, have to make use also of the precise submodule structure, which turns out to be rather rigid. We use the recent result of Kleshchev and Sheth in [@KS99] that describes this structure precisely. In order to state it we need to introduce some more conventions and definitions. For a Young diagram $\tau=[a,b]$ we set $c=a-b+1$, and consider the $p$-adic expansion \[eq-cpadic\] c=c\_0+c\_1p+c\_2p\^2+…+c\_rp\^r c\_j{0,1,…,p-1}. As in [@KS99] we denote by $\hat A_{\tau}$ the family of sets of integers of the form \[eq-defIA\] I=\[i\_1,i\_2)\[i\_3,i\_4)…\[i\_[2u-1]{},i\_[ut]{}) such that i\_1<i\_2<…< i\_[2u]{},c\_[i\_[2j-1]{}]{}0 c\_[i\_[2j]{}]{}p-1. For such a set $I\in A_{\tau}$ we define as in [@KS99]the number \[eq-defdelta\] \^\_I=\_[iI]{} (p-1-c\_i)p\^i +\_[j=1]{}\^up\^[i\_[2j-1]{}]{} =\_[j=1]{}\^u\^\_[\[i\_[2j-1]{},i\_[2j]{})]{}, where $\displaystyle\;\delta_{[u,w)}^{\tau}=\sum_{i=u}^{w-1}(p-1-c_i)p^i\,+\,p^u\,$. Also as in [@KS99] we introduce the smaller set $A_{\tau}\subset \hat A_{\tau}$ given by \[eq-defAb\] A\_={IA\_:\^\_Ib}, as well as for any $I\in A_{\tau}$ the function $\nu_I$ on Young diagrams $\tau=[a,b]$ defined as \[eq-defnu\] \_I()=\[a+\_I\^, b-\_I\^\]. As before we denote by ${\cal D}^{\mu}_p$ the irreducible quotient of the Specht module ${\cal S}^{\mu}_p$ over ${{\mathbb F}}_p$ for any two row diagram $\mu$. Also denote by ${\cal F}(M)$ the set of irreducible factors that occur in a composition series of a representation $M$. Further, let ${\cal M}_{\mu}^{\lambda}$ be the smallest sub module of $S^{\lambda}_p$ such that ${\cal D}^{\mu}_p\in {\cal F}({\cal M}_{\mu}^{\lambda})$. The description of the submodule structure in [@KS99] uses the partial order on ${\cal F}({\cal S}^{\tau}_p)$ defined as ${\cal D}_p^{\mu_1}\leq_{\tau}{\cal D}_p^{\mu_2}$ if and only if ${\cal M}_{{\mu}_1}^{\tau}\subseteq {\cal M}_{{\mu}_2}^{\tau}$, i.e., if and only if ${\cal D}_p^{\mu_1}\subseteq {\cal M}_{{\mu}_2}^{\tau}$. \[thm-KS99\] 1. All multiplicities of ${\cal D}^{\mu}$ in $S^{\tau}$ are zero or one. 2. ${\cal F}({\cal S}_p^{\tau})=\{{\cal D}_p^{\nu_I(\tau)}\,:\;I\in A_{\tau}\}$ 3. ${\cal D}_p^{\nu_J(\tau)}\geq_{\tau} {\cal D}_p^{\nu_I(\tau)}$ if and only if $J\subseteq I$. We define now a submodules ${\cal S}_p^{\tau}$ by choosing a special subset of of $A_{\tau}$ and $\hat A_{\tau}$. It is defined as \[eq-defAsets\] A\^0\_={IA\_: 0I} A\^+\_ = A-A\^0\_={IA\_: 0I} as well as $ A^0_{\tau}=\hat A^0_{\tau}\cap A_{\tau} \qquad\quad \mbox{and}\qquad\quad A^+_{\tau}=\hat A^+_{\tau}\cap A_{\tau} $. Let us also introduce the number $k_{\tau}= min\{j\geq 1:\;c_j\neq 0\}$ with $k_{\tau}=\infty$ if $c=c_0$. Hence \[eq-Atauplus\] c=c\_0+c\_[k\_]{}p\^[k\_]{}+…+c\_rp\^r \[0,k\_)I=IA\_\^+. The latter follows since if $0\not\in I$ we must have $i_1>0$. But with $c_{i_1}\neq 0$ we find $i_1\geq k_{\tau}$. For the following we assume $c\not\equiv 0\mod p$, i.e., $c_0\neq 0$. Given this we introduce for a diagram $\tau=[a,b]$ with $a-b=c-1\geq 2c_0$ (so that $k_{\tau}<\infty$) the notation \[eq-tauprime\] ’==\_[\[0,k\_)]{}(’) For $\tau'$ we thus have $c'=c-2c_0$ and for the $p$-adic expansion of $c'=\sum_jc'_jp^j$ we obtain \[eq-ccoeff\] [rcll]{} c\_0’ &=& p-c\_0 &\ c\_i’ &=& p-1 & 1i <k\_\ [rcll]{} c’\_[k\_]{} &=& c\_[k\_]{}-1&\ c’\_i &=& c\_i & i>[k\_]{}\ From these equations it is clear that $c_0'\neq 0$ and $c'_{k_{\tau}}\neq p-1$ so that $[0,k_{\tau})$ is an admissible interval for $\hat A_{\tau'}$. In fact, it is the unique minimal interval of the special subset \[eq-AprimeInt\] A\_[’]{}\^0={IA\_[’]{}:\[0,k\_)I} \[0,k\_)A\_[’]{}\^0. This is obvious since we must have $i_1=0$ for $0\in I$ and then the next possible $i_2$ is $k_{\tau}$. It is also easy to see that $\delta_{[0,k_{\tau}]}^{\tau'}=c_0$, confirming the relation in (\[eq-tauprime\]) between $\tau$ and $\tau'$ via (\[eq-defnu\]). We have the following simple but crucial observation. \[lm-hatbij\] With $\tau$ and $\tau'$ as above we have a well defined bijection with inverse $$\phi:\;\hat A^0_{\tau'}\,\longrightarrow\, \hat A^+_{\tau}\;\; :\;I\;\mapsto\;I-[0,k_{\tau})\quad\quad\mbox{and}\quad\quad \phi^{-1}:\;\hat A^+_{\tau}\,\longrightarrow\, \hat A^0_{\tau'}\;\; :\;J\;\mapsto\;J\cup [0,k_{\tau})$$ [Proof:]{} We first show that $\phi$ is well defined. From (\[eq-AprimeInt\]) we know that every $I\in A_{\tau'}^0$ contains the special interval and is thus of the form $$I=[0,i_2)\cup [i_3,i_4)\cup\ldots \cup [i_{2u-1}, i_{2u})\qquad\quad\mbox{with} \quad i_2\geq k_{\tau}\;.$$ Clearly, $0\not\in\phi(I)$. We further distinguish the following two cases: #### Case $i_2={k_{\tau}}$ : In this situation $\phi(I)= [i_3,i_4)\cup\ldots\;$. Now $\phi(I)\in \hat A^+_{\tau}$ since by (\[eq-ccoeff\]) the coefficients of $c$ and $c'$ and hence the conditions on the $i_j$ for $j\geq 3$ are the same. #### Case $i_2>{k_{\tau}}$ : In this situation $\phi(I)= [{k_{\tau}},i_2)\cup R$ with $R=\,[i_3,i_4)\cup\ldots\;$. We have $c_{k_{\tau}}\neq 0$ by assumption, and $c_{i_2}=c'_{i_2}\neq q-1$ as $i_2>{k_{\tau}}$. Also all intervals in $R$ are admissible. Hence $\phi(I)\in \hat A^+_{\tau}$ again. In a similar fashion we show that $\phi^{-1}$ is well defined using the fact that any $L\in \hat A^+_{\tau}$ is by (\[eq-Atauplus\]) of the form $L=[j_1,j_2)\cup[j_3,j_4)\cup\ldots\,$ with $j_1\geq k_{\tau}$. For $j_1> k_{\tau}$ we have $\phi^{-1}(L)=[0,k_{\tau})\cup [j_1,j_2)\cup[j_3,j_4)\cup\ldots\,$. This is $\hat A^0_{\tau'}$ as $c'_0\neq 0$, $c'_{k_{\tau}}\neq p-1$, and $c'_j=c_j$ for $j\geq j_1$. In case that $j_1= k_{\tau}$ we have $\phi^{-1}(L)=[0,j_2)\cup[j_3,j_4)\cup\ldots\,$, which is again in $\hat A^0_{\tau'}$ since $c'_j=c_j$ for $j\geq j_2>j_1$. It is obvious that $\phi$ and $\phi^{-1}$ are inverses of each other given (\[eq-Atauplus\]) and (\[eq-AprimeInt\]). Next we consider how the differential from (\[eq-defdelta\]) changes under this bijection. \[lm-delta\] We have $$\delta^{\tau'}_I \;=\;\delta^{\tau}_{\phi(I)}\,+\,c_0$$ [Proof:]{} This is a computation done according to the same cases as in the proof of Lemma \[lm-hatbij\]: #### Case $i_2=k_{\tau}$ : Then $\delta_{I}^{\tau'}=\delta_{[0,{k_{\tau}})}^{\tau'}+\sum_{j\geq 2} \delta^{\tau'}_{[i_{2j-1},i_{2j})}$. Now, since $c_i'=p-1$ for $i=1,\ldots,{k_{\tau}}-1$ we have $\delta_{[0,{k_{\tau}})}^{\tau'}=(p-1-c'_0)+p^0=c_0$. We also have $\delta^{\tau'}_{[i_{2j-1},i_{2j})}=\delta^{\tau}_{[i_{2j-1},i_{2j})}$ since the $c_i$ are all the same with $i_j>{k_{\tau}}$. But $\delta^{\tau}_{I-[0,{k_{\tau}})}= \sum_{j\geq 2} \delta^{\tau}_{[i_{2j-1},i_{2j})}$ which implies the assertion. #### Case $i_2>k_{\tau}$ : As before we write $I=[0,i_2)\cup[i_3,i_4)\cup\ldots\,=[0,i_2)\cup R$ and $\phi(I)= [k_{\tau},i_2)\cup R$ $$\begin{aligned} \delta_{I}^{\tau'}&=& \sum_{i=0}^{i_2-1}(p-1-c_i')p^i + 1\;+\;\delta_R^{\tau'}\\ &=& (p-1-c'_0)+1 \,+\,(p-1-c_{k_{\tau}}')p^{k_{\tau}}\,+ \,\sum_{i>{k_{\tau}}}^{i_2-1}(p-1-c_i')p^i\;+\;\delta_R^{\tau}\\ &=& c_0\,+\,(p-1-c_{k_{\tau}})p^{k_{\tau}}\,+\,p^{k_{\tau}}\,+ \,\sum_{i>{k_{\tau}}}^{i_2}(p-1-c_i)p^i\;+\;\delta_R^{\tau}\\ &=& c_0\,+\, \sum_{i={k_{\tau}}}^{i_2-1}(p-1-c_i)p^i\,+\,p^{k_{\tau}} \;+\;\delta_R^{\tau}\\ &=& c_0\,+\, \delta_{[{k_{\tau}},i_2)}^{\tau}\;+\;\delta_R^{\tau} \;\;\;\;\;\;=\;\;\;\;\;\; c_0\,+\,\;\delta_{[{k_{\tau}},i_2)\cup R}^{\tau} \;\;\;\;\;\;=\;\;\;\;\;\; c_0\,+\,\;\delta_{I-[0,{k_{\tau}})}^{\lambda}\end{aligned}$$ which proves the assertion. \[cor-bij\] The map $\phi$ restricts to a bijection $ \phi:\; A^0_{\tau'}\,\longrightarrow\, A^+_{\tau}\;\;\; $ with the properties that it is monotonous with respect to inclusions and $$\label{eq-nuform} {\nu}_I(\tau')\;=\;\nu_{\phi(I)}(\tau)\;.$$ [Proof:]{} From (\[eq-tauprime\]) we have $b'=b+c_0$. Moreover, Lemma \[lm-hatbij\] and Lemma \[lm-delta\] imply that $I\in A^0_{\tau'}$ iff $\delta^{\tau'}_I\leq b'$ iff $\delta^{\tau}_{\phi(I)}+c_0\leq b+c_0$ iff $\delta^{\tau}_{\phi(I)}\leq b$ iff $\phi(I)\in A^+_{\tau}$. The fact that $\phi(A)\subset \phi(B)$ iff $A\subset B$ is obvious. Relation (\[eq-nuform\]) is immediate from Lemma \[lm-delta\], (\[eq-defnu\]) and (\[eq-tauprime\]). Let us turn now to the analogous relations between the modules. The set $A^0_{\tau}$ contains a minimal element, namely $I^0_{\tau}=[0,h_{\tau})$, which is the smallest interval with $h_{\tau}>0$ and $c_{h_{\tau}}\neq p-1$. We denote the modules $$\zer {\cal S}_p^{\tau}\,=\,{\cal M}_{\nu_{I^0_{\tau}}(\tau)}^{\tau} \qquad\qquad\mbox{and}\qquad\qquad \ste {\cal S}_p^{\tau}\;=\; \raise4pt\hbox{${\cal S}_p^{\tau}$}\Big/ \raise-4pt\hbox{$\zer {\cal S}_p^{\tau}$}\;.$$ By [3.]{} of Theorem \[thm-KS99\] we know that the modules in the composition series of ${\cal M}_{\nu_{I^0_{\tau}}(\tau)}^{\tau}$ are of the form ${\cal D}_p^{\nu_J(\tau)}$, where $I^0_{\tau}\subseteq J$, which is equivalent to $J\in A_{\tau}^0$. Hence (\_p\^)={[D]{}\_p\^[\_I()]{}:IA\_\^0} (\_p\^)={[D]{}\_p\^[\_I()]{}:IA\_\^+}. With $h_{\tau'}=k_{\tau}$ and $I_{\tau'}^0=\phi(\emptyset)$ we thus have $\zer {\cal S}_p^{\tau'}={\cal M}^{\tau'}_{\tau}$. The bijection from Corollary \[cor-bij\] allows us now to show that the composition series of $\zer {\cal S}_p^{\tau'}$ and $\ste {\cal S}_p^{\tau}$ yield exactly the same set of irreducibles via the identification \[eq-phistar\] \^\:(\_p\^[’]{}) (\_p\^):\^[\_I(’)]{}\_p \^[\_[(I)]{}()]{}\_p. \[lm-maps\] Let $\tau$ and $\tau'$ be as above. Suppose there is a non-zero map $$\xi\;:\;\;{\cal S}_p^{\tau}\;\longrightarrow\;{\cal S}_p^{\tau'}\qquad\mbox{with} \qquad\zer {\cal S}_p^{\tau}\subseteq ker(\xi)\;.$$ Then we have $$im(\xi)=\zer {\cal S}_p^{\tau'}\qquad\qquad\mbox{and}\qquad\qquad ker(\xi)=\zer {\cal S}_p^{\tau}\;.$$ [Proof:]{} Consider the map $\overline{\xi}:\ste {\cal S}_p^{\tau}\longrightarrow {\cal S}_p^{\tau'}$ defined on the quotient by $\zer {\cal S}_p^{\tau}$ as well as the composite $\overline{\overline{\xi}}:\ste {\cal S}_p^{\tau}\longrightarrow \ste {\cal S}_p^{\tau'}$ with the projection. Now, since by [1.]{} of Theorem \[thm-KS99\] we have a [disjoint]{} union ${\cal F}({\cal S}_p^{\tau'})={\cal F}(\zer {\cal S}_p^{\tau'})\cup {\cal F}(\ste {\cal S}_p^{\tau'})$ so that by (\[eq-phistar\]) ${\cal F}(\ste {\cal S}_p^{\tau'})\cap {\cal F}(\ste {\cal S}_p^{\tau}) =\emptyset$. Hence $\overline{\overline{\xi}}$ is a map between modules with no common irreducibles in their composition series so that $\overline{\overline{\xi}}=0$. Thus the image of $\xi$ or $\overline{\xi}$ must lie in $\zer {\cal S}_p^{\tau'}$. Consider the sequence of maps $$\xi\;:\;\;\;{\cal M}^{\tau}_{\tau}={\cal S}_p^{\tau}\; \stackrel{\gamma}{\longrightarrow}\; \ste {\cal S}_p^{\tau}\; \stackrel{\overline{\xi}}{\longrightarrow}\; \zer {\cal S}_p^{\tau'}={\cal M}^{\tau'}_{\tau}\;$$ Since all multiplicities are one we have again a disjoint union ${\cal F}({\cal M}^{\tau}_{\tau})={\cal F}(ker(\xi))\cup {\cal F}(im(\xi))$ so that either ${\cal D}_p^{\tau}\in {\cal F}(ker(\xi))$ or ${\cal D}_p^{\tau}\in {\cal F}(ker(\xi))$. In the first case this would mean that $ker(\xi)\subseteq {\cal M}^{\tau}_{\tau}$ is a submodule that contains ${\cal D}_p^{\tau}$ in its composition series so that by minimality $ker(\xi)= {\cal M}^{\tau}_{\tau}$. We had, however, assumed that $\xi\neq 0$. Hence ${\cal D}_p^{\tau}$ must be contained in the composition series of $im(\xi)\subseteq {\cal M}^{\tau'}_{\tau}$. Again it follows by minimality that $im(\xi)={\cal M}^{\tau'}_{\tau}$. This also means that $\overline{\xi}$ is a surjective map. The fact that all multiplicities are one together with (\[eq-phistar\]) implies that $dim(\ste {\cal S}_p^{\tau})=dim(\zer {\cal S}_p^{\tau'})$. Hence $\overline{\xi}$ must be an isomorphism. We are finally in the position to establish following resolution of irreducible modules by Specht modules. \[thm-Snexact\] The sequences (\[eq-Cseq\]) from Corollary \[cor-Cseq\] are exact. [Proof:]{} We begin by deriving from Theorem \[thm-KS99\] that the first term ${\cal S}^{\{n+1-2l\}}_p$ in the sequence is irreducible. Here we have $b=l$ and $c=n+1-2l=2ph+2q-k-2l$ so that by (\[eq-lastindex\]) $$b=\left\{\begin{array}{ll} q-k\;&\;\mbox{if}\;q\geq k\\ q \;&\;\mbox{if}\;q<k\\ \end{array}\right. \qquad\mbox{and}\qquad c_0=\left\{\begin{array}{ll} k\;&\;\mbox{if}\;q\geq k\\ p-k \;&\;\mbox{if}\;q<k\\ \end{array}\right.$$ Now for a set $I\in A_{\tau}$ with $I\neq\emptyset$ we have $\delta_I\geq p-c_0$. We also need $ \delta_I\leq b$ and hence $p-c_0\leq b$. For $q\geq k$ this condition reduces to $p\leq q$, which is not possible, and for $q< k$ we find $q\geq k$, a contradiction as well. Hence $A_{\tau}=\{\emptyset\}$ so that ${\cal S}^{\{n+1-2l\}}_p=\ste {\cal S}^{\{n+1-2l\}}_p={\cal D}^{\{n+1-2l\}}_p$ and $\zer {\cal S}^{\{n+1-2l\}}_p=0$. Next, we observe that all maps $E^{k_i}:{\cal S}^{\{(i+1)p+k_{i+1}\}}_q\to {\cal S}^{\{ip+k_i\}}_q$ in the sequence are of the type of the ones in Lemma \[lm-maps\] with diagrams related by (\[eq-tauprime\]). More precisely, we have $c_0=k_{i+1}$ so that $c'=c-2c_0=(i+1)p+k_{i+1}-2k_{i+1}=jp+(p-k_{i+1})=ip+k_i$. Corollary \[cor-Cseq\] also implies that these maps are non-zero and equivariant. It now follows by induction, going from large to small $i$, that $$ker(E^{k_i})\;=\; \zer {\cal S}_q^{\{(i+1)p+k_{i+1}\}} \qquad\quad\mbox{and}\qquad\quad im(E^{k_i})\;=\; \zer {\cal S}_q^{\{ip+k_i\}} \;.$$ For the first two maps in the sequence this is clear by irreducibility of the first module and the fact that the next map is non-zero. Once we have proved the relation for $E^{k_{i+1}}$ we know from $E^{k_i}E^{k_{i+1}}=0$ that $\zer {\cal S}_q^{\{(i+1)p+k_{i+1}\}}=im(E^{k_{i+1}}) \subseteq ker(E^{k_i})$. We can thus apply Lemma \[lm-maps\] to $E^{k_i}$ and infer the statement for $i$ from $i+1$. Hence exactness holds for the terms before ${\cal S}^{\{k\}}_p$. For this last Specht module we have $c_0=k$ and $c_j=0$ for $j>0$. Thus $A_{\tau}^+=\{\emptyset\}$ and $A_{\tau}^0=\{[0,j): p^j\leq\frac {n+k+1}2\}$ using that $\delta_{[0,j)}=p^j-k$ and $b=\frac {n-k+1}2$. Particularly, we have that $\ste {\cal S}^{\{k\}}_p={\cal D}^{\{k\}}_p$ is already irreducible. The kernel of the last map already contains $\zer {\cal S}^{\{k\}}_p$, the image of the previous map, in its kernel. If the kernel was bigger the map would thus have to be zero, which is not the case. Hence exactness also holds at ${\cal S}^{\{k\}}_p$. Finally, the last map is by irreducibility onto so that exactness holds through out the sequence. As a result we obtain that the sequence of TQFT’s in (\[eq-seqTQFT\]) of Corollary \[cor-seqTQFT\] is exact and this yields a resolutions of the TQFT’s $\dov{\cal V}_{p}^{(k)}$ with $0<k<p$, thus proving Theorem \[thm-TQFTreso\]. The kernels and cokernels of these sequences also define TQFT’s, which we denote in analogy to the symmetric group representations by $\zer {\cal V}^{(j)}_p$ and $\ste {\cal V}^{(j)}_p$ so that we have short exact sequences \[eq-zersteV\] 0\^[(j)]{}\_p\^[(j)]{}\_p \^[(j)]{}\_p 0. [6. Characters, Dimensions, and the Alexander Polynomial]{} \[S6\] An obvious application of Theorem \[thm-Snexact\] is that we can express the characters $\varphi^{\tau_k}_p$ and dimensions of the ${\cal D}_p^{\tau_k}= {\cal D}_p^{\{k\}}$ for diagrams $\tau_k=[\frac {n+k-1}2, \frac {n-k+1}2]$ with $k\equiv n+1\mod 2$ and $0<k<p$ in terms of the ordinary characters $\chi^{\tau}$ of Specht modules ${\cal S}^{\tau}$. \[cor-characters\] For $k\equiv n+1\mod 2$ and $0<k<p$ we have the following identity of $S_n$-characters \[eq-characters\] \^[\_k]{}\_p=\_[i0]{} (-1)\^[i]{} \^[\_[j\_i]{}]{}, where $j_i=ip+k_i$ and $k_i=k$ for $i$ even and $k_i=p-k$ for $i$ odd. As an example we consider the special case $p=5$, where $k=1$ or $3$ if $n$ is even and $k=2$ or $4$ if $n$ odd. In [@Ryba94] Ryba constructs a family of irreducible so called Fibonacci representations $R_n$ and $R_n'$ of $S_n$ over ${{\mathbb F}}_5$ with Brauer characters $\varphi_n$ and $\varphi'_n$ respectively. It follows by straight forward computation from (\[eq-characters\]) and the formulae in Definition 2 of [@Ryba94] that $\phi'_n=\phi^{[r,r]}_5$ and $\phi_n=\phi^{[r+1,r-1]}_5$ if $n=2r$ is even, and $\phi'_n=\phi^{[r+2,r-1]}_5$ and $\phi_n=\phi^{[r+1,r]}_5$ if $n=2r+1$ is odd. \[cor-ryba\] The Fibonacci representations from [@Ryba94] are $$R_n\;\cong\;\left\{\begin{array} {cl} {\cal D}_5^{[r+1,r-1]}&\;\;\mbox{if}\;\;n=2r\\ {\cal D}_5^{[r+1,r]}&\;\;\mbox{if}\;\;n=2r+1\\ \end{array}\right. \qquad \mbox{and}\qquad R_n'\;\cong\;\left\{\begin{array} {cl} {\cal D}_5^{[r,r]}&\;\;\mbox{if}\;\;n=2r\\ {\cal D}_5^{[r+2,r-1]}&\;\;\mbox{if}\;\;n=2r+1\\ \end{array}\right.$$ We expect similar relations with the generalizations of these representations obtained by Kleshchev in [@Kle96]. The dimensions of the Specht modules are naturally given by the [Catalan numbers]{} $C(n,j)={n\choose j}-{n\choose {j-1}}$. More precisely, $dim({\cal S}^{[n-b,b]})=C(n,b)$. Particularly, we find for the components of the sequence in (\[eq-Cseq\]) that $dim({\cal S}^{\{c_{2s}\}})=C(n,b -sp)$ and $dim({\cal S}^{\{c_{2s+1}\}})=-C(n,b +(s+1)p)$, with $b=\frac {n+1-k} 2\,$, where we also use that $C(n,j)=-C(n,n+1-j)$. The alternating sum of the Specht module dimensions comes out to be \[eq-DdimCat\] d\^n\_k=dim([D]{}\^[\[n-b,b\]]{}\_p)=\_[s[[Z]{}]{}]{} C(n,b+sp) 0<k= n-2b+1 < p. Note that if we extend the above formula for $d^n_b$ to the next indices we find \[eq-dimbd\] d\^n\_0=0n,d\^n\_p=0 n In order to describe generating functions for these dimensions we introduce some notation. First we write $\displaystyle [n]_x=\frac {x^n-x^{-n}}{x-x^{-1}}\in{{\mathbb Z}}[x,x^{-1}]$ for the usual quantum integers. Moreover we consider now the ring of cyclotomic integers ${{\mathbb Z}}[\zeta_p]$ obtained ${{\mathbb Z}}[x,x^{-1}]$ by imposing the relation $\sum_{j=0}^{p-1}x^j=0$. It is free as a ${{\mathbb Z}}$-module of rank $p-1$. We also denote by ${{\mathbb A}}[\zeta_p]\subset {{\mathbb Z}}[\zeta_p]$ the subring invariant under conjugation $\zeta_p\mapsto\zeta_p^{-1}$. This is a free ${{\mathbb Z}}$-module of rank $\frac {p-1} 2$. It is not hard to see that the set of $[k]_{\zeta_p}$ restricted to either even $k$ or to odd $k$ yields a ${{\mathbb Z}}$-basis for ${{\mathbb A}}[\zeta_p]$. \[lm-dimgenfct\] We have the following identity in ${{\mathbb A}}[\zeta_p]$: \[eq-dimgenfct\] \[2\]\_[\_p]{}\^n=\_[ ]{}d\^n\_k\[k\]\_[\_p]{} [Proof:]{} We have by Schur Weyl duality that $L^n_{{{\mathbb Z}}}\cong V_2^{\otimes n}\cong\bigoplus_{j\equiv n-1\mod 2} V_j\otimes {\cal S}^{[\frac {n+j-1}2,\frac {n-j+1}2 ]}$. The operator $x^H$ with $H\in\sl(2,{{\mathbb Z}})$ is well defined and has trace $tr_{V_j}(x^H)=[j]_x$. It thus follows that $[2]_x^n=\sum_{l\geq 1, l\equiv n+1\mod 2}[l]_xC(n,\frac{n-l+1}2)$. Now any such $l$ can be uniquely written in the form $l=k+2sp$ or $l=-k+2(s+1)p$ with $s\geq 0$ and $k\equiv n+1\mod 2$ and $0<k<p$. Specializing to a root of unity $x=\zeta_p$ we have then that $[l]_{\zeta_p}=[k]_{\zeta_p}$ and $[l]_{\zeta_p}=-[k]_{\zeta_p}$ respectively. Also $C(n,\frac{n-l+1}2)= C(n,\frac{n-k+1}2-sp)$ and $C(n,\frac{n-l+1}2)= C(n,\frac{n+k+1}2-(s+1)p)=-C(n,\frac{n-k+1}2+(s+1)p)$ respectively. Hence, the terms for a fixed $k$ are given by $[k]_{\zeta_p} C(n,b-sp)$ and $[k]_{\zeta_p} C(n,b+(s+1)p)$ for $s\geq 0$ and $b=\frac {n+1-k} 2\,$, which with $s\in{{\mathbb Z}}$ adds up to the expression in (\[eq-DdimCat\]). Using $[2]_x[k]_x=[k-1]_x+[k+1]_x$ and the bases of ${{\mathbb A}}[\zeta_p]\subset {{\mathbb Z}}[\zeta_p]$ we readily derive from (\[eq-dimgenfct\]) the recursion $d^{n+1}_k=d^{n}_{k-1}+d^{n}_{k+1}$. This translates for $0\leq a-b\leq p-2$ to \[eq-dimrec\] dim([D]{}\_p\^[\[a,b\]]{})={ [cl]{} dim([D]{}\_p\^[\[a,a-1\]]{})&a=b\ dim([D]{}\_p\^[\[a-1,b\]]{})+dim([D]{}\_p\^[\[a,b-1\]]{})&0< a-b< p-2\ dim([D]{}\_p\^[\[a-1,b\]]{})&a-b=p-2\ .. It is easy to see from this form that the dimensions are indeed given by the number of paths through the set of diagrams with $a-b\leq p-2$ as described in in [@Mat96] for general diagrams. In the case $p=5$ this recursion reduces to $dim(R_n)=dim(R_{n-1})+dim(R_{n-1}')$ and $dim(R_n')=dim(R_{n-1})$ so that the dimensions are given by Fibonacci numbers. More precisely, we have $dim(R_n)=f_n$ and $dim(R_n')=f_{n-1}$, where $f_n$ are the Fibonacci numbers defined by $f_0=0$, $f_1=1$ and $f_{n+1}=f_n+f_{n-1}$. Note that together with (\[eq-DdimCat\]) we find interesting presentations of Fibonacci numbers in terms of alternating, 5-periodic sums of Catalan numbers: $$\begin{aligned} f_{2r}&=&C(2r,r-1)-C(2r,r-3)+C(2r,r-6)-C(2r,r-8)+\ldots\label{eq-fibcatideven}\\ &=&C(2r+1,r-1)-C(2r+1,r-2)+C(2r+1,r-6)-C(2r+1,r-7)+\ldots\nonumber \\ \mbox{and}\quad&&\nonumber\\ f_{2r+1}&=&C(2r+1,r)-C(2r+1,r-3)+C(2r+1,r-5)-C(2r+1,r-8)+\ldots\label{eq-fibcatidodd}\\ &=&C(2r+2,r+1)-C(2r+2,r-3)+C(2r+2,r-4)-C(2r+2,r-8)+\ldots\nonumber \end{aligned}$$ The reader is invited to check these identities independently via recursion relations such as $C(n+1,j)=C(n,j)+C(n,j-1)$ or via the well known generating functions of Catalan and Fibonacci numbers. Another useful tool in the determination of dimensions are fusion algebras or Verlinde algebras, see [@FK93]. For the quantum group $U_q(\sl_2)$ at a $p$-th root of unity the ring of irreducible representations yields the fusion algebra $\Phi_p$ generated by the irreducibles $\lb 1\rb, \lb 2\rb, \ldots, \lb p-1\rb$. Let us denote by $mult(k,R)\in{{\mathbb N}}\cup\{0\}$ the multiplicity of $\lb j\rb$ in an element $R\in\Phi_p$ so that $R=\sum_jmult(k,R)\lb k\rb$. We have relations $\lb 1\rb\circ \lb k\rb= \lb k\rb$, $\lb p-1\rb\circ\lb k\rb= \lb p-k\rb$ and $\lb 2\rb\circ \lb k\rb= \lb k+1\rb+\lb k-1\rb$ for $1<k<p-1$. Comparing this to the recursion in (\[eq-dimrec\]) we find that \[eq-dimmult\] dim([D]{}\_p\^[\[2,2\]]{})=mult(k,2\^[n]{}) Now, as the ${\cal D}_p^{\tau}$ are isomorphic to the weight spaces $\dov {\cal W}^{(k)}_p(\lambda, g)$ with $n=n(\lambda)$ we find $ dim(\dov {\cal V}^{(k)}_p(\Sigma_g))=\;\sum_{n=0}^g2^{g-n}{g\choose n} dim({\cal D}_p^{[\frac{n+k-1}2,\frac{n-k+1}2]})$ so that we find the following Verlinde type formula. \[lm-dimspaces\] \[eq-dimspaces\] dim(\^[(k)]{}\_p(\_g))= mult(k,[[f-4.4mu f]{}]{}\^[g]{}\_p) [[f-4.4mu f]{}]{}\_p=21+ 2\_p. Compare this to the TQFT’s ${\cal V}_p^{RT}$ and ${\cal V}_p^{RT}$ of Reshetikhin Turaev for the quantum groups $U_q(\sl_2)$ and $U_q({\mathfrak s}{\mathfrak 0}_3)$ respectively at a $p$-th root of unity. (${\cal V}_p^{RT}$ is really a factor TQFT obtained from ${\cal V}_p^{RT}$ by restricting to odd dimensional representations). With ${{\sf F}\mkern -8mu{\sf F}}_p=\sum_j\lb 2j+1\rb^2$ and ${{\sf F}\mkern -8mu{\sf F}}^_p=2{{\sf F}\mkern -8mu{\sf F}}_p=\sum_k\lb k\rb^2$ we have \[eq-verlinde\] dim([V]{}\_p\^[\RT]{}(\_g))=mult(1,[[[F]{}-8mu[F]{}]{}\^\\_p]{}\^g )=2\^g mult(1,[[F]{}-8mu[F]{}]{}\^g\_p )=2\^gdim([V]{}\_p\^[RT]{}(\_g)) In the case $p=5$ these formulae allow us to efficiently compute and compare dimensions. We have for the dimensions $D^{(k)}_g=dim(\dov {\cal V}_5^{(k-1)}(\Sigma_g))\,$. 1. For even $g$: $$\begin{array}{rcl} D^{(1)}_g&=&\frac 12\Bigl(5^{\frac g2} f_{g-1}+f_{2g+1} \Bigr) \\ D^{(4)}_g&=&\frac 12\Bigl(5^{\frac g2} f_{g-1}-f_{2g+1} \Bigr)\\ \end{array} \qquad\qquad \begin{array}{rcl} D^{(2)}_g&=&\frac 12\Bigl( 5^{\frac g2} f_{g}+f_{2g} \Bigr)\\ D^{(3)}_g&=&\frac 12\Bigl( 5^{\frac g2} f_{g}-f_{2g} \Bigr)\\ \end{array}$$ 2. For odd $g$: $$\begin{array}{rcl} D^{(1)}_g&=&\frac 12\Bigl(5^{\frac {g-1}2}(f_{g-2}+f_{g}) +f_{2g+1} \Bigr)\\ D^{(4)}_g&=&\frac 12\Bigl(5^{\frac {g-1}2}(f_{g-2}+f_{g}) -f_{2g+1} \Bigr)\\ \end{array} \qquad\qquad \begin{array}{rcl} D^{(2)}_g&=&\frac 12\Bigl( 5^{\frac {g-1}2}(f_{g-1}+f_{g+1}) +f_{2g} \Bigr)\\ D^{(3)}_g&=&\frac 12\Bigl( 5^{\frac {g-1}2}(f_{g-1}+f_{g+1}) -f_{2g} \Bigr)\\ \end{array}$$ and \[eq-dimRT5FN\] dim([V]{}\_5\^[RT]{}(\_g))=D\^[(1)]{}\_g+D\^[(4)]{}\_gg0. [Proof:]{} Let us use a more convenient notation $1=\lb 1\rb$, $\rho=\lb 3\rb$, $\sigma=\lb 4\rb$, and $\sigma\circ \rho=\lb 2\rb$ subject to relations $\rho\circ \rho=1+\rho$ and $\sigma^2=1$. These relations imply $\rho^n=f_{n-1}1+f_n\rho$, ${{\sf f\mkern -4.4mu f}}_5=2+\sigma\circ \rho$ and ${{\sf F}\mkern -8mu{\sf F}}_5=2+\rho$. We note now that ${{\sf F}\mkern -8mu{\sf F}}_5^2=(2+\rho)^2=4+4\rho+\rho^2=5(1+\rho)=5\rho^2$ so that ${{\sf F}\mkern -8mu{\sf F}}_5^{g}=5^{\frac g2}(f_{g-1}+f_g\rho)$ if $g$ is even. From there we compute directly that for odd $g$ we have ${{\sf F}\mkern -8mu{\sf F}}_p^{g}=5^{\frac {g-1}2}((f_{g-2}+f_g)+(f_{g-1}+f_{g+1})\rho)$. Consider also $\eta=1-\rho$. We have $\eta^2=1-2\rho+\rho^2=2-\rho=1+\eta$ and thus again $\eta^n=f_{n-1}+f_n\eta$. We find $(2-\rho)^g=\eta^{2g}= f_{2g-1}+f_{2g}\eta=f_{2g+1}-f_{2g}\rho$. Now, we have $(1+\sigma)\circ{{\sf f\mkern -4.4mu f}}_5=(1+\sigma)\circ{{\sf F}\mkern -8mu{\sf F}}_5$ and thus $(1+\sigma)\circ{{\sf f\mkern -4.4mu f}}_5^g=(1+\sigma)\circ{{\sf F}\mkern -8mu{\sf F}}_5^g$. Similarly, we find $(1-\sigma)\circ{{\sf f\mkern -4.4mu f}}_5^g=(1+\sigma)\circ(2-\rho)^g$ so that ${{\sf f\mkern -4.4mu f}}_5^g=\frac 12 (1+\sigma)\circ{{\sf F}\mkern -8mu{\sf F}}_5^g+\frac 12(1-\sigma)\circ(2-\rho)^g$. With the previous results on ${{\sf F}\mkern -8mu{\sf F}}_5^g$ and $(2-\rho)^g$ we thus find a formula for ${{\sf f\mkern -4.4mu f}}_5^g$, which inserted into (\[eq-dimspaces\]) yields the asserted formulae. The formula in (\[eq-dimRT5FN\]) reflects the fact that $\dov {\cal V}_5^{(1)}\oplus \dov {\cal V}_5^{(4)}$ is the sum of the irreducible constituents of the ${{\mathbb F}}_5$-reductions of ${\cal V}_5^{RT}$ as TQFT’s, see [@KerFib]. For larger primes $p$ it is, however, not possible that a ${{\mathbb F}}_p$-reduction ${\cal V}_p^{RT}$ has only the ${\cal V}_p^{(j)}$ as irreducible components. This is easily see by looking at the large $g$ asymptotics of the dimension expressions in (\[eq-dimspaces\]) and (\[eq-verlinde\]). The operations on $\Phi_p$ given by multiplication by ${{\sf F}\mkern -8mu{\sf F}}_g$ or ${{\sf f\mkern -4.4mu f}}_g$ are represented by matrices with non-negative integer coefficents. Perron-Frobenius theory thus implies that the matrix elements of of ${{\sf F}\mkern -8mu{\sf F}}_g^g$ or ${{\sf f\mkern -4.4mu f}}_g^g$, such as those in (\[eq-dimspaces\]) and (\[eq-verlinde\]), grow like $\sim\\| {{\sf f\mkern -4.4mu f}}_p\\|^g$ and $\sim\\| {{\sf F}\mkern -8mu{\sf F}}_p\\|^g$, respectively, where $\\| {{\sf F}\mkern -8mu{\sf F}}_p\\|=\frac p{4\sin^2(\frac \pi p)}$ and $\\| {{\sf f\mkern -4.4mu f}}_p\\|=4\cos^2(\frac \pi{2p})$ are the largest eigenvalues of the associate matrices. We thus obtain (\[eq-PFasy\]). Note, that $\\| {{\sf F}\mkern -8mu{\sf F}}_5\\|=\\| {{\sf f\mkern -4.4mu f}}_5\\|$ but that $\\| {{\sf F}\mkern -8mu{\sf F}}_p\\|>\\| {{\sf f\mkern -4.4mu f}}_p\\|$ if $p>5$. Thus, a linear relation as Theorem \[thm-KerFib\] cannot generalize to $p>5$. Instead, we can find polynomials $\,R_p(f)\in{{\mathbb Z}}[f]\,$ of with degree $\displaystyle {\rm deg}(R_p)=\frac {p-3}2$ such that ${{\sf F}\mkern -8mu{\sf F}}_p\;=\;R_p({{\sf f\mkern -4.4mu f}}_p)\,$ by using recursive relations in $\Phi_p$, which may be identified with a ${\mathbb F}_2$-extension of the real part of ${{\mathbb Z}}[\zeta_p]$. Using modified Tschebycheff polynomials, which we define by the recursion $\,P_{j+1}(x)+P_{j-1}\,=\,x\cdot P_j(x)\,$, with $P_0(x)=1$ and $P_1(x)=x$, the $R_p$ can be written as follows. R\_p(f)=\_[j=0]{}\^[2]{}n\_jP\_j(f-2), n\_j= { [cl]{} 2&\ 2& .. For example, $R_5(f)=f$, $R_7(f)=2f^2-7f+7$, $R_9(f)=2f^3-9f^2+9f+3$, $R_{11}=3f^4-22f^3+55f^2-55f+22$, and $R_{13}(f)=3f^5-26f^4+78f^3-91f^2+26f+13$. These polynomial appear to play an important rôle in representation theoretic aspects of Conjecture \[conj-sums\]. The analogs of the character expansions given in Corollary \[cor-characters\] and relations as in Lemma \[lm-dimgenfct\] in the context of the corresponding TQFT’s attain a topological interpretation via the Alexander Polynomial $\Delta_{\varphi}(M)\in{{\mathbb Z}}[x,x^{-1}]$ for a compact, oriented 3-manifold $M$ with a selected epimorphism $\varphi:H_1(M,{{\mathbb Z}})\to\mkern-15mu\to{{\mathbb Z}}$. Up to $S$-equivalence the cocycle $\varphi$ defines a two-sided embedded surface $\Sigma\subset M$ and cobordism $C_{\Sigma}:\Sigma\to\Sigma$ obtained by removing a neighborhood of $\Sigma$ from $M$. The Alexander Polynomial is then given (up to a sign, which is determined by the additional framing structure on $M$) by the following identity extracted in Section 11 of [@Ker01]: \[eq-AlexClass\] \_(M)=trace(x\^[-H]{}[V]{}\_[[[Z]{}]{}]{}(C\_))= \_[j1]{}\[j\]\_[-x]{}trace([V]{}\^[(j)]{}\_[[[Z]{}]{}]{}(C\_)) By inserting $x=\zeta_p$ or $x=-\zeta_p$, a $p$-th root of unity, and reducing the integer coefficients to ${{\mathbb F}}_p$ we consider the image of the Alexander Polynomial under the two so defined natural maps \[eq-defredAlex\] [[Z]{}]{}\[x,x\^[-1]{}\][[F]{}]{}\_p\[\_p\]:x\_p: \_(M)\_[, p]{}\^(M). Now, if we insert $x=\zeta_p$ into (\[eq-AlexClass\]) and use $[j_i]_{\pm\zeta_p}=(\pm 1)^{k-1}(-1)^i[k]_{\zeta_p}$, where $j_i=ip+k_i$ as before, we infer from the resolutions of TQFT’s in (\[eq-seqTQFT\]) the following identities $$\begin{aligned} \label{eq-alexalt} \dov \Delta_{\varphi, p}^{+}(M)\;\;&=&\;\;\sum_{k=1}^{\frac{p-1}2}(-1)^{k-1}[k]_{\zeta_p}\Bigl( trace(\dov {\cal V}^{(k)}_p(C_{\Sigma}))+ trace(\dov {\cal V}^{(p-k)}_p(C_{\Sigma}))\Bigr)\;, \\ \dov \Delta_{\varphi, p}^{-}(M)\;\;&=&\;\;\sum_{k=1}^{\frac{p-1}2}\,[k]_{\zeta_p}\Bigl( trace(\dov {\cal V}^{(k)}_p(C_{\Sigma}))- trace(\dov {\cal V}^{(p-k)}_p(C_{\Sigma}))\Bigr)\;. \end{aligned}$$ This implies Theorem \[thm-pAlexander\]. At a 5-th root of unity $\dov \Delta_{\varphi, p}^{+}(M)$ depends only on traces of $C_{\Sigma}$ under ${\cal V}^{(1)}_5\oplus {\cal V}^{(4)}_5$ and ${\cal V}^{(2)}_5\oplus {\cal V}^{(3)}_5$. The former is in [@KerFib] identified with the integral semisimple reduction of the Reshetikhin Turaev theory ${\cal V}^{RT}_5$ and implies (\[eq-Lescop\]). [7. Johnson-Morita Extensions]{} \[S7\] As before we consider for $H=H_1(\Sigma)$ a standard basis $\{a_1,\ldots,a_g,b_1,\ldots,b_g\}$ that is symplectic with respect to the standard skew form $(\,,)$ and orthonormal with respect to the inner form $\lz\,,\rz$. Let $J\in {\rm Sp}(2g,{{\mathbb Z}})$ be the special element defined by $Ja_i=b_i$ and $Jb_i=-a_i$ so that $\lz x,y\rz=(x,Jy)$ and $Jg^{-1}J^{-1}=g^$. Also denote by $\omega=\sum_ja_j\wedge b_j$ the standard invariant 2-form. For any $x\in{\mbox{\raisebox{.4ex}{$\bigwedge^{\!m}$}}\mkern-1mu}H$ we can now define a degree-$m$ map $\nu(x):{\mbox{\raisebox{.4ex}{$\bigwedge^{\!}$}}\mkern-1mu}H\to {\mbox{\raisebox{.4ex}{$\bigwedge^{\!+m}$}}\mkern-1mu}H$ by $\nu(x).y=x\wedge y$. From this we define another map as $\mu(x)=\nu(Jx)^: {\mbox{\raisebox{.4ex}{$\bigwedge^{\!}$}}\mkern-1mu}H\to {\mbox{\raisebox{.4ex}{$\bigwedge^{\!-m}$}}\mkern-1mu}H$ of degree $-m$. The maps $\nu$ and $\mu$ have the following properties: 1. Covariance: $g\nu(x)g^{-1}=\nu(gx)$ and $g\mu(x)g^{-1}=\mu(gx)$ for all $x\in{\mbox{\raisebox{.4ex}{$\bigwedge^{\!}$}}\mkern-1mu} H$ and $g\in {\rm Sp}(2g,{{\mathbb Z}})$. 2. Homomorphy: $\nu(x\wedge y)=\nu(x)\nu(y)$ and $\mu(x\wedge y)=\mu(y)\mu(x)$. 3. Generators: $\nu(\omega)=E$ and $\mu(\omega)=F$ so that $[E,\nu(x)]=[F,\mu(x)]=0$ for all $x\in{\mbox{\raisebox{.4ex}{$\bigwedge^{\!}$}}\mkern-1mu}H$. 4. Anticommutator: $\mu(x)\nu(y)+\nu(y)\mu(x)=(x,y){{\mathbb I}}$ for all $x,y\,\in\,H\,.$. 5. Commutators: $\displaystyle \qquad[E,\mu(x)]=\nu(x)\qquad\mbox{and}\qquad [F,\nu(x)]=-\mu(x) \qquad\mbox{for}\;\;x\,\in\,H\,.$ [Proof:]{} Covariance for $\nu$ is obvious. For $\mu$ consider $g\mu(x)g^{-1}=(g^{-1}\nu(Jx)g^)^=\nu(g^{-1}Jx)^= \nu(Jgx)^=\mu(gx)$. Also homomorphy is obvious and the fact that $\nu(\omega)=E$ follows by definition. This implies the zero-commutators since $\omega$ is even, hence central, and $F=E^$. The anticommutator relation is readily translated to $\nu(x)^\nu(y)+\nu(y)\nu(x)^=\lz x,y\rz{{\mathbb I}}$. As a symmetric bilinear relation it suffices to prove this for a system of two orthonormal vectors $v$ and $w$. Then ${\mbox{\raisebox{.4ex}{$\bigwedge^{\!}$}}\mkern-1mu}H={\mbox{\raisebox{.4ex}{$\bigwedge^{\!}$}}\mkern-1mu}L\otimes {\mbox{\raisebox{.4ex}{$\bigwedge^{\!}$}}\mkern-1mu}L^{\perp}$, where $L$ is the space spanned by $v$ and $w$ and $ L^{\perp}$ its orthogonal complement. Clearly, $\nu(v)$ and $\nu(w)$ act only on the ${\mbox{\raisebox{.4ex}{$\bigwedge^{\!}$}}\mkern-1mu}L\cong({{{\mathbb Z}}}^2)^{\otimes 2}$ with basis $\{ 1,v,w,v\wedge w\}=\{e_-\otimes e_-, e_+\otimes e_-, e_-\otimes e_+, e_+\otimes e_+\}\,$. In this form the operators take on the form $\nu(v)=E\otimes{{\mathbb I}}$ and $\nu(w)=-H\otimes E$, where the $\sl_2$-generators $E$ and $H$ act as usual. The relation follows now from $HE+EH=0$ and $E^E+EE^=FE+EF={{\mathbb I}}$. The commutators follow by direct calculation. $E\mu(x)=\nu(\omega)\mu(x)=\sum_i\nu(a_i)\nu(b_i)\mu(x)= -\sum_i\nu(a_i)\mu(x)\nu(b_i)+\sum_i\nu(a_i)(x,b_i)= \sum_i\mu(x)\nu(a_i)\nu(b_i) -\sum_i(x,a_i)\nu(b_i) +\sum_i\nu(a_i)(x,b_i) =\mu(x)E+\nu\Bigl( \sum_ia_i(x,b_i)-b_i(x,a_i)\Bigr) =\mu(x)E+\nu\Bigl( \sum_ia_i\lz x,a_i\rz+b_i\lz x,b_i\rz\Bigr) =\mu(x)E+\nu(x)$. The second relation is just the conjugate of the first. Equipped with these relations we can now construct maps between the ${\rm Sp}(2g,{{\mathbb Z}})$-representation spaces from previous sections. \[lm-muall\] The map $\mu$ restricts and factors into an ${\rm Sp}(2g,{{\mathbb Z}})$-covariant map $$\mu\;:\;\;\frac {{\mbox{\raisebox{.4ex}{$\bigwedge^{\!m}$}}\mkern-1mu} H}{\omega\wedge{\mbox{\raisebox{.4ex}{$\bigwedge^{\!m-2}$}}\mkern-1mu} H} \;\;\longrightarrow\;\;Hom({\cal V}^{(j)}_{{{\mathbb Z}}}(\Sigma), {\cal V}^{(j+m)}_{{{\mathbb Z}}}(\Sigma))\;.$$ Moreover, for any $x\in{\mbox{\raisebox{.4ex}{$\bigwedge^{\!m}$}}\mkern-1mu} H$ we have that \[eq-muEmap\] (x): im(E\^[l+m]{}) im(E\^[l]{}). [Proof:]{} Recall that ${\cal V}^{(j)}(\Sigma_g)=ker(F)\cap {\mbox{\raisebox{.4ex}{$\bigwedge^{\!(g-j+1)}$}}\mkern-1mu}H$. Now $\mu(x)$ commutes with $F$ and thus maps $ker(F)$ to itself and thus by counting degrees to ${\cal V}^{(j+m)}(\Sigma_g)=ker(F)\cap {\mbox{\raisebox{.4ex}{$\bigwedge^{\!(g-j-m+1)}$}}\mkern-1mu}H$. Moreover, $\mu(\omega\wedge x)=\mu(x)\mu(\omega)=\mu(x)F=0$ if restricted to $ker(F)$. By the homomorphism property it suffices to show the second relation for $m=1$. In this case $\mu(x)E^{l+1}z=E^{l+1}\mu(x)z+(l+1)E^l\nu(x)z \in im(E^l)$ by iteration of the commutator relation. Consider now the following family of extensions of the symplectic group defined as a semidirect product. \[eq-defJMgroup\] JM\_[a]{}(m,g)=( 1 a )(2g,[[Z]{}]{}) \[propos-Umods\] For $a\not\equiv 0\mod p$ we have well defined representations of $JM_{a}(m,g)$ on $${\cal U}^{(j)}_p(m,g)\;\;=\;\; {\cal V}_p^{(j)}(\Sigma_g)\stackrel{\mu}{\oplus}{\cal V}_p^{(j+m)}(\Sigma_g)$$ which decomposes as indicated if restricted to ${\rm Sp}(2g,{{\mathbb Z}})$ and is given by $\mu$ if restricted to the abelian part. For $j\equiv k\mod p$ with $0<k<p-m$ and using the notation from (\[eq-zersteV\]) we have that $\zer {\cal U}^{(j)}_p(m,g)\;\;=\;\; \zer {\cal V}_p^{(j)}(\Sigma_g)\stackrel{\mu}{\oplus} \zer {\cal V}_p^{(j+m)}(\Sigma_g)$ is a proper submodule with subquotient \[eq-Umods\] \^[(j)]{}\_p(m,g)= \_p\^[(j)]{}(\_g) \_p\^[(j+m)]{}(\_g) / -3pt. [Proof:]{} In general if $V$, $W$ and $M$ are $G$-modules, and $\mu:M\to Hom(V,W)$ a covariant map, we construct a module $V\stackrel{\mu}{\oplus}W$ of $M\rtimes G$ by letting $(m,g)$ act on the sum $V\oplus W$ by the block matrix $\scriptstyle \left[\matrix {g & 0\cr \mu(m)g &g\cr}\right]$ so that $W\subset V\stackrel{\mu}{\oplus}W$ is a submodule with subquotient $V$. Thus the map $\mu$ from Lemma \[lm-muall\] defines such a module for $JM_1(m,g)$ over ${{\mathbb Z}}$. In the ${{\mathbb F}}_p$-reduction $a$ is invertible so that $\mu$ can be extended to $JM_a(m,g)$. By exactness of the sequences in (\[eq-seqTQFT\]) we have that $\zer {\cal V}_p^{(j)}(\Sigma_g)={\cal V}_p^{(j)}(\Sigma_g)\cap im(E^{p-k})$. Thus by (\[eq-muEmap\]) of Lemma \[lm-muall\] we have that $\mu(x)(\zer {\cal V}_p^{(j)}(\Sigma_g))\subseteq {\cal V}_p^{(j+m)}(\Sigma_g)\cap im(E^{p-k-m})= \zer {\cal V}_p^{(j+m)}(\Sigma_g)$. This implies, by construction, that $\zer {\cal V}_p^{(j)}(\Sigma_g){\oplus} \zer {\cal V}_p^{(j+m)}(\Sigma_g)$ is indeed a submodule. Note that for $0<k<p$ the factors in (\[eq-Umods\]) are irreducible. We also write \[eq-Uirred\] \^[(j)]{}\_p(m,g)= \_p\^[(j)]{}(\_g) \_p\^[(j+m)]{}(\_g). The case that is topologically relevant is $m=3$. For this case Morita constructed in [@Mor93] a homomorphism $\tilde k:\Gamma_g\to JM_2(3,g)$ on the mapping class group $\Gamma_g$. Its kernel is the group ${\cal K}_g$ generated by bounding cycles and its restriction to the Torelli group coincides with the Johnson homomorphism $\tau_2:{\cal I}_g\to\frac {{\mbox{\raisebox{.2ex}{$\scriptstyle \bigwedge^{\!3}$}}\mkern-1mu} H}{ H}$ from [@Joh80]. There is a finite index subgroup ${\bf Q}_g\subset JM_2(3,g)= \frac 1 2 \frac { {\mbox{\raisebox{.2ex}{$\scriptstyle \bigwedge^{\!3}$}}\mkern-1mu} H} {\,\, H\,} \rtimes {\rm Sp}(2g,{{\mathbb Z}})$, and homomorphisms $\tau_2$ and $\tilde k$ such that the following diagram is commutative with all rows and the last column exact. &&&&&&&0 &&\ &&&&&&&&&\ 0 & &[K]{}\_g & &[I]{}\_g& \^[ \_2 ]{}& & &&0\ && && & && &&\ 0 & &[K]{}\_g & &\_g& \^[ k ]{}& & [Q]{}\_g &&0\ && && && &\ 0 & &[I]{}\_g & &\_g& & & [Sp]{}(2g,[[Z]{}]{}) &&0\ &&&&&&&&&\ &&&&&&&0 &&\ Combining this result with Proposition \[propos-Umods\] now yields Theorem \[thm-JMext\], where we denoted the special module $\dov {\cal U}^{(j)}_p(\Sigma_g)=\dov {\cal U}^{(j)}_p(3,g)$. Note, that we obtain from the sequences in (\[eq-seqTQFT\]) similar resolutions. Particularly, for $g\geq 0$, $p\geq 5$ a prime and $0<k<p-3$ there is an exact sequence of maps of ${{\mathbb F}}_p$-modules as follows. $$\begin{aligned} \label{eq-seqU} \ldots\,\longrightarrow&&{\cal U}^{((i+1)p+k_{i+1})}_p(\Sigma_g)\,\longrightarrow\, {\cal U}^{(ip+k_i)}_p(\Sigma_g)\,\longrightarrow\,\ldots\,\longrightarrow\, {\cal U}^{(2p+k)}_p(\Sigma_g)\,\longrightarrow\nonumber \\ &\,&\longrightarrow\,\,{\cal U}^{(2p-k-3)}_p(\Sigma_g)\, \longrightarrow\, {\cal U}^{(k)}_p(\Sigma_g)\,\longrightarrow \,\dov {\cal U}^{(k)}_p(\Sigma_g)\,\longrightarrow\,0\;, \end{aligned}$$ where now $k_i=k$ for $i$ is even, and $k_i=p-i-3$ if $i$ is odd. The maps are alternatingly given by $E^k\oplus E^{k+3}$ and $E^{p-k}\oplus E^{p-k-3}$. Note, however, that the module extensions work alternatingly in opposite ways so that the maps cannot be ${\bf Q}_g$-equivariant. It is true that by setting $\tilde R(g)=R(\hat Jg^{-1}\hat J^{-1})^$, where $\hat J\in\Gamma_g$ is a representative of the $J\in {\rm Sp}(2g,{{\mathbb Z}})$ above we can reverse the exact sequences that define an extension. Yet, even by flipping every second extension in (\[eq-seqU\]) still does not yield equivariant maps. [99]{} Fröhlich, J., Kerler, T.: Quantum groups, quantum categories and quantum field theory. [Lecture Notes in Mathematics]{} [1542]{} Springer-Verlag, Berlin, 1993. Frohman, C., Nicas, A.: The Alexander Polynomial via topological quantum field theory, [Differential Geometry, Global Analysis, and Topology, Canadian Math. Soc. Conf. Proc.]{} Vol. 12, Amer. Math. Soc. Providence, RI, (1992) 27–40. Gilmer, P.: Integrality for TQFT’s. Preprint 2001. [math.QA/0105059]{}. Goodman, R., Wallach, N.R.: Representations and invariants of the classical groups. [Encyclopedia of Mathematics and its Applications]{} [68]{} Cambridge University Press, 1998. James, G.D.: The representation theory of the symmetric groups. [Lecture Notes in Mathematics]{} [682]{} Springer, Berlin, 1978. Johnson, D.: An abelian quotient of the mapping class group ${\cal I}\sb{g}$. [Math. Ann.]{} [249]{} (1980), no. 3, 225–242. Kerler, T., Lyubashenko, V.V.: Non-semisimple topological quantum field theories for 3-manifolds with corners. [Lecture Notes in Mathematics]{} [1765]{}, xi + 376 pp.. Springer Verlag 2001. Kerler, T.: “Homology TQFT’s and the Alexander-Reidemeister Invariant of 3-Manifolds via Hopf Algebras and Skein Theory". [Canad. J. Math.]{} 50 pages 38 figures. [math.GT/0008204 ]{} Kerler, T.: “$p$-Modular TQFT’s, Milnor torsion and the Casson-Lescop invariant”. Invariants of knots and 3-manifolds (Kyoto 2001). [Geom. Topol. Monogr.]{} [4]{} (2002) 119-141. ([http://www.maths.warwick.ac.uk/gt/gtmcontents4.html]{}) Kerler, T.: On the Structure of the Fibonacci TQFT. (in preparation). Kleshchev, A.: Completely splittable representations of symmetric groups. [J. Algebra]{} [181]{} (1996), no. 2, 584–592. Kleshchev, A. S., Sheth, J.: On extensions of simple modules over symmetric and algebraic groups. [J. Algebra]{} [221]{} (1999), no. 2, 705–722. Masbaum, G.; Roberts, J. D.: A simple proof of integrality of quantum invariants at prime roots of unity. [Math. Proc. Cambridge Philos. Soc.]{} [121]{} (1997), no. 3, 443–454. Mathieu, O.: On the dimension of some modular irreducible representations of the symmetric group. [Lett. Math. Phys.]{} [38]{} (1996), no. 1, 23–32. Morita, S.: Casson’s invariant for homology $3$-spheres and characteristic classes of surface bundles. I. [Topology]{} [28]{} (1989), no. 3, 305–323. Morita, S.: The extension of Johnson’s homomorphism from the Torelli group to the mapping class group. [Invent. Math.]{} [111]{} (1993), no. 1, 197–224. Murakami, H.: Quantum ${\rm SO}(3)$-invariants dominate the ${\rm SU}(2)$-invariant of Casson and Walker. [Math. Proc. Cambridge Philos. Soc.]{} [117]{} (1995), no. 2, 237–249. Murakami, J., Ohtsuki, T.: Topological quantum field theory for the universal quantum invariant. [Comm. Math. Phys.]{} [188]{} (1997), no. 3, 501–520. Ohtsuki, T.: A polynomial invariant of rational homology $3$-spheres. [Invent. Math.]{} [123]{} (1996), no. 2, 241–257. Ohtsuki, T. (Editor): Problems on Invariants of Knots and 3-Manifolds. [Geom. Topol. Monogr.]{} (2002) (to appear) (see, [http://www.ms.u-tokyo.ac.jp/$\sim$ tomotada/proj01/]{}) Reshetikhin,N.Yu., Turaev, V.G.: [Invariants of 3-manifolds via link polynomials and quantum groups]{}, [Invent. Math.]{} [103]{} (1991), 547–597. Ryba, A. J. E.: Fibonacci representations of the symmetric groups. [J. Algebra]{} [170]{} (1994), no. 2, 678–686. The Ohio State University, Department of Mathematics, 231 West 18th Avenue, Columbus, OH 43210, U.S.A. [E-mail: ]{}[ [email protected]]{} [^1]: 2000 Mathematics Subject Classification: Primary 57R56, 57M27; Secondary 17B10, 17B37, 17B50, 20B30, 20C30, 20C20.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We compute the Lyapunov exponents and the Kolmogorov-Sinai (KS) entropy for a self-bound $N$-body system that is realized as a convex billiard. This system exhibits truly high-dimensional chaos, and $2N-4$ Lyapunov exponents are found to be positive. The KS entropy increases linearly with the numbers of particles. We examine the chaos generating defocusing mechanism and investigate how high-dimensional chaos develops in this system with no dispersing elements.' address: 'Institute for Nuclear Theory, Department of Physics, University of Washington, Seattle, WA 98195, USA' author: - Thomas Papenbrock title: 'Lyapunov exponents and Kolmogorov-Sinai entropy for a high-dimensional convex billiard' --- Introduction ============ Billiards are simple yet nontrivial examples of systems that display classically chaotic motion. Of special importance are the Sinai billiard [@SinaiB] and the Bunimovich stadium [@BunimS] since they are known to be completely chaotic. Interestingly, these two systems exhibit two different mechanisms that generate chaos. While dispersion is the chaos-generating mechanism in the Sinai billiard it is defocusing that leads to chaotic dynamics in the Bunimovich stadium. Dispersing yields a permanent divergence of neighbored trajectories. Defocusing may occur upon reflection at a focusing boundary element. Provided the free path is sufficiently long nearby trajectories start to diverge after passing through the focusing point, and on average the divergence might exceed the convergence thus leading to exponential instability. Dispersing billiards are well known also in higher dimensions. Popular examples are the three-dimensional Sinai billiard and the hard sphere gas. However, it was not until recently that completely chaotic billiards were constructed in more than two spatial dimensions that rely entirely on the defocusing mechanism [@RehaBu96; @BCG96]. These billiards use spherical caps as the focusing elements of the boundary. A trajectory diverges mainly in a two-dimensional plane that is defined by the points of consecutive reflections with the spherical cap, and focusing may be very weak in the transversal directions. This makes it more difficult to create truly high-dimensional chaos in focusing billiards than in dispersing ones. Sufficient conditions for the construction of high-dimensional focusing billiards were given in ref.[@RehaBu96], but it was found that these are not necessary ones[@BCG96]. Besides of their intrinsic interest billiards are important model systems in the field of quantum chaos [@GMW] and statistical and fluid mechanics [@Gaspard]. Questions related to chaos, ergodicity, transport and equilibration are often studied in billiard models, see e.g. refs.[@DellPoschHoov96; @vZvBD]. While two-dymensional chaos is fairly well understood by now, much less is known in high-dimensional systems. Recently, a high-dimensional billiard model has been proposed in the context of nuclear physics [@TP99] and quantum chaos [@PP99]. Within this model a self-bound $N$-body system is realized as a convex billiard. Numerical computations yielded a positive largest Lyapunov exponent and showed that this system is predominantly chaotic. In this article we want to extend previous calculations and compute the full Lyapunov spectrum and the KS entropy for this chaotic $N$-body system. These quantities characterize the degree of hyperbolic instability in dynamical systems and may be related to transport coefficients in non-equilibrium situations [@Ruelle]. Since the studied billiard is convex, defocusing is the only possible source causing this instability [@Bunim91]. This makes it interesting to examine this mechanism in more detail and compare to the situation of defocusing billiards with spherical caps. The results of such an investigation are not only of theoretical interest but may also be useful for further applications. We have in mind general questions concerning chaos in self-bound many-body systems like nuclei or atomic clusters and its influence on equilibration, damping or transport processes. This paper is organized as follows. In the next section we describe the model system and the techniques used to compute the Lyapunov exponents. The third section contains the results of our numerical computations for various system sizes $N$. In section four we investigate the defocusing mechanism in more detail. We finally give a summary. High-dimensional billiard and Lyapunov exponents ================================================ Let us consider a classical system of $N$ particles with Hamiltonian \[ham\] H=\_[i=1]{}\^N[p\_i\^22m]{} + \_[i<j]{}V(\|\_i-\_j\|), where $\vec{r}_i$ is a two-dimensional position vector of the $i$-th particle and $\vec{p}_i$ is its conjugate momentum. The interaction is given by \[int\] V(r)={ [ll]{} 0 & ,\ & . . Thus, the particles move freely and interact whenever the distance between a pair of particles reaches its maximum value $a$. Hamiltonian (\[ham\]) defines a self-bound, interacting many-body system. Energy, total momentum and total angular momentum are conserved quantities. For large numbers of particles the points of interactions are close to a circle of diameter $a$ and therefore define a rather thin surface. Therefore, this system is a simple classical model for nuclei or atomic clusters. For finite values of the binding potential the system is amenable to a mean field description [@BPR99]. Hamiltonian (\[ham\]) may also be viewed as a special case of the square well gas [@SWG] with infinite binding potential. However, to the best of our knowledge, the square well gas has not been investigated for such parameter values. In what follows we restrict ourselves to the case of vanishing total momentum and angular momentum. In the limit $N\to\infty$ the number density diverges for the self-bound many-body system (\[ham\],\[int\]). A constant density may be obtained once the parameter $a$ is rescaled as $a\to a N^{1/3}$, thus turning the Hamiltonian (\[ham\],\[int\]) into an effective Hamiltonian. In what follows we work with a $N$-independent parameter $a$. Since the billiard is a scaling system one may easily rescale the results obtained below to adapt for different values of $a$. The time evolution of a many-body system with billiard like interactions requires an effort ${\cal{O}}(N\ln{N})$ to be compared with the effort ${\cal{O}}(N^2)$ for a generic two-body interaction [@Prosen]. Initially one computes the $N(N-1)/2$ times at which pairs of particles may interact and organizes these in a partially ordered binary tree, keeping the shortest time at its root. Immediately after an interaction of particles labelled $i$ and $j$ one has to recompute $2N-3$ times corresponding to future interactions between particles $i$ and $j$ and the remaining ones. The insertion of each new time into the partially ordered tree requires only an effort ${\cal{O}}(\ln{N})$. Between consecutive interactions particles move freely. Upon an interaction of particles labelled by $i$ and $j$, respectively the momenta change accordingly to \[reflect\] ’= - 2[a\^2]{}. Here, $\vec{p}'$ and $\vec{p}\equiv\vec{p}_i-\vec{p}_j$ are the relative momentum vectors immediately before and after the interaction, respectively, and $\vec{r}\equiv\vec{r}_i-\vec{r}_j$ is the relative position vector with magnitude $\|\vec{r}\|=a$ at the interaction. Obviously, eq. (\[reflect\]) describes a reflection in the center of mass system of the two interacting particles. We now turn to the computation of the Lyapunov exponents. We describe the used techniques rather briefly since a large body of literature exists on the subject, see e.g. refs.[@Gaspard; @LichtLieb; @Reichl]. A Hamiltonian system with $f$ degrees of freedom possesses $f$ independent Lyapunov exponents $\lambda_1,\ldots,\lambda_f$ ordered such that $0\le\lambda_1\le\ldots\le\lambda_f$. Since the Hamiltonian flow preserves phase space volume there are also $f$ non-positive Lyapunov exponents with $\lambda_{-j}=-\lambda_j$. A system with $n$ integrals of motion has $n$ vanishing Lyapunov exponents $\lambda_1=\ldots=\lambda_n=0$, while a chaotic system has a positive largest Lyapunov exponent $\lambda_f>0$. This exponent $\lambda_f$ is the rate at which neighbored trajectories diverge under the time evolution. Benettin [et al.]{} [@Benettin76] gave a method to compute the largest Lyapunov exponent from following the time evolution of a reference trajectory and a second one that is initially slightly displaced. The displacement vector has to be rescaled after some finite evolution in a compact phase space. To compute the full spectrum of Lyapunov exponents one has to follow $f$ trajectories besides the reference trajectory [@Benettin79]. This defines $f$ independent displacement vectors, and finite numerical precision requires their reorthogonalization besides the rescaling during the time evolution. Rather than following the time evolution of finite displacement vectors one may also use infinitesimal displacements (tangent vectors) in the computation of the Lyapunov exponents. In tangent space the time evolution is given by a linear mapping. Details about the tangent map in high-dimensional billiards can be found in refs. [@DellPoschHoov96; @Sieber]. In a completely chaotic system the KS entropy is given by the sum of all positive Lyapunov exponents [@Piesin76], i.e. $h_{\rm KS}=\sum_{j=1}^f \lambda_j$. The KS entropy measures at which rate information about the initial state of a system is lost. Results ======= In what follows we consider the $N$-body system at vanishing total momentum and angular momentum. We use units such that $a=m=E/N=1$. Times are then given in units of $a(mN/E)^{1/2}$. We choose initial conditions at random and follow a trajectory for at least $10^6$ collisions. This ensures a good convergence of the numerically computed Lyapunov spectra. We have checked our results as follows: The time evolution was checked by comparing forward with backward propagation; the Lyapunov spectra were checked by comparing the results obtained from the tangent map with those obtained by Benettin’s method involving finite displacements; the computation of all Lyapunov exponents showed that $\lambda_{-j}+\lambda_j$ vanishes within our numerical accuracy; we found four pairs of vanishing Lyapunov exponents corresponding to the conserved quantities. The Lyapunov spectra for systems of sizes $N=10,30,100,300$ particles are plotted in Fig. \[fig1\]. We note that the $N$-body system possesses $2N-4$ positive Lyapunov exponents. This shows that there are no further integrals of motion besides energy, momentum and angular momentum, and that truly high-dimensional chaos is developed. We discuss this finding in detail in the following section. The Lyapunov exponent $\lambda_i$ is a smooth function of its index with a rather small smallest positive Lyapunov exponent $\lambda_5$. This behavior is similar to the case of the Lennard-Jones fluid [@LJ] or the Fermi-Pasta-Ulam model [@FPU] but differs from the hard sphere gas where a rather large smallest positive Lyapunov exponent was found [@DellPoschHoov96]. Note that the spectra seem to converge somehow with increasing $N$. Table \[tab1\] displays the largest and smallest positive Lyapunov exponents, collision rates and the KS entropies. It is interesting to examine the $N$-dependence in more detail. Fig. \[fig2\] shows that the KS entropy $h_{\rm KS}$ and the collision rate $\tau^{-1}$ depend linearly on the system size $N$. The case of the collision rate is easily understood since the constant single particle energy keeps the collision rate of each particle with the surface constant, too. The KS entropy is roughly given by the area under the corresponding spectrum presented in Fig. \[fig1\]. Since the spectra converge approximately with increasing $N$ this area increases linearly with the number of particles. The $N$-dependence of the largest Lyapunov exponent $\lambda_{2N}$ is shown in Fig. \[fig3\] and may be approximated by an logarithmicly increasing curve. In the case of the hard sphere gas the $N$-dependence could be understood for sufficiently low densities using techniques borrowed from kinetic theory [@vZon]. Unfortunately, these ideas can not directly be transferred to our system since the density is not a small parameter. Note however, that the largest Lyapunov exponent decreases with increasing $N$ once the density is kept constant after rescaling $a\to a N^{1/3}$. This is interesting with view on nuclear physics since this result differs qualitatively from simple billiard (mean-field) models. Scaling arguments for such models show that the largest Lyapunov exponent increases with $N$ at constant density and single-particle energy. The numerical results obtained in this work indicate that the considered billiard systems exhibit truly high-dimensional chaos. We recall that the system is convex and does not possess any dispersing elements. Furthermore, it differs in construction from the high-dimensional focusing billiards with spherical caps studied in refs.[@RehaBu96; @BCG96]. Thus, a closer examination of the chaos generating defocusing mechanism is of interest and presented in the following section. Defocusing mechanism ==================== Let us examine the defocusing mechanism in the billiard considered in this work. We do not try to proof that the considered system is completely chaotic – which seems difficult at least – but rather want to understand the numerically observed phenomenon of chaotic motion in more detail. To this purpose and based on our numerical results we assume that the system is (predominantly) chaotic, and that chaos is generated by the only possible mechanism, namely defocusing [@Bunim91]. We may then clarify how high-dimensional chaos develops and thus understand why we observe $2N-4$ positive Lyapunov exponents. This investigation may hopefully serve also as a starting point and a motivation for further research. For simplicity let us consider the three-body system first. It is useful to study this system as a billiard in full six-dimensional configuration space. This is possible since the change in relative momentum (\[reflect\]) caused by an interaction of two particles corresponds to a specular reflection in the billiard. We denote vectors in configuration space by capital letters as $\vec{R}=(\vec{r}_1,\vec{r}_2,\vec{r}_3)$, where $\vec{r}_i=(x_i,y_i)$ is the two-dimensional position vector of the $i^{\rm th}$ particle. The part of the boundary where particles labelled $i=1, 2$ interact may be parametrized as \[b12\] \_[(12)]{}&=&(+[a2]{}\_, -[a2]{}\_,\_3),\ \_&=&(,). The (outwards pointing) normal vector $\partial_a\vec{X}_{(12)}$ and the tangent vector $\partial_\alpha\vec{X}_{(12)}$ span the two-dimensional planes where divergence due to defocusing might be generated. These planes come in a four-dimensional family due to the parameters $\vec{r}$ and $\vec{r}_3$ in eq. (\[b12\]). Basis vectors for these planes may be chosen as \_1&=&((1,0),(-1,0),(0,0))/,\ \_2&=&((0,1),(0,-1),(0,0))/. Similar arguments show that there are two further planes where defocusing might be generated corresponding to interactions between particles $(1,3)$ and $(2,3)$, respectively. These planes are spanned by the basis vectors \_3&=&((1,0),(0,0),(-1,0))/,\ \_4&=&((0,1),(0,0),(0,-1))/ and \_5&=&((0,0),(1,0),(-1,0))/,\ \_6&=&((0,0),(0,1),(0,-1))/, respectively. Four of the six basis vectors $\vec{E}_i$ are linearly independent. The vectors $\vec{X}=((1,0),(1,0),(1,0))$ and $\vec{Y}=((0,1),(0,1),(0,1))$ correspond to displacements of the center of mass and are orthogonal to the vectors $\vec{E}_i$. This is expected since the center of mass motion moves freely. It is important to note that the boundary is neutral (i.e. neither focusing nor dispersing) in the transverse directions. It is straight forward to generalize these considerations to $N$ bodies. In the case of the $N$-body billiard there are $N(N-1)/2$ families of two-dimensional planes where defocusing might possibly occur. These families are related by those permutations that involve two out of $N$ particles, i.e. transpositions. $2(N-1)$ out of the $N(N-1)$ basis vectors $\vec{E}_i$ are linearly independent. The two vectors corresponding to the displacement of the center of mass are orthogonal to the vectors $\vec{E}_i$. It would be interesting to relate the number of positive Lyapunov exponents and the number of linearly independent basis vectors $\vec{E}_i$. Clearly, the former cannot exceed the latter. Assume that defocusing causes divergence in the directions of all linearly independent $\vec{E}_i$. Then there would be exactly $2(N-1)$ positive Lyapunov exponents. However, the conservation of energy and angular momentum puts two additional constraints, and $2N-4$ is the number of positive Lyapunov exponents. This reasoning is consistent with the numerical results presented in the previous section. The following picture thus arises. The boundary of the billiard considered in this work consists of several equivalent elements each of which cause a reflected trajectory to diverge [only]{} in a two-dimensional plane. The orientation of this plane is determined by the reflecting boundary element. In transverse directions the reflection is neutral, i.e. neither focusing nor dispersing. A trajectory that gets reflected from sufficiently many different boundary elements may exhibit divergence in all directions. It is interesting to note that this mechanism differs from the one investigated by Bunimovich [et al.]{} [@RehaBu96; @BCG96]. The neutral behavior in the transverse directions has the advantage that it avoids the problems caused by the weak convergence occurring in the transversal directions upon reflections from higher-dimensional spherical caps. It has the disadvantage that several focusing elements are needed to produce high-dimensional chaos while a single spherical cap may be sufficient. Conclusions =========== We have computed the Lyapunov spectrum and the KS entropy for an interacting $N$-body system in two spatial dimensions which is realized as a convex billiard in $2N$-dimensional configuration space. In presence of four conserved quantities we find the maximal number of $2N-4$ positive Lyapunov exponents. Thus, the system exhibits high-dimensional chaos. At fixed single particle energy the largest Lyapunov exponent grows with $\ln{N}$ while the KS entropy grows and the collision rate increase linearly with $N$. In an attempt to understand the chaotic nature of the billiard we have identified several symmetry related two-dimensional planes where defocusing might be generated. Their number and orientation in configuration space is such that a long trajectory may exhibit divergence in $2N-4$ directions of phase space. This mechanism of focusing differs from the one proposed recently by Bunimovich and Rehacek. Let us finally comment on chaos in realistic many-body systems. Though the considered model is a crude approximation of realistic self-bound many-body systems like nuclei or clusters it incorporates the important ingredient of an attractive two-body interaction that acts mainly at the surface of the system. This is the basic picture we have for nuclei and clusters, where the complicated two-body force creates a rather flat mean-field potential, and particles experience mainly a surface interaction. However, unlike in the model system, an interaction at the nuclear surface involves more than just two nucleons, and the two-dimensional planes where defocusing is generated in the model system are replaced by some higher-dimensional ones. This might also introduce the problem caused by weak focusing in transverse directions. It is fair to assume that truly high-dimensional chaos may develop upon several collisions with the surface. Though the detailed analysis seems much more complicated than in the studied model system, the basic picture developed in this work should be applicable to some extend also in the case of more realistic two-body interactions. Ya. G. Sinai, Sov. Math. Dokl. [4]{}, 1818 (1963) L. A. Bunimovich, Funct. Anal. Appl. [8]{}, 254 (1974) L. A. Bunimovich and J. Rehacek, Commun. Math. Phys. [189]{}, 729 (1997) L. Bunimovich, G. Casati, and I. Guarneri, , 2941 (1996) T. Guhr, A. Müller-Groeling, and H. A. Weidenmüller, Phys. Rep. [299]{}, 189 (1998) P. Gaspard, [Chaos, Scattering and Statistical Mechanics]{}, Cambridge University Press (Cambridge, 1998) Ch. Dellago, H. A. Posch, and W. G. Hoover, 1485, (1996) R. van Zon, H. van Beijeren, and J. R. Dorfman, e-print chao-dyn/9906040 T. Papenbrock, e-print chao-dyn/9905007 T. Papenbrock and T. Prosen, e-print chao-dyn/9905008 J.-P. Eckmann and D. Ruelle, Rev. Mod. Phys. [57]{}, 617 (1985) L. A. Bunimovich, Chaos [1]{}, 187 (1991) G. F. Bertsch, T. Papenbrock, and S. Reddy, e-print nucl-th/9906054 For a review see e.g. K. D. Luks and J. J. Kozak, Adv. Chem. Phys. [37]{}, 139 (1978) T. Prosen (private communication) A. J. Lichtenberg and M. A. Lieberman, [Regular and Stochastic Motion]{}, Springer-Verlag, (New York 1983) L. E. Reichl, [The Transition to Chaos in Conservative Classical Systems: Quantum Manifestations]{}, Springer-Verlag, (New York, 1992) G. Benettin, L. Galgani, and J. M. Strelcyn, , 2338 (1976) G. Benettin, C. Froeschle, and J. P. Scheidecker, , 2454 (1979) M. Sieber, Nonlinearity [11]{}, 1607 (1998) Ya. G. Piesin, Math. Dokl. [17]{}, 196 (1976) H. A. Posch and W. G. Hoover, , 473 (1994) R. Livi, A. Politi, and S. Ruffo, J. Phys. A [19]{}, 2033 (1986) R. van Zon, H. van Beijeren, and Ch. Dellago, , 2035 (1998) [\|r\|\|d\|d\|d\|d\|]{} $N$ & $\lambda_{5}$ & $\lambda_{2N}$ & $\tau^{-1}$ & $ h_{\rm KS}$\ 3 & 0.70 & 0.96 & 2.60 & 1.66\ 10 & 0.22 & 1.14 & 10.6 & 11.5\ 30 & 0.083 & 1.33 & 33.1 & 40.2\ 100 & 0.025 & 1.51 & 112. & 141.\ 300 & 0.008 & 1.65 & 338. & 428.\	{ "pile_set_name": "ArXiv" }
--- abstract: 'The cross sections for $D^0 {\bar D}^0$ and $D^+D^-$ production at 3.773 GeV have been measured with BES-II detector at BEPC. These measurements are made by analyzing a data sample of about $17.3$ $\rm pb^{-1}$ collected at the center-of-mass energy of 3.773 GeV. Observed cross sections for the charm pair production are radiatively corrected to obtain the tree level cross section for $D\bar D$ production. A measurement of the total tree level hadronic cross section is obtained from the tree level $D \bar D$ cross section and an extrapolation of the $R_{uds}$ below the open charm threshold.' author: - \| M. Ablikim$^{1}$, J. Z. Bai$^{1}$, Y. Ban$^{10}$, J. G. Bian$^{1}$, X. Cai$^{1}$, J. F. Chang$^{1}$, H. F. Chen$^{15}$, H. S. Chen$^{1}$, H. X. Chen$^{1}$, J. C. Chen$^{1}$, Jin Chen$^{1}$, Jun Chen$^{6}$, M. L. Chen$^{1}$, Y. B. Chen$^{1}$, S. P. Chi$^{2}$, Y. P. Chu$^{1}$, X. Z. Cui$^{1}$, H. L. Dai$^{1}$, Y. S. Dai$^{17}$, Z. Y. Deng$^{1}$, L. Y. Dong$^{1}$, S. X. Du$^{1}$, Z. Z. Du$^{1}$, J. Fang$^{1}$, S. S. Fang$^{2}$, C. D. Fu$^{1}$, H. Y. Fu$^{1}$, C. S. Gao$^{1}$, Y. N. Gao$^{14}$, M. Y. Gong$^{1}$, W. X. Gong$^{1}$, S. D. Gu$^{1}$, Y. N. Guo$^{1}$, Y. Q. Guo$^{1}$, K. L. He$^{1}$, M. He$^{11}$, X. He$^{1}$, Y. K. Heng$^{1}$, H. M. Hu$^{1}$, T. Hu$^{1}$, L. Huang$^{6}$, X. P. Huang$^{1}$, X. B. Ji$^{1}$, Q. Y. Jia$^{10}$, C. H. Jiang$^{1}$, X. S. Jiang$^{1}$, D. P. Jin$^{1}$, S. Jin$^{1}$, Y. Jin$^{1}$, Y. F. Lai$^{1}$, F. Li$^{1}$, G. Li$^{1}$, H. H. Li$^{1}$, J. Li$^{1}$, J. C. Li$^{1}$, Q. J. Li$^{1}$, R. B. Li$^{1}$, R. Y. Li$^{1}$, S. M. Li$^{1}$, W. G. Li$^{1}$, X. L. Li$^{7}$, X. Q. Li$^{9}$, X. S. Li$^{14}$, Y. F. Liang$^{13}$, H. B. Liao$^{5}$, C. X. Liu$^{1}$, F. Liu$^{5}$, Fang Liu$^{15}$, H. M. Liu$^{1}$, J. B. Liu$^{1}$, J. P. Liu$^{16}$, R. G. Liu$^{1}$, Z. A. Liu$^{1}$, Z. X. Liu$^{1}$, F. Lu$^{1}$, G. R. Lu$^{4}$, J. G. Lu$^{1}$, C. L. Luo$^{8}$, X. L. Luo$^{1}$, F. C. Ma$^{7}$, J. M. Ma$^{1}$, L. L. Ma$^{11}$, Q. M. Ma$^{1}$, X. Y. Ma$^{1}$, Z. P. Mao$^{1}$, X. H. Mo$^{1}$, J. Nie$^{1}$, Z. D. Nie$^{1}$, H. P. Peng$^{15}$, N. D. Qi$^{1}$, C. D. Qian$^{12}$, H. Qin$^{8}$, J. F. Qiu$^{1}$, Z. Y. Ren$^{1}$, G. Rong$^{1}$, L. Y. Shan$^{1}$, L. Shang$^{1}$, D. L. Shen$^{1}$, X. Y. Shen$^{1}$, H. Y. Sheng$^{1}$, F. Shi$^{1}$, X. Shi$^{10}$, H. S. Sun$^{1}$, S. S. Sun$^{15}$, Y. Z. Sun$^{1}$, Z. J. Sun$^{1}$, X. Tang$^{1}$, N. Tao$^{15}$, Y. R. Tian$^{14}$, G. L. Tong$^{1}$, D. Y. Wang$^{1}$, J. Z. Wang$^{1}$, K. Wang$^{15}$, L. Wang$^{1}$, L. S. Wang$^{1}$, M. Wang$^{1}$, P. Wang$^{1}$, P. L. Wang$^{1}$, S. Z. Wang$^{1}$, W. F. Wang$^{1}$, Y. F. Wang$^{1}$, Zhe Wang$^{1}$, Z. Wang$^{1}$,Zheng Wang$^{1}$, Z. Y. Wang$^{1}$, C. L. Wei$^{1}$, D. H. Wei$^{3}$, N. Wu$^{1}$, Y. M. Wu$^{1}$, X. M. Xia$^{1}$, X. X. Xie$^{1}$, B. Xin$^{7}$, G. F. Xu$^{1}$, H. Xu$^{1}$, Y. Xu$^{1}$, S. T. Xue$^{1}$, M. L. Yan$^{15}$, F. Yang$^{9}$, H. X. Yang$^{1}$, J. Yang$^{15}$, S. D. Yang$^{1}$, Y. X. Yang$^{3}$, M. Ye$^{1}$, M. H. Ye$^{2}$, Y. X. Ye$^{15}$, L. H. Yi$^{6}$, Z. Y. Yi$^{1}$, C. S. Yu$^{1}$, G. W. Yu$^{1}$, C. Z. Yuan$^{1}$, J. M. Yuan$^{1}$, Y. Yuan$^{1}$, Q. Yue$^{1}$, S. L. Zang$^{1}$,Yu. Zeng$^{1}$, Y. Zeng$^{6}$, B. X. Zhang$^{1}$, B. Y. Zhang$^{1}$, C. C. Zhang$^{1}$, D. H. Zhang$^{1}$, H. Y. Zhang$^{1}$, J. Zhang$^{1}$, J. Y. Zhang$^{1}$, J. W. Zhang$^{1}$, L. S. Zhang$^{1}$, Q. J. Zhang$^{1}$, S. Q. Zhang$^{1}$, X. M. Zhang$^{1}$, X. Y. Zhang$^{11}$, Y. J. Zhang$^{10}$, Y. Y. Zhang$^{1}$, Yiyun Zhang$^{13}$, Z. P. Zhang$^{15}$, Z. Q. Zhang$^{4}$, D. X. Zhao$^{1}$, J. B. Zhao$^{1}$, J. W. Zhao$^{1}$, M. G. Zhao$^{9}$, P. P. Zhao$^{1}$, W. R. Zhao$^{1}$, X. J. Zhao$^{1}$, Y. B. Zhao$^{1}$, H. Q. Zheng$^{10}$, J. P. Zheng$^{1}$, L. S. Zheng$^{1}$, Z. P. Zheng$^{1}$, X. C. Zhong$^{1}$, B. Q. Zhou$^{1}$, G. M. Zhou$^{1}$, L. Zhou$^{1}$, N. F. Zhou$^{1}$, K. J. Zhu$^{1}$, Q. M. Zhu$^{1}$, Y. C. Zhu$^{1}$, Y. S. Zhu$^{1}$, Yingchun Zhu$^{1}$, Z. A. Zhu$^{1}$, B. A. Zhuang$^{1}$, B. S. Zou$^{1}$,\ (BES Collaboration)\ title: 'Measurement of Cross Sections for $D^0 {\bar D}^0$ and $D^+D^-$ Production in $e^+e^-$ Annihilation at $\sqrt{s}=3.773$ GeV' --- Introduction ============ Around the center-of-mass energy of 3.770 GeV, the $\psi(3770)$ resonance is produced in $e^+e^-$ annihilation, and open charm pairs, $D^0 {\bar D}^0$ and $D^+D^-$, are mainly produced in $\psi(3770)$ decays. So the measurement of the cross sections for $D^0 {\bar D}^0$ and $D^+D^-$ production at an energy point around 3.770 GeV is very important in the understanding of $\psi(3770)$ decays. A coupled-channel model [@Eichten] predicts that the cross section for $\psi(3770)$ production is about $3~ \rm nb$ and that the $\psi(3770)$ decay exclusively into $D^0 {\bar D}^0$ and $D^+D^-$. Experimental results on the measurement of the $D^0 {\bar D}^0$, $D^+D^-$ and $D\bar D$ cross sections can be used to test the theoretical prediction. Also the measured values of the cross sections can be used to determine the branching fraction for $\psi(3770) \rightarrow {\rm non} D\bar D$ using the measured cross section for $\psi(3770) \rightarrow {\rm hadrons}$ at the same energy point. The determination of the partial width of $\psi(3770) \rightarrow {\rm non} D\bar D$ has great interest since it would be helpful for investigating the mixing between S and D waves in its wave function [@JLRosner], and in turn to help in developing the Potential Model [@Eichten]. In addition, by adding the tree level open charm cross section to an extrapolation of the tree level hadronic cross section for the light hadron production in the region below the open charm threshold, the total tree level hadronic cross section can be obtained [@bes4030][@coles]. The tree level cross section for inclusive hadronic event production in the $e^+e^-$ annihilation at all energies is needed to calculate the effects of vacuum polarization on the parameters of the Standard Model. The largest uncertainty in this calculation arises from the uncertainties in the measured inclusive hadronic cross sections in the open charm threshold region. Traditionally, the tree level hadronic cross sections are measured by counting inclusive hadronic events. Using the measured cross sections for the $D \bar D$ production in the charm threshold region, we can also measure the tree level inclusive hadronic cross sections. The BES-II Detector =================== BES-II is a conventional cylindrical magnetic detector that is described in detail in Ref. [@BES-II]. A 12-layer vertex chamber (VC) surrounding the beryllium beam pipe provides input to the event trigger, as well as coordinate information. A forty-layer main drift chamber (MDC) located just outside the VC yields precise measurements of charged particle trajectories with a solid angle coverage of $85\%$ of $4\pi$; it also provides ionization energy loss ($dE/dx$) measurements which are used for particle identification. Momentum resolution of $1.7\%\sqrt{1+p^2}$ ($p$ in GeV/$c$) and $dE/dx$ resolution of $8.5\%$ for Bhabha scattering electrons are obtained for the data taken at $\sqrt{s}=3.773$ GeV. An array of 48 scintillation counters surrounding the MDC measures the time of flight (TOF) of charged particles with a resolution of about 180 ps for electrons. Outside the TOF, a 12 radiation length, lead-gas barrel shower counter (BSC), operating in limited streamer mode, measures the energies of electrons and photons over $80\%$ of the total solid angle with an energy resolution of $\sigma_E/E=0.22/\sqrt{E}$ ($E$ in GeV) and spatial resolutions of $\sigma_{\phi}=7.9$ mrad and $\sigma_Z=2.3$ cm for electrons. A solenoidal magnet outside the BSC provides a 0.4 T magnetic field in the central tracking region of the detector. Three double-layer muon counters instrument the magnet flux return and serve to identify muons with momentum greater than 500 MeV/c. They cover $68\%$ of the total solid angle. Data sample and method to determine $\sigma_{D\bar D}$ ====================================================== The data used for this analysis were collected at the center-of-mass energy of 3.773 GeV with the Beijing Spectrometer [@BES-II] at Beijing Electron Positron Collider. The total integrated luminosity of the data set is 17.3 $\rm pb^{-1}$, which is obtained based on analysis of large angle Bhabha scattering from the same data set. The measurements of the cross sections for the $D^0 {\bar D}^0$ and $D^+D^-$ production are made based on the analysis of singly tagged $D^0$ and $D^+$ events. At the center-of-mass energy $\sqrt{s}=3.773$ GeV, the $D^0$ (through this paper, charge conjugation is implied) and $D^+$ are produced in pair via the process of $$e^+e^- \rightarrow D^0 {\bar D}^0, D^+D^-.$$ The total observed number $N_{D^0_{\rm tag}}$ ($N_{D^+_{\rm tag}}$) of $D^0$ ($D^+$) meson and the observed cross section $\sigma^{\rm obs}_{D^0 {\bar D}^0}$ ($\sigma^{\rm obs}_{D^+D^-}$) are related as $$\sigma_{D^0{\bar D}^0}^{\rm obs} = \frac {N_{D^0_{\rm tag}}} {2 \times L \times B \times \epsilon },$$ and $$\sigma_{D^+ D^-}^{\rm obs} = \frac {N_{D^+_{\rm tag}}} {2 \times L \times B \times \epsilon },$$ where $L$ is the integrated luminosity of the data set used in the analysis, $B$ is the branching fraction for decay mode in question, and $\epsilon$ is the efficiency determined from Monte Carlo for reconstruction of this decay mode. In the measurements of the cross sections, the singly tagged neutral and charged $D$ mesons are observed in the invariant mass spectra of the daughter particles from the $D$ decay. Data Analysis ============= Event selection --------------- The neutral and charged $D$ mesons are reconstructed in the final states of $K^-\pi^+$, $K^-\pi^+\pi^+\pi^-$ and $K^-\pi^+\pi^+$. Events which contain at least two reconstructed charged-tracks with good helix fits are selected. In order to ensure well-measured 3-momentum vectors and reliable charged-particle identification, the charged tracks used in the single tag analysis are required to be within $\|cos\theta\|<$0.85, where $\theta$ is the angle with respect to beam direction. All tracks must originate from the interaction region, which require that the closest approach of the charged track in $xy$ plane is less than 2.0 cm and the $z$ position of the charged track is less than 20.0 cm. Pions and kaons are identified by means of TOF and $dE/dx$ measurements. The pion identification requires a consistency with the pion hypothesis at a confidence level greater than 0.1%. In order to reduce misidentification, the kaon candidate is required to have a larger confidence level for a kaon hypothesis than that for a pion hypothesis. Analysis of inclusive $D$ meson events -------------------------------------- Taking advantage of the $D\bar D$ pair production, we use a kinematic fit to candidate $D^0$ or $D^+$ decays to improve the ratio of signal to noise and mass resolution in the invariant mass spectrum. The energy-constraint is imposed on the measured momenta of the $D$ daughter particles via the kinematic fit to improve the measured charged track three-momenta. Events with a kinematic fit probability greater than $1\%$ are accepted. If more than one combination satisfies the fit probability greater than $1\%$, the combination with the largest fit probability is retained. The resulting distributions of the fitted masses of $Km\pi$ ($m=1$, or $2$, or $3$) combinations, which are calculated using the fitted momentum vectors from the kinematic fit, are shown in figure 1. The signals for $D^0$ and $D^+$ production are clearly observed in the fitted mass spectra. A maximum likelihood fit is performed to the mass spectrum, a Gaussian is used to model the signal shape and a special function \[6\] is used to describe the background shape. The $D^0$ and $D^+$ yields obtained from this fit is given in table I. ![Distribution of the fitted masses of the $Km\pi$ (m=1, or 2, or 3) combinations for three singly tagged modes, where figure (a) and figure (b) are for the decay modes of $D^0\rightarrow K^-\pi^+$ and $D^0\rightarrow K^-\pi^+\pi^+\pi^-$, respectively, and the figure (c) is for the decay mode of $D^+\rightarrow K^-\pi^+\pi^+$. ](figure1.eps "fig:"){width="9.0cm" height="11.0cm"} (-200,280.0)[(a)]{} (-200,180.0)[(b)]{} (-200,85.0)[(c)]{} ---------- --------------------------------------------------------------------------------- -- -- Tag Mode $M_{\rm fit}$ \[MeV/$c^2$\] & $N^{\rm obs}_{D_{\rm tag}}$ & $N_{D_{\rm tag}}$\ $K^-\pi^+$ & $1865.5\pm 0.1$ & $1642.8 \pm 49.9$ & $1627.4 \pm 49.9$\ $K^-\pi^+\pi^+\pi^-$ & $1865.4\pm 0.1$ & $1327.2 \pm 48.6$ & $1299.1\pm 48.6$\ $K^-\pi^+\pi^+$ & $1870.2\pm 0.1$ & $2029.3 \pm 57.4$ & $2010.8\pm 57.4$\ ---------- --------------------------------------------------------------------------------- -- -- : Singly tagged $D^0$ and $D^+$ samples. Where the $M_{\rm fit}$ is the fitted mass of singly tagged $D$ meson, the $N^{\rm obs}_{D_{\rm tag}}$ is the observed number of singly tagged $D$ meson and the $N_{D_{\rm tag}}$ is the ”true” number of the singly tagged $D$ meson after correcting the contamination from other decay modes. The same $Km\pi$ combinations from other decay modes can also pass the above selection criteria and the distributions of the fitted masses of the combinations are with a small peak around the masses of $D$ mesons. The rates of the contaminations are evaluated to be 0.0094, 0.0212 and 0.0091 for the $D^0 \rightarrow K^-\pi^+$, $D^0 \rightarrow K^-\pi^+\pi^+\pi^-$ and $D^+ \rightarrow K^-\pi^+\pi^+$, respectively. After correcting the observed numbers of the singly tagged $D$ events for these combinations, the ”true” numbers of the $D$ signal events for the three singly tagged $D$ modes are obtained to be $1627\pm 50$, $1299 \pm 49$ and $2011 \pm 57$. Table I summarizes the results of the inclusive $D$ analysis. Observed cross sections for $D^0 {\bar D}^0$ and $D^+D^-$ production ==================================================================== Monte Carlo Efficiency ---------------------- To estimate the reconstructed efficiencies of $D^0 \rightarrow K^-\pi^+$, $D^0 \rightarrow K^-\pi^+\pi^+\pi^-$ and $D^+ \rightarrow K^-\pi^+\pi^+$, the Monte Carlo samples of $D\bar D$ production and decays are generated according to equation (1), where the ratio of the neutral over the total $D\bar D$ production cross section is set to be 0.58. Both $D$ and $\bar D$ mesons decay to all possible modes according to the branching fractions quoted from PDG[@PDG02]. The generated events are simulated with a GEANT based Monte Carlo package. All decay processes which contribute to the decay modes in question are considered in estimating the efficiencies. Detailed Monte Carlo studies give the efficiencies to be $(35.26\pm0.19)\%$, $(13.73 \pm 0.09)\%$ and $(25.00\pm 0.13)\%$ for the reconstruction of $D^0 \rightarrow K^-\pi^+$, $D^0 \rightarrow K^-\pi^+\pi^+\pi^-$ and $D^+ \rightarrow K^-\pi^+\pi^+$ decay modes, respectively. Observed Cross Sections ----------------------- Inserting the number of the singly tagged $D$ events and the efficiencies for each of the three decay modes, the $\sigma_{D\bar D}\times B$ of $D^0{\bar D}^0$ and $D^+D^-$ are obtained and the results are shown in table II. The first error is statistical and second systematic which arise from the uncertainty in the measured luminosity ($\sim 3\%$), tracking ($\sim 2\%$ per track), particle identification ($\sim 0.5\%$/track), kinematic fit ($\sim 1\%$), fitting parameters ($\sim 3\%$) and Monte Carlo statistics ($\sim 0.6\%$). The total systematic uncertainty is obtained by adding all systematic uncertainties in quadrature. Mode $\sigma_{D\bar D}\times B$ \[nb\] -------------------------------------- ----------------------------------- $D^0 \rightarrow K^-\pi^+$ $0.133\pm0.004\pm0.008$ $D^0 \rightarrow K^-\pi^+\pi^+\pi^-$ $0.273\pm0.010\pm0.025$ $D^+ \rightarrow K^-\pi^+\pi^+$ $0.233\pm0.007\pm0.018$ : Summary of the observed cross section times branching fraction. ---------------------- --------------------------- --------------------------- --------------------------- $\sigma_D\times B$ \[nb\] $\sigma_D\times B$ \[nb\] $\sigma_D\times B$ \[nb\] Mode (This experiment) (MARK-II) (MARK-I) $E_{cm}=3.773$ GeV $E_{cm}=3.771$ GeV $E_{cm}=3.774$ GeV $K^-\pi^+$ $0.27 \pm 0.02$ $0.24\pm0.02$ $0.25\pm0.05$ $K^-\pi^+\pi^+\pi^-$ $0.55 \pm 0.05$ $0.68\pm0.11$ $0.36\pm0.10$ $K^-\pi^+\pi^+$ $0.47 \pm 0.04$ $0.38\pm0.05$ $0.36\pm0.06$ ---------------------- --------------------------- --------------------------- --------------------------- : A comparison of $\sigma_D \times B$ measured by this experiment, MARK-I and MARK-II experiments. A comparison of our measured $\sigma_D\times B$ with that measured by MARK-II [@mark2_dxsct] and MARK-I [@mark1_dxsct] is given in Table III. The observed cross section for $D^+D^-$ production is obtained by dividing the $\sigma_{D\bar D} \times B$ by branching fraction quoted from PDG[@PDG02], which gives, $$\sigma_{D^+ {D^-} }^{\rm obs} = (2.56 \pm 0.08 \pm 0.26)~~ {\rm nb}.$$ The observed cross sections for $D^0 {\bar D}^0$ production are $ \sigma_{D^0 \bar {D^0} }^{\rm obs} = (3.50 \pm 0.11)~{\rm nb}$ and $\sigma_{D^0 \bar {D^0} }^{\rm obs} = (3.66 \pm 0.13)~{\rm nb}$, which are determined from the analysis of the singly tagged modes of $D^0 \rightarrow K^-\pi^+$ and $D^0 \rightarrow K^- \pi^+\pi^+\pi^-$ respectively; where the errors are statistical. Averaging the two observed cross sections for $D^0 {\bar D}^0$ production gives the average of the observed cross section for $D^0 {\bar D}^0$ production to be $$\sigma_{D^0 \bar {D^0} }^{\rm obs} = (3.58 \pm 0.09 \pm 0.31)~~ {\rm nb},$$ where the first error is statistical and second systematic which is estimated based on the averaged three charged tracks in the two modes of the single tags. Adding the observed cross sections of the neutral and charged modes together gives the observed cross section for $D \bar D$ production to be $$\sigma_{D \bar {D} }^{\rm obs} = (6.14 \pm 0.12 \pm 0.50)~~ {\rm nb},$$ where the first error is statistical and the second systematic. In the estimation of the systematic error of the $\sigma_{D\bar{D}}^{\rm obs}$, sources of systematic uncertainty are segregated into components that are common or independent for $D^0$ and $D^+$ measurements. The common components are the uncertainty in the measured integrated luminosity, the uncertainty in tracking and the uncertainty in particle identification. Since the absolute branching fraction scale for $D^+ \rightarrow K^-\pi^+\pi^+$ and $D^0 \rightarrow K^-\pi^+\pi^+\pi^-$ depend on the branching fraction scale for $D^0 \rightarrow K^-\pi^+$, the total percentage uncertainty for the two channel branching fractions ($6.6\%$ and $4.1\%$) are split into a common component that matches the percentage uncertainty for the $D^0 \rightarrow K^-\pi^+$ branching fraction ($2.3\%$) and independent components ($6.2\%$ and $3.4\%$). All other systematic uncertainties are treated as independent and added in quadrature. The common uncertainties are added linearly. --------------------------- --------------------------------------- ---------------------------------------- $\sigma^{\rm obs}_{D \bar D}$ \[nb\] $\sigma^{\rm obs}_{D \bar D}$ \[nb\] (This experiment) (MARK-III) $E_{cm}=3.773$ GeV $E_{cm}=3.768$ GeV $\sigma_{D^0 {\bar D}^0}$ $3.58 \pm 0.09 \pm 0.31$ $2.90 \pm 0.25 \pm 0.30$ $\sigma_{D^+D^-}$ $2.56 \pm 0.08 \pm 0.26$ $2.10\pm0.30\pm0.15$ --------------------------- --------------------------------------- ---------------------------------------- : Comparison of the observed cross section with that measured by MARK-III [@mark3_dbltag] experiment. As a comparison, Table IV lists the observed cross sections for $D^0\bar D^0$ and $D^+D^-$ production at the c.m. energies of 3.773 GeV and 3.768 GeV, which were measured by this experiment and MARK-III [@mark3_dbltag]. Radiative Corrections ===================== In any $e^+e^-$ colliding beam experiment, the electron (positron) always radiates at the interaction point because of the potential of the positron (electron). Since this radiation (Bremsstrahlung) carries energy away, the actual center-of-mass energy for the $e^+e^-$ annihilation is reduced by Bremsstrahlung to $\sqrt{s(1-x)}$, where $xE_{\rm beam}$ is the total energy of the emitted photons. The Bremsstrahlung is principally responsible for the distortions to the tree level resonance line shape, while the self energy of the electron and positron and the vertex corrections to the initial state affect the overall factors to change the scale of the cross section. All of these corrections are called Initial State Radiation (ISR) corrections. The tree level cross section for $D \bar D$ production at the energy of 3.773 GeV can be obtained by correcting the observed cross section for the effects of the ISR and vacuum polarization. The observed cross section, $\sigma^{\rm obs}$, at the nominal energy $\sqrt{s}$ can be written as a convolution of the Born cross section $\sigma^B(s(1-x))$ and a sampling function $f(x,s)$, $$\sigma^{\rm obs}(s)= \int^1_0 dx \cdot f(x,s)\sigma^{B}(s(1-x))(1+\delta_{VP}(s(1-x))) .$$ The vacuum polarization correction $(1+\delta_{VP})$ includes both leptonic and hadronic terms. It varies from charm threshold to 4.14 GeV by less than $\pm 2\%$ [@bes4030]. In this data analysis, we treat it as a constant of $$(1+\delta_{VP})=1.047\pm0.024.$$ Since we are interested in the $\psi(3770)$ resonance in this analysis, we take the $\sigma^B$ to be the bare Breit-Wigner cross section $$\sigma^B(E) = \frac{12 \pi \Gamma^0_{ee}\Gamma_{\rm tot}(E)} {{(E^2-M^2)^2 + M^2\Gamma^2_{\rm tot}(E)}},$$ where $\Gamma^0_{ee} = \Gamma_{ee}/(1+\delta_{\rm {vp}})$, $M$ and $\Gamma_{ee}$ are the mass and leptonic width of the $\psi(3770)$ resonance respectively; $E$ is the center-of-mass energy; $\Gamma_{\rm tot}(E)$ is chosen to be energy dependent and normalized to the total width $\Gamma_{\rm tot}$ at the peak of the resonance [@PDG02][@mark2]. The $\Gamma_{\rm tot}(E)$ is defined as $$\Gamma_{\rm tot}(E) = \Gamma_0 \frac{ \frac{p^3_{D^0}}{1+(r {p_{D^0}})^2} + \frac{p^3_{D^{\pm}}}{1+(r {p_{D^{\pm}}})^2} } { \frac{{p^0}^3_{D^0}}{1+(r {{p^0}_{D^0}})^2} + \frac{{p^0}^3_{D^{\pm}}}{1+(r {{p^0}_{D^{\pm}}})^2} },$$ where $p^0_D$ is the momentum of the $D$ mesons produced at the peak of $\psi(3770)$, $p^{ }_D$ is the momentum of the $D$ mesons produced at the c.m. energy $\sqrt{s}$, $\Gamma_0$ is the width of the $\psi(3770)$ at the peak, and $r$ is the interaction radius which was set to be 0.5 fm. In the calculation of the Born order cross section, the $\psi(3770)$ resonance parameters $M=3769.9\pm 2.5$ MeV; $\Gamma_0=23.6\pm2.7$ MeV and $\Gamma_{ee}=0.26 \pm 0.04$ keV [@PDG02] were used. In the structure function approach introduced by Kuraev and Fadin [@kuraev][@altarelli], the sampling function can be written as $$f(x,s)\;=\;\beta x^{\beta-1}\delta^{V+S}+\delta^{H} ,$$ where $\beta$ is the electron equivalent radiator thickness, $$\beta\;=\;\frac{2\alpha}{\pi} \left(\ln \frac{s}{m^{2}_{e}}-1\right) ,$$ $$\delta^{V+S}\;=\;1+\frac{3}{4}\beta+\frac{\alpha}{\pi} \left(\frac{\pi^{2}}{3}-\frac{1}{2}\right)+ \frac{\beta^2}{24}\left(\frac{1}{3}\ln \frac{s}{m^2_e}+2\pi^2-\frac{37}{4}\right),$$ $$\delta^{H}\;=\;\delta^{H}_{1}+\;\delta^{H}_{2} ,$$ $$\delta^{H}_{1}\;=\;-\beta\left(1-\frac{x}{2}\right) ,$$ $$\delta^{H}_{2}\;=\;\frac{1}{8}\beta^{2}\left[4(2-x)\ln\frac{1}{x}- \frac{1+3(1-x)^{2}}{x}\ln(1-x)-6+x\right] .$$ In the above formula, $m_e$ is the electron mass and $\alpha$ is the fine structure constant. The $\psi(3770)$ width ($\sim 24$ MeV) is much large than the energy spread ($\sim 1.37 $ MeV) of BEPC. So the effect of the beam energy spread on the cross section could be ignored. The $\psi(3770)$ is generally assumed to decay exclusively into $D \bar D$. Taking these considerations, the observed cross section of Equation (7) should be replaced by $$\sigma^{\rm obs}(s)=(1+\delta_{VP}) \int^{1-4M_D^2/s}_0 dx \cdot f(x,s)\sigma^{B}(s(1-x))$$ in calculation of the radiative corrections. The correction factor for the radiative effects is given by $$g = \frac{\sigma^{\rm obs}}{\sigma^{B}}.$$ Figure 2 shows the factor of the radiative corrections as a function of the nominal center-of-mass energy. At the center-of-mass energy $\sqrt{s}=3.773~ \rm GeV$, the factor is $$g = 0.779\pm 0.031,$$ where the error is the uncertainty arising from the errors on the $\psi(3770)$ resonance parameters. The uncertainty is mainly due to the error of the mass of the resonance. ![The factor of radiative corrections as a function of the nominal center-of-mass energy.](figure2.eps "fig:"){width="9.0cm" height="9.0cm"} (-170.0,-4.0)[***]{} (-260.0,170.0) Cross section for $D \bar D$ production ======================================= The tree level cross section for $D \bar D$ production is obtained by dividing the observed cross section by the factor of the radiative corrections. At $\sqrt{s}=3.773$ GeV, the charged, neutral and total tree level $D$ pair production cross sections are $$\sigma_{D^0 {\bar D}^0} = (4.60 \pm 0.12 \pm 0.45)~~\rm nb,$$ $$\sigma_{D^+D^-} = (3.29 \pm 0.10 \pm 0.37)~~\rm nb,$$ and $$\sigma_{D \bar D} = (7.88 \pm 0.15 \pm 0.74)~~\rm nb,$$ where the first error is statistical and the second systematic which include the uncertainty in the factor of the radiative corrections. These results are compared to the coupled-channel model prediction in Table V. Experiment coupled-channel Model --------------------------- ----------------------------- ----------------------- $\sigma_{D^0 {\bar D}^0}$ $4.60 \pm 0.12 \pm 0.45$ nb $1.80$ nb $\sigma_{D^+D^-}$ $3.29 \pm 0.10 \pm 0.37$ nb $1.28$ nb $\sigma_{D\bar D}$ $7.88 \pm 0.15 \pm 0.74$ nb $3.08$ nb : Comparison of tree level cross section measurements with prediction of the coupled-channel model at $\sqrt{s}=3.773$ GeV. Measurement of $R_D$ and $R$ ============================ The tree level cross section for $\mu^+\mu^-$ production in QED is given by $$\sigma_{e^+e^- \rightarrow \mu^+\mu^-} = \frac{86.8~ {\rm nb}}{E^2_{cm}},$$ where the $E_{cm}$ is the center-of-mass energy in GeV. A measurement of $R_D$ [@coles] is obtained by dividing $2\sigma_{D \bar D}$ by the tree level muon pair cross section, which gives $$R_D = 2.58 \pm 0.05 \pm 0.24.$$ BES-II experiment measured $R_{uds}$[@bes2R], which is the ratio of the tree level light hadron (containing the u,d and s light quarks) cross section over that for $\mu^+\mu^-$ production in the energy region from 2.0 to 3.0 GeV. Theoretical expectation is that $R_{uds}$ is approximately independent of center-of-mass energy in this region [@bes4030]. Fitting to the $R_{uds}$ values at 9 energy points in the energy region, we obtain $R_{uds}=2.26\pm0.14$. Assuming that $\psi(3770)$ decays exclusively into $D \bar D$, the value of $R$ is evaluated using $R=R_D/2 + R_{uds}$, which gives $$R=3.55 \pm 0.03 \pm 0.18.$$ Summary ======= In summary, using the 17.3 $\rm pb^{-1}$ of data collected with the BES-II detector at BEPC at center-of-mass energy $\sqrt{s}=3.773$ GeV, the observed cross sections for $D^0 {\bar D}^0$, $D^+D^-$ and $D\bar D$ production have been measured. Those are $\sigma_{D^0 \bar {D^0} }^{\rm obs} = (3.58 \pm 0.09 \pm 0.31)~{\rm nb}$, $\sigma_{D^+ {D^-} }^{\rm obs} = (2.56 \pm 0.08 \pm0.26)~{\rm nb}$ and $\sigma_{D \bar {D} }^{\rm obs} = (6.14 \pm 0.12 \pm 0.50)~{\rm nb}$. The tree level cross sections for the $D^0 {\bar D}^0$, $D^+D^-$ and $D\bar D$ production are determined to be $\sigma_{D^0 {\bar D}^0} = (4.60 \pm 0.12 \pm 0.45)~\rm nb$, $\sigma_{D^+D^-} = (3.29 \pm 0.10 \pm 0.37)~\rm nb$ and $\sigma_{D \bar D} = (7.88 \pm 0.15 \pm 0.74)~\rm nb$, which are about a factor 2.5 times larger than that predicted by the coupled-channel model. Using the measured $R_{uds}$ in the energy region from 2.0 to 3.0 GeV from BES-II experiment and assuming that $\psi(3770)$ only decays to $D \bar D$, the total tree level cross section for inclusive hadronic event production at 3.773 GeV is obtained to be $R=3.55 \pm 0.03 \pm 0.18$. The BES collaboration thanks the staff of BEPC for their hard efforts. This work is supported in part by the National Natural Science Foundation of China under contracts Nos. 19991480,10225524,10225525, the Chinese Academy of Sciences under contract No. KJ 95T-03, the 100 Talents Program of CAS under Contract Nos. U-11, U-24, U-25, and the Knowledge Innovation Project of CAS under Contract Nos. U-602, U-34(IHEP); by the National Natural Science Foundation of China under Contract No.10175060(USTC),and No.10225522(Tsinghua University). [\\]{} E. Eichten [et al.]{}, Phys. Rev. [D21]{}, 203 (1980). J.L. Rosner, hep-ph/0405196 v2. J.Z. Bai [et al.]{}, (BES Collaboration), Phys. Rev. [D62]{}, 012002-1 (2000). M.W. Coles et al. Phys. Rev. [D26]{}, 2190 (1982). J.Z. Bai [et al.]{} (BES Collaboration), Nucl. Instr. Meth. [A458]{}, 627 (2001). BES Collaboration, M. Ablikim [et al.]{}, Phys. Lett. B597(2004)39. A Gaussian function was assumed for the signal. The background shape was $$(1.0+p_1 y + p_2 y^2) N\sqrt{1-(\frac{x}{E_b})^2} x e^{-f(1-\frac{x}{E_b})^2},$$ where $N\sqrt{1-(\frac{x}{E_{\rm beam}})^2}xe^{-f(1-\frac{x}{E_{\rm beam}})^2}$ is ARGUS background shape, $x$ is the fitted mass, $E_{\rm beam}$ is the beam energy, $y=(E_{\rm beam}-x)/(E_{\rm beam} -1.82)$, $N$, $f$, $p_1$ and $p_2$ are the fit parameters; The ARGUS background shape was used by ARGUS experiment to parametrize the background for fitting $B$ mass peaks. For detail, see: Ian C. Brock, Mn-Fit, a Fitting and plotting package using MINUIT, Version 4.07, December 22, 2000. K. Hagiwara [et al.]{} (Particle Data Group), Phys. Rev. [D66]{} (2002). R.H. Schindler [et al.]{}, Phys. Rev. [D24]{}, 78(1981). I. Peruzzi [et al.]{}, Phys. Rev. Lett. [39]{}, 1301(1977), D.L. Scharre [et al.]{}, Phys. Rev. Lett. [40]{}, 74(1978). J. Adler [et al.]{}, Phys. Rev. Lett. [60]{}, 89(1988). R.H. Schindler [et al.]{}, Phys. Rev. [D21]{} 2716(1980). E.A. Kuraev and V.S. Fadin, Sov. J. Nucl. Phys. 41(1985)466. G. Altarelli and G. Martinelli, CERN Yellow Report 86-02(1986)47; O. Nicrosini and L. Trentadue, Phys. Lett. B196(1987)551. J.Z. Bai [et al.]{}, (BES Collaboration), Phys. Rev. Lett. [88*]{}, 101802-1 (2002).	{ "pile_set_name": "ArXiv" }
--- abstract: 'We investigate the nonlinear instability of a smooth Rayleigh-Taylor steady-state solution (including the case of heavier density with increasing height) to the three-dimensional incompressible nonhomogeneous magnetohydrodynamic (MHD) equations of zero resistivity in the presence of a uniform gravitational field. We first analyze the linearized equations around the steady-state solution. Then we construct solutions of the linearized problem that grow in time in the Sobolev space $H^k$, thus leading to the linear instability. With the help of the constructed unstable solutions of the linearized problem and a local well-posedness result of smooth solutions to the original nonlinear problem, we establish the instability of the density, the horizontal and vertical velocities in the nonlinear problem. Moreover, when the steady magnetic field is vertical and small, we prove the instability of the magnetic field. This verifies the physical phenomenon: instability of the velocity leads to the instability of the magnetic field through the induction equation.' address: - 'College of Mathematics and Computer Science, Fuzhou University, Fuzhou, 350108, China.' - 'Institute of Applied Physics and Computational Mathematics, Beijing, 100088, China.' author: - Fei Jiang - Song Jiang - Weiwei Wang title: \| Nonlinear Rayleigh-Taylor Instability for Nonhomogeneous\ Incompressible Viscous Magnetohydrodynamic Flows --- Incompressible MHD flows, steady solutions, Rayleigh-Taylor instability, Navier-Stokes equations. 35Q35,76D03. \[section\] \[section\] \[section\] \[section\] \[section\] \[section\] Introduction {#Intro} ============ This paper is concerned with nonlinear instability of a smooth Rayleigh-Taylor (RT) steady-state solution to the following three-dimensional (3D) nonhomogeneous incompressible magnetohydrodynamic (MHD) equations with zero resistivity (i.e. without magnetic diffusivity) in the presence of a uniform gravitational field (see, for example, [@CHTA1; @kulikovskiy1965magnetohydrodynamics; @landau1984electrodynamics; @LFZPGC] on the derivation of the equations): $$\label{0101}\left\{\begin{array}{l} \rho_t+{\bf v}\cdot\nabla \rho=0,\\[1mm] \rho\mathbf{v}_t+\rho {\bf v}\cdot\nabla {\bf v}+\nabla p=(\nabla\times {\mathbf}{M})\times {\mathbf}{M}+\mu\Delta {\bf v}-\rho{g}{\mathbf}{e}_3,\\[1mm] {\mathbf}{M}_t-\nabla\times ({\mathbf}{v}\times {\mathbf}{M})={\mathbf}{0},\\[1mm] \mathrm{div}\mathbf{v}=0,\quad \mathrm{div}\mathbf{M}=0.\end{array}\right.$$ Here the unknowns $\rho:=\rho(t,{\mathbf}{x})$, $\mathbf{v}:=\mathbf{v}(t,{\mathbf}{x})$, ${\mathbf}{M}:={\mathbf}{M}(t,{\mathbf}{x})$ and $p:=p(t,{\mathbf}{x})$ denote the density, velocity, magnetic field and pressure of the incompressible fluid, respectively; $\mu>0$ stands for the coefficient of shear viscosity, $g>0$ for the gravitational constant, ${\mathbf}{e}_3=(0,0,1)$ for the vertical unit vector, and $-\rho g{\mathbf}{e}_3$ for the gravitational force. In the system the equation $_1$ is the continuity equation, while $_2$ describes the balance law of momentum. It is well-known that the electromagnetic field is governed by the Maxwell equations. In MHD, the displacement current can be neglected [@kulikovskiy1965magnetohydrodynamics; @landau1984electrodynamics]. As a consequence, $_3$ is called the induction equation. As for the constraint ${\mathrm}{div}\,{\mathbf}{M}=0$, it can be seen just as a restriction on the initial value of ${\mathbf}{M}$ since $({\mathrm}{div}\,{\mathbf}{M})_t=0$. We remark that, the resistivity in $_3$ is zero, which arises in the physics regime with negligible electrical resistance, see [@cowling1957magnetohydrodynamics]. In addition, if ${\mathbf}{M}\equiv0$, the system reduces to the incompressible Navier-Stokes equations in the presence of a uniform gravitational field. In this paper we consider the problem of the RT instability in a horizontally periodic domain $\Omega:=(2\pi L\mathbb{T})^2\times \mathbb{R}$, where $2\pi L \mathbb{T}$ stands for the 1D-torus of length $2\pi L$. We assume that a smooth RT (steady-state) density profile $\bar{\rho}:=\bar{\rho}({x}_3)\in L^\infty(\mathbb{R})$ exists and satisfies $$\begin{aligned} \label{0102} &&\bar{\rho}'\in C_0^{\infty}(\mathbb{R}),\quad \inf_{x_3\in \mathbb{R}}\bar{\rho}>0,\\[0em] && \label{0103}\bar{\rho}'(x_3^0)>0\;\;\mbox{ for some point }x_3^0\in \mathbb{R},\end{aligned}$$ where $'=d/dx_3$. We refer to [@NJTSC2 Remark 1.1] for the construction of such $\bar{\rho}$. Let $\bar{{\mathbf}{M}}\in \mathbb{R}^3$ be a constant magnetic field, then the RT density profile $\bar{\rho}$ with $(\mathbf{v},\mathbf{M})(t,\mathbf{x}) \equiv(\mathbf{0},\bar{\mathbf{M}})$ defines a steady state solution to , provided the steady pressure $\bar{p}$ is determined by $$\nabla \bar{p}=-\bar{\rho}g {\mathbf}{e}_3,\;\mbox{ i.e., }\;\frac{d\bar{p}}{dx_3}=-\bar{\rho}g.$$ We point out that by virtue of the condition , there is at least a region in which the RT density profile has larger density with increasing $x_3$ (height), thus leading to the classical RT instability as shown in Theorem \[thm:0101\] below. Now, we denote the perturbation to the RT steady state by $$\varrho=\rho -\bar{\rho},\quad \mathbf{u}=\mathbf{v}-\mathbf{0},\quad \mathbf{N}={\mathbf}{M}-\bar{{\mathbf}{M}},\quad q=p-\bar{p},$$ then, $(\varrho ,\mathbf{u},q)$ satisfies the perturbed equations $$\label{0105}\left\{\begin{array}{l} \varrho_t+{\bf u}\cdot\nabla (\varrho+\bar{\rho})=0, \\[1mm] (\varrho+\bar{\rho}){\bf u}_t+(\varrho+\bar{\rho}){\bf u}\cdot\nabla {\bf u}+\nabla q+g \varrho {\mathbf}{e}_3=\mu \Delta \mathbf{u}+(\nabla\times {\mathbf}{N})\times ({\mathbf}{N}+\bar{{\mathbf}{M}}),\\[1mm] {\mathbf}{N}_t=\nabla\times ({\mathbf}{u}\times ({\mathbf}{N}+\bar{{\mathbf}{M}})),\\[1mm] \mathrm{div}\mathbf{u}=0,\ \mathrm{div}\mathbf{N}=0.\end{array}\right.$$ For (\[0105\]) we impose the initial and boundary conditions: $$\label{0106} (\varrho,{\bf u},\mathbf{N} )\|_{t=0}=(\varrho_0,{\bf u}_0,{\mathbf}{N}_0)\quad\mbox{in } \Omega$$ and $$\label{0107} \lim_{\|{x}_3\|\rightarrow +\infty}(\varrho,{\bf u},{\mathbf}{N})(t,\mathbf{x}',x_3)={\bf 0}\quad \mbox{ for any }t>0,$$where we have written ${\mathbf}{x}'=x_1{\mathbf}{e}_1+x_2{\mathbf}{e}_2$, ${\mathbf}{e}_1:=(1,0,0)$ and ${\mathbf}{e}_2:=(0,1,0)$. Moreover, the initial data should satisfy the compatibility conditions $$\mathrm{div}\mathbf{u}_0=0\quad\mbox{ and }\quad\mathrm{div}\mathbf{N}_0=0.$$ If we linearize the equations around the steady state $(\bar{\rho},\mathbf{0},\bar{{\mathbf}{M}})$, then the resulting linearized equations read as $$\label{0108} \left\{\begin{array}{ll} \varrho_t+\bar{\rho}'{u}_3=0, \\[1mm] \bar{\rho}\mathbf{u}_t +\nabla q+g\varrho {\mathbf}{e}_3=\mu \Delta \mathbf{u} +(\nabla\times {\mathbf}{N})\times \bar{{\mathbf}{M}},\\[1mm] {\mathbf}{N}_t=\nabla\times ({\mathbf}{u}\times \bar{{\mathbf}{M}}),\\[1mm] \mathrm{div}\mathbf{u}=0,\quad \mathrm{div}\mathbf{N}=0. \end{array}\right.$$ The RT instability is well-known as gravity-driven instability in fluid dynamics when heavy fluid is on top of light one. The linear instability for an incompressible fluid was first introduced by Rayleigh in 1883 [@RLAP]. The analogue of the RT instability arises when fluids are electrically conducting and a magnetic field is present, and the growth of the instability will be influenced by the magnetic field due to the generated electromagnetic induction and the Lorentz force. Some authors have extended the partial results concerning the RT instability to the case of MHD fluids by overcoming difficulties induced by presence of the magnetic field. For example, Kruskal and Schwarzchild in 1954 first showed that a horizontal magnetic field has no effect on the development of the linear RT instability [@KMSMSP]. Then the influence of a vertical magnetic field was investigated by Hide in [@HRWP] where the effect of finite viscosity and resistivity was included and his analysis was encumbered with many parameters. By a variational approach, Hwang in 2008 studied the nonlinear RT instability of – for the inviscid case (i.e. $\mu=0$) in a 2D periodic domain [@HHVQ]. To our best knowledge, however, it is still open mathematically whether there exists an unstable solution to the nonlinear RT problem – of 3D viscous MHD fluids. The aim of this article is to rigorously verify the instability for the nonlinear RT problem –. Moreover, the impact of the magnetic filed on the instability will be analyzed, for example, we shall show that if the steady magnetic field is vertical and small, then the magnetic field is unstable, thus verifying the physical phenomenon: instability of the velocity leads to the instability of the magnetic field through the induction equation. The main result of this paper reads as follows. \[thm:0101\] Let $$\begin{aligned}\label{0109} M_\mathrm{c}:=\sqrt{\sup_{\psi\in H^1(\mathbb{R})\atop \psi \equiv\!\!\!\!/~0} \frac{\int_{\mathbb{R}}g\bar{\rho}'\|\psi\|^2{\mathrm}{d}x}{\int_{\mathbb{R}}\|\psi'\|^2{\mathrm}{d}x}}>0. \quad\qquad\mbox{(critical number)} \end{aligned}$$ Assume that the RT density profile $\bar{\rho}\in L^\infty(\Omega)$ satisfies and , and $$\label{mageticfields}{\mathbf}{\bar{M}}=\left\{ \begin{array}{ll} M{\mathbf}{e}_1, & M\neq 0, \mbox{ constant}, \\[1mm] M{\mathbf}{e}_3, & \|M\|\in (0,M_{\mathrm}{c}), \mbox{ constant}. \end{array} \right.$$ Then, the steady state $(\bar{\rho},\mathbf{0},{\mathbf}{\bar{M}})$ of – is unstable, that is, there exist positive constants $\Lambda^$, $\varepsilon$, $m_0$ and $\iota$, and a triple $(\bar{\varrho}_0,\bar{{\mathbf}{u}}_0,\bar{{\mathbf}{N}}_0)\in H^\infty(\Omega):=\cap_{k=0}^\infty H^k(\Omega)$, such that for any $\delta\in (0,\iota)$ and the initial data $(\varrho_0,\mathbf{u}_0,\mathbf{N}_0):=(\delta\bar{\varrho}_0,\delta\bar{{\mathbf}{u}}_0,\delta\bar{{\mathbf}{N}}_0)$, there is a unique classical solution $({\varrho},\mathbf{u},{\mathbf}{{N}})$ of – on $[0,T^{\max})$, but $$\label{0111}\\|\varrho(T^\delta)\\|_{L^2(\Omega)},\ \\|({u}_1,u_2)(T^\delta)\\|_{L^2(\Omega)},\ \\|{u}_3(T^\delta)\\|_{L^2(\Omega)}\geq {\varepsilon}\;$$ for some escape time $T^\delta:=\frac{1}{\Lambda^}{\mathrm}{ln}\frac{2\varepsilon}{m_0\delta}\in (0,T^{\max})$, where $\mathrm{div}\bar{\mathbf{u}}_0=0$, $\mathrm{div}\bar{\mathbf{N}}_0=0$, and $T^{\max}$ denotes the maximal time of existence of the solution $(\varrho,\mathbf{u},{\mathbf}{N})$. Moreover, $$\\|N_3(T^\delta)\\|_{L^2(\Omega)}\geq \varepsilon\quad \mbox{ for the case }\;\bar{{\mathbf}{M}}=M{\mathbf}{e}_3.$$ We should point out that since the above Sobolev spaces $H^k(\Omega)$ are defined on the horizontally periodic domain $\Omega=(2\pi L\mathbb{T})^2\times \mathbb{R}$, the solution $({\varrho},\mathbf{u},{\mathbf}{{N}})$ constructed in Theorem \[thm:0101\] is horizontally periodic. One should keep this in mind in what follows. Throughout the rest of this article we shall repeatedly use the abbreviations: $$\begin{aligned} & I_T:=(0,T),\ \bar{I}_T:=[0,T],\ W^{m,p}:=W^{m,p}(\Omega),\;\; H^m:=H^m(\Omega),\ H^\infty:=\cap_{m=1}^\infty H^{m}(\Omega),\nonumber\\ &L^p:=L^p(\Omega),\ \\|\cdot\\|_{W^{m,p}}:=\\|\cdot\\|_{W^{m,p}(\Omega)},\;\; \\|\cdot\\|_{H^m}:=\\|\cdot\\|_{H^m(\Omega)},\;\; \\|\cdot\\|_{L^p}:=\\|\cdot\\|_{L^p(\Omega)},\quad \mbox{etc.}\nonumber\end{aligned}$$ The conclusion in Theorem \[thm:0101\] also holds for the general horizontal magnetic field $\bar{{\mathbf}{M}}=({M}_1,{M}_2,0)$. In fact, rotating the $o$-$xy$ coordinates properly so that $\bar{{\mathbf}{M}}=({M},0,0)$ with ${M}=\sqrt{{M}_1^2+{M}_1^2}$, we have the same case as in Theorem \[thm:0101\] under the rotated coordinates, since the $L^2$-norms of the density, horizontal velocity, and vertical velocity are invariant under the horizontal rotation. We should point out here that for the both cases of the horizontal and vertical magnetic fields, the ordinary differential equation (ODE) associated with the normal mode solutions enjoys good variational structure. For a general constant magnetic field $\bar{{\mathbf}{M}}=({M}_1,{M}_2,M_3)$, similarly to that in the derivation of , we can also deduce a ODE corresponding to $\bar{{\mathbf}{M}}=({M}_1,{M}_2,M_3)$. The resulting ODE, however, possesses the terms ${\mathrm}{i}\psi'$ and ${\mathrm}{i}\psi'''$ which destroy the good variational structure, and consequently, one could not directly construct the growing mode solutions to the linearized problem. In addition, Theorem \[thm:0101\] also holds when $\bar{{\mathbf}{M}}={\mathbf}{N}={\mathbf}{0}$, hence the weak RT instability for incompressible viscous fluids given in [@NJTSC2] can be further shown to be strongly RT unstable as in in the case of the horizontally periodic domain. In [@HHJGY] Hwang and Guo proved the nonlinear RT instability for 2D nonhomogeneous incompressible inviscid flows (i.e. $\mu=0$ and $\bar{{\mathbf}{M}}={\mathbf}{N}={\mathbf}{0}$ in , ) with boundary condition $\mathbf{u}\cdot \mathbf{n}\|_{\partial\Omega'}=0$ where $\Omega'=\{(x_1,x_2)\in \mathbb{R}^2~\|~-l<x_2<m\}$ and $\mathbf{n}$ denotes the outer normal vector to $\partial\Omega'$. Later, Hwang [@HHVQ] further investigated the nonhomogeneous incompressible inviscid MHD fluid on a periodic domain, and get the instability of the norm $\\|(\varrho,\mathbf{u},\mathbf{N})\\|_{L^2(\Omega')}$. Our result is more precise than those in [@HHJGY; @HHVQ] in the sense that Theorem \[thm:0101\] reveals that the vertical velocity induces the instability of the density and horizontal velocity; and moreover, we can show the instability of the magnetic field for the case $\bar{{\mathbf}{M}}=M{\mathbf}{e}_3$. This mathematically verifies the physical phenomenon: instability of the velocity further leads to the instability of the magnetic field through the induction equation. \[criticalmget\] The number $M_{\mathrm}{c}$ in is infinite for $\int_{\mathbb{R}}\bar{\rho}'{\mathrm}{d}x> 0$ (i.e., $\bar{\rho}(+\infty)>\bar{\rho}(-\infty)$) and is finite for $\int_{\mathbb{R}}\bar{\rho}'{\mathrm}{d}x< 0$ (i.e., $\bar{\rho}(+\infty)<\bar{\rho}(-\infty)$), see Proposition \[pro:n0201\] for the detailed proof. This means that any vertical steady magnetic field cannot restrain growth of the nonlinear RT instability for the case $\int_{\mathbb{R}}\bar{\rho}'{\mathrm}{d}x> 0$. Now it is still not clear to us that whether a sufficiently large vertical steady magnetic field has impact on growth of the nonlinear RT instability for the case $\int_{\mathbb{R}}\bar{\rho}'{\mathrm}{d}x< 0$ due to some technical difficulties. However, if we consider $\Omega =(2\pi L\mathbb{T})^2\times (-l,m)$, then, similarly to the derivation of (3.71) in [@WYC], we can show the stability of the velocity for any classical solution of the linearized problem satisfying boundary conditions ${\mathbf}{u}\|_{x_3=-l}={\mathbf}{u}\|_{x_3=m}=0$, provided the vertical steady magnetic field is sufficiently large. This result does not contradict that in [@HHVQ] where for any vertical steady magnetic field, the nonlinear RT instability for a nonhomogeneous incompressible inviscid MHD flow in a periodic domain (the vertical direction is also periodic) is shown. These mathematical results reveal that the domain and boundary conditions of the velocity do have impact on the instability of MHD flows. The proof of Theorem \[thm:0101\] is divided into four steps given in Sections \[sec:02\]–\[sec:05\]: (i) First, we make an ansatz to seek the “normal mode” solutions of the linearized equations , which grow exponentially in time by the factor $e^{\lambda(\xi)t}$ with $\xi\in\mathbb{R}^2$ being the horizontal spatial frequency and $\lambda (\xi)>0$. This reduces the equations to a system of ODEs defined on $\mathbb{R}$ with $\lambda(\xi)>0$ for some $\xi$ (see , ). All such points $\xi$ constitute a solvable domain. Because of presence of the magnetic field, the solvable domain is not a ball for the horizontal case as in [@GYTI2], resulting in some difficulties in constructing a solvable domain. In order to circumvent such difficulties, similarly to [@WYC; @JFJSWWWOA], we introduce the notions of the critical frequency function $S(\xi)$ and critical frequency constant $\|\xi\|_{\mathrm{vc}}^M$ to define the solvable domain $\mathbb{A}^{{\mathrm}{g}}$ (cf. , ). Thus, by careful constructing the solutions and adapting the modified variational method in [@GYTI2], we can also solve the ODEs for any given $\xi\in \mathbb{A}^{{\mathrm}{g}}$ and obtain a normal mode with $\lambda(\xi)>0$, thus leading to a mechanism for the global linear RT instability. Using the normal modes, we can further construct real-valued solutions of the linearized problem – that grow in time, when measured in $H^k(\Omega)$ for any $k\geq 0$. In particular, the density, horizontal velocity and vertical velocity in the solutions of the linearized problem are not zero, this fact will play a key role in the nonlinear instability in . (ii) In Section \[sec:04\], we state a local well-posedness result of the perturbed problem –, which will be proved in Section \[appendix\]. Then we derive the integrand form of Gronwall’s inequality of high-order energy estimate $\mathcal{E}(\varrho,\mathbf{u},\mathbf{N})$ for the perturbed problem, and this makes the escape time occurring before break-down of the classical solutions. Since the equilibrium state of the magnetic field $\bar{{\mathbf}{M}}$ is a no-zero vector, we shall introduce a simple technique to deal with the terms including $\bar{{\mathbf}{M}}$ in the energy estimates, see Subsection \[sec:0404\]. (iii) Finally, in Section \[sec:05\], with the help of the results established in Sections \[sec:02\]–\[sec:04\], we adapt a careful bootstrap argument as in [@GYSWIC] to establish the instability of the nonlinear problem. We mention that although the approach in [@GYSWIC] has been also used to treat the instability of other problems (see [@GYSWIC; @GYHCSDDC; @JJTIIA; @wang2011viscous] for examples), but Duhamel’s principle in the standard bootstrap argument can not be directly applied to our problem, since the nonlinear terms in do not satisfy the compatibility condition of divergence-free. To circumvent this obstacle, we shall use some specific energy estimates to replace Duhamel’s principle. Moreover, we can also find in the proof that $\Lambda$ is indeed a sharp exponential growth rate for general solutions to the linearized problem (see Remark \[rem0401n\]). We end this section by briefly reviewing some of the previous results on the nonlinear RT instability for two layer incompressible fluids separated by a free interface (stratified fluids), where the RT steady-state solution is a denser fluid lying above a lighter one separated by a free interface. When the densities of two layer fluids are two constants, Wang and Tice [@wang2011viscous] proved the (local) existence of nonlinear unstable solutions in a horizontally periodic domain $\mathbb{T}^2\times (-b,1)$, where the instability term is described by the sum of $L^{2}$-norm of the velocity and the moving internal interface. Prüss and Simonett used the $C^0$-semigroup theory and the Henry instability theorem to show the existence of nonlinear unstable solutions in the Sobolev-Slobodeckii spaces in $\mathbb{R}^3$ [@PJSGOI5], where the instability term is described by the sum of $\\|\mathbf{u}\\|_{W_p^{2-2/p}(\mathbb{R}^3)}$ and $\\|\mathbf{h}\\|_{W_p^{3-2/p}(\mathbb{R}^3)}$ (see [@PJSGOI5 Theorem 1.2] for details). When densities of two layer fluids are variable, to our best knowledge, the (local) existence of solutions to the nonlinear problem – still remains open, consequently, the strong nonlinear instability is still open. For compressible fluids, there are very few results on the nonlinear RT instability. Guo and Tice proved the instability of immiscible compressible inviscid fluids in the frame of Lagrangian coordinates under the existence assumption of solutions [@GYTI1]. This is in some sense a compressible analogue to the local ill-posedness of the RT problem for incompressible fluids given in [@EDGTC1111]. Finally we mention related results on the instability for stratified MHD fluids. Wang [@WYC] introduced the critical number of stratified MHD fluids (denoted by $M_\mathrm{c}^{\mathrm{s}}$) to investigate the linear RT instability of stratified MHD fluids in a infinite slab domain $\mathbb{R}^2\times (-l,l)$. Later, Jiang et al. [@JFJSWWWOA] further showed the weak nonlinear RT instability of stratified MHD fluids and found that $M_\mathrm{c}^{\mathrm{s}}=\sqrt{gl(\varrho_+-\varrho_-)/2}$, where $\varrho_+> \varrho_-$, and $\varrho_+ $ resp. $\varrho_-$ denotes the density of the upper- resp. lower-layer fluid. Obviously, $M_\mathrm{c}^{\mathrm{s}}\to\infty$ as the height $2l\rightarrow \infty$. Hence, the critical number of stratified MHD fluids in $\mathbb{R}^3$ is infinite. This means that the vertical magnetic field may have no impact on growth of the RT instability in $\mathbb{R}^3$. Construction of solutions to the linearized problem {#sec:02} =================================================== We wish to construct a solution to the linearized equations that has growing $H^k$-norm for any $k$. We will construct such solutions via some synthesis as in [@GYTI2] by first constructing a growing mode for any but fixed spatial frequency. Moreover, we shall introduce the techniques of the critical frequency function and critical frequency constant to carefully analyze the terms involving with the magnetic field. Linear growing modes -------------------- To start with, we make a growing mode ansatz of solutions, i.e., for some $\lambda>0$, $${\varrho}(t,\mathbf{x})=\tilde{\rho} (\mathbf{x})e^{\lambda t},\; \mathbf{u}(t,\mathbf{x})=\tilde{\mathbf{v}}(\mathbf{x})e^{\lambda t},\; {q}(t,\mathbf{x})=\tilde{p}(\mathbf{x})e^{\lambda t},\; {\mathbf}{N}(t,{\mathbf}{x})=\tilde{{\mathbf}{M}}(\mathbf{x})e^{\lambda t}.$$ Substituting this ansatz into , and then eliminating $\tilde{\rho}$ and $\tilde{{\mathbf}{M}}$ by using the first and third equations, we arrive at the time-independent system for $\tilde{\mathbf{v}}=(\tilde{v}_1,\tilde{v}_2,\tilde{v}_3)$ and $\tilde{p}$: $$\label{0201} \left\{ \begin{array}{ll} \lambda^2\bar{\rho}\tilde{\mathbf{v}}+\lambda\nabla \tilde{{p}} = (\nabla \times (\nabla\times (\tilde{{\mathbf}{v}}\times \bar{{\mathbf}{M}})))\times \bar{{\mathbf}{M}}+\lambda\mu \Delta \tilde{\mathbf{v}}+g\bar{\rho}'\tilde{v}_3{\mathbf}{e}_3,\\[1mm] \mathrm{div}\, \tilde{\mathbf{v}}=0 \end{array} \right.$$ with $$\label{0202} \lim_{\|x_3\|\rightarrow +\infty}(\tilde{\mathbf{v}},\tilde{\mathbf{M}})(\mathbf{x'},x_3)=\mathbf{0},$$where $\bar{{\mathbf}{M}}$ is given by . After a straightforward calculation, we find that $$\label{n0203}\begin{aligned} \nabla \times (\nabla\times (\tilde{{\mathbf}{v}}\times \bar{{\mathbf}{M}}))\times \bar{{\mathbf}{M}} &=\nabla \times (\bar{{\mathbf}{M}}\cdot\nabla\tilde{{\mathbf}{v}} )\times \bar{{\mathbf}{M}}\\ &=M^2\left\{ \begin{array}{ll} (0,\partial_{1}^2 \tilde{v}_2-\partial_{21}^2\tilde{v}_1,\partial_{11}^2\tilde{v}_3 -\partial_{31}^2\tilde{v}_1), & \hbox{for }\bar{{\mathbf}{M}}=M{\mathbf}{e}_1; \\[1mm] (\partial_{3}^2 \tilde{v}_1-\partial_{13}^2\tilde{v}_3,\partial_{33}^2\tilde{v}_2 -\partial_{23}^2\tilde{v}_3,0), & \hbox{ for }\bar{{\mathbf}{M}}=M{\mathbf}{e}_3. \end{array} \right. \end{aligned}$$ We fix a spatial frequency $\xi=(\xi_1,\xi_2)\in \mathbb{R}^2$, and define the new unknowns $$\tilde{v}_1({\mathbf}{x})=-\mathrm{i}\varphi(x_3)e^{{\mathrm}{i}{\mathbf}{x}'\cdot\xi}, \;\; \tilde{v}_2({\mathbf}{x})=-\mathrm{i}\theta(x_3)e^{{\mathrm}{i}{\mathbf}{x}'\cdot\xi},\;\;\tilde{v}_3({\mathbf}{x})=\psi(x_3)e^{{\mathrm}{i}{\mathbf}{x}'\cdot\xi},\;\; {{\tilde{p}}}({\mathbf}{x})=\pi(x_3)e^{{\mathrm}{i}{\mathbf}{x}'\cdot\xi}.$$Then, in view of –, we see that $\varphi$, $\theta$, $\psi$ and $\lambda$ satisfy the following system of ODEs: $$\label{0205} \left\{ \begin{array}{ll} \lambda^2\bar{\rho} \varphi-\lambda\xi_1\pi+\lambda\mu (\|\xi\|^2\varphi-\varphi'')=M^2B_1,\\[1mm] \lambda^2\bar{\rho} \theta-\lambda\xi_2\pi+\lambda\mu (\|\xi\|^2\theta-\theta'')=M^2B_2,\\[1mm] \lambda^2\bar{\rho} \psi+\lambda\pi'+\lambda\mu (\|\xi\|^2\psi-\psi'')-g\bar{\rho}'\psi=M^2B_3,\\[1mm] \xi_1\varphi+\xi_2\theta+\psi'=0 \end{array} \right.$$ with $$\begin{aligned} \label{0206} \varphi(-\infty)=\theta(-\infty)=\psi(-\infty)=\varphi(+\infty)=\theta(+\infty)=\psi(+\infty)=0,\end{aligned}$$ where $$\left( \begin{array}{c} B_1\\[1mm] B_2\\[1mm] B_3 \end{array}\right)=\left\{\begin{array}{ll} \left( \begin{array}{c} 0 \\[1mm] (\xi_1\xi_2\varphi-\xi_1^2\theta)\\[1mm] -(\|\xi_1\|^2\psi+\xi_1\varphi') \end{array} \right) & \mbox{ for }\bar{{\mathbf}{M}}=M{\mathbf}{e}_1, \\[10mm] \left( \begin{array}{c} (\varphi''+\xi_1\psi')\\[1mm] (\theta''+\xi_2\psi')\\[1mm] 0 \end{array} \right)&\mbox{ for }\bar{{\mathbf}{M}}=M{\mathbf}{e}_3.\end{array}\right.$$ Eliminating $\pi$ from the third equation in , we obtain the following ODE for $\psi$ $$\label{0207}\begin{aligned} &-\lambda^2[\bar{\rho}\|\xi\|^2\psi-(\bar{\rho}\psi')']\\ =&\lambda\mu (\|\xi\|^4\psi-2\|\xi\|^2\psi''+\psi'''') -g\|\xi\|^2\bar{\rho}'\psi + M^2\left\{ \begin{array}{ll} \xi_1^2(\|\xi\|^2\psi-\psi''), & \hbox{ for }\bar{{\mathbf}{M}}=M{\mathbf}{e}_1, \\[1mm] (\psi''''-\|\xi\|^2\psi''), & \hbox{ for }\bar{{\mathbf}{M}}=M {\mathbf}{e}_3, \end{array} \right. \end{aligned}$$ with $$\begin{aligned} \label{0208} &&\psi(-\infty)=\psi'(-\infty)=\psi(+\infty)=\psi'(+\infty)=0.\end{aligned}$$ Next we use the modified variational method to construct a solution of , . This idea can be found in Guo and Tice’s paper on compressible viscous stratified flows [@GYTI2], and has been adapted by other authors to study the instability for other fluid models [@JFJSWWWO; @JJTIIA; @WYC; @DRJFJSRS]. To this end, we now fix a non-zero vector $\xi\in \mathbb{R}^2$ and $s>0$. From and we get a family of the modified problems $$\label{0209}\begin{aligned} &-\lambda^2[\bar{\rho}\|\xi\|^2\psi-(\bar{\rho}\psi')']\\ & = s\mu (\|\xi\|^4\psi-2\|\xi\|^2\psi''+\psi'''')-g\|\xi\|^2\bar{\rho}'\psi+M^2\left\{ \begin{array}{ll} \xi_1^2(\|\xi\|^2\psi-\psi''), & \hbox{ for }\bar{{\mathbf}{M}}=M{\mathbf}{e}_1, \\[1mm] (\psi''''-\|\xi\|^2\psi''), & \hbox{ for }\bar{{\mathbf}{M}}=M{\mathbf}{e}_3, \end{array} \right. \end{aligned}$$ coupled with the condition . We define the energy functional of by $$\label{0210}E(\psi)=\|\xi\|^2E_0(\psi)+s{E}_1(\psi)$$ with an associated admissible set $$\label{0211} \mathcal{A}=\left\{\psi\in H^2(\mathbb{R})~\bigg\|~J(\psi):=\int_{\mathbb{R}}\bar{\rho}(\|\xi\|^2\|\psi\|^2+\|\psi'\|^2)\mathrm{d}x_3=1\right\},$$ where $$\begin{aligned} \label{0212} &&E_0(\psi)=\left\{ \begin{array}{ll} \int_{\mathbb{R}}M^2\xi_1^2\left(\|\psi\|^2 +\frac{\|\psi'\|^2}{\|\xi\|^2}\right) -g\bar{\rho}'\psi^2{\mathrm}{d}x_3, & \hbox{ for }\bar{{\mathbf}{M}}=M {\mathbf}{e}_1, \\[3mm] \int_{\mathbb{R}}M^2\left(\|\psi'\|^2+\frac{\|\psi''\|^2}{\|\xi\|^2}\right) -g\bar{\rho}'\psi^2{\mathrm}{d}x_3, & \hbox{ for }\bar{{\mathbf}{M}}=M {\mathbf}{e}_3, \end{array} \right.\\[2mm] &&\label{n03010212} E_1(\psi)= \mu \int_{\mathbb{R}}(4\|\xi\|^2\|\psi'\|^2 +\|\|\xi\|^2\psi+\psi''\|^2){\mathrm}{d}x_3.\end{aligned}$$ Thus we can find a $-\lambda^2$ (depending on $\xi$) by minimizing $$\label{0214} -\lambda^2(\xi)=\alpha(\xi):=\inf_{\psi\in \mathcal{A}}E(\psi).$$ In order to emphasize the dependence on $s\in (0,\infty)$ we will sometimes write $$E(\psi,s):=E(\psi)\mbox{ and } \alpha(s):=\inf_{\psi\in \mathcal{A}}E(\psi,s)<+\infty.$$ Before constructing the growth solutions, we shall introduce some preliminary results, which will be used in Subsections \[sec:0202\]–\[sec:0204\]. Let the critical frequency function $S(\xi)$ for the horizontal case and the critical frequency constant $\|\xi\|_{\mathrm{vc}}^M$ for the vertical case be given by the following variational forms $$\label{0217} S(\xi):=\sqrt{\sup_{\psi\in H^1({\mathbb{R}}),\,\psi\not\equiv 0} \frac{g\|\xi\|^2\int_{\mathbb{R}}\bar{\rho}'\psi^2{\mathrm}{d}x_3/(M\xi_1)^2- \int_{{\mathbb{R}}}\|\psi'\|^2\mathrm{d}x_3}{\int_{\mathbb{R}}\|\psi\|^2\mathrm{d}x_3}} >0$$ for $0<\|M\xi_1\|/\|\xi\|<M_\mathrm{c}$, and $$\begin{aligned}\label{0216} \|\xi\|_{\mathrm{vc}}^M:=\sqrt{\inf_{\psi\in \mathcal{M}_{\mathrm{vc}} }\frac{M^2\int_{\mathbb{R}}\|\psi''\|^2{\mathrm}{d}x_3}{ \int_{\mathbb{R}}(g\bar{\rho}'\psi^2-M^2\|\psi'\|^2){\mathrm}{d}y}}\quad\mbox{ for }\|M\|\in (0,M_\mathrm{c}), \end{aligned}$$ respectively, where $M_\mathrm{c}$ is given by , and $$\mathcal{M}_{\mathrm{vc}}:=\Big\{\psi\in H^2(\mathbb{R})~\Bigg\|~\int_{\mathbb{R}} (g\bar{\rho}'\psi^2-M^2\|\psi'\|^2){\mathrm}{d}x_3>0\Big\}.$$ We remark that it depends on the choice of $\bar{\rho}$ whether $M_\mathrm{c}$ becomes infinite or finite. More precisely, we have the following conclusions: \[pro:n0201\] Assume that $\bar{\rho}'\in C_0^0(\mathbb{R})$, 1. if $\int_{\mathbb{R}}\bar{\rho}'{\mathrm}{d}x>0$, then $M_\mathrm{c}$ is infinite, 2. if $\int_{\mathbb{R}}\bar{\rho}'{\mathrm}{d}x<0$, then $M_\mathrm{c}$ is finite. \(1) We first show the first assertion. Let ${\mathrm}{sup}\bar{\rho}'\subset (-l,l)$ with $l>0$, and $$\psi_n(x):=\left\{ \begin{array}{ll} 1,& x\in [-l,l],\\ 1+(x+l)/n, & x\in (-l-n,-l),\\ 1-(x-l)/{n}, & x\in (l,l+n),\\ 0& x\in\!\!\!\!\!/~ (-l-n,l+n). \end{array} \right.$$ Then $\psi_n(x)\in H^1(\mathbb{R})$ and $$\begin{aligned} \frac{\int_{\mathbb{R}}g\bar{\rho}'\|\psi_n\|^2{\mathrm}{d}y}{\int_{\mathbb{R}}\|\psi'_n\|^2{\mathrm}{d}y} \geq g n\int_{\mathbb{R}}\bar{\rho}'{\mathrm}{d}x \rightarrow \infty\mbox{ as }n\rightarrow\infty, \end{aligned}$$ which implies $M_\mathrm{c}=\infty$. We mention that, in such a case, we can infer that the critical frequency constant $\|\xi\|_{\mathrm{vc}}^M$ defined by is equal to zero. \(2) We turn to show the second assertion by contradiction. Assume that $M_\mathrm{c}$ is infinite, i.e., there exists a sequence of functions $\{\psi_n\}_{n=1}^\infty$ such that $$\psi_n\in H^1(\mathbb{R}), \mbox{ and }0<\frac{\int_{\mathbb{R}}g\bar{\rho}'\|\psi_n\|^2{\mathrm}{d}x} {\int_{\mathbb{R}}\|\psi'_n\|^2{\mathrm}{d}x} \rightarrow \infty\mbox{ as }n\rightarrow \infty.$$ Denote $$\varphi_n = \frac{\psi_n} {\sqrt{\int_{\mathbb{R}}g\bar{\rho}'\|\psi_n\|^2{\mathrm}{d}x}}\in H^1(\mathbb{R}),$$ then $$0<\frac{1} {\int_{\mathbb{R}}\|\varphi'_n\|^2{\mathrm}{d}x} \rightarrow \infty\mbox{ as }n\rightarrow \infty,$$ which yields that $${\int_{\mathbb{R}}\|\varphi'_n\|^2{\mathrm}{d}x} \rightarrow 0\mbox{ as }n\rightarrow \infty.$$ Now, using Newton-Leibniz’s formula, we immediately get that, for any given $\varepsilon\in (0,1)$, there exists a $N_\varepsilon>0$ (may depend on $\varepsilon$), such that, for any $n>N_\varepsilon$ and for any $x\in (-l,l)$, $$\label{A010216} \|\varphi_n(x)-\varphi_n(0)\|=\left\|\int_{0}^x \varphi'_n(y){\mathrm}{d}y\right\|\leq \left(\int_{0}^x \|\varphi'_n(y)\|^2{\mathrm}{d}y\right)^{\frac{1}{2}}x^{\frac{1}{2}}\leq \sqrt{l}\varepsilon.$$ Thus, recalling the conditions $\int_{\mathbb{R}}\bar{\rho}'{\mathrm}{d}x< 0$ and $\bar{\rho}\in C_0^0(\mathbb{R})$, and using the following relation $$\label{A02} \begin{aligned} 1=&\int_{\mathbb{R}}g\bar{\rho}'\|\varphi_n\|^2{\mathrm}{d}x =\int_{\mathbb{R}}g\bar{\rho}'(\varphi_n- \varphi_n(0)+\varphi_n(0))^2 {\mathrm}{d}x\\ =&\varphi_n^2(0)\int_{\mathbb{R}}g\bar{\rho}'{\mathrm}{d}x +2\varphi_n(0)\int_{\mathbb{R}}g\bar{\rho}' (\varphi_n(x)-\varphi_n(0)){\mathrm}{d}x\\ &+ \int_{\mathbb{R}}g\bar{\rho}'(\varphi_n- \varphi_n(0))^2{\mathrm}{d}x:=R(n), \end{aligned}$$ we infer that there exists a constant $c$, independent of $n$ and $\varepsilon$, such that $$\label{A030217}\|\varphi_n(0)\|\leq c\mbox{ for any }n>N.$$ In fact, if this is not true, then there exists a subsequence, denoted by $\{\varphi_{n_m}(0)\}_{m=1}^\infty$, satisfying $\|\varphi_{n_m}(0)\|\geq m$, which implies that $1=R(n_m)\rightarrow\infty$. This is a contradiction. Finally, making use of – and $\int_{\mathbb{R}}\bar{\rho}'{\mathrm}{d}x< 0$, we have that for any $n>N_\varepsilon$, $$R(n) \leq (2l^{\frac{3}{2}}c + l^2)g\\|\bar{\rho}'\\|_{L^\infty}\varepsilon.$$ Consequently, letting $\varepsilon\rightarrow 0$, we immediately see that $R(n)$ can be sufficiently small for some $n$, and this contradicts with $R(n)\equiv 1$. Hence the second assertion holds. $\Box$ Now we define a solvable domain for a growing solution $$\label{0215} \mathbb{A}^{{\mathrm}{g}}=\left\{ \begin{array}{l} \big\{\xi\in \mathbb{R}^2~\|~ \|\xi_1M\|/\|\xi\|\in (0,M_{\mathrm}{c}),\ \|\xi\|<S(\xi)\big\}\cup \{\xi_1=0\} \setminus\{{\mathbf}{0}\}\ \hbox{ for }\bar{{\mathbf}{M}}=M{\mathbf}{e}_1, \\[2mm] \big\{\xi\in \mathbb{R}^2~\|~ \|\xi\|_{\mathrm{vc}}^M<\|\xi\|\big\} \hbox{ for }\bar{{\mathbf}{M}}={M}{\mathbf}{e}_3 \mbox{ with }\|M\|\in (0,M_\mathrm{c}). \end{array} \right.$$ By virtue of the definition of $M_\mathrm{c}$, it is easy to verify that the above two definitions , make sense. In addition, it is easy to see that $\|\xi\|<S(\xi)$ in for any $\|\xi\|>0$ with sufficiently small $\xi_1$ (the smallness of $\xi_1$ depends on $\|\xi\|$). Hence, the set $ \big\{\xi\in \mathbb{R}^2~\|~\|\xi_1M\|/\|\xi\|\in (0,M_{\mathrm}{c}),\ \|\xi\|<S(\xi)\big\}$ is not empty. Moreover, $\|\xi\|_{\mathrm{vc}}^M$ and $S(\xi )$ are finite. Next we introduce some properties concerning with the critical frequency and the solvable domain $\mathbb{A}^{{\mathrm}{g}}$. The supremum in is achieved for each $\xi$ and $M$ with $\xi_1 M\neq 0$. Moreover, $S(\xi)$ is continuous on $\mathbb{D}: =\{\xi\in\mathbb{R}^2~\|~0<\|M\xi_1\|/\|\xi\|<M_{\mathrm}{c}\}$ for any given $M\neq 0$, and $$\label{behavior} S( \xi)\rightarrow +\infty\;\;\mbox{ as }\;\frac{\xi_1}{\|\xi\|}\to 0\;\mbox{ and }\;\xi\in \mathbb{D}.$$ We rewrite as $$\begin{aligned} {S^2( \xi)}=\sup_{\psi\in \mathcal{A}_{L^2}}Q(\psi), \end{aligned}$$ where $$\label{0219} Q(\psi):=\frac{g\|\xi\|^2}{(M\xi_1)^2}\int_{\mathbb{R}}\bar{\rho}'\psi^2{\mathrm}{d}x_3- \int_{{\mathbb{R}}}\|\psi'\|^2\mathrm{d}x_3,\quad \mathcal{A}_{L^2}:=\left\{ H^1(\mathbb{R})~\bigg\|~ \\|\psi\\|^2_{L^2(\mathbb{R})}=1\right\}.$$ Then, it is easy to see that ${S}^2(\xi )>0$, since $0<\|M\xi_1\|/\|\xi\|<M_\mathrm{c}$. Let $\psi_n\in {\mathcal{A}}_{L^2}$ be a minimizing sequence of $S^2(\xi)$, i.e., $\limsup_{n\to\infty} Q(\psi_n)=\sup_{\psi\in\mathcal{A}_{L^2}}Q(\psi)$, we have from that $\psi_n$ is uniformly bounded in $H^1(\mathbb{R})$ on $n$. Hence there exists a $\psi_0\in H^1(\mathbb{R})$, such that $\psi_n\to\psi_0$ weakly in $H^1(\mathbb{R})$ and strongly in $L^2_{\mathrm{loc}}(\mathbb{R})$. Hence, $$\begin{aligned} 0<\sup_{\psi\in\mathcal{A}_{L^2}}Q(\psi)=\limsup_{n\rightarrow \infty} Q(\psi_n)\leq Q(\psi_0)\mbox{ and }0<\\|\psi_0 \\|_{L^2(\mathbb{R})}\leq \liminf_{n\rightarrow \infty} \\|\psi_n\\|_{L^2(\mathbb{R})}=1. \end{aligned}$$ We proceed to verify that $\\|\psi_0\\|_{L^2(\mathbb{R})}=1$. Suppose by contradiction that $\\|\psi_0\\|_{L^2(\mathbb{R})}<1$, then we may scale up $\psi_0$ by $\alpha>1$ so that $\alpha \psi_0\in{\mathcal{A}}_{L^2}$. From this we deduce that $$\begin{aligned} \sup\limits_{\psi\in{\mathcal{A}}_{L^2}}Q(\psi) \geq Q(\alpha \psi_0)=\alpha^2Q(\psi_0)\geq \alpha^2\sup\limits_{\psi\in{\mathcal{A}}_{L^2}}Q(\psi)>\sup\limits_{\psi\in{\mathcal{A}}_{L^2}}Q(\psi),\end{aligned}$$ which is a contradiction. Hence $\\|\psi\\|_{L^2(\mathbb{R})}=1$, which shows that $Q(\psi)$ achieves its infinimum on $\mathcal{A}_{L^2}$. Next, we prove the continuity of ${S}^2(\xi_0)$ for each given $\xi_0$ in $\mathbb{D}$. Letting $\xi\in \mathbb{D}\to\xi_0$, we have ${\|\xi\|^2}/{\xi_{1}^2}\to {\|\xi_0\|^2}/{\xi_{01}^2}$, where $\xi_{01}$ represents the first component of $\xi_0$. Without loss of generality, we assume that ${\|\xi\|^2}/{\xi_{1}^2}$ and ${\|\xi_0\|^2}/{\xi_{01}^2}$ belong to a finite interval $(a,b)$ with $a>0$. Denote $\delta:={\|\xi\|^2}/{\xi_{1}^2}-{\|\xi_0\|^2}/{\xi_{01}^2}$, then $\delta\rightarrow 0$ as $\xi\rightarrow \xi_0$. On the other hand, $Q(\psi)$ achieves its infinimum on $\mathcal{A}_{L^2}$, i.e. for any $\xi\in \mathbb{D}$, there is $\psi_{\xi}\in \mathcal{A}_{L^2}$, such that $$\begin{aligned}\label{0221} S^2(\xi):=\frac{g\|\xi\|^2}{(M\xi_1)^2}\int_{\mathbb{R}}\bar{\rho}'\psi^2_\xi{\mathrm}{d}x_3- \int_{{\mathbb{R}}}\|\psi'_\xi\|^2\mathrm{d}x_3. \end{aligned}$$ Substituting ${\|\xi_0\|^2}/{\xi_{01}^2}={\|\xi\|^2}/{\xi_{1}^2}-\delta$ into , one has $$\begin{aligned}\label{0222} {S}^2(\xi)=&\frac{g\|\xi_0\|^2}{(M\xi_{01})^2}\int_{\mathbb{R}}\bar{\rho}'\psi^2_{\xi}{\mathrm}{d}x_3- \int_{{\mathbb{R}}}\|\psi'_{\xi}\|^2\mathrm{d}x_3+ \frac{g\delta}{M^2} \int_{\mathbb{R}}\bar{\rho}'\psi^2_{\xi}{\mathrm}{d}x_3\\ \leq &{S}(\xi_0)+ \frac{g\delta}{M^2}\int_{\mathbb{R}}\bar{\rho}'\psi^2_{\xi}{\mathrm}{d}x_3. \end{aligned}$$ Similarly to , we obtain $$\begin{aligned} {S}^2(\xi_0)=&\frac{g\|\xi\|^2}{(M\xi_{ 1})^2}\int_{\mathbb{R}}\bar{\rho}'\psi^2_{\xi_0}{\mathrm}{d}x_3- \int_{{\mathbb{R}}}\|\psi'_{\xi_0}\|^2\mathrm{d}x_3- \frac{g\delta}{M^2} \int_{\mathbb{R}}\bar{\rho}'\psi^2_{\xi_0}{\mathrm}{d}x_3\\ \leq &{S}(\xi)- \frac{g\delta}{M^2}\int_{\mathbb{R}}\bar{\rho}'\psi^2_{\xi_0}{\mathrm}{d}x_3, \end{aligned}$$ which, together with , yields that $$\begin{aligned} \frac{g\delta}{M^2} \int_{\mathbb{R}}\bar{\rho}'\psi^2_{\xi_0}{\mathrm}{d}x_3 \leq{S}^2(\xi)-{S}^2(\xi_0) \leq \frac{g\delta}{M^2}\int_{\mathbb{R}}\bar{\rho}'\psi^2_{\xi}{\mathrm}{d}x_3. \end{aligned}$$ Hence, $$\begin{aligned} \|{S}^2(\xi)-{S}^2(\xi_0)\|\leq \frac{g\delta}{M^2}\\|\bar{\rho}'\\|_{L^\infty(\mathbb{R})} \to 0\mbox{ as }\delta\rightarrow 0, \end{aligned}$$ and ${S}^2(\xi)$ is continuous at each given point $\xi_0\in \mathbb{D}$. Finally, obviously holds by the definition . The completes the proof.$\Box$ \[pro:0202\] Let $\mathbb{A}^{{\mathrm}{g}}$ be defined by , then 1. the set $\mathbb{A}^{{\mathrm}{g}}$ is symmetric on $x$-axis and $y$-axis in $\mathbb{R}^2$, respectively; 2. there exist countably infinite lattice points of $(L^{-1}\mathbb{Z})^2$ belongs to $\mathbb{A}^{{\mathrm}{g}}$; 3. the set $\mathbb{A}^{{\mathrm}{g}}$ is a nonempty open set in $\mathbb{R}^2$. The first two assertions obviously hold by virtue of the definition of $\mathbb{A}^{{\mathrm}{g}}$. It suffices to show the last assertion. Here we only give the proof of the horizontal case for the reader’s convenience. Let $0\neq \xi_0\in \mathbb{A}^{{\mathrm}{g}}$, then $\xi_0\in \{\xi_1=0\}$ or $\xi_0\in \big\{\xi\in\mathbb{R}^2~\|~ \|\xi_1M\|/\|\xi\|\in (0,M_{\mathrm}{c}),\ \|\xi\|<S(\xi)\big\}$. For the first case, noting that $\|\xi_1M\|/\|\xi\|\in (0,M_{\mathrm}{c})$ and $\|\xi\|<S(\xi)$ hold for any $\|\xi\|>0$ with sufficiently small $\xi_1$ there exists a sufficiently small disk $B_{\xi_0}^\delta:=\{\xi\in\mathbb{R}^2~\|~\|\xi-\xi_0\|<\delta\} \subset \mathbb{A}^{{\mathrm}{g}}$. For the latter case, noting that $0<\|\xi_{01}M\|/\|\xi_0<M_{\mathrm}{c}$ and $\|\xi_0\|<S(\xi_0)$, we use the continuity of $\|\xi_1 M\|/\|\xi\|$, $\|\xi\|$ and $S(\xi)$ as $\xi\rightarrow \xi_0\neq 0$ to deduce that there also exists a sufficiently small disk $B_{\xi_0}^\alpha:=\{\xi\in\mathbb{R}^2~\|~\|\xi-\xi_0\|<\alpha\}\subset \big\{\xi\in \mathbb{R}^2~\|~ \|\xi_1M\|/\|\xi\|\in (0,M_{\mathrm}{c}),\ \|\xi\|<S(\xi)\big\}$. Summing up the previous discussions, we immediately conclude that the set $\mathbb{A}^{{\mathrm}{g}}$ is a nonempty open set in $\mathbb{R}^2$ by the definition of open sets. $\Box$ \[pro:0203\]Let $\xi\neq{\mathbf}{0}$. Then, - if $\xi\in \mathbb{A}^{{\mathrm}{g}}$, there exists a $$\psi_0\in \left\{ \begin{array}{ll} H^1(\mathbb{R})&\;\hbox{ for the horizontal case,} \\ \mathcal{M}_{\mathrm{vc}}&\;\hbox{ for the vertical case,} \end{array} \right.$$ such that $E_0(\psi_0)<0$; - else, $\|\xi\|^2E_0(\psi)$ $\geq 0$ for any $$\psi\in \left\{ \begin{array}{ll} H^1(\mathbb{R})&\;\hbox{ for the horizontal case; } \\[1mm] H^2(\mathbb{R})&\;\hbox{ for the vertical case.} \end{array} \right.$$ The above assertions in fact follow from the definitions , , , and , here we only give the proof of the horizontal case for the reader’s convenience. Let $\xi\in \mathbb{A}^{{\mathrm}{g}}$. If $\xi_1=0$, then obviously, there exists a $\psi_0$, such that $$E_0= \int_{\mathbb{R}}\Big[(M\xi_1)^2\Big(\|\psi_0\|^2+\frac{\|\psi'_0\|^2}{\|\xi\|^2}\Big) -g\bar{\rho}'\psi^2_0\Big]{\mathrm}{d}x_3= \int_{\mathbb{R}} -g\bar{\rho}'\psi^2_0{\mathrm}{d}x_3<0.$$ If $\xi$ satisfies $\|M\xi_1\|/\|\xi\|\in (0,M_{\mathrm}{c})$ and $\|\xi\|<S(\xi)$, then by virtue of the definition of , there also exists a $\psi_0$, such that $$E_0= \int_{\mathbb{R}}\Big[(M\xi_1)^2\Big(\|\psi_0\|^2+\frac{\|\psi'_0\|^2}{\|\xi\|^2}\Big) -g\bar{\rho}'\psi^2_0\Big] {\mathrm}{d}x_3<0.$$ Summing up the above discussions, we see that there exists a $\psi_0$, such that $E_0(\psi_0)<0$ for $\xi\in \mathbb{A}^{{\mathrm}{g}}$. Let $\xi\not\in\mathbb{A}^{{\mathrm}{g}}\cup\{{\mathbf}{0}\}$, then $\xi$ can be divided by two cases: (i) $\|M\xi_1\|/\|\xi\|\in (0, M_{\mathrm}{c})$ with $\|\xi\|\geq S(\xi)$, and (ii) $\|M\xi_1\|/\|\xi\|\geq M_{\mathrm}{c}$ if $M_{\mathrm}{c}<\infty$ (noting that this case will not appear if $M_{\mathrm}{c}=\infty$). For the first case, in view of the definition of , we find that $$\|\xi\|^2E_0(\psi)= \|\xi\|^2 \int_{\mathbb{R}}\Big[(M\xi_1)^2\Big(\|\psi\|^2+\frac{\|\psi'\|^2} {\|\xi\|^2}\Big)-g\bar{\rho}'\psi^2\Big]{\mathrm}{d}x_3\geq 0\;\;\mbox{ for any }\psi\in H^1(\mathbb{R}).$$ Finally, for the second case, by the definition of , we have $$\int_{\mathbb{R}}\Big(\frac{(M\xi_1)^2}{\|\xi\|^2}\|\psi'\|^2 - g\bar{\rho}'\psi^2\Big){\mathrm}{d}x_3\geq 0\;\;\mbox{ for any }\psi\in H^1(\mathbb{R}),$$ which yields $$\|\xi\|^2E_0= \|\xi\|^2 \int_{\mathbb{R}}\Big[(M\xi_1)^2\Big(\|\psi\|^2+\frac{\|\psi'\|^2}{\|\xi\|^2}\Big) -g\bar{\rho}'\psi^2\Big]{\mathrm}{d}x_3\geq 0\;\;\mbox{ for any }\psi \in H^1(\mathbb{R}).$$ Summarizing the above discussions, we see that $\|\xi\|^2E_0(\psi)$ $\geq 0$ for any $\psi\in H^1(\mathbb{R})$ if $\xi \in\!\!\! \! \! /~ \mathbb{A}^{{\mathrm}{g}}\cup\{{\mathbf}{0}\}$. This completes the proof for the horizontal case. $\Box$ Solutions to the variational problem {#sec:0202} ------------------------------------ In this seusection we show that a minimizer of exists for the case of $\inf_{\psi\in \mathcal{A}}E(\psi,s)<0$ which will be shown to be true for sufficiently small $s$ in Proposition \[pro:0205\] below, and that the corresponding Euler-Lagrange equations are equivalent to , . \[pro:0204\] For any fixed $s>0$ and $\xi$ with $\|\xi\|\neq 0$, we have 1. $\inf_{\psi\in\mathcal{A}}E(\psi,s)>-\infty$. 2. if there exists a $\bar{\psi}\in \mathcal{A}$, such that $E(\bar{\psi})<0$, then $E({\psi})$ achieves its infinimum on $\mathcal{A}$. 3. let $\tilde{\psi}$ be a minimizer and $-\lambda^2:=E(\tilde{\psi})$, then the pair $(\tilde{\psi}$, $\lambda^2)$ satisfies , . Moreover, $\tilde{\psi}\in H^\infty(\mathbb{R}):=\cap_{k=0}^\infty H^k(\mathbb{R})$. We only show the proposition for the horizontal case, the vertical case can be dealt with in the same manner. \(1) Noticing that for any $\psi\in \mathcal{A}$, $$\label{0226}E(\psi)\geq -{g\|\xi\|^2}\int_{\mathbb{R}} \bar{\rho}'\psi^2\mathrm{d}x_3\geq -{g}\left\\|\frac{\bar{\rho}'}{\bar{\rho}}\right\\|_{L^\infty(\mathbb{R})}\int_{\mathbb{R}} \bar{\rho}\|\xi\|^2\psi^2\mathrm{d}x_3\geq -{g}\left\\|\frac{\bar{\rho}'}{\bar{\rho}}\right\\|_{L^\infty(\mathbb{R})},$$ we see that $E$ is bounded from below on $\mathcal{A}$ by virtue of . This proves (1). \(2) We proceed to show (2). Let $\psi_n\in\mathcal{A}$ be a minimizing sequence, then $E(\psi _n)$ is bounded. This together with and implies that $\psi_n$ is bounded in $H^2(\mathbb{R})$. So, there exists a $\psi_0\in H^2(\mathbb{R})$, such that $\psi_n\rightarrow \psi_0$ weakly in $H^2(\mathbb{R})$ and strongly in $H^1_{\mathrm{loc}}(\mathbb{R})$. Moreover, by the lower semi-continuity, and the assumption that $E(\bar{\psi})<0$ for some $\bar{\psi}\in \mathcal{A}$, we deduce that $$E(\psi_0)\leq \liminf_{n\rightarrow \infty}E(\psi_n)=\inf_{\psi\in \mathcal{A}}E(\psi)<0,\quad\mbox{and }\; 0<J(\psi_0)\leq 1.$$ Suppose by contradiction that $J(\psi_0)<1$. By the homogeneity of $J$ we may find an $\alpha>1$ so that $J(\alpha \psi_0)=1$, i.e., we may scale up $\psi_0$ so that $\alpha \psi_0\in{\mathcal{A}}$. From this we deduce that $$\begin{aligned} E(\alpha\psi_0)=\alpha^2E(\psi_0)\leq\alpha^2\inf\limits_{\mathcal{A}} E<\inf\limits_{\mathcal{A}} E<0,\end{aligned}$$ which is a contradiction since $\alpha\psi_0\in{\mathcal{A}}$. Hence $J(\psi_0)=1$ and $\psi_0\in{ \mathcal{A}}$. This shows that $E(\psi)$ achieves its infinimum on $\mathcal{A}$. \(3) Finally we prove (3). Notice that since $E(\psi)$ and $J(\psi)$ are homogeneous of degree $2$, is equivalent to $$\label{0227}\alpha(s)=\inf_{\psi\in H^2(\mathbb{R}),\psi \equiv\!\!\!\!/~0}\frac{E(\psi)}{J(\psi)}.$$ For any $\tau\in \mathbb{R}$ and $\psi\in H^2(\mathbb{R})$ we take $\psi(\tau):=\tilde{\psi}+\tau\psi$, then implies $$E(\psi(\tau))+\lambda^2J(\psi(\tau))\geq 0.$$ If we set $I(\tau)=E(\psi(\tau))+\lambda^2J(\psi(\tau))$, then $I(\tau)\geq 0$ for all $\tau\in \mathbb{R}$ and $I(0)=0$. This implies $I'(0)=0$. By virtue of and , a direct computation leads to $$\label{0228}\begin{aligned} & s\mu \int_{\mathbb{R}}\Big(4\|\xi\|^2\tilde{\psi}'\psi'+(\|\xi\|^2\tilde{\psi}+\tilde{\psi}'') (\|\xi\|^2\psi +\psi'') +(M\xi_1)^2(\|\xi\|^2\tilde{\psi}\psi+\tilde{\psi}'\psi')\Big)\mathrm{d}x_3\\ & = g\|\xi\|^2\int_\mathbb{R}\bar{\rho}'\tilde{\psi}\psi\mathrm{d}x_3 -\lambda^2\int_{\mathbb{R}}\bar{\rho}(\|\xi\|^2\tilde{\psi}\psi+\tilde{\psi}'\psi')\mathrm{d}x_3, \end{aligned}$$ where we have used the upper boundedness of $\bar{\rho}$. By further assuming that $\psi$ is compactly supported in $\mathbb{R}$, we find that $\tilde{\psi}$ satisfies the equation in the weak sense on $\mathbb{R}$ for the horizontal case. In order to improve the regularity of $\tilde{\psi}$, we rewrite as $$\label{0229}\begin{aligned} \int_\mathbb{R} \tilde{\psi}''\psi''\mathrm{d}x_3 =&\frac{1}{s\mu}\int_{\mathbb{R}}\left(g\|\xi\|^2\bar{\rho}'\tilde{\psi} -\lambda^2(\|\xi\|^2\bar{\rho}\tilde{\psi}-(\bar{\rho}\tilde{\psi}')')\right.\\ & \left. +s\mu(2\|\xi\|^2\tilde{\psi}'' -\|\xi\|^4\tilde{\psi})+(M\xi_1)^2(\tilde{\psi}''-\|\xi\|^2\tilde{\psi} )\right)\psi\mathrm{d}x_3\\ :=&\int_{\mathbb{R}}f \psi\mathrm{d}x_3. \end{aligned}$$ For any $n\geq 1$, let $\psi_{1,n}$, $\psi_2\in C_0^\infty(\mathbb{R})$ satisfy $\psi_{1,n}(x_3)\equiv 1$ for $\|x_3\|\leq n$. If one takes $\psi=\psi_{1,n}\int_{-\infty}^{x_3}\psi_2\mathrm{d}y$ in , then one has $$\begin{aligned}\int_\mathbb{R} (\psi_{1,n}\tilde{\psi}'')\psi_2'\mathrm{d}x_3=& \int_{\mathbb{R}}\left(f\psi_{1,n}\int_{-\infty}^{x_3}\psi_2\mathrm{d}y -\psi_{1,n}''\psi''\int_{-\infty}^{x_3}\psi_2\mathrm{d}y-2\psi'_{1,n}\tilde{\psi}'' \psi_2\right)\mathrm{d}x_3 \\ = & \int_{\mathbb{R}}\left(\int_{x_3}^{+\infty}(f\psi_{1,n}-\psi_{1,n}''\tilde{\psi}'') \mathrm{d}y-2\psi'_{1,n}\tilde{\psi}'' \right)\psi_2\mathrm{d}x_3, \end{aligned}$$ which, recalling $\tilde{\psi}\in H^2(\mathbb{R})$, implies $\tilde{\psi}''\in H^1_{\mathrm{loc}}(\mathbb{R})$ and $$\begin{aligned} \tilde{\psi}'''=(\psi_{1,n}\psi'')'=\int_{x_3}^{+\infty}(f\psi_{1,n}-\psi_{1,n}''\tilde{\psi}'')\mathrm{d}y\quad \mbox{ for any }x_3 \mbox{ with }\|x_3\|\leq n. \end{aligned}$$ Integrating by parts, we can rewrite as $$\begin{aligned} -\int_\mathbb{R} \tilde{\psi}'''\psi'\mathrm{d}x_3 =&\frac{1}{s\mu }\int_{\mathbb{R}}\left(g\|\xi\|^2\bar{\rho}'\tilde{\psi} - \lambda^2(\|\xi\|^2\bar{\rho}\tilde{\psi}-(\bar{\rho}\tilde{\psi}')')\right.\\ &\left.\qquad \qquad +s\mu(2\|\xi\|^2\tilde{\psi}'' -\|\xi\|^4\tilde{\psi})+(M\xi_1)^2(\tilde{\psi}''-\|\xi\|^2\tilde{\psi} ) \right)\psi\mathrm{d}x_3, \end{aligned}$$ which, keeping in mind that $\tilde{\psi}\in H^2(\mathbb{R})$, yields $\tilde{\psi}''''\in L^2(\mathbb{R})$. Hence $\tilde{\psi}\in H^4_{\mathrm{loc}}(\mathbb{R})\cap C^{3,1/2}_{\mathrm{loc}}(\mathbb{R})$, and $\tilde{\psi}^{'}(\infty)=\tilde{\psi}^{''}(\infty)=\tilde{\psi}^{'''}(\infty)=0$. Using these facts, Hölder’s inequality, and integration by parts, we conclude that $$\begin{aligned} \\|\tilde{\psi}'''\\|_{L^2(\mathbb{R})}^2=\int_{\mathbb{R}}\|\tilde{\psi}'''\|^2\mathrm{d}x_3 =-\int_{\mathbb{R}}\tilde{\psi}''\tilde{\psi}''''\mathrm{d}x_3\leq\\|\tilde{\psi}''\\|_{L^2(\mathbb{R})}\\| \tilde{\psi}''''\\|_{L^2(\mathbb{R})}, \end{aligned}$$ i.e., $\tilde{\psi}'''\in L^2(\mathbb{R})$. Consequently, $\tilde{\psi}\in H^4(\mathbb{R})$ solves , . This immediately gives $\tilde{\psi}\in H^\infty(\mathbb{R})$. $\Box$ Next, we want to show that there is a fixed point such that $\lambda=s$. To this end, we first give some properties of $\alpha(s)$ as a function of $s> 0$. \[pro:0205\] The function $\alpha(s)$ defined on $(0,\infty)$ enjoys the following properties: 1. $\alpha(s)\in C_{\mathrm{loc}}^{0,1}(0,\infty)$ is nondecreasing. 2. for any $\xi\in \mathbb{A}^{{\mathrm}{g}}$, there exist constants $c_1$, $c_2>0$ depending on $g$, $M$, $\bar{\rho}$, $\mu$, and $\xi$, such that $$\label{0231}\alpha(s)\leq -c_1+sc_2 .$$ We still give the proof for the horizontal case only, and the vertical case can be dealt with in the same way. \(1) Let $\{\psi^n_{s_2}\}\subset\mathcal{A}$ be a minimizing sequence of $\inf_{\psi\in \mathcal{A}}E(\psi,s_2)$. From and it follows that $$\alpha(s_1)\leq \limsup_{n\rightarrow\infty}E(\psi_{s_2}^n,s_1) \leq \limsup_{n\rightarrow\infty}E(\psi_{s_2}^n,s_2)=\alpha(s_2)\; \mbox{ for any }0<s_1<s_2<\infty.$$ Hence $\alpha(s)$ is nondecreasing on $(0,\infty)$. Next we shall use this fact to show the continuity of $\alpha(s)$. Let $I:=[a,b]\subset\mathbb{R}^+$ be a bounded interval. In view of and the monotonicity of $\alpha(s)$, we know that $$\label{02321} \|\alpha(s)\|\leq \max\left\{\|\alpha(b)\|,{g} \left\\| {\bar{\rho}'}/{\bar{\rho}} \right\\|_{L^\infty(\mathbb{R})}\right\}<\infty\;\;\mbox{ for any }s\in I.$$ On the other hand, for any $s\in I$, there exists a minimizing sequence $\{\psi^n_{s}\}\subset \mathcal{A}$ of $\inf_{\psi\in \mathcal{A}}E(\psi,s)$, such that $$\|\alpha(s)-E(\psi_{s}^n,s)\|<1.$$ Making use of – and , we infer that $$\begin{aligned} 0\leq E_1(\psi_s^n,s)=&\frac{E(\psi_s^n,s)}{s} + \frac{g\|\xi\|^2}{s}\int_{\mathbb{R}} \bar{\rho}'\psi^2\mathrm{d}x_3-\frac{(M\xi_1)^2}{s} \int_{\mathbb{R}}\left({\|\xi\|^2}\|\psi\|^2+ {\|\psi'\|^2}\right)\mathrm{d}x_3 \\ \leq & \frac{1+\max\{\|a(b)\|,{g}\left\\|{\bar{\rho}'}/{\bar{\rho}} \right\\|_{L^\infty(\mathbb{R})}\}}{a}+\frac{g}{a}\left\\|\frac{\bar{\rho}'}{\bar{\rho}} \right\\|_{L^\infty(\mathbb{R})}:=K. \end{aligned}$$ For $s_i\in I$ ($i=1,2$), we further find that $$\begin{aligned}\label{0235}\alpha(s_1)\leq \limsup_{n\rightarrow \infty}E(\varphi_{s_2}^n,s_1)\leq & \limsup_{n\rightarrow \infty}E(\psi_{s_2}^n,s_2)+\|s_1-s_2\|\limsup_{n\rightarrow \infty}E_1(\psi_{s_2}^n)\\ \leq & \alpha(s_2)+K\|s_1-s_2\|. \end{aligned}$$ Reversing the role of the indices $1$ and $2$ in the derivation of the inequality , we obtain the same boundedness with the indices switched. Therefore, we have $$\begin{aligned}\|\alpha(s_1)-\alpha(s_2)\|\leq K\|s_1-s_2\|, \end{aligned}$$ which yields $\alpha(s)\in C_{\mathrm{loc}}^{0,1}(0,\infty)$. \(2) Since $\xi\in \mathbb{A}^{{\mathrm}{g}}$, by virtue of Proposition \[pro:0203\], there exists a $\bar{\psi}\in H^1(\mathbb{R})$, such that $$\label{0236} E_0(\psi)= \int_{\mathbb{R}}(M\xi_1)^2\left(\|\bar{\psi}\|^2 +\frac{\|\bar{\psi}'\|^2}{\|\xi\|^2}\right)-g\bar{\rho}'\bar{\psi}^2{\mathrm}{d}x_3<0.$$ On the other hand, $H^2(\mathbb{R})$ is dense in $H^1(\mathbb{R})$, thus there is a function sequence $\bar{\psi}_n\in H^2(\mathbb{R})$, so that $$\label{0237} \bar{\psi}_n\rightarrow\bar{\psi}\;\;\mbox{ stronly in }H^1(\mathbb{R}).$$ Putting and together, we see that there is a subsequence $\bar{\psi}_{n_0}\in H^2(\mathbb{R})$, such that $\bar{\psi}_{n_0}\not\equiv 0$ and $$E_0(\bar{\psi}_{n_0})=\int_{\mathbb{R}}\Big[(M\xi_1)^2\Big( \|\bar{\psi}_{n_0}\|^2 +\frac{\|\bar{\psi}_{n_0}'\|^2}{\|\xi\|^2}\Big) -g\bar{\rho}'\bar{\psi}_{n_0}^2\Big]{\mathrm}{d}x_3<0.$$ Thus, we have $$\alpha(s)= \inf_{\psi\in \mathcal{A}}E(\psi)=\inf_{\psi\in H^2(\mathbb{R})}\frac{E(\psi)}{J(\psi)} \leq \frac{E(\bar{\psi}_{n_0})}{J(\bar{\psi}_{n_0})}=\|\xi\|^2\frac{E_0(\bar{\psi}_{n_0})}{J(\bar{\psi}_{n_0})} +s\frac{E_1(\bar{\psi}_{n_0})}{J(\bar{\psi}_{n_0})}:= -c_1+sc_2$$ for two positive constants $c_1: =c_1(g,M,\bar{\rho},\|\xi\|)$ and $c_2:=c_2(g,M,\mu,\bar{\rho},\|\xi\|)$. This completes the proof for the vertical case. $\Box$ Given $\xi\in \mathbb{A}^g$, by virtue of , there exists a $s_0>0$ depending on the quantities $g$, $M$, $\mu$, $\bar{\rho}$ and $\xi$, such that for any $s\in (0, s_0]$, $\alpha(s)<0$. Let $$\label{0240}\mathfrak{S}_{\xi}:=\sup\{s~\|~\alpha(\tau)<0\mbox{ for any }\tau\in (0,s)\},$$ then $\mathfrak{S}_{\xi}>0$. This allows us to define $\lambda(s)=\sqrt{-\alpha(s)}>0$ for any $s\in \mathcal{S}_{\xi}:=(0,\mathfrak{S}_{\xi})$. Therefore, as a result of Proposition \[pro:0204\], we have the following existence for the modified problem , . \[pro:0206\] For each $\xi\in \mathbb{A}^g$ and $s\in \mathcal{S}_{\xi}$ there is a solution $\psi=\psi(\xi,x_3)\not\equiv 0$ with $\lambda=\lambda(\xi,s)>0$ to the problem , . Moreover, $\psi\in H^\infty(\mathbb{R})\cap \mathcal{A}$. Now, we can use Proposition \[pro:0205\], and to find that $\lambda(s)\in C_{\mathrm{loc}}^{0,1}(\mathcal{S}_{\xi})$ is nonincreasing, $\lambda(s)\leq \sqrt{g}\\|\sqrt{{\bar{\rho}'}/{\bar{\rho}}}\\|_{L^\infty(\mathbb{R})}$, $\lim_{s\rightarrow 0}\lambda(s)>0$ and $\lim_{s\rightarrow \mathfrak{S}_{\xi}}\lambda(s)=0$ if $\mathfrak{S}_{\|\xi\|}<+\infty$. Hence, we can employ a fixed-point argument to find $s\in\mathcal{S}_{\xi}$ so that $s=\lambda(\xi, s)$, and thus obtain a solution to the original problem , . \[pro:0207\] Let $\xi\in \mathbb{A}^g$, then there exists a unique $s\in \mathcal{S}_{\xi}$, such that $\lambda(\xi,s)=\sqrt{-\alpha(s)}>0$ and $s=\lambda(\xi,s)$. We refer to [@GYTI2 Theorem 3.8] (or [@WYC Lemma 3.7]) for a proof.$\Box$ Moreover, in view of Propositions \[pro:0206\] and \[pro:0207\], we conclude the following existence for the problem –. \[thm:0201\] For each $\xi\in \mathbb{A}^g$, there exist $\psi=\psi(\xi,x_3)\equiv\!\!\!\!\!\!/\ 0$ and $\lambda(\xi)>0$ satisfying and . Moreover, $\psi\in H^\infty(\mathbb{R})\cap \mathcal{A}$. We end this subsection by giving additional properties of the solutions established in Theorem \[thm:0201\] in terms of $\lambda(\xi)$, which show that $\lambda$ is a bounded, continuous function of $\xi$. \[pro:0208\] The positive function $\lambda :\mathbb{A}^{\mathrm}{g}\rightarrow \mathbb{R}^+$ is continuous and satisfies $$\label{0241} \sup_{\xi\in \mathbb{A}^{{\mathrm}{g}}}\lambda(\xi)\leq \sqrt{g\left\\|{\bar{\rho}'}/{\bar{\rho}}\right \\|_{L^\infty(\mathbb{R})}}.$$ The boundedness of $\lambda$ in follows from . As for the proof of the continuity of $\lambda$, we still give the proof for the horizontal case only. First, let $\xi_0\in \mathcal{A}$ be arbitrary but fixed, and $\xi\rightarrow \xi_0$. Without loss of generality, assume $\xi\in \mathbb{A}^{\mathrm}{g}$, since $\mathbb{A}^{\mathrm}{g}$ is an open set by Proposition \[pro:0202\]. Then there exists an interval $[a,b]\subset \mathbb{R}^+$ so that $\|\xi\|$ and $\|\xi_0\|\in (a,b)$. Let $\delta=\|\xi\|^2-\|\xi_0\|^2$, then $\delta\rightarrow 0 $ as $\|\xi\|\rightarrow \|\xi_0\|$. \(i) We begin with the proof of the following conclusion: $$\label{0242} \lim_{\xi\rightarrow \xi_0} \alpha(\xi,s)=\alpha(\xi_0,s)\;\;\mbox{ for any }s\in \mathcal{S}_{\xi}.$$ By virtue of Proposition \[pro:0206\], for any $\xi\in \mathbb{A}^{{\mathrm}{g}}$, there exists a function $\psi_{\xi}\in \mathcal{A}$, such that $$\label{0243} \alpha(\xi)=\int_{\mathbb{R}}\Big[s\mu(4\|\xi\|^2\|\psi'_{\xi}\|^2 +\|\|\xi\|^2\psi_{\xi}+\psi''_{\xi}\|^2) +(M\xi_1)^2\left(\|\xi\|^2\|\psi_{\xi}\|^2 +{\|\psi'_{\xi}\|^2}\right) - g\|\xi\|^2\bar{\rho}'\psi^2_{\xi}\Big] \mathrm{d}x_3<0,$$ which, together with , yields $$\label{0244}\\|\psi_{\xi}\\|_{H^2(\mathbb{R})}\leq c_3,$$ where $c_3$ depends on $g$, $M$, $\mu$, $\bar{\rho}$, $a$, $b$ and $s$. To deal with the term involved with $\|\xi_1\|$, we denote $\delta_1:=\|\xi_1\|^2-\|\xi_{01}\|^2$. Substitution of $\|\xi\|^2=\|\xi_0\|^2+\delta$ and $\|\xi_1\|^2=\|\xi_{01}\|^2+\delta_1$ into results in $$\label{0245}\begin{aligned} \alpha(\xi)=&\int_{\mathbb{R}}\Big[ s\mu(4\|\xi_0\|^2\|\psi'_{\xi}\|^2 +\|\|\xi_0\|^2\psi_{\xi}+\psi''_{\xi}\|^2) +(M\xi_{01})^2\left(\|\xi_0\|^2\|\psi_{\xi}\|^2 +{\|\psi'_{\xi}\|^2}\right) - g\|\xi_0\|^2\bar{\rho}'\psi^2_{\xi}\Big] \mathrm{d}x_3\\ & +\delta f(\delta,\psi_{\xi}) +\delta_1 h(\psi_{\xi})\geq\alpha(\xi_0) +\delta f(\delta,\psi_{\xi}) +\delta_1 h(\psi_{\xi}),\end{aligned}$$ where $$f(\delta,\psi_{\xi})=\int_{\mathbb{R}}\Big[ s\mu(4\|\psi'_{\xi}\|^2 +2\delta \psi_{\xi}(\|\xi_0\|^2\psi_{\xi}+\psi''_{\xi})+\delta\psi^2_{\xi}+(M\xi_{01})^2\|\psi_{\xi}\|^2- g\bar{\rho}'\psi^2_{\xi}\Big] \mathrm{d}x_3$$ and $$h(\delta,\psi_{\xi})= \int_{\mathbb{R}}M^2\left((\|\xi_0\|^2+\delta)\|\psi_{\xi}\|^2+{\|\psi'_{\xi}\|^2}\right)\mathrm{d}x_3.$$ By Hölder’s inequality and , we can bound $$\label{0246}\begin{aligned}\| f(\delta,\psi_{\xi})\|+\|h(\delta,\psi_{\xi})\|\leq c_4\;\;\mbox{ for some constant }c_4. \end{aligned}$$ Similarly to and , we also have $$\label{0247}\begin{aligned}\alpha(\xi_0)\geq\alpha(\xi)-\delta f(-\delta,\psi_{\xi_0})-\delta_1 h(\psi_{\xi_0})\end{aligned}$$ [ and ]{}$$\label{0248}\begin{aligned}\|f(-\delta,\psi_{\xi_0})\|+\| h(-\delta,\psi_{\xi_0})\|\leq c_5.\end{aligned}$$ Combining with , we get $$\delta f(-\delta,\psi_{\xi_0}) +\delta_1 h(\psi_{\xi_0}) \geq \alpha(\xi)-\alpha(\xi_0)\geq \delta f(\delta,\psi_{\xi})+\delta_1 h(\psi_{\xi}),$$ which, together with and , implies . Hence, $$\label{0249nn} \lim_{\xi\rightarrow \xi_0} \lambda(\xi,s)=\lambda(\xi_0,s)\;\;\mbox{ for any }s\in \mathcal{S}_{\xi},$$ because of $\lambda(\xi,s)=\sqrt{-\alpha(\xi,s)}$. \(ii) In view of and Proposition \[pro:0207\], we see that for any $\varepsilon>0$, there exists a $\delta>0$ such that $\|\lambda(\xi,s_{\xi_0})- \lambda(\xi_0,s_{\xi_0})\|<\varepsilon$ and $s_{\xi_0}=\lambda({\xi_0},s_{\xi_0})=\sqrt{-\alpha({\xi_0,s_{\xi_0}})}$ for any $\xi\in B_{\xi_0}^\delta:=\{{\mathbf}{y}\in \mathbb{R}^2~\|~\|{\mathbf}{y}-\xi_0\|<\delta\}\subset\mathbb{A}^{\mathrm}{g}$. On the other hand, for each $\|\xi\|>0 $, $\lambda(s)$ is nonincreasing and continuous on $\mathcal{S}_{\xi}$, and there exists a unique $s_{\xi}\in \mathcal{S}_{\xi}$ satisfying $\lambda(\xi,s_{\xi})=s_{\xi}>0$ by Proposition \[pro:0207\]. Consequently, we immediately infer that $\|\lambda(\xi,s_{\xi})-\lambda(\xi_0,s_{\xi_0})\|<\varepsilon$ with $s_{\xi}=\lambda(\xi,s_{\xi})$. Therefore, $\lambda(\xi)$ is continuous. This completes the proof of the proposition. $\Box$ Construction of a solution to the system , ------------------------------------------- A solution to , gives rise to a solution of the ODEs , for the growing mode velocity $\mathbf{u}$ as well. \[thm:0302\] For each $\xi\in \mathbb{A}^{\mathrm}{g}$, there exists a solution $({\varphi},{\theta},{\psi},{\pi}):= ({\varphi},{\theta},{\psi},{\pi})(\xi,x_3)$ with $\lambda =\lambda(\xi)>0$ to , , and the solution belongs to $H^\infty(\mathbb{R})$. Moreover $\psi\in \mathcal{A}$. We still give the proof for the horizonal case only. First, in view of Theorem \[thm:0201\], we have a solution $(\psi ,\lambda ):=(\psi (\xi,x_3),\lambda (\xi))$ satisfying , . Moreover, $\lambda >0$ and $\psi\in\mathcal{A}\cap H^\infty(\mathbb{R}^2)$. Then, multiplying $_1$ and $_2$ by $\xi_1$ and $\xi_2$ respectively, adding the resulting equations, and utilizing $_4$, we find that $\pi$ can be expressed by $\psi$, i.e., $$\begin{aligned} &&\pi =\pi(\xi,x_3)\nonumber\\ &&\quad =[-{\lambda\mu (\xi_1\varphi+\xi_2\theta)''+(\lambda^2\bar{\rho}+\lambda\mu \|\xi\|^2+M^2\xi_1^2)(\xi_1\varphi+\xi_2\theta)}-M^2\xi_1\|\xi\|^2\varphi]/({\lambda\|\xi\|^{2}})\nonumber\\ &&\label{0249} \quad =[{\lambda\mu \psi'''-(\lambda^2\bar{\rho}+\lambda\mu \|\xi\|^2+M^2\xi_1^2)\psi'}-M^2\xi_1\|\xi\|^2\varphi]/({\lambda\|\xi\|^{2}}). \end{aligned}$$ Thus $_1$ can be rewritten as $$\label{0250}-\varphi''+\sigma \varphi=\omega$$ with boundary conditions $$\label{0251} \varphi(-\infty)=\varphi(+\infty)=0,$$ where $\sigma=(\lambda^2\bar{\rho}+\lambda\mu \|\xi\|^2+M^2\xi_1^2)/{(\lambda\mu)}>0$ and $\omega=\xi_1[{\lambda\mu \psi'''-(\lambda^2\bar{\rho}+\lambda\mu \|\xi\|^2+M^2\xi_1^2)\psi'}]/({\lambda\mu\|\xi\|^{2}})$ $\in H^\infty(\mathbb{R})$. By the ODE theory on a bounded interval and the domain expansion technique, we can obtain a unique solution $\varphi:=\varphi(\xi,x_3)\in H^\infty(\mathbb{R})$ to , . In view of , thus $\pi$ can be uniquely determined by the known functions $\psi$ and $\varphi$. Employing arguments similar to those in the construction of $\varphi$, we can obtain a unique solution $\theta:=\theta(\xi,x_3)\in H^\infty(\mathbb{R})$ to $_3$ with $\theta(-\infty)=\theta(+\infty)=0$, where the solution $\theta$ depends on the known functions $\pi$ and $\varphi$. Finally, by a simple computation, we can check that $({\varphi},{\theta},{\psi},{\pi})$ with $\lambda$ constructed as above indeed is a solution to , . $\Box$ \[rem:0201\] For each $x_3$, it is easy to see that the solution $({\varphi}(\xi,\cdot),{\theta}(\xi,\cdot),{\psi}(\xi,\cdot),{\pi}(\xi,\cdot),\lambda(\xi))$ constructed in Theorem \[thm:0302\] possesses the following properties: 1. $\lambda(\xi)$, ${\psi}(\xi,\cdot)$ and ${\pi}(\xi,\cdot)$ are even on $\xi_1$ or $\xi_2$, when another variable is fixed; 2. ${\varphi}(\xi,\cdot)$ is odd on $\xi_1$, but even on $\xi_2$, when another variable is fixed; 3. ${\theta}(\xi,\cdot)$ is even on $\xi_1$, but odd on $\xi_2$, when another variable is fixed. We mention that the system , for the vertical case enjoys rotational structure. Hence, we can also use the rotation method as in [@NJTSC2] to construct a solution of , , which is simper than the above construction process. The next lemma provides an estimate for the $H^k(\mathbb{R})$-norm of the solution $({\varphi},{\theta},{\psi},{\pi})$ with $\xi$ varying, which will be useful in the next subsection. To emphasize the dependence on $\xi$, we write these solutions as $(\varphi ,\theta ,\psi ,\pi )(\xi):=(\varphi ,\theta ,\psi ,\pi )(\xi,x_3)$. \[lem:0201\] Let $k\geq 0$, and $\mathbb{D}\subset \mathbb{A}^{{\mathrm}{g}}$ be a nonempty bounded set satisfying the closure $\bar{\mathbb{D}}\subset \mathbb{A}^{{\mathrm}{g}}$. Then for any $\xi\in \mathbb{D}$ there are positive constants $A_k$, $B_k$, $C_k$ and $D_k$, which may depend on $g$, $M$, $\mu$, $\bar{\rho}$ and $\mathbb{D}$, such that $$\begin{aligned} &&\label{0252n} \\|\psi(\xi)\\|_{H^k(\mathbb{R})}\leq A_k,\\ &&\label{0253n}\\|\varphi(\xi)\\|_{H^k(\mathbb{R})}\leq B_k,\\ &&\label{0254n}\\|\theta(\xi)\\|_{H^k(\mathbb{R})}\leq C_k,\\ &&\label{0255n} \\|\pi(\xi)\\|_{H^k(\mathbb{R})}\leq D_k,\end{aligned}$$ where $(\varphi,\theta,\psi,\pi)(\xi)$ and $\lambda(\xi)$ be constructed as in Theorem \[thm:0302\]. Moreover, $$\label{0255} \\|\psi(\xi)\\|_{L^2(\mathbb{R})}^2> 0.$$ We still give the proof for the horizonal case only. Throughout this proof, we denote by $\tilde{c}$ a generic positive constant which may vary from line to line, and may depend on $g$, $M$, $\mu$, $\bar{\rho}$ and $\mathbb{D}$. We begin with the estimate . We first rewrite as $$\label{0256n}\begin{aligned} \psi''''(\xi)=&\big[(\lambda^2\bar{\rho}+2\lambda\mu \|\xi\|^2+(M\xi_1)^2)\psi''(\xi)+\lambda^2\bar{\rho}'\psi'(\xi)\\ &\ -\|\xi\|^2(\lambda^2\bar{\rho} +\lambda\mu \|\xi\|^2+(M\xi_1)^2-g\bar{\rho}')\psi(\xi)\big]/\lambda\mu, \end{aligned}$$ which yields $$\label{0257n}\begin{aligned} \\|\psi''(\xi)\\|_{L^2(\mathbb{R})}=&(\lambda\mu)^{-1}\int_\mathbb{R}\big[ \lambda(\lambda\bar{\rho}+2\mu \|\xi\|^2+(M\xi_1)^2)\psi''(\xi)+\lambda^2\bar{\rho}'\psi'(\xi)\\ &\ -\|\xi\|^2(\lambda^2\bar{\rho} +\lambda\mu \|\xi\|^2+(M\xi_1)^2-g\bar{\rho}')\psi(\xi)\big]\psi(\xi) {\mathrm}{d}x_3 \end{aligned}$$ and $$\label{0257nn}\begin{aligned} \\|\psi'''(\xi)\\|_{L^2(\mathbb{R})}=&-(\lambda\mu)^{-1}\int_\mathbb{R}\big[ \lambda(\lambda\bar{\rho}+2\mu \|\xi\|^2+(M\xi_1)^2)\psi''(\xi)+\lambda^2\bar{\rho}'\psi'(\xi)\\ &\ -\|\xi\|^2(\lambda^2\bar{\rho} +\lambda\mu \|\xi\|^2+(M\xi_1)^2-g\bar{\rho}')\psi(\xi)\big]\psi''(\xi) {\mathrm}{d}x_3. \end{aligned}$$ Noting that since $\psi\in \mathcal{A}$, the inequality holds, and there is a constant $\tilde{c}$, such that $$\label{0258n} \\|\psi(\xi)\\|_{H^1(\mathbb{R})}\leq \tilde{c}.$$ On the other hand, in view of Proposition \[pro:0208\], we have $$\label{0259n} \lambda(\xi) \geq \tilde{c}>0 \quad\mbox{for any }\xi\in \mathbb{D}.$$ Thus, using Cauchy-Schwarz’s inequality, and , we deduce from and that $$\begin{aligned} \\|\psi''(\xi)\\|_{H^1(\mathbb{R})}\leq \tilde{c}, \end{aligned}$$ whence, by , $$\label{0262n}\begin{aligned} \\|\psi(\xi)\\|_{H^4(\mathbb{R})}\leq \tilde{c}. \end{aligned}$$ Consequently, differentiating with respect to $x_3$ and using , we find, by induction on $k$, that holds for any $k\geq 0$. Now we turn to the estimate of . From , we have $$\\|\varphi'\\|_{L^2(\mathbb{R})}^2+\sigma\\| \varphi\\|_{L^2(\mathbb{R})}^2=\int_\mathbb{R}\omega\varphi{\mathrm}{d}x_3,$$ which, together with Cauchy-Schwarz’s inequality and , yields $$\\|\varphi\\|_{H^1(\mathbb{R})}^2\leq \tilde{c}.$$ We further deduce from that $$\label{0263n}\\|\varphi\\|_{H^2(\mathbb{R})}^2\leq \tilde{c}.$$ Thus, we can employ and to deduce that holds for any $k\geq 0$. Using this fact, the estimates and , the expression of $\pi$ in implies . Finally, similarly to the deduction of , we can show that holds. This completes the proof. $\Box$ Exponential growth rate {#sec:0204} ----------------------- In this subsection we will construct a linear real-valued solution to the linearized problem along with which grows in-time in the Sobolev space of order $k$. \[thm:0203\] Under the assumptions of Theorem \[thm:0101\], let $$\label{0267} \Lambda=\sup_{\xi\in \mathbb{A}^{\mathrm}{g}\cap (L^{-1}\mathbb{Z})^2}\lambda(\xi),$$ then there exist a positive constant $\Lambda^\in (2\Lambda/3, \Lambda]$, and a real-valued solution $(\varrho,\mathbf{u},\mathbf{N},q)$ to the linearized problem – defined on the horizontally periodic domain $\Omega$, such that 1. For every $k\in \mathbb{N}$, $$\label{0262} \\|(\varrho,\mathbf{u},\mathbf{N},q)(0)\\|_{H^k}<\infty¡£$$ 2. For every $t>0$, $(\varrho,\mathbf{u},\mathbf{N},q)(t)\in H^k$, and $$\begin{aligned} &&\label{0263} e^{\Lambda^ t}\\|(\varrho,{\mathbf}{N},q)(0)\\|_{H^k} =\\|(\varrho,{\mathbf}{N},q)(t)\\|_{H^k}= e^{t\Lambda^}\\|(\varrho,{\mathbf}{N},q)(0)\\|_{H^k},\\[1mm] &&\label{0264}e^{\Lambda^ t}\\|u_i(0)\\|_{H^k}= \\|u_i(t)\\|_{H^k}= e^{t\Lambda^}\\|u_i(0)\\|_{H^k},\quad i=1,2,3.\end{aligned}$$ 3. Moreover, $$\label{02div} {\mathrm}{div}\,{\mathbf}{u}(0)={\mathrm}{div}\,{\mathbf}{N}(0)=0,$$ and $$\label{0268} \\|({u}_1,u_2)(0)\\|_{L^2}\\|{u}_3(0)\\|_{L^2}>0.$$ 4. In addition, if ${\mathbf}{\bar{M}}=M{\mathbf}{e}_3$, then $$\label{0268withN3} \\|N_3(0)\\|_{L^2}>0.$$ In view of the third conclusion in Proposition \[pro:0202\] and , we see that $$0<\Lambda\leq \sup_{\xi\in \mathbb{A}^{{\mathrm}{g}}}\lambda(\xi)\leq \sqrt{ g\left\\|{\bar{\rho}'}/{\bar{\rho}}\right \\|_{L^\infty(\mathbb{R})} }.$$ Let $$\mathbf{v}(\xi,x_3)=-\mathrm{i}\varphi(\xi,x_3){\mathbf}{e}_1-\mathrm{i}\theta(\xi,x_3) {\mathbf}{e}_2+\psi(\xi, x_3){\mathbf}{e}_3,$$ where $(\varphi,\theta,\psi)$ with an associated growth rate $\lambda(\xi)$ is constructed in Theorem \[thm:0302\] for any given $\xi\in\mathbb{A}^{{\mathrm}{g}}$. Recalling the definition of supremum, and using Proposition \[pro:0202\], we find that there exist $\xi^1$ and $\xi^2:=-{\xi}^1$ such that $$\Lambda^:=\lambda(\xi^1)=\lambda(\xi^2)\in (2\Lambda/3, \Lambda]\mbox{ and } \xi^i\in \mathbb{A}^{{\mathrm}{g}}\cap (L^{-1}\mathbb{Z})^2\mbox{ for }i=1\mbox{ and }2.$$ In view of Remark \[rem:0201\], $$\begin{aligned} \label{linearsolution}\left\{\begin{array}{l} {\varrho} (t,\textbf{x})=-e^{\Lambda^t}\bar{\rho}'\sum_{m=1}^2{v}_3(\xi^m,x_3) e^{\mathrm{i}{\mathbf}{x}'\xi^m},\\[2mm] {\mathbf{u}}(t,\textbf{x})=\Lambda^e^{\Lambda^t}\sum_{m=1}^2 \mathbf{v}(\xi^m,x_3)e^{\mathrm{i}{\mathbf}{x}'\xi^m},\\[2mm] {q}(t,\textbf{x})=\Lambda^e^{\Lambda^t}\sum_{m=1}^2{\pi}(\xi^m,x_3)e^{\mathrm{i}{\mathbf}{x}'\xi^m},\\[2mm] {{\mathbf}{N}}(t,{\mathbf}{x})=e^{\Lambda^t}\sum_{m=1}^2\bar{{\mathbf}{M}}\cdot\nabla_{{\mathbf}{x}}( \mathbf{v}(\xi^m,x_3)e^{\mathrm{i}{\mathbf}{x}'\xi^m}) \end{array}\right.\end{aligned}$$ gives a horizontally periodic, real-value solution to – satisfying –. Lemma \[lem:0201\] immediately implies that $(\varrho,\mathbf{u},\mathbf{N},q)$ constructed in also satisfies . Finally, thanks to , we get $$\begin{aligned} \\|{u}_3(0)\\|_{L^2}^2=&4(\Lambda^)^2\int_{\mathbb{R}}\psi^2(\xi^1,x_3) {\mathrm}{d}x_3\int_{(2\pi L\mathbb{T})^2}\cos^2({\mathbf}{x}'\cdot\xi^1)\mathrm{d}{\mathbf}{x}'>0, \end{aligned}$$ which, together with , implies . When ${\mathbf}{\bar{M}}=M{\mathbf}{e}_3$, recalling the expression of ${\mathbf}{M}$, one gets from a simple computation that $$\\|N_3(0)\\|_{L^2}^2=4M^2 \int_{\mathbb{R}}(\partial_{x_3}\psi(\xi^1,x_3))^2{\mathrm}{d}x_3 \int_{(2\pi L \mathbb{T})^2}\cos^2({\mathbf}{x}'\cdot\xi^1){\mathrm}{d}{\mathbf}{x}'.$$ On the other hand, thanks to and $\psi\in H^\infty(\mathbb{R})$, we have $\partial_{x_3}\psi(\xi^1,x_3)\not\equiv 0$. Consequently, follows. This completes the proof of Theorem \[thm:0203\]. $\Box$ Nonlinear energy estimates for the perturbed problem {#sec:04} ==================================================== In this section, we derive some nonlinear energy estimates for the perturbed problem –, and the integrand form of Gronwall’s inequality of the high-order energy estimates. First we shall give a local well-posedness result of the perturbed problem. We mention that the local existence of classical solutions and global existence of classical small solutions to the non-resistive MHD equations have been established by many authors, see [@CLFDSMCRJLR; @KSS1P; @LXLSNWDHL; @LFZPGC; @HXPLFHG] for example. To our best knowledge, there is no existence result for the MHD equations in the horizontally periodic domain $\Omega$. However, with the help of the usual approaches in the proof of local existence for fluid dynamical equations and some mathematical techniques to deal with the horizontally periodic domain, we can establish the following local well-posedness result, the proof of which will be offered in Section \[appendix\] for the completeness. \[pro:0401new\] For any given initial data $(\varrho_0,{\mathbf}{u}_0,{\mathbf}{N}_0)\in H^3\times H^4\times H^3$ satisfying $\inf_{ \mathbf{x}\in\Omega}\{(\varrho_0+\bar{\rho})(\mathbf{x})\} >0$ and ${\mathrm}{div}{\mathbf}{u}_0={\mathrm}{div}{\mathbf}{N}_0=0$, there exist a $T>0$ and a classical unique solution $(\varrho,{\mathbf}{u},{\mathbf}{N})\in C^0(\bar{I}_T,H^3\times H^4\times H^3)$ to the perturbed problem – with an associated pressure $q$. With Proposition \[pro:0401new\] in hand, we further derive the integrand form of Gronwall’s inequality of the high-order energy estimates. It should be pointed out that the solution $(\varrho,{\mathbf}{u},{\mathbf}{N},q)$ constructed in Proposition \[pro:0401new\] possesses more regularity, see the proof in Section \[appendix\]. This additional regularity makes it sense to derive some identities and high-order energy estimates later. In particular, we omit the standard regularization argument in the derivation of some identities, for example, we refer to or in Section \[appendix\] for the proof. In what follows, the notation $a\lesssim b$ means that $a\leq Cb$ for a universal constant $C>0$, which may depend on some physical parameters in . $C(\delta_0)$ means that the positive constant $C$ further depends on $\delta_0$. We denote $$\mathcal{E}:=\mathcal{E}(t):=\mathcal{E}(\varrho,\mathbf{u},\mathbf{N}):= ({\\|(\varrho,\mathbf{N})\\|^2_{H^3}+\\|\mathbf{u}\\|_{H^4}^2})^{{1}/{2}}$$ and $\mathcal{E}_0={\mathcal{E}}(\varrho_0,\mathbf{u}_0,\mathbf{N}_0)$. $\alpha=(\alpha_1,\alpha_2,\alpha_3)$ denotes a multi-index of order $\|\alpha\|=\alpha_1+\alpha_2+\alpha_3$. Estimate of the perturbation density ------------------------------------ We first note that by the classical Sobolev embedding results (see [@ARAJJFF Chapter 5]), we have $$\begin{aligned} &&\label{0403} \\|u\\|_{L^4}\lesssim \\|u\\|_{L^2}^{\frac{1}{4}}\\| u\\|_{H^1}^{\frac{3}{4}}\lesssim \\|u\\|_{H^{1}},\\ && \label{n0404}\\|u\\|_{L^\infty}\lesssim \\| u\\|_{L^2}^{\frac{1}{4}}\\| u\\|_{H^2}^{\frac{3}{4}}\lesssim\\| u\\|_{H^2},\\ && \label{0405} \\|u\\|_{H^{j}}\lesssim \\|u\\|_{L^2}^{\frac{1}{j+1}}\\|u\\|_{H^{j+1}} ^{\frac{j}{j+1}}.\end{aligned}$$ With the help of the above estimates, we can bound the perturbation density $\varrho$. In fact, applying $\partial^\alpha$ to $_1$, multiplying the resulting identity by $\partial^\alpha\varrho$ in $L^2(\Omega)$, we get $$\label{0406}\begin{aligned} \frac{1}{2}\frac{d}{dt}\sum_{0\leq \|\alpha \|\leq 3}\int_{\Omega}(\partial^\alpha\varrho)^2{\mathrm}{d}{\mathbf}{x}=& -\sum_{0\leq \|\alpha \|\leq 3}\int_{\Omega}\partial^\alpha(\bar{\rho}' u_3)\partial^\alpha \varrho{\mathrm}{d}{\mathbf}{x}+ \sum_{0\leq \|\alpha \|\leq 3}\int_{\Omega} \partial^\alpha ({\mathbf}{u}\cdot\nabla \varrho) \partial^\alpha\varrho{\mathrm}{d}{\mathbf}{x}\\ :=& I_1+I_2, \end{aligned}$$ where we have defined $$\partial^\alpha ({\mathbf}{u}\cdot\nabla \varrho)= \sum_{\alpha=\beta+\gamma\atop \|\gamma\|\leq 2} \partial^\beta{\mathbf}{u}\cdot\partial^\gamma\nabla \varrho\mbox{ for }\|\alpha\|=3,$$ so that makes sense (cf. Lemma \[lem:0501\]). Using , Hölder’s and Young’s inequalities, $I_1$ can be bounded as follows: $$\label{0407}\begin{aligned} I_1\lesssim &\sum_{0\leq \|\alpha \|\leq 3}\\|\partial^\alpha(\bar{\rho}'u_3)\\|_{L^2}\\|\partial^\alpha \varrho\\|_{L^2}\\ \lesssim &\\|\varrho\\|_{L^2}\\|u_3\\|_{L^2} +\\|\nabla \varrho\\|_{L^2}(\\| u_3\\|_{L^2} +\\|\nabla u_3\\|_{L^2})\\ &+\\|\nabla^2 \varrho\\|_{L^2}(\\| u_3\\|_{L^2}+\\|\nabla u_3\\|_{L^2}+ \\|\nabla^2 u_3\\|_{L^2}) \\ &+\\|\nabla^3 \varrho\\|_{L^2}( \\| u_3\\|_{L^2}+\\|\nabla u_3\\|_{L^2}+ \\|\nabla^2 u_3\\|_{L^2}+\\|\nabla^3 u_3\\|_{L^2})\\ \lesssim &\\|\varrho\\|_{H^3}\\|u_3\\|_{L^2}+\\|\nabla \varrho\\|_{H^2} \\| u_3\\|_{L^2}^{\frac{1}{2}}\\| u_3\\|_{H^2}^{\frac{1}{2}} +\\|\nabla^2 \varrho\\|_{H^1}\\| u_3\\|_{L^2}^{\frac{1}{3}} \\| u_3\\|_{H^3}^{\frac{2}{3}}\\ &+\\|\nabla^3 \varrho\\|_{L^2}\\| u_3\\|_{L^2}^{\frac{1}{4}} \\| u_3\\|_{H^3}^{\frac{3}{4}}\lesssim C(\delta_0)\\|u_3\\|_{L^2}^2 +\delta_0\mathcal{E}^2. \end{aligned}$$ On the other hand, we use , and Hölder’s inequality to control $I_2$ as follows. $$\label{0408}\begin{aligned} I_2\lesssim &\\|\nabla{\mathbf}{u}\\|_{L^4}\\|\nabla \varrho \\|_{L^4}\\|\nabla \varrho\\|_{L^2}+\\|\nabla^2{\mathbf}{u}\\|_{L^4}\\|\nabla\varrho\\|_{L^4} \\|\nabla^2\varrho\\|_{L^2}+\\|\nabla {\mathbf}{u}\\|_{L^\infty}\\|\nabla^2\varrho\\|_{L^2}^2 \\ & +\\|\nabla^3{\mathbf}{u}\\|_{L^4}\\|\nabla\varrho\\|_{L^4}\\|\nabla^3\varrho\\|_{L^2} +\\|\nabla^2 {\mathbf}{u}\\|_{L^\infty}\\|\nabla^2\varrho\\|_{L^2}\\|\nabla^3\varrho\\|_{L^2} +\\|\nabla {\mathbf}{u}\\|_{L^\infty}\\|\nabla^3\varrho\\|_{L^2}^2 \\ \lesssim &\\|\nabla{\mathbf}{u}\\|_{H^1}\\|\nabla \varrho\\|_{H^1}\\|\nabla \varrho\\|_{L^2}+\\|\nabla{\mathbf}{u}\\|_{H^2} \\|\nabla\varrho\\|_{H^1}\\|\nabla^2\varrho\\|_{L^2}+\\|\nabla {\mathbf}{u}\\|_{H^2}\\|\nabla^2\varrho\\|_{L^2}^2\\ &+\\|\nabla {\mathbf}{u}\\|_{H^3}\\|\nabla\varrho\\|_{H^2}^2\lesssim \mathcal{E}^3. \end{aligned}$$ Hence, one gets from – that $$\label{density}\begin{aligned} \frac{1}{2}\frac{d}{dt}\\|\varrho(t)\\|^2_{H^3}\lesssim C(\delta_0)\\|u_3(t)\\|_{L^2}^2+ \delta_0\mathcal{E}^2(t)+ \mathcal{E}^3(t). \end{aligned}$$ In addition, we can deduce from $_1$ that $$\begin{aligned} && \label{0410}\\|\varrho_t\\|_{L^2} \lesssim \\| {u}_3\\|_{L^2} + \mathcal{E}^2,\\ && \label{0411} \\|\nabla \varrho_t\\|_{L^2}\lesssim \\| {u}_3\\|_{L^2} +\\|\nabla {u}_3\\|_{L^2} +\mathcal{E}^2 \lesssim C(\delta_0)\\| {u}_3\\|_{L^2} + \delta_0\mathcal{E}+\mathcal{E}^2,\\ && \label{0411nn}\\|\varrho_{tt}\\|_{L^2}\lesssim \\|{\mathbf}{u}_t\\|_{L^2}(1+\mathcal{E})+\mathcal{E}^2+\mathcal{E}^3.\end{aligned}$$ Estimate of the perturbation magnetic field ------------------------------------------- We continue to bound the perturbation magnetic field ${\mathbf}{N}$. Applying $\partial^\alpha$ to $_3$, multiplying the resulting identity by $\partial^\alpha\mathbf{N}$ in $L^2$, we obtain $$\begin{aligned} &\frac{1}{2}\frac{d}{dt}\sum_{0\leq \|\alpha \|\leq 3}\int_{\Omega}\|\partial^\alpha {\mathbf}{N}\|^2{\mathrm}{d}{\mathbf}{x}\\ &= \sum_{0\leq \|\alpha \|\leq 2}\int_{\Omega}\partial^\alpha(\bar{{\mathbf}{M}} \cdot\nabla{\mathbf}{u})\cdot\partial^\alpha {\mathbf}{N} {\mathrm}{d}{\mathbf}{x}+ \sum_{0\leq \|\alpha \|\leq 3}\int_{\Omega} [\partial^\alpha ({\mathbf}{N}\cdot\nabla {\mathbf}{u})-\partial^\alpha({\mathbf}{u}\cdot\nabla {\mathbf}{N})]\cdot\partial^\alpha{\mathbf}{N}{\mathrm}{d}{\mathbf}{x}\\ &\quad +\sum_{ \|\alpha \|= 3}\int_{\Omega}\partial^\alpha(\bar{{\mathbf}{M}} \cdot\nabla{\mathbf}{u})\cdot\partial^\alpha {\mathbf}{N} {\mathrm}{d}{\mathbf}{x}:= J_1+J_2+J_3, \end{aligned}$$ where we have used the facts $\nabla \times ({\mathbf}{u}\times \bar{{\mathbf}{M}})=\bar{{\mathbf}{M}}\cdot\nabla {\mathbf}{u}$ and $\nabla \times ({\mathbf}{u}\times {\mathbf}{N})={\mathbf}{N}\cdot\nabla {\mathbf}{u}-{\mathbf}{u}\cdot\nabla {\mathbf}{N}$, and defined $$\partial^\alpha ({\mathbf}{u}\cdot\nabla{\mathbf}{N})= \sum_{\alpha=\beta+\gamma\atop \|\gamma\|\leq 2} \partial^\beta{\mathbf}{u}\cdot\partial^\gamma\nabla {\mathbf}{N}\mbox{ for }\|\alpha\|=3.$$ Similarly to the derivation of and , we can control $J_1$ and $J_2$ as follows. $$\begin{aligned} J_1\lesssim &\sum_{0\leq \|\alpha \|\leq 2} \\|\partial^\alpha\nabla {\mathbf}{u}\\|_{L^2}\\|\partial^\alpha{\mathbf}{N}\\|_{L^2}\\ \lesssim & \\|\nabla{\mathbf}{u}\\|_{L^2}\\|{\mathbf}{N}\\|_{L^2}+ \\|\nabla^2{\mathbf}{u}\\|_{L^2}\\|\nabla {\mathbf}{N}\\|_{L^2}+ \\|\nabla^3{\mathbf}{u}\\|_{L^2}\\|\nabla^2 {\mathbf}{N}\\|_{L^2}\\ \lesssim & \\|{\mathbf}{u}\\|_{L^2}^{\frac{1}{2}}\\|{\mathbf}{u}\\|_{H^2}^{\frac{1}{2}} \\|{\mathbf}{N}\\|_{L^2}+\\|{\mathbf}{u}\\|_{L^2}^{\frac{1}{3}}\\|{\mathbf}{u}\\|_{H^3}^{\frac{2}{3}}\\|\nabla {\mathbf}{N}\\|_{L^2}+ \\|{\mathbf}{u}\\|_{L^2}^{\frac{1}{4}}\\|{\mathbf}{u}\\|_{H^4}^{\frac{3}{4}}\\|\nabla^2 {\mathbf}{N}\\|_{L^2}\\ \lesssim & C(\delta_0)\\|{\mathbf}{u}\\|_{L^2}^2+\delta_0\mathcal{E}^2, \end{aligned}$$ and $$\begin{aligned} J_2\lesssim &\\|\nabla {\mathbf}{u}\\|_{L^4}\\|{\mathbf}{N}\\|_{L^4}\\|{\mathbf}{N}\\|_{L^2} +(\\|\nabla{\mathbf}{u}\\|_{L^4}\\|\nabla {\mathbf}{N} \\|_{L^4}+\\|\nabla^2{\mathbf}{u}\\|_{L^4}\\|{\mathbf}{N} \\|_{L^4})\\|\nabla{\mathbf}{N}\\|_{L^2}\\ & +(\\|\nabla{\mathbf}{u}\\|_{L^\infty}\\|\nabla^2 {\mathbf}{N}\\|_{L^2}+\\|\nabla^2{\mathbf}{u}\\|_{L^4}\\|\nabla {\mathbf}{N} \\|_{L^4}+\\|\nabla^3{\mathbf}{u}\\|_{L^2}\\|{\mathbf}{N} \\|_{L^\infty})\\|\nabla^2{\mathbf}{N}\\|_{L^2}\\ & +(\\|\nabla{\mathbf}{u}\\|_{L^\infty}\\|\nabla^3 {\mathbf}{N} \\|_{L^2}+\\|\nabla^2{\mathbf}{u}\\|_{L^4}\\|\nabla^2 {\mathbf}{N} \\|_{L^4}\\ &\qquad +\\|\nabla^3{\mathbf}{u}\\|_{L^2}\\|\nabla {\mathbf}{N}\\|_{L^\infty}+\\|\nabla^4{\mathbf}{u}\\|_{L^2}\\| {\mathbf}{N} \\|_{L^\infty})\\|\nabla^3{\mathbf}{N}\\|_{L^2}\\ \lesssim &\\|\nabla {\mathbf}{u}\\|_{H^1}\\|{\mathbf}{N}\\|_{H^1}\\|{\mathbf}{N}\\|_{L^2}+\\|\nabla{\mathbf}{u}\\|_{H^2}\\| {\mathbf}{N} \\|_{H^2}\\|\nabla{\mathbf}{N}\\|_{L^2} +\\|\nabla{\mathbf}{u}\\|_{H^2}\\| {\mathbf}{N} \\|_{H^2}\\|\nabla^2{\mathbf}{N}\\|_{L^2}\\ &+\\|\nabla{\mathbf}{u}\\|_{H^3}\\| {\mathbf}{N} \\|_{H^3}\\|\nabla^3{\mathbf}{N}\\|_{L^2} \lesssim \mathcal{E}^3. \end{aligned}$$ Hence, one gets $$\label{0415}\begin{aligned} \frac{1}{2}\frac{d}{dt}\\|{\mathbf}{N}(t)\\|^2_{H^3}\lesssim C(\delta_0)\\|{\mathbf}{u}(t)\\|_{L^2}^2+ \delta_0\mathcal{E}^2(t) +\mathcal{E}^3(t) +J_3. \end{aligned}$$ In addition, we infer from $_4$ that $$\begin{aligned} && \label{0416} \begin{aligned} \\| {\mathbf}{N}_t\\|_{L^2}\lesssim & \\|\nabla{\mathbf}{u} \\|_{L^2}+ \mathcal{E}^2 \lesssim C(\delta_0)\\| {\mathbf}{u} \\|_{L^2}+\delta_0\mathcal{E} +\mathcal{E}^2,\end{aligned}\\ &&\label{0417} \\|\nabla{\mathbf}{N}_t\\|_{L^2}\lesssim \\|\nabla^2{\mathbf}{u} \\|_{L^2}+ \mathcal{E}^2 \lesssim C(\delta_0)\\|{\mathbf}{u} \\|_{L^2}+\delta_0\mathcal{E}+\mathcal{E}^2,\\ &&\label{ttN0417} \\|{\mathbf}{N}_{tt}\\|_{L^2}\lesssim \\|{\mathbf}{u}_t \\|_{H^1}(1+ \mathcal{E})+\mathcal{E}^2+\mathcal{E}^3.\end{aligned}$$ Estimate of the perturbation velocity ------------------------------------- First we restrict the density $\rho=\varrho+\bar{\rho}$ so that it has a positive lower bound. By virtue of , there is a constant $\delta_0'\in (0,1)$, such that $$\label{conditiondensity} \\|\varrho_0\\|_{L^\infty}\leq \frac{\inf_{{x}_3\in\mathbb{R}}\{\bar{\rho}{(x_3)}\}}{2}\;\; \mbox{ for any }\\|\varrho_0\\|_{H^2}\leq \delta_0'.$$ Consequently, in view of $_1$ and ${\mathrm}{div}{\mathbf}{u}=0$, we find that for any $(t,{\mathbf}{x})\in [0,T)\times \Omega$ by the method of characteristics, $$\label{0418}\frac{\inf_{{x}_3\in\mathbb{R}}\{\bar{\rho}({x_3})\}}{2} \leq \inf_{\mathbf{x}\in\Omega}\{\rho_0(t,\mathbf{x})\} \leq \rho(t,{\mathbf}{x})\leq \sup_{\mathbf{x}\in\Omega}\{{\rho}_0(t,{\mathbf{x}})\} \leq \frac{3\sup_{{x}_3\in\mathbb{R}}\{\bar{\rho}({x_3})\}}{2},$$ where $\rho_0=\varrho_0+\bar{\rho}$. From now on, we always assume that $\mathcal{E}\leq \delta_0\leq \delta_0'<1$. We remark that the upper and lower boundedness of $\rho$ will be repeatedly used below. We now estimate the time-derivative of the perturbation velocity. Multiplying $_2$ by ${\mathbf}{u}$ in $L^2$, utilizing $_1$ and integrating by parts, we obtain $$\begin{aligned} \frac{1}{2}\frac{{d}}{{d}t} \int_{\Omega}(\varrho+\bar{\rho} )\|{\mathbf}{u}\|^2{\mathrm}{d}{\mathbf}{x}+ \mu\int_{\Omega}\|\nabla {\mathbf}{u}\|^2{\mathrm}{d}{\mathbf}{x} =&\int_{\Omega}[(\nabla\times {\mathbf}{N})\times ({\mathbf}{N}+\bar{{\mathbf}{M}}) -g \varrho {\mathbf}{e}_3]\cdot{\mathbf}{u}{\mathrm}{d}{\mathbf}{x} \\ :=& \int_{\Omega}{\mathbf}{Y}_0\cdot{\mathbf}{u}{\mathrm}{d}{\mathbf}{x}, \end{aligned}$$ where ${\mathbf}{Y}_0:=(\nabla\times {\mathbf}{N})\times ({\mathbf}{N}+\bar{{\mathbf}{M}})-g \varrho {\mathbf}{e}_3$. Applying $\partial_t$ to $_2$, multiplying the resulting identity by ${\mathbf}{u}_t$ in $L^2$, we make use of $_1$, and integrate by parts to deduce $$\begin{aligned}&\frac{1}{2}\frac{{d}}{{d}t} \int_{\Omega}(\varrho+\bar{\rho} )\|{\mathbf}{u}_t\|^2{\mathrm}{d}{\mathbf}{x}+ \mu\int_{\Omega}\|\nabla {\mathbf}{u}_t\|^2{\mathrm}{d}{\mathbf}{x}\\ & = \int_{\Omega}\Big[ (\nabla\times {\mathbf}{N}_t)\times ({\mathbf}{N}+\bar{{\mathbf}{M}}) +(\nabla\times {\mathbf}{N})\times {\mathbf}{N}_t-\varrho_t{\mathbf}{u}\cdot\nabla {\mathbf}{u}\\ &\quad -(\varrho+\bar{\rho}){\mathbf}{u}_t\cdot\nabla {\mathbf}{u}-g \varrho_t {\mathbf}{e}_3 -\varrho_t{\mathbf}{u}_t\Big]\cdot{\mathbf}{u}_t {\mathrm}{d}{\mathbf}{x}:=\int_{\Omega}{\mathbf}{Y}_1\cdot{\mathbf}{u}_t{\mathrm}{d}{\mathbf}{x}. \end{aligned}$$ Applying $\partial_{t}^2$ to $_2$ again, multiplying the resulting identity by ${\mathbf}{u}_t$, employing $_1$ and $$(\nabla\times {\mathbf}{N})\times ({\mathbf}{N}+\bar{{\mathbf}{M}})=({\mathbf}{N}+\bar{{\mathbf}{M}}) \cdot \nabla {\mathbf}{N}-\frac{1}{2}\nabla\|({\mathbf}{N}+\bar{{\mathbf}{M}})\|^2,$$ we have $$\begin{aligned} &\frac{1}{2}\frac{{d}}{{d}t} \int_{\Omega}(\varrho+\bar{\rho} )\|{\mathbf}{u}_{tt}\|^2{\mathrm}{d}{\mathbf}{x}+ \mu\int_{\Omega}\|\nabla {\mathbf}{u}_{tt}\|^2{\mathrm}{d}{\mathbf}{x}\nonumber\\ &=\int_{\Omega}\Big\{-{\mathbf}{N}_{tt}\cdot\nabla{\mathbf}{u}_{tt}\cdot{\mathbf}{N} -{\mathbf}{N}_{t}\cdot\nabla{\mathbf}{u}_{tt}\cdot{\mathbf}{N}_t-({\mathbf}{N} +\bar{{\mathbf}{M}})\cdot\nabla {\mathbf}{u}_{tt}\cdot{\mathbf}{N}_{tt}-\Big[g\varrho_{tt} {\mathbf}{e}_3-\varrho_{tt}{\mathbf}{u}_t \nonumber\\ &\quad -\varrho_{t}{\mathbf}{u}_{tt}-\varrho_{tt}{\mathbf}{u}\cdot\nabla {\mathbf}{u} -\varrho_{t}{\mathbf}{u}_t\cdot\nabla {\mathbf}{u} -\varrho_{t}{\mathbf}{u}\cdot\nabla {\mathbf}{u}_t -(\varrho+\bar{\rho}){\mathbf}{u}_t\cdot\nabla {\mathbf}{u}_t -(\varrho+\bar{\rho}){\mathbf}{u}_{tt}\cdot\nabla {\mathbf}{u}\Big]\cdot{\mathbf}{u}_{tt}\Big\} {\mathrm}{d}{\mathbf}{x}\nonumber\\ &\label{uttinform}:=-\sum_{i=0}^2 \int_{\Omega}\partial_t^i({\mathbf}{N}+\bar{{\mathbf}{M}})\cdot\nabla{\mathbf}{u}_{tt}\cdot\partial_t^{2-i}{\mathbf}{N} {\mathrm}{d}{\mathbf}{x}+\int_{\Omega}{\mathbf}{Y}_2\cdot{\mathbf}{u}_{tt}{\mathrm}{d}{\mathbf}{x}.\end{aligned}$$ Adding the above three equalities, we get $$\label{0421}\begin{aligned}&\frac{1}{2}\frac{{d}}{{d}t}\sum_{i=0}^2 \int_{\Omega}(\varrho+\bar{\rho} )\|\partial_t^i{\mathbf}{u}\|^2{\mathrm}{d}{\mathbf}{x}+\mu\sum_{i=0}^2 \int_{\Omega}\|\nabla \partial_t^i{\mathbf}{u}\|^2{\mathrm}{d}{\mathbf}{x}\\ =&\sum_{i=0}^2 \int_{\Omega}{\mathbf}{Y}_i\cdot\partial_t^i{\mathbf}{u}{\mathrm}{d}{\mathbf}{x}-\sum_{i=0}^2 \int_{\Omega}\partial_t^i({\mathbf}{N}+\bar{{\mathbf}{M}})\cdot\nabla{\mathbf}{u}_{tt}\cdot\partial_t^{2-i}{\mathbf}{N} {\mathrm}{d}{\mathbf}{x}. \end{aligned}$$ Obviously, $$\label{0422}\begin{aligned}\\|{\mathbf}{Y}_0\\|_{L^2}\lesssim & \\|\varrho\\|_{L^2} +\\|{\mathbf}{N}\\|_{L^2}^{\frac{1}{2}}\\| {\mathbf}{N}\\|_{H^2}^{\frac{1}{2}}\\|{\mathbf}{N}\\|_{L^\infty} \lesssim C(\delta_0)\\|(\varrho,{\mathbf}{N})\\|_{L^2}+\delta_0\mathcal{E}^2. \end{aligned}$$ Recalling that $\mathcal{E}\leq \delta<1$, we use , , , , and to arrive at $$\label{0423}\begin{aligned}\\|{\mathbf}{Y}_1\\|_{L^2}\lesssim & \\|\varrho_t\\|_{L^2} +\\|\varrho_t\\|_{L^2}\\|{\mathbf}{u}\\|_{L^\infty}\\|\nabla {\mathbf}{u}\\|_{L^\infty} + \\|\varrho+\bar{\rho}\\|_{L^\infty}\\| {\mathbf}{u}_t\\|_{L^2}\\|\nabla {\mathbf}{u}\\|_{L^\infty} \\ & +(1+\\| {\mathbf}{N}\\|_{L^\infty})\\| \nabla {\mathbf}{N}_t\\|_{L^2}+\\|\nabla {\mathbf}{N}\\|_{L^4}\\|{\mathbf}{N}_t\\|_{L^4} +\\|\varrho_t\\|_{L^4}\\|{\mathbf}{u}_t\\|_{L^4}\\ \lesssim &\\|\varrho_t\\|_{L^2} +\\| {\mathbf}{u}\\|_{H^3}^2\\|\varrho_t\\|_{L^2} +\\|\varrho_t\\|_{H^1}\\| {\mathbf}{u}_t\\|_{H^1}+\\|\nabla {\mathbf}{u}\\|_{H^2}\\|{\mathbf}{u}_t\\|_{L^2} \\ & +\\| \nabla {\mathbf}{N}_t\\|_{L^2}+ \\|{\mathbf}{N}\\|_{H^2}\\| {\mathbf}{N}_t\\|_{H^1} \\ \lesssim & C(\delta_0)\\|{\mathbf}{u} \\|_{L^2}+\delta_0 \mathcal{E}+\mathcal{E}\\| {\mathbf}{u}_t\\|_{H^1}. \end{aligned}$$ To bound ${\mathbf}{Y}_2$, we argue in a way similar to that used for , with additional estimates and , to infer that $$\label{0423nb}\begin{aligned}\\|{\mathbf}{Y}_2\\|_{L^2}\lesssim & \\|\varrho_{tt}\\|_{L^2}(1+\mathcal{E}^2+\\|{\mathbf}{u}_t\\|_{H^2}) + \\|\varrho_t\\|_{H^1}(\\|{\mathbf}{u}_{tt}\\|_{L^2}^{\frac{1}{4}}\\| {\mathbf}{u}_{tt}\\|_{H^1}^{\frac{3}{4}} +\mathcal{E}\\|{\mathbf}{u}_{t}\\|_{H^2}) \\[1mm] & +\\| {\mathbf}{u}_t\\|_{H^1}\\| \nabla {\mathbf}{u}_t\\|_{H^1}+\mathcal{E}\\| {\mathbf}{u}_{tt}\\|_{L^2}\\ \lesssim & [\\|{\mathbf}{u}_t\\|_{H^1}(1+\mathcal{E})+\mathcal{E}^2](1+\mathcal{E}^2 +\\|{\mathbf}{u}_t\\|_{H^2})+ \mathcal{E} (\\|{\mathbf}{u}_{tt}\\|_{L^2}^{\frac{1}{4}} \\| {\mathbf}{u}_{tt}\\|_{H^1}^{\frac{3}{4}} +\mathcal{E}\\|{\mathbf}{u}_{t}\\|_{H^2}). \end{aligned}$$ Finally, using –, we obtain $$\label{0423nb1}\begin{aligned} &\left\| \sum_{i=0}^2 \int_{\Omega}\partial_t^i({\mathbf}{N}+\bar{{\mathbf}{M}})\cdot\nabla{\mathbf}{u}_{tt}\cdot\partial_t^{2-i}{\mathbf}{N} {\mathrm}{d}{\mathbf}{x}\right\| \\[1mm] &\lesssim [(1+\mathcal{E})\\|{\mathbf}{N}_{tt}\\|_{L^2}+\\|{\mathbf}{N}_t\\|_{H^1}^2] \\|\nabla{\mathbf}{u}_{tt}\\|_{L^2}\lesssim [(1+\mathcal{E})\\|{\mathbf}{u}_t\\|_{H^1} + \mathcal{E}^2]\\| {\mathbf}{u}_{tt}\\|_{H^1}. \end{aligned}$$ Putting – together and using Cauchy-Schwarz’s inequality, we conclude $$\label{04041n}\begin{aligned} \frac{1}{2} & \frac{d}{dt} \sum_{i=0}^2 (\varrho+\bar{\rho}) \\|\partial_t^i\mathbf{u}(t)\\|^2_{L^2}+\sum_{i=0}^2\\|\nabla \partial_t^i \mathbf{u}(t)\\|^2_{L^2} \lesssim C(\delta_0)\\|(\varrho,{\mathbf}{u},{\mathbf}{N})(t) \\|_{L^2}^2+\delta_0\mathcal{E}^2(t) \\ & +(C(\delta_0)\\|{\mathbf}{u}(t) \\|_{L^2}+\delta_0 \mathcal{E}(t) +\mathcal{E}^2(t) +\mathcal{E}(t)\\|{\mathbf}{u}_t(t)\\|_{H^1})\\|{\mathbf}{u}_t(t) \\|_{L^2} \\ & + [\\|{\mathbf}{u}_t(t)\\|_{H^1}(1+\mathcal{E}(t))+\mathcal{E}^2(t)](1+\mathcal{E}^2(t)+\\|{\mathbf}{u}_t(t)\\|_{H^2})\\| {\mathbf}{u}_{tt}(t) \\|_{L^2} \\ & + \mathcal{E}(t) (\\|{\mathbf}{u}_{tt}(t)\\|_{L^2}^{\frac{1}{4}}\\| {\mathbf}{u}_{tt}(t)\\|_{H^1}^{\frac{3}{4}} +\mathcal{E}(t)\\|{\mathbf}{u}_{t}(t)\\|_{H^2})\\|{\mathbf}{u}_{tt}(t) \\|_{L^2} \\ & +[(1+\mathcal{E}(t))\\|{\mathbf}{u}_t(t)\\|_{H^1}+ \mathcal{E}^2(t)]^2. \end{aligned}$$ We proceed to estimate higher derivatives of the perturbation velocity. Rewriting $_2$ as $$\label{0424} -\mu \Delta \mathbf{u}+ (\varrho+\bar{\rho}){\bf u}_t+\nabla q =(\nabla\times {\mathbf}{N})\times ({\mathbf}{N}+\bar{{\mathbf}{M}})-g \varrho {\mathbf}{e}_3-(\varrho+\bar{\rho}){\bf u}\cdot\nabla {\bf u},$$ and multiplying by ${\mathbf}{u}_t$ in $L^2$, we have $$\frac{\mu}{2}\frac{d}{dt} \int_{\Omega}\|\nabla \mathbf{u}(t)\|^2{\mathrm}{d}{\mathbf}{x}+\int_{\Omega}(\varrho+\bar{\rho})\| \mathbf{u}_t\|^2 {\mathrm}{d}{\mathbf}{x}=\int_{\Omega}[{\mathbf}{Y}_0-(\varrho+\bar{\rho}){\bf u}\cdot\nabla {\bf u}]\cdot\mathbf{u}_t{\mathrm}{d}{\mathbf}{x}.$$ Differentiating with respect to $t$ and multiplying the resulting equations by ${\mathbf}{u}_{tt}$ in $L^2$, we obtain $$\begin{aligned} & \frac{\mu}{2}\frac{d}{dt} \int_{\Omega}\|\nabla\mathbf{u}_t(t)\|^2{\mathrm}{d}{\mathbf}{x} +\int_{\Omega}(\varrho+\bar{\rho})\| \mathbf{u}_{tt}\|^2 {\mathrm}{d}{\mathbf}{x} =\int_{\Omega}[{\mathbf}{Y}_1-(\varrho+\bar{\rho}){\bf u}\cdot\nabla {\bf u}_t]\cdot\mathbf{u}_{tt} {\mathrm}{d}{\mathbf}{x}. \end{aligned}$$ Adding the above two equalities, we get $$\label{nn425}\begin{aligned} & \frac{\mu}{2}\frac{d}{dt}\sum_{i=0}^1 \int_{\Omega}\|\nabla \partial_t^i\mathbf{u}(t)\|^2{\mathrm}{d}{\mathbf}{x} +\sum_{i=0}^1\int_{\Omega}(\varrho+\bar{\rho})\|\partial_t^{i+1} \mathbf{u}\|^2{\mathrm}{d}{\mathbf}{x}\\ & =\int_{\Omega}[{\mathbf}{Y}_0-(\varrho+\bar{\rho}){\bf u}\cdot\nabla {\bf u}]\cdot \mathbf{u}_t{\mathrm}{d}{\mathbf}{x}+\int_{\Omega}\left[{\mathbf}{Y}_1 - (\varrho+\bar{\rho}){\bf u}\cdot\nabla {\bf u}_t\right]\cdot\mathbf{u}_{tt} {\mathrm}{d}{\mathbf}{x}. \end{aligned}$$ Clearly, $$\label{0422n}\begin{aligned} \\|(\varrho+\bar{\rho}){\bf u}\cdot\nabla {\bf u}\\|_{L^2}+\\| (\varrho+\bar{\rho}){\bf u}\cdot\nabla {\bf u}_t\\|_{L^2}\lesssim \mathcal{E}^2+\mathcal{E}\\|\nabla{\mathbf}{u}_t\\|_{L^2}. \end{aligned}$$ Consequently, in view of , and , using Cauchy-Schwarz’s inequality, we get from that $$\label{highvelcocity}\begin{aligned} & \frac{d}{dt}\sum_{i=0}^1 \\|\nabla \partial_t^i\mathbf{u}(t)\\|^2_{L^2} +\sum_{i=0}^1\\|\partial_t^{i+1} \mathbf{u}(t)\\|^2_{L^2} \\ &\lesssim C(\delta_0)\\|(\varrho,{\mathbf}{u},{\mathbf}{N})(t)\\|_{L^2}^2 +\delta_0\mathcal{E}^2(t)+\mathcal{E}^2(t)\\|{\mathbf}{u}_t(t)\\|_{H^1}^2 . \end{aligned}$$ Next, we continue to derive more estimates of higher-order derivatives of the perturbation velocity. Multiplying $_2$ by ${\mathbf}{u}_t$, integrating and using , one gets $$\label{l2uestimate}\begin{aligned} \\|{\mathbf}{u}_t\\|_{L^2}^2\lesssim&\\|\sqrt{\varrho+\bar{\rho}}{\mathbf}{u}_t\\|_{L^2}^2\lesssim \\|\varrho\\|_{L^2}^2+\\|\Delta{\mathbf}{u}\\|_{L^2}^2+\\|\nabla{\mathbf}{N}\\|_{L^2}^2+\mathcal{E}^4\\ \lesssim &C(\delta_0)\\|(\varrho,{\mathbf}{u},{\mathbf}{N})\\|_{L^2}^2+\delta_0\mathcal{E}^2, \end{aligned}$$ while applying $\partial_{i}$ to $_2$, multiplying the resulting equations by $\partial_{i}{\mathbf}{u}_t$ in $L^2$, and using , we have $$\label{utone}\begin{aligned} \\|\partial_{i}{\mathbf}{u}_t\\|_{L^2}^2\lesssim &\\|\sqrt{\varrho+\bar{\rho}} \partial_{i}{\mathbf}{u}_t\\|_{L^2}^2\lesssim C(\delta_0)\\|(\varrho,{\mathbf}{u},{\mathbf}{N})\\|_{L^2}^2+\delta_0\mathcal{E}^2+(1+\mathcal{E}^2) \\|{\mathbf}{u}_t\\|_{L^2}^2\\ \lesssim & C(\delta_0)\\|(\varrho,{\mathbf}{u},{\mathbf}{N})\\|_{L^2}^2+\delta_0\mathcal{E}^2. \end{aligned}$$ If we take $\partial_{i}\partial_{j}$ to $_2$ and multiply the resulting equations with $\partial_{i}{\mathbf}{u}_t$ in $L^2$, we see that $$\begin{aligned} \\|\partial_{i}\partial_{j}{\mathbf}{u}_t\\|_{L^2}^2\lesssim \\|\sqrt{\varrho +\bar{\rho}} \partial_{i}\partial_{j}{\mathbf}{u}_t\\|_{L^2}^2 \lesssim\mathcal{E}^2+(1 +\mathcal{E}^2)\\|{\mathbf}{u}_t\\|_{H^1}^2\lesssim\mathcal{E}^2. \end{aligned}$$ In particular, summing up the above three estimates, we conclude $$\label{partialu} \\|{\bf u}_t\\|^2_{H^2} \lesssim \mathcal{E}^2.$$ Finally, adding to , and utilizing –, we arrive at $$\label{0404}\begin{aligned} &\frac{d}{dt}\left(\sum_{i=0}^1\\|\nabla \partial_t^i\mathbf{u}(t)\\|^2_{L^2}+\sum_{i=0}^2 \\|\sqrt{\varrho +\bar{\rho}}\partial_t^i{\mathbf}{u}(t)\\|^2_{L^2}\right) \\ &\qquad +\sum_{i=0}^1 \\|\partial_t^i(\nabla{\mathbf}{u},{\mathbf}{u}_t,\nabla{\mathbf}{u}_t)(t)\\|^2_{L^2}\lesssim C(\delta_0)\\| (\varrho,{\mathbf}{u},{\mathbf}{N})(t) \\|_{L^2}^2+\delta_0\mathcal{E}^2(t), \end{aligned}$$ provided $\delta_0$ is sufficiently small. Energy estimates {#sec:0404} ---------------- Now, we sum up the previous estimates , and , and use Cauchy-Schwarz’s inequality to find that $$\label{densityenegr}\begin{aligned} &\frac{d}{dt}\left[\\|(\varrho,{\mathbf}{N})(t)\\|^2_{H^3} +\sum_{i=0}^1\\|\nabla \partial_t^i\mathbf{u}(t)\\|^2_{L^2} +\sum_{i=0}^2 \\|\sqrt{\varrho+\bar{\rho}}\,\partial_t^i{\mathbf}{u}(t)\\|^2_{L^2}\right] \\ &\qquad +\sum_{i=0}^1 \\| \partial_t^i(\nabla{\mathbf}{u},{\mathbf}{u}_t,\nabla{\mathbf}{u}_t)(t)\\|^2_{L^2}\lesssim C(\delta_0)\\| (\varrho,{\mathbf}{u},{\mathbf}{N})(t) \\|_{L^2}^2 +J_3+\delta_0\mathcal{E}^2(t), \end{aligned}$$ provided $\delta_0$ is sufficiently small. To deal with the term $J_3$, we shall make use of the momentum equations $_2$. Let $\alpha:=(\alpha_1,\alpha_2,\alpha_3)$ be given and satisfy $\|\alpha\|=3$. Without loss of generality, assume $\alpha_1\neq 0$. By employing a partial integration, one sees $$\label{mageequality} \int_{\Omega}\partial^\alpha(\bar{{\mathbf}{M}}\cdot\nabla{\mathbf}{u})\cdot\partial^\alpha{\mathbf}{N} {\mathrm}{d}{\mathbf}{x}=\int_{\Omega}\bar{{\mathbf}{M}}\cdot \nabla \partial^\gamma{\mathbf}{N}\cdot \partial^\beta {\mathbf}{u}{\mathrm}{d}{\mathbf}{x},$$ where $\gamma:=(\alpha_1-1, \alpha_2,\alpha_3)$ and $\beta:=(\alpha_1+1, \alpha_2,\alpha_3)$ satisfying $\|\gamma\|=2$ and $\|\beta\|=4$. Letting $\partial^\gamma:=\partial_{i}\partial_{j}$, and applying $\partial^\gamma$ to $_2$, keeping in mind that $$\begin{aligned} (\nabla\times {\mathbf}{N})\times\bar{{\mathbf}{M}}=\bar{{\mathbf}{M}}\cdot \nabla {\mathbf}{N} -\nabla ({\mathbf}{N}\cdot \bar{{\mathbf}{M}}), \end{aligned}$$ we find that $$\begin{aligned}\label{0426} &({\varrho}+\bar{\rho})\partial^\gamma\mathbf{u}_t +\nabla \partial^\gamma q +g \partial^\gamma\varrho {\mathbf}{e}_3-\mu \Delta \partial^\gamma\mathbf{u}\\ & = -\mathbf{u}_t \partial^\gamma({\varrho}+\bar{\rho})- \partial_{i}({\varrho}+\bar{\rho}) \partial_{j}\mathbf{u}_t- \partial_{j}({\varrho}+\bar{\rho})\partial_{i}\mathbf{u}_t -\partial^\gamma[(\varrho+\bar{\rho}){\bf u}\cdot\nabla {\bf u}] \\ &\quad +\partial^\gamma[(\nabla\times {\mathbf}{N})\times {\mathbf}{N}]-\nabla (\partial^\gamma{\mathbf}{N}\cdot \bar{{\mathbf}{M}})+\bar{{\mathbf}{M}}\cdot \nabla \partial^\gamma{\mathbf}{N}. \end{aligned}$$ Multiplying by $\partial^\beta{\mathbf}{u}$ with $\|\beta\|=4$ in $L^2$, we get $$\begin{aligned} &\frac{1}{2}\frac{d}{dt}\int_{\Omega}({\varrho}+\bar{\rho})\|\partial^\alpha\mathbf{u}\|^2{\mathrm}{d}{\mathbf}{x} +\mu \int_{\Omega}\|\nabla \partial^\alpha\mathbf{u}\|^2{\mathrm}{d}{\mathbf}{x} \\ =&-\int_{\Omega}\partial_{1}({\varrho}+\bar{\rho})\partial^\gamma\mathbf{u}_t \cdot \partial^\alpha\mathbf{u}{\mathrm}{d}{\mathbf}{x} +\int_{\Omega}\mathbf{u}_t \partial^\gamma({\varrho}+\bar{\rho})\cdot \partial^\beta\mathbf{u}{\mathrm}{d}{\mathbf}{x} \\ &+\int_{\Omega} \partial_{i}({\varrho}+\bar{\rho})\partial_{j}\mathbf{u}_t \cdot\partial^\beta\mathbf{u}{\mathrm}{d}{\mathbf}{x}+ \int_{\Omega} \partial_{j}({\varrho}+\bar{\rho})\partial_{i}\mathbf{u}_t \cdot\partial^\beta\mathbf{u}{\mathrm}{d}{\mathbf}{x} \\ & +\int_{\Omega} \partial^\gamma[(\varrho+\bar{\rho}){\bf u}\cdot\nabla {\bf u}]\cdot\partial^\beta\mathbf{u}{\mathrm}{d}{\mathbf}{x} -\int_{\Omega} \partial^\gamma[(\nabla\times {\mathbf}{N})\times {\mathbf}{N}]\cdot\partial^{\beta}\mathbf{u}{\mathrm}{d}{\mathbf}{x}\\ &-g\int_{\Omega}\partial^\alpha\varrho \partial^\alpha{u}_3{\mathrm}{d}{\mathbf}{x} +\int_{\Omega}(\varrho+\bar{\rho}){\mathbf}{u}\cdot\nabla\partial^\alpha{\mathbf}{u}\cdot\partial^\alpha{\mathbf}{u}{\mathrm}{d}{\mathbf}{x}-\int_{\Omega}\bar{{\mathbf}{M}}\cdot \nabla \partial^\gamma{\mathbf}{N}\cdot\partial^{\beta} \mathbf{u}{\mathrm}{d}{\mathbf}{x}=\sum_{n=1}^{9} I_n \end{aligned}$$ Recalling that $\mathcal{E}(t)<\delta_0\ll 1$, it is easy to verify that $$\begin{aligned} \left\|\sum_{n=1}^8 I_n\right\|\lesssim & \\|\partial^\alpha\varrho\\|_{L^2}\\|\partial^\alpha u_3\\|_{L^2}+(\\|\partial^\alpha {\mathbf}{u}\\|_{L^2}+\mathcal{E}^2)\\|{\mathbf}{u}_t\\|_{H^2}+ \mathcal{E}^3 \\ \lesssim & C(\delta_0) \\|{\mathbf}{u}\\|_{L^2}^2+\delta_0\mathcal{E}^{\frac{6}{7}}\\|{\mathbf}{u}_t\\|_{H^2}^{\frac{8}{7}} +\mathcal{E}^2\\|{\mathbf}{u}_t\\|_{H^2}+ \delta_0\mathcal{E}^2. \end{aligned}$$ Therefore, $$\begin{aligned}\label{0427} \frac{1}{2} & \frac{d}{dt}\int_{\Omega}({\varrho}+\bar{\rho})\|\partial^\alpha\mathbf{u}(t)\|^2 {\mathrm}{d}{\mathbf}{x}+\mu \int_{\Omega}\|\nabla \partial^\alpha\mathbf{u}(t)\|^2{\mathrm}{d}{\mathbf}{x} \\ & \lesssim C(\delta_0) \\|{\mathbf}{u}(t)\\|_{L^2}^2+\delta_0\mathcal{E}^{\frac{6}{7}}(t) \\|{\mathbf}{u}_t(t)\\|_{H^2}^{\frac{8}{7}} +\mathcal{E}^2(t)\\|{\mathbf}{u}_t(t)\\|_{H^2}+ \delta_0\mathcal{E}^2(t)+I_{9} \\ & \lesssim C(\delta_0) \\|{\mathbf}{u}(t)\\|_{L^2}^2+\delta_0\mathcal{E}^2(t)+I_{9}, \end{aligned}$$ where we have used in the last inequality. Adding to and using , we have $$\begin{aligned} &\frac{d}{dt}\Big[\\|(\varrho,{\mathbf}{N})(t)\\|^2_{H^3} +\sum_{\|\alpha\|=3} \\|\sqrt{\varrho+\bar{\rho}}\partial^\alpha{\mathbf}{u}(t) \\|_{L^2}^2+\sum_{i=0}^1\\|\nabla \partial_t^i\mathbf{u}(t)\\|^2_{L^2}+\sum_{i=0}^2 \\|\sqrt{\varrho+\bar{\rho}}\partial_t^i{\mathbf}{u}(t)\\|^2_{L^2}\Big]\\ &+\sum_{i=0}^1\\|\partial_t^i(\nabla \mathbf{u}, \mathbf{u}_t,\nabla\mathbf{u}_t)(t)\\|^2_{L^2} +\\|\nabla^4{\mathbf}{u}(t)\\|_{L^2}^2 \lesssim C(\delta_0)\\| (\varrho,{\mathbf}{u},{\mathbf}{N})(t) \\|_{L^2}^2+ \delta_0\mathcal{E}^2(t). \end{aligned}$$ In particular, integrating the above inequality over $(0,t)$, we get immediately $$\label{enggyn}\begin{aligned} &\\|(\varrho,{\mathbf}{N})(t)\\|^2_{H^3}+\\|\mathbf{u}(t)\\|_{H^1}^2+\\|\nabla^3\mathbf{u}(t)\\|_{L^2}^2 +\\|\mathbf{u}_t(t)\\|_{H^1}^2+\\|\mathbf{u}_{tt}(t)\\|^2_{L^2} \\ &\quad +\int_0^t\Big[\sum_{i=0}^1\\|\partial_s^i(\nabla \mathbf{u}, \mathbf{u}_s,\nabla{\mathbf}{u}_s)(s)\\|^2_{L^2} +\\|\nabla^4{\mathbf}{u}(s)\\|_{L^2}^2\Big]{\mathrm}{d}s \\ & \lesssim \\|(\varrho_0,{\mathbf}{N}_0)\\|^2_{H^3}+\\|\mathbf{u}_0\\|_{H^3}^2+\\| \partial_t\mathbf{u}_0\\|^2_{H^2}+\\| \partial_{t}^2\mathbf{u}_0\\|^2_{L^2} \\ &\qquad +\int_0^t\Big[C(\delta_0)\\| (\varrho,{\mathbf}{u},{\mathbf}{N})(s) \\|_{L^2}^2 + \delta_0\mathcal{E}^2(s)\Big]{\mathrm}{d}s. \end{aligned}$$ Applying the classical regularity theory on the Stokes equations to $_2$ (referring to the estimate ), one finds that $$\label{regularity}\begin{aligned}\\|\nabla^2{\mathbf}{u}_t\\|_{L^2}+\\|\nabla q_t\\|_{L^2} \lesssim & \\|\partial_t\big[(\nabla\times {\mathbf}{N})\times ({\mathbf}{N}+\bar{{\mathbf}{M}}) -g \varrho {\mathbf}{e}_3-(\varrho+\bar{\rho}){\bf u}\cdot\nabla {\bf u}-(\varrho+\bar{\rho}){\bf u}_t\big]\\|_{L^2} \\ \lesssim & (1+\mathcal{E}^2)\\|\varrho_t\\|_{L^2}+(\mathcal{E}+\\|\varrho_t\\|_{L^2})\\|{\mathbf}{u}_t\\|_{H^1}+(1+\mathcal{E})\\|{\mathbf}{N}_t\\|_{H^1}+ \\|{\mathbf}{u}_{tt}\\|_{L^2}\\ \lesssim & \mathcal{E}+\\|{\mathbf}{u}_t\\|_{H^1}+ \\|{\mathbf}{u}_{tt}\\|_{L^2}, \end{aligned}$$ and $$\begin{aligned} \\|\nabla^2{\mathbf}{u}\\|_{H^2}+\\|\nabla q\\|_{H^2} \lesssim & \\|(\nabla\times {\mathbf}{N})\times ({\mathbf}{N}+\bar{{\mathbf}{M}})-g \varrho {\mathbf}{e}_3-(\varrho+\bar{\rho}){\bf u}\cdot\nabla {\bf u}-(\varrho+\bar{\rho}){\bf u}_t\\|_{H^2}\\ \lesssim &\mathcal{E} +\\|{\bf u}_t\\|_{H^2}. \end{aligned}$$ On the other hand, letting $t\rightarrow 0$ in , we infer that $$\label{initial}\\| \partial_t{\bf u}_0\\|^2_{H^2} \lesssim \mathcal{E}^2_0.$$ Similarly, it is easy to infer that $$\label{initial123}\\| \partial_t^2{\bf u}_0\\|^2_{L^2} \lesssim \mathcal{E}^2_0.$$ Consequently, combining with –, we conclude $$\begin{aligned} & \mathcal{E}^2(t)+\\|\mathbf{u}_t(t)\\|^2_{H^2}+\\|\nabla q_t\\|_{L^2}^2 +\\|\nabla q\\|_{H^2}^2+\\|{\mathbf}{u}_{tt}\\|_{L^2}^2 \\ & \quad +\int_0^t\left[\sum_{i=0}^1\\|\partial_s^i(\nabla \mathbf{u},\mathbf{u}_s,\nabla \mathbf{u}_s)(s)\\|^2_{L^2}+\\|\nabla^4{\mathbf}{u}(s)\\|_{L^2}^2\right]{\mathrm}{d}s \\ & \lesssim \mathcal{E}^2_0 +\int_0^t\left[C(\delta_0)\\|(\varrho,{\mathbf}{u},{\mathbf}{N})(s) \\|_{L^2}^2 + \delta_0\mathcal{E}^2(s)\right]{\mathrm}{d}s,\ \mbox{ provided }\delta_0\mbox{ is sufficiently small}. \end{aligned}$$ In addition, from , , , and we get the following estimate, which will be used in Section \[sec:05\]. $$\\|F\\|_{L^2}^2+ \\|{\mathbf}{G}_t\\|_{L^2}^2 + \\|{\mathbf}{H}\\|_{H^1}^2\lesssim \mathcal{E}^2+\\|{\mathbf}{u}_{tt}\\|^2_{L^2},$$ where $F=-{{\mathbf{u}}}\cdot \nabla\varrho$, $\mathbf{G}:= (\nabla\times {\mathbf}{N})\times {\mathbf}{N} - (\varrho+\bar{\rho})\mathbf{u}\cdot\nabla \mathbf{u}-\varrho{\mathbf{u}}_t$, and ${\mathbf}{H}:=\nabla\times ({\mathbf}{u}\times{\mathbf}{N})$. Finally, summing up the above estimates and the existence statement, we arrive at the following conclusion: \[pro:0401\] Let $(\varrho,{\mathbf}{u},{\mathbf}{N})\in C^0(\bar{I}_T,H^3\times H^4\times H^3)$ be constructed in Proposition \[pro:0401new\], and $\delta_0'$ be chose as in . If $\varrho_0$ further satisfies $\\|\varrho_0\\|_{H^2}\leq \delta_0'$, then there is a constant $\delta_0\in (0,\delta_0']$, such that if $\mathcal{E}(t)\leq \delta_0$ on $\bar{I}_T$, then the classical solution satisfies $$\label{energyinequality}\begin{aligned} &\mathcal{E}^2(t)+\\|\mathbf{u}_t(t)\\|^2_{H^2}+\\|(\nabla q,{\mathbf}{u}_t)_t\\|_{L^2}^2+\\|\nabla q\\|_{H^2}^2 \\ & \qquad +\int_0^t\Big[\sum_{i=0}^1\\|\partial_s^i(\nabla \mathbf{u}, \mathbf{u}_s, \nabla\mathbf{u}_s)(s)\\|^2_{L^2} +\\|\nabla^4{\mathbf}{u}(s)\\|_{L^2}^2\Big]{\mathrm}{d}s \\ &\quad \leq C(\delta_0)\mathcal{E}^2_0 +\int_0^t\left[C(\delta_0)\\| (\varrho,{\mathbf}{u},{\mathbf}{N})(s) \\|_{L^2}^2+\Lambda^\mathcal{E}^2(s)\right]{\mathrm}{d}s, \end{aligned}$$ and $$\label{energyinequality1}\begin{aligned} \\|F\\|_{L^2}^2+ \\|{\mathbf}{G}_t\\|_{L^2}^2 + \\|{\mathbf}{H}\\|_{H^1}^2\leq C(\delta_0) (\mathcal{E}^2+\\|{\mathbf}{u}_{tt}\\|^2_{L^2}), \end{aligned}$$ where $\Lambda^$ is provided by Theorem \[thm:0203\]. Proof of Theorem \[thm:0101\] {#sec:05} ============================= Now we are in a position to prove Theorem \[thm:0101\]. First, in view of Theorem \[thm:0203\], we can construct a linear solution $$\label{0501} \left(\varrho^{\mathrm}{l}, {\mathbf{u}}^{\mathrm}{l},{\mathbf{N}}^{\mathrm}{l} \right)= e^{\Lambda^ t} \left(\bar{\varrho}_0, \bar{\mathbf{u}}_0, \bar{\mathbf{N}}_0 \right)\mbox{ for some } \Lambda^\in (2\Lambda/3,\Lambda],$$ to with the initial data $(\bar{\varrho}_0,\bar{\mathbf{u}}_0,\bar{\mathbf{N}}_0) \in H^\infty$ and ${\mathrm}{div}\bar{{\mathbf}{u}}_0={\mathrm}{div}\bar{{\mathbf}{N}}_0=0$; moreover, the linear solution satisfies $$\label{n0502}\begin{aligned} &0< m_0:=\left\{ \begin{array}{l} \min\{\\|{\varrho}_0\\|_{L^2},\\|(\bar{u}_{01},\bar{u}_{02}) \\|_{L^2},\\|\bar{u}_{03}\\|_{L^2}\}, \\ \min\{\\|{\varrho}_0\\|_{L^2},\\|({\bar{u}}_{01},{\bar{u}}_{02}) \\|_{L^2},\\|{\bar{u}}_{03}\\|_{L^2},\\|{\bar{N}}_{03}\\|_{L^2}\} , \hbox{ if } \bar{\bf{M}}=M{\mathbf}{e}_3, \end{array} \right.\end{aligned}$$ where the sharp linear growth rate $\Lambda>0$ is defined by , and $\bar{u}_{0i}$ and $\bar{N}_{03}$ stand for the $i$-th component of $\bar{{\mathbf}{u}}_0$ and the third component of $\bar{{\mathbf}{N}}_0$, respectively. Denote $C_0:=\\|(\bar{\varrho}_0,\bar{{\mathbf}{u}}_0,\bar{{\mathbf}{N}}_0)\\|_{\mathcal{E}} = \mathcal{E}(\bar{\varrho_0},\bar{{\mathbf}{u}}_0,\bar{{\mathbf}{N}}_0)$, $(\varrho_0^\delta,{\mathbf}{u}_0^\delta,{\mathbf}{N}^\delta_0) :=\delta (\bar{\varrho}_0,\bar{{\mathbf}{u}}_0$, $\bar{{\mathbf}{N}}_0)$, and $C_1:=\\|(\bar{\varrho_0},\bar{{\mathbf}{u}}_0,\bar{{\mathbf}{N}}_0)\\|_{L^2}$. By virtue of Proposition \[pro:0401\], there exists a unique local classical solution $(\varrho^\delta,\mathbf{u}^\delta,\mathbf{N}^\delta)\in C^0(\bar{I}_T,H^3\times H^4\times H^3)$ to , emanating from the initial data $(\varrho_0^\delta,{\mathbf}{u}_0^\delta,{\mathbf}{N}^\delta_0)$ with $\\|(\varrho_0^\delta,{\mathbf}{u}_0^\delta,{\mathbf}{N}^\delta_0)\\|_{\mathcal{E}}=C_0\delta$. Let us now choose $\delta_0\in (0,1)$ as small as in Proposition \[pro:0401\] and let $C(\delta_0)>0$ be the constant appearing in Proposition \[pro:0401\] for the fixed choice of $\delta_0$. Let $\iota=\min\{\delta_0,\varepsilon_0\}$, and $\delta\in (0,\iota)$ satisfy $$\label{times} T^{\delta}=\frac{1}{\Lambda^}{\mathrm}{ln}\frac{2\varepsilon_0}{\delta}>0,\quad\mbox{i.e. }\; \delta e^{\Lambda^* T^\delta}=2\varepsilon_0,$$ where the value of $\varepsilon_0$ will be fixed later to be sufficiently small and independent of $\delta$. We then define $$T^=\sup\left\{t\in (0,T^{\max})\left\|~\left\\|\left(\varrho^\delta, {\mathbf{u}}^\delta,{\mathbf{N}}^\delta \right)(t)\right\\|_{\mathcal{E}}\leq {\delta_0}\right\}\right.$$ and $$T^{}=\sup\left\{t\in (0,T^{\max})\left\|~\left\\|\left(\varrho^\delta, {\mathbf{u}}^\delta,{\mathbf{N}}^\delta \right)(t)\right\\|_{{L}^2}\leq 2\delta C_1e^{\Lambda^ t}\right\}\right.,$$ where $T^{{\mathrm}{max}}$ denotes the maximal time of existence. Obviously, $T^T^{}>0$, and furthermore, $$\label{0502n1} \begin{aligned}&\left\\|\left(\varrho^\delta, {\mathbf{u}}^\delta,{\mathbf{N}}^\delta \right)(T^)\right\\|_{\mathcal{E}}={\delta_0} \mbox{ if }T^<\infty,\\ &\left\\|\left(\varrho^\delta, {\mathbf{u}}^\delta,{\mathbf{N}}^\delta \right)(T^{})\right\\|_{{L}^2}=2\delta C_1e^{\Lambda^ T^{}} \mbox{ if }T^{}<T^{\max}. \end{aligned}$$ Then for all $t\leq \min\{T^\delta,T^,T^{}\}$, we deduce from the estimate , and the definitions of $T^$ and $T^{*}$ that $$\begin{aligned}\label{0503} &\left\\|\left(\varrho^\delta, {\mathbf{u}}^\delta,{\mathbf{N}}^\delta \right)(t)\right\\|_{\mathcal{E}}^2+\\|{\mathbf}{u}_{tt}^\delta\\|^2_{L^2}\\ \leq &C(\delta_0) \delta^2\left\\|\left( \bar{\varrho}_0, \bar{\mathbf{u}}_0,\bar{\mathbf{N}}_0 \right)\right\\|_{\mathcal{E}} ^2+\Lambda^\int_0^t \left\\|\left(\varrho^\delta, {\mathbf{u}}^\delta,{\mathbf{N}}^\delta \right)(s)\right\\|_{\mathcal{E}}^2{\mathrm}{d}s \\ &\ +C(\delta_0)\int_0^t\left\\|\left(\varrho^\delta, {\mathbf{u}}^\delta,{\mathbf{N}}^\delta \right)(s)\right\\|_{L^2}^2{\mathrm}{d}s\\ \leq & {\Lambda^}\int_0^t \left\\|\left( \varrho^\delta, {\mathbf{u}}^\delta,{\mathbf{N}}^\delta \right)(s)\right\\|_{\mathcal{E}}^2{\mathrm}{d}s +C_0^2C(\delta_0 )\delta^2+2 C(\delta_0 )C_1^2\delta^2e^{2\Lambda^ t}/\Lambda^\\ \leq & \Lambda^\int_0^t \left\\|\left( \varrho^\delta, {\mathbf{u}}^\delta,{\mathbf{N}}^\delta \right)(s)\right\\|_{\mathcal{E}}^2{\mathrm}{d}s + C_2\delta^2e^{2\Lambda^* t} \end{aligned}$$ for some constant $C_2:=C_2(\delta_0)>0$ independent of $\delta$. Then applying Gronwall’s inequality to the above estimate, one obtains $$\begin{aligned}\label{0504} \left\\|\left( \varrho^\delta, {\mathbf{u}}^\delta, {\mathbf{N}}^\delta \right)(t)\right\\|_{\mathcal{E}}^2+\\|{\mathbf}{u}_{tt}^\delta\\|^2_{L^2} \leq &C_2\delta^2e^{2\Lambda^* t}+C_2\delta^2 e^{\Lambda^* t}\int_0^t \Lambda^* e^{\Lambda^* s}{\mathrm}{d}s\\ \leq & C_2\delta^2e^{2\Lambda^* t}+C_2\delta^2 e^{2\Lambda^* t}= 2C_2\delta^2e^{2\Lambda^* t} \end{aligned}$$ for any $t\leq \min\{T^\delta,T^,T^{}\}$. Let $(\varrho^{\mathrm{d}}, {\mathbf{u}}^{\mathrm{d}},{\mathbf{N}}^{\mathrm{d}})=(\varrho^{\delta}, {\mathbf{u}}^{\delta},{\mathbf{N}}^{\delta})-\delta(\varrho^{\mathrm{l}}, {\mathbf{u}}^{\mathrm{l}},{\mathbf{N}}^{\mathrm{l}})$. Then $(\varrho^{\mathrm}{a}_\delta,{\mathbf}{u}^{{\mathrm}{a}}_\delta,{\mathbf}{N}^{{\mathrm}{a}}_\delta):= \delta(\varrho^{{\mathrm}{l}},{\mathbf}{u}^{{\mathrm}{l}},{\mathbf}{N}^{{\mathrm}{l}})$ is also a linear solution to with the initial data $(\varrho_0^\delta,\mathbf{u}_0^\delta,\mathbf{N}_0^\delta)\in H^\infty$. Moreover, $(\varrho^{\mathrm{d}}, {\mathbf{u}}^{\mathrm{d}},{\mathbf{N}}^{\mathrm{d}})$ satisfies the following non-homogenous equations: $$\label{newfor1232}\left\{\begin{array}{ll} \varrho_t^{\mathrm{d}}+\bar{\rho}'{u}_3^{\mathrm{d}}=F^\delta, \\[1mm] \bar{\rho}\mathbf{u}_t^{\mathrm{d}}+\nabla q^{\mathrm{d}}+g\varrho^{\mathrm{d}} {\mathbf}{e}_3 -\mu \Delta \mathbf{u}^{\mathrm{d}}-(\nabla\times {\mathbf}{N}^{\mathrm{d}})\times\bar{{\mathbf}{M}}={\mathbf}{G}^\delta, \\[1mm] {\mathbf}{N}_t^{\mathrm{d}}- \nabla\times ({\mathbf}{u}^{\mathrm{d}}\times \bar{{\mathbf}{M}})={\mathbf}{H}^\delta,\\[1mm] \mathrm{div}\mathbf{u}^{\mathrm{d}}={0},\ \mathrm{div}\mathbf{N}^{\mathrm{d}}={0}, \end{array}\right.$$ where $$\displaystyle\left(\begin{array}{c} F^\delta\\ \mathbf{G}^\delta\\ {\mathbf}{H}^\delta \end{array}\right):=\left(\begin{array}{c} \displaystyle -{{\mathbf{u}}}^{\delta}\cdot \nabla\varrho^{\delta} \\ \displaystyle (\nabla\times {\mathbf}{N}^{\delta})\times {\mathbf}{N}^{\delta} -( \varrho^{\delta}+\bar{\rho})\mathbf{u}^{\delta}\cdot\nabla \mathbf{u}^{\delta}-\varrho^\delta\mathbf{u}^{\delta}_t\\[1mm] \nabla\times ({\mathbf}{u}^{\delta}\times{\mathbf}{N}^{\delta}) \end{array}\right)$$ with initial conditions $$(\varrho^{\mathrm{d}}(0), {\mathbf{u}}^{\mathrm{d}}(0), {\mathbf{N}}^{\mathrm{d}}(0))= {\mathbf{0}},\quad \mathrm{div}{\mathbf}{u}_0^{\mathrm{d}}=0\mbox{ and }\mathrm{div}{\mathbf}{N}_0^{\mathrm{d}}=0.$$ From the estimates and , it follows that $$\label{FGHestimates} \\|F^\delta\\|_{L^2}^2+ \\|{\mathbf}{G}_t^\delta\\|_{L^2}^2 + \\|{\mathbf}{H}^\delta\\|_{H^1}^2 \leq 2C_2C(\delta_0)\delta^2e^{2\Lambda^ t}\mbox{ for any }t\leq \min\{T^\delta,T^,T^{}\}.$$ Moreover, we have the following error estimate for $\left(\varrho^{\mathrm{d}}, {\mathbf{u}}^{\mathrm{d}},{\mathbf{N}}^{\mathrm{d}}\right)$: There exists a constant $C_3$ such that $$\begin{aligned}\label{0508} \sqrt{\left\\|\left(\varrho^{\mathrm{d}}, {\mathbf{u}}^{\mathrm{d}}\right)(t)\\|_{H^1}^2+\\|{\mathbf{N}}^{\mathrm{d}} (t)\right\\|_{L^2}^2} \leq C_3(\delta e^{\Lambda^ t})^{\frac{3}{2}}. \end{aligned}$$ The proof is divided into two steps, in which we omit the superscript “${\mathrm}{d}$" in $\left(\varrho^{\mathrm{d}}, {\mathbf{u}}^{\mathrm{d}},{\mathbf{N}}^{\mathrm{d}} \right)$ for simplicity. Step 1: We use the definition of $\Lambda$ to deduce the following estimate $$\label{0302}\int_{\Omega} \left(g\bar{\rho}'\|u_3 \|^2 -M^2\|\partial_i{\mathbf}{u} \|^2\right){\mathrm}{d}{\mathbf}{x}\leq {\Lambda^2}\int_{\Omega}\bar{\rho}\|{\mathbf}{u} \|^2 {\mathrm}{d}{\mathbf}{x}+ {\Lambda}\mu\int_{\Omega} \|\nabla\mathbf{u}\|^2 {\mathbf}{x}\quad\mbox{ for }i=1,3.$$ Let $\hat{f}$ be the horizontal Fourier transform of $f$, i.e., $$\hat{f}(\xi,x_3)=\int_{(2\pi L\mathbb{T})^2}f({\mathbf}{x}',x_3)e^{-{\mathrm}{i}{\mathbf}{x}'\cdot\xi} {\mathrm}{d}{\mathbf}{x}',$$ and $$\hat{u}_1(\xi,x_3)=-i\varphi(\xi,x_3),\;\; \hat{u}_2(\xi,x_3)=-i\theta(\xi,x_3),\;\; \hat{u}_3(\xi,x_3)=\psi(\xi,x_3).$$ Then $\xi_1\varphi+\xi_2\theta+\psi'=0$ because of $\mathrm{div}\mathbf{u}=0$. Moreover, $$\widehat{\nabla \mathbf{u}}=(\widehat{\partial_i u_j})=\left( \begin{array}{ccc} \xi_1\varphi & \xi_1\theta & i\xi_1\psi \\ \xi_2\varphi & \xi_2\theta & i\xi_2\psi \\ -i\varphi' & -i\theta' & \psi' \\ \end{array} \right).$$ By the Fubini and Parseval theorems (see [@grafakos2008classical Proposition 3.1.16]), we have (see [@GYTI2 Section 3.3]) $$\label{0305}\begin{aligned} \int_{\Omega} g\bar{\rho}'\|u_3\|^2{\mathrm}{d}{\mathbf}{x} =&\frac{1}{4\pi^2 L^2}\sum_{\xi\in (L^{-1}\mathbb{Z})^2}\int_{\mathbb{R}}g(\bar{\rho}'1_{\{\bar{\rho}'\geq 0\}} +\bar{\rho}'1_{\{\bar{\rho}'< 0\}})\|\hat{u}_3\|^2{\mathrm}{d}x_3\\ = &\frac{1}{4\pi^2 L^2}\sum_{\xi\in (L^{-1}\mathbb{Z})^2}\int_{\mathbb{R}}g(\bar{\rho}'1_{\{\bar{\rho}'\geq 0\}} +\bar{\rho}'1_{\{\bar{\rho}'< 0\}})\|\psi\|^2{\mathrm}{d}x_3\\ =&\frac{1}{4\pi^2 L^2}\sum_{\xi\in (L^{-1}\mathbb{Z})^2}\int_{\mathbb{R}}g\bar{\rho}'\|\psi\|^2{\mathrm}{d}x_3, \end{aligned}$$ where $1_{\{\bar{\rho}'\geq 0\}}$ and $1_{\{\bar{\rho}'< 0\}}$ denote the characteristic functions. Obviously, $$\label{additional0305}\begin{aligned} \lim_{\|\xi\|\rightarrow 0}\int_{\mathbb{R}}g\bar{\rho}'\|\psi\|^2{\mathrm}{d}x_3=\int_{\mathbb{R}}g\bar{\rho}'\|\psi(0,x_3)\|^2{\mathrm}{d}x_3. \end{aligned}$$ In addition, we can control the following terms involved with $\psi$ as follows. $$\begin{aligned} &\frac{1}{4\pi^2 L^2}\left[\liminf_{\|\xi\|\rightarrow 0}\int_{\mathbb{R}}\frac{\xi_1^2}{\|\xi\|^2}(\|{\xi}\|^2\|\psi\|^2 +\|\psi'\|^2){\mathrm}{d}x_3+\sum_{\xi\in (L^{-1}\mathbb{Z})^2\setminus\{{\mathbf}{0}\}}\int_{\mathbb{R}}\frac{\xi_1^2}{\|\xi\|^2}(\|{\xi}\|^2\|\psi\|^2 +\|\psi'\|^2){\mathrm}{d}x_3\right]\\ &=\frac{1}{4\pi^2 L^2}\sum_{\xi\in (L^{-1}\mathbb{Z})^2\setminus\{{\mathbf}{0}\}}\int_{\mathbb{R}}\frac{\xi_1^2}{\|\xi\|^2}(\|\xi\|^2\|\psi\|^2+ \|\xi_1\varphi+\xi_2\theta\|^2){\mathrm}{d}x_3\\ &\leq\frac{2}{4\pi^2 L^2}\sum_{\xi\in (L^{-1}\mathbb{Z})^2}\int_{\mathbb{R}}\|\xi_1\|^2 (\|\varphi\|^2+\|\theta\|^2+\|\psi\|^2){\mathrm}{d}x_3\\ &=\frac{2}{4\pi^2 L^2}\sum_{\xi\in (L^{-1}\mathbb{Z})^2}\sum_{j=1}^3\int_{\mathbb{R}} \|\widehat{\partial_1{u}_j}\|^2{\mathrm}{d}x_3= 2\int_{\Omega} \|{\partial_1 \mathbf{u}}\|^2{\mathrm}{d}{\mathbf}{x}, \end{aligned}$$ $$\begin{aligned} &\frac{1}{4\pi^2 L^2}\left[\liminf_{\|\xi\|\rightarrow 0}\int_{\mathbb{R}}\frac{(\|{\xi}\|^2\|\psi'\|^2 +\|\psi''\|^2)}{\|\xi\|^2} {\mathrm}{d}x_3+ \sum_{\xi\in (L^{-1}\mathbb{Z})^2\setminus\{{\mathbf}{0}\}}\int_{\mathbb{R}}\frac{(\|{\xi}\|^2\|\psi'\|^2 +\|\psi''\|^2)}{\|\xi\|^2} {\mathrm}{d}x_3\right]\\ &\leq\frac{2}{4\pi^2 L^2}\sum_{\xi\in (L^{-1}\mathbb{Z})^2}\int_{\mathbb{R}} (\|\varphi'\|^2+\|\theta'\|^2+\|\psi'\|^2) {\mathrm}{d}x_3= 2\int_{\Omega} \|{\partial_3\mathbf{u}}\|^2{\mathrm}{d}{\mathbf}{x}, \end{aligned}$$ $$\begin{aligned} & \frac{1}{4\pi^2 L^2}\left[\limsup_{\|\xi\|\rightarrow 0}\int_{\mathbb{R}}\frac{\bar{\rho} (\|{\xi}\|^2\|\psi\|^2+\|\psi'\|^2)}{\|\xi\|^2}{\mathrm}{d}x_3 +\sum_{\xi\in (L^{-1}\mathbb{Z})^2\setminus\{{\mathbf}{0}\}}\int_{\mathbb{R}}\frac{\bar{\rho} (\|{\xi}\|^2\|\psi\|^2+\|\psi'\|^2)}{\|\xi\|^2}{\mathrm}{d}x_3 \right] \\ & \leq \frac{2}{4\pi^2 L^2}\sum_{\xi\in (L^{-1}\mathbb{Z})^2}\int_{\mathbb{R}}\bar{\rho} (\|\psi\|^2+\|\varphi\|^2+\|\theta\|^2){\mathrm}{d}x_3 =2\int_{\Omega}\bar{\rho} \|{\mathbf}{u}\|^2 \mathrm{d}{\mathbf}{x},\end{aligned}$$ and $$\label{0308}\begin{aligned} &\frac{1}{4\pi^2 L^2}\left[\limsup_{\|\xi\|\rightarrow 0}\int_{\mathbb{R}}\frac{(4\|{\xi}\|^2\|\psi'\|^2 +\|\|{\xi}\|^2\psi+\psi''\|^2)}{\|\xi\|^2}{\mathrm}{d}x_3 \right.\\ &\qquad\qquad \left.+\sum_{\xi\in (L^{-1}\mathbb{Z})^2\setminus\{{\mathbf}{0}\}}\int_{\mathbb{R}}\frac{(4\|{\xi}\|^2\|\psi'\|^2 +\|\|{\xi}\|^2\psi+\psi''\|^2)}{\|\xi\|^2}{\mathrm}{d}x_3\right]\\ &=\frac{1}{4\pi^2 L^2} \left[\limsup_{\|\xi\|\rightarrow 0}\int_{\mathbb{R}} \frac{2\|{\xi}\|^2\|\psi'\|^2+\|{\xi}\|^4\|\psi\|^2+\|\psi''\|^2}{\|\xi\|^2}{\mathrm}{d}x_3 \right.\\ &\qquad\qquad \left.+\sum_{\xi\in (L^{-1}\mathbb{Z})^2\setminus\{{\mathbf}{0}\}}\int_{\mathbb{R}} \frac{2\|{\xi}\|^2\|\psi'\|^2+\|{\xi}\|^4\|\psi\|^2+\|\psi''\|^2}{\|\xi\|^2}{\mathrm}{d}x_3\right] \\ &\leq \frac{2}{4\pi^2 L^2}\sum_{\xi\in (L^{-1}\mathbb{Z})^2}\int_{\mathbb{R}} \left[\|\xi\|^2(\|\varphi\|^2+\|\theta\|^2+\|\psi\|^2)+\|\varphi'\|^2+\|\theta'\|^2+\|\psi'\|^2\right]{\mathrm}{d}x_3\\ & = \frac{2}{4\pi^2 L^2}\sum_{\xi\in (L^{-1}\mathbb{Z})^2} \sum_{1\leq i,j\leq 3} \int_{\mathbb{R}} \|\widehat{\partial_i {u}_j}\|^2{\mathrm}{d}x_3= 2\int_{\Omega} \|{\nabla\mathbf{u}}\|^2{\mathrm}{d}{\mathbf}{x}. \end{aligned}$$ Now, let $$Z(\psi,\xi)= \int_{\mathbb{R}}g\bar{\rho}'\|\psi\|^2\mathrm{d}x_3-M^2\left\{ \begin{array}{ll} \int_{\mathbb{R}}\xi_1^2\left(\|\psi\|^2+\frac{\|\psi'\|^2}{\|\xi\|^2}\right){\mathrm}{d}x_3, & \hbox{ for }\bar{{\mathbf}{M}}={{M}}e_1; \\[3mm] \int_{\mathbb{R}}\left(\|\psi'\|^2+\frac{\|\psi''\|^2}{\|\xi\|^2}\right){\mathrm}{d}x_3, & \hbox{ for }\bar{{\mathbf}{M}}={M}e_3. \end{array}\right.$$ By splitting $Z(\psi,\xi)=Z(\mathfrak{R}(\psi),\xi)+Z(\mathfrak{I}(\psi),\xi)$, we may reduce the boundedness of $Z(\psi,\xi)$ to the boundedness of $Z(\mathfrak{R}(\psi),\xi)$, since the imaginary part can be dealt with in the same manner. Hence, without loss of generality, assume that $\psi$ is real-valued. For ${\xi}\in \mathbb{A}^{\mathrm}{g}\cap (L^{-1}\mathbb{Z})^2\neq \emptyset$, we make use of and with $s=\lambda({\xi})$ to deduce that $$\label{0311}\begin{aligned} Z(\psi,\xi)\leq & \frac{\lambda^2({\xi})}{\|{\xi}\|^2} \int_{\mathbb{R}}\bar{\rho}(\|{\xi}\|^2\|\psi\|^2+ \|\psi'\|^2)\mathrm{d}x_3+\frac{\mu\lambda({\xi})}{\|{\xi}\|^2} \int_{\mathbb{R}}\left(4\|{\xi}\|^2\|\psi'\|^2 +\|\|{\xi}\|^2\psi+\psi''\|^2\right)\mathrm{d}x_3\\ \leq & \frac{\Lambda}{\|{\xi}\|^2}\left[\Lambda\int_{\mathbb{R}}\bar{\rho}(\|{\xi}\|^2\|\psi\|^2+ \|\psi'\|^2)\mathrm{d}x_3 +\mu\int_{\mathbb{R}}(4\|{\xi}\|^2\|\psi'\|^2 +\|\|{\xi}\|^2\psi+\psi''\|^2)\mathrm{d}x_3\right]. \end{aligned}$$ On the other hand, by Proposition \[pro:0203\], we infer that for any $\xi\not\in\mathbb{A}^{{\mathrm}{g}}\setminus\{{\mathbf}{0}\}$, $$\label{xion0312} Z(\psi,\xi)\leq 0\leq \frac{\Lambda}{\|{\xi}\|^2}\left[\Lambda\int_{\mathbb{R}}\bar{\rho}(\|{\xi}\|^2\|\psi\|^2+ \|\psi'\|^2)\mathrm{d}x_3 +\mu\int_{\mathbb{R}}(4\|{\xi}\|^2\|\psi'\|^2 +\|\|{\xi}\|^2\psi+\psi''\|^2)\mathrm{d}x_3\right].$$ Consequently, putting – together, we can conclude that holds. In fact, taking $i=1$ for example, we have $$\begin{aligned} & 2\int_{\Omega}\left( g\bar{\rho}'\|u_3\|^2-M^2\|\partial_1{\mathbf}{u}\|^2\right){\mathrm}{d}{\mathbf}{x}\leq \frac{1}{4\pi^2 L^2}\sum_{\xi\in (L^{-1}\mathbb{Z})^2}\int_{\mathbb{R}}g\bar{\rho}'\|\psi\|^2{\mathrm}{d}x_3\\ &- \frac{M^2 }{4\pi^2 L^2}\left[\liminf_{\|\xi\|\rightarrow 0}\int_{\mathbb{R}}\frac{\xi_1^2}{\|\xi\|^2}(\|{\xi}\|^2\|\psi\|^2 +\|\psi'\|^2){\mathrm}{d}x_3+\sum_{\xi\in (L^{-1}\mathbb{Z})^2\setminus\{{\mathbf}{0}\}}\int_{\mathbb{R}}\frac{\xi_1^2}{\|\xi\|^2}(\|{\xi}\|^2\|\psi\|^2 +\|\psi'\|^2){\mathrm}{d}x_3\right] \\& =\frac{1}{4\pi^2 L^2}\left\{\limsup_{\|\xi\|\rightarrow 0}\int_{\mathbb{R}}\left[g\bar{\rho}'\|\psi\|^2-\frac{\xi_1^2}{\|\xi\|^2}(\|{\xi}\|^2\|\psi\|^2 +\|\psi'\|^2)\right]{\mathrm}{d}x_3+\sum_{\xi\in (L^{-1}\mathbb{Z})^2\setminus\{{\mathbf}{0}\}} Z(\psi,\xi)\right\} \\ &\leq \frac{1}{4\pi^2 L^2} \left\{\limsup_{\|\xi\|\rightarrow 0}\frac{\Lambda}{\|{\xi}\|^2}\left[\Lambda\int_{\mathbb{R}}\bar{\rho}(\|{\xi}\|^2\|\psi\|^2+ \|\psi'\|^2)\mathrm{d}x_3 +\mu\int_{\mathbb{R}}(4\|{\xi}\|^2\|\psi'\|^2 +\|\|{\xi}\|^2\psi+\psi''\|^2)\mathrm{d}x_3\right]\right.\\ &\left.+\sum_{\xi\in (L^{-1}\mathbb{Z})^2\setminus\{{\mathbf}{0}\}}\frac{\Lambda}{\|{\xi}\|^2} \left[\Lambda\int_{\mathbb{R}}\bar{\rho}(\|{\xi}\|^2\|\psi\|^2 + \|\psi'\|^2)\mathrm{d}x_3 +\mu\int_{\mathbb{R}}(4\|{\xi}\|^2\|\psi'\|^2 +\|\|{\xi}\|^2\psi+\psi''\|^2)\mathrm{d}x_3\right] \right\} \\ & \leq 2{\Lambda^2} \int_{\mathbb{R}^2}\rho\|{\mathbf}{u}\|^2{\mathrm}{d}{\mathbf}{x}+ 2{\Lambda} \mu\int_{\Omega}\|\nabla\mathbf{u}\|^2{\mathrm}{d}{\mathbf}{x},\end{aligned}$$ which gives for $i=1$. Step 2: Proof of the error estimate In what follows, we still denote by $C$ a generic positive constant which may depend on physical parameters and the known functions $\bar{\varrho}_0$, $\bar{\mathbf{u}}_0$, and $\bar{\mathbf{N}}_0$. We differentiate $_2$ with $\bar{{\mathbf}{M}}={M}e_i$ ($i=1,3$) in time, multiply the resulting equation by ${\mathbf}{u}_t$ and integrate by part over $\Omega$. Then, using the first and third equations in , we obtain $$\label{stpe20420} \begin{aligned} &\frac{d}{dt} \int_{\Omega}\left(\bar{\rho}\|\mathbf{u}_t\|^2+M^2\|\partial_i{\mathbf}{u}\|^2 -g\bar{\rho}'{u}_3^2\right)\mathrm{d}\mathbf{x}+ 2\mu\int_{\Omega}\|\nabla\partial_t\mathbf{u}\|^2{\mathrm}{d}{\mathbf}{x}\\ &=\int_\Omega \{{\mathbf}{G}_t^\delta+[(\nabla \times {\mathbf}{H}^\delta)\times \bar{{\mathbf}{M}}]-gF^\delta {\mathbf}{e}_3\}\cdot {\mathbf}{u}_t^{{\mathrm}{d}}{\mathrm}{d}{\mathbf}{x}. \end{aligned}$$ Here, utilizing , and , we see that $$\begin{aligned} \left\|\int_\Omega \{{\mathbf}{G}_t^\delta+[(\nabla \times {\mathbf}{H}^\delta)\times \bar{{\mathbf}{M}}] -gF^\delta {\mathbf}{e}_3\}\cdot {\mathbf}{u}_t^{{\mathrm}{d}}{\mathrm}{d}{\mathbf}{x}\right\| \leq C(\delta e^{\Lambda^* t})^{3}. \end{aligned}$$ Noting that $\\|{\mathbf}{u}_t(0)\\|\leq C\delta^3$, we integrate from $0$ to $t$ and use to infer that $$\label{0314} \begin{aligned} \\|\sqrt{\bar{\rho}} \mathbf{u}_t(t)\\|^2_{L^2}+2\mu\int_0^t\\|\nabla\partial_s\mathbf{u}(s)\\|^2_{L^2}{\mathrm}{d}s \leq {\Lambda^2} \\|\sqrt{\bar{\rho}}{\mathbf}{u}(t)\\|_{L^2}^2+ {\Lambda}\mu\\|\nabla\mathbf{u}(t)\\|^2_{L^2}+C(\delta e^{\Lambda^* t})^{3}. \end{aligned}$$ The following inequality follows easily from integration in time and Cauchy-Schwarz’s inequality. $$\begin{aligned}\label{0316} \Lambda\\|\nabla\mathbf{u}(t)\\|^2_{L^2}=& 2\Lambda\int_0^t \int_{\Omega}\nabla\partial_s\mathbf{u}(s):\nabla\mathbf{u}(s){\mathrm}{d}{\mathbf}{x}\mathrm{d}s\\ \leq &\int_0^t\\|\nabla\partial_s\mathbf{u}(s)\\|^2_{L^2}\mathrm{d}s +\Lambda^2\int_0^t\\|\nabla\mathbf{u}(s)\\|^2_{L^2}\mathrm{d}s. \end{aligned}$$ On the other hand, $$\begin{aligned}\label{0317} \Lambda\partial_t\\|\sqrt{\bar{\rho}}\mathbf{u}(t)\\|^2_{L^2}=2\Lambda\int_{\Omega} \bar{\rho}\mathbf{u}\cdot\partial_t\mathbf{u}(t){\mathrm}{d}{\mathbf}{x}\leq\\|\sqrt{\bar{\rho}}\partial_t\mathbf{u}(t)\\|^2_{L^2} +\Lambda^2\\|\sqrt{\bar{\rho}}\mathbf{u}(t)\\|^2_{L^2}. \end{aligned}$$ Hence, putting – together, we obtain the differential inequality: $$\partial_t\\|\sqrt{\bar{\rho}}\mathbf{u}(t)\\|^2_{L^2}+\mu\\|\nabla\mathbf{u}(t)\\|^2_{L^2} \leq 2\Lambda\left(\\|\sqrt{\bar{\rho}}\mathbf{u}(t)\\|^2_{L^2}+\mu\int_0^t\\|\nabla\mathbf{u}(s)\\|^2_{L^2} \mathrm{d}s\right)+C(\delta e^{\Lambda^* t})^{{3}}.$$ An application of Gronwall’s inequality then implies that $$\begin{aligned}\label{0319} \\|\sqrt{\bar{\rho}}\mathbf{u}(t)\\|^2_{L^2}+\mu\int_0^t\\|\nabla\mathbf{u}(s)\\|^2_{L^2}{\mathrm}{d}s \leq C \int_0^t(\delta e^{\Lambda^* s})^{{3}}e^{2\Lambda(t-s)} {\mathrm}{d}s\leq C(\delta e^{\Lambda^* t})^{{3}} \end{aligned}$$ for any $t\geq 0$, where we have used the fact that $3\Lambda^-2\Lambda>0$. Thus, making use of , and , we deduce that $$\label{nablaestime1357}\begin{aligned} \frac{1}{\Lambda}\\|\sqrt{\bar{\rho}} \mathbf{u}_t(t)\\|^2_{L^2}+ \mu\\|\nabla\mathbf{u}(t)\\|^2_{L^2} \leq {\Lambda}\\|\sqrt{\bar{\rho}}\mathbf{u}(t)\\|^2_{L^2}+2 {\Lambda}\mu\int_0^t\\|\nabla\mathbf{u}(s)\\|^2_{L^2}{\mathrm}{d}s \leq C(\delta e^{\Lambda^ t})^{{3}}. \end{aligned}$$ Now we use , and $_1$ to arrive at $$\begin{aligned}\label{0323} \\|\varrho(t)\\|_{X}\leq \int_0^t\\|\varrho_s(s)\\|_{X}{\mathrm}{d}s \leq \\|\bar{\rho}'\\|_{L^\infty}\int_0^t\\|u_3(s)\\|_{X}{\mathrm}{d}s\leq C (\delta e^{\Lambda^* t})^{\frac{3}{2}}, \end{aligned}$$ where $X=L^2$ or $H^1$. Similarly to the derivation of , one obtains $$\begin{aligned} \\|{\mathbf}{N}(t)\\|_{L^2}\leq& \int_0^t\\|{\mathbf}{N}_s(s)\\|_{L^2}{\mathrm}{d}s \leq \\|\bar{{\mathbf}{M}}\\|_{L^\infty}\int_0^t\\|\nabla {\mathbf}{u}(s)\\|_{L^2}{\mathrm}{d}s\leq C(\delta e^{\Lambda^* t})^{\frac{3}{2}}. \end{aligned}$$ Thus, putting the above four estimates together, we immediately get . This completes the proof of the lemma. $\Box$ \[rem0401n\] Slightly modifying the above proof, we can easily show that any linear solution $(\varrho,{\mathbf}{u},{\mathbf}{N},q)$, enjoying proper regularity, of with $\bar{{\mathbf}{M}}={M}e_i$ ($i=1,3$) satisfies the following estimates for any $t\geq 0$: $$\begin{aligned} & \\|\varrho(t)\\|_{X}^2\leq Ce^{2\Lambda t}(\\|\varrho_0\\|_{X}^2+\\|\mathbf{u}_0\\|_{H^2}^2+\\|\nabla{\mathbf}{N}_0\\|_{L^2}),\quad X=L^2\mbox{ or }H^1,\nonumber\\[1mm] & \\|\mathbf{u}(t)\\|_{H^1 }^2+\\|\mathbf{u}_t(t)\\|^2_{L^2 }+ \int_0^t\\|\nabla\mathbf{u}(s)\\|^2_{L^2}{\mathrm}{d}s\leq Ce^{2\Lambda t}(\\|\varrho_0\\|_{L^2}^2+\\|\mathbf{u}_0\\|_{H^2}^2+\\|\nabla {\mathbf}{N}_0\\|_{L^2}^2),\nonumber\\[1mm] &\\|{\mathbf}{N}(t)\\|_{L^2}\leq Ce^{\Lambda t}(\\|\varrho_0\\|_{L^2}+\\|\mathbf{u}_0\\|_{H^2}+\\|{\mathbf}{N}_0\\|_{H^1}),\nonumber\end{aligned}$$ where the constant $C$ depends on $\bar{\rho}$, $\mu$, $\Lambda$ and $\bar{{\mathbf}{M}}$. Now, we claim that $$\label{n0508} T^\delta=\min\left\{T^\delta,T^,T^{}\right\},$$ provided that small $\varepsilon_0$ is taken to be $$\label{defined} \varepsilon_0=\min\left\{\frac{{\delta_0}} {4\sqrt{2C_2}},\frac{C_1^2}{3C_3^2}, \frac{m_0^2}{8C_3^2} \right\}.$$ Indeed, if $T^=\min\{T^{\delta},T^,T^{}\}$, then $T^<\infty$; moreover, by and we have $$\left\\|\left(\varrho^\delta, {\mathbf{u}}^\delta,{\mathbf{N}}^\delta \right)(T^)\right\\|_{\mathcal{E}}\leq \sqrt{2C_2}\delta e^{\Lambda^ T^} \leq \sqrt{2C_2}\delta e^{\Lambda^ T^\delta}=2\sqrt{2C_2}\varepsilon_0<{\delta_0},$$ which contradicts with $_1$. On the other hand, if $T^{*}=\min\{T^{\delta},T^,T^{}\}$, then $T^{}<T^{\max}$. Moreover, in view of , and , we have $$\begin{aligned} \left\\|\left( \varrho^\delta, {\mathbf{u}}^\delta,{\mathbf{N}}^\delta \right)(T^{})\right\\|_{L^2} \leq & \left\\|\left( \varrho^{\mathrm}{a}_{\delta}, {\mathbf{u}}^{\mathrm}{a}_{\delta},{\mathbf{N}}^{\mathrm}{a}_{\delta} \right)(T^{})\right\\|_{L^2} +\left\\|\left( \varrho^{\mathrm{d}}, {\mathbf{u}}^{\mathrm{d}},{\mathbf{N}}^{\mathrm{d}} \right)(T^{})\right\\|_{L^2} \\ \leq &\delta \left\\|\left( \varrho^{\mathrm}{l}, {\mathbf{u}}^{{\mathrm}{l}},{\mathbf{N}}^{\mathrm}{l} \right)(T^{})\right\\|_{L^2}+C_3(\delta e^{\Lambda^* T^{*}})^{\frac{3}{2}} \\ \leq & \delta C_1e^{\Lambda^ T^{*}}+C_3(\delta e^{\Lambda^ T^{*}})^{\frac{3}{2}} \leq \delta e^{\Lambda^ T^{*}}(C_1+C_3\sqrt{2\varepsilon_0})\\ <&2\delta C_1 e^{\Lambda^ T^{*}}, \end{aligned}$$ which also contradicts with $_2$. Therefore, holds. Finally, we again use , and to deduce that $$\begin{aligned} \\|u_3^{\delta}(T^\delta)\\|_{L^2}\geq & \\|u^{\mathrm{a}}_{\delta 3}(T^{\delta})\\|_{L^2}-\\|u_3^{{\mathrm}{d}}(T^{\delta})\\|_{L^2} = \delta\\|u_3^{\mathrm{l}}(T^{\delta})\\|_{L^2}-\\|u_3^{{\mathrm}{d}}(T^{\delta})\\|_{L^2} \\ \geq & m_0\delta e^{\Lambda^ T^\delta} - C_3(\delta e^{\Lambda^* T^{\delta}})^{\frac{3}{2}} \geq 2m_0\varepsilon_0-C_3(2\varepsilon_0)^{\frac{3}{2}} \geq m_0{\varepsilon_0}, \end{aligned}$$ where $u^{\mathrm{a}}_{\delta 3}$ and $u_3^{{\mathrm}{d}}(T^{\delta})$ denote the third component of ${\mathbf}{u}^{\mathrm{a}}_{\delta}$ and ${\mathbf}{u}^{{\mathrm}{d}}(T^{\delta})$, respectively. Similarly, we have $$\\|\varrho(T^\delta)\\|_{L^2(\Omega)},\ \\|({u}_1,u_2)(T^\delta)\\|_{L^2(\Omega)} \geq m_0{\varepsilon_0}.$$ Moreover, $$\\|N_3(T^\delta)\\|_{L^2(\Omega)}\geq m_0\varepsilon_0\quad\mbox{ if }\; \bar{\bf{M}}=M{\mathbf}{e}_3.$$ Finally, we take $\varepsilon=m_0\varepsilon_0$ and thus obtain Theorem \[thm:0101\] immediately. Proof of the local well-posedness {#appendix} ================================= In this section, we adapt the standard method to briefly show the local well-posedness for the problem – by three steps, in which we shall exploit some mathematical techniques to deal with the horizontally periodic domain. Firstly we solve the following linearized problem for given $\mathbf{v}$: $$\label{0501problem}\left\{\begin{array}{l} \varrho_t+{\mathbf}{v}\cdot\nabla \varrho=-{\mathbf}{v}\cdot\nabla \bar{\rho}, \\[1mm] (\varrho+\bar{\rho}){\bf u}_t+\nabla\tilde{q}-\mu \Delta \mathbf{u}= ({\mathbf}{N}+\bar{{\mathbf}{M}})\cdot\nabla {\mathbf}{N} -g \varrho {\mathbf}{e}_3-(\varrho+\bar{\rho}){\bf v}\cdot\nabla {\mathbf}{v},\\[1mm] {\mathbf}{N}_t+{\mathbf}{v}\cdot\nabla {\mathbf}{N}-\nabla {\mathbf}{v}{\mathbf}{N}=\bar{{\mathbf}{M}}\cdot\nabla {\mathbf}{v},\\[1mm] \mathrm{div}\mathbf{u}=0,\ \mathrm{div}\mathbf{N}=0\end{array}\right.$$ with initial and boundary conditions: $$\label{bounndar0502} \begin{aligned} & (\varrho,\mathbf u,\mathbf{N} )\|_{t=0}=(\varrho_0,{\mathbf u}_0,\mathbf{N}_0)\in H^3\times H^4\times H^3,\\ & \lim_{\|{x}_3\|\rightarrow +\infty}(\varrho,{\bf u},{\mathbf}{N})(t,\mathbf{x}',x_3)={\bf 0}\;\mbox{ for any }t>0, \end{aligned}$$ where $({\mathbf}{u}_0,{\mathbf}{N}_0)$ satisfies the compatibility conditions $\mathrm{div}\mathbf{u}_0=0$ and $\mathrm{div}\mathbf{N}_0=0$. Secondly, we use the technique of iteration to construe a sequence of solutions $\{(\varrho^k,{\mathbf}{u}^k,{\mathbf}{N}^k)\}_{k=1}^\infty$, in which $(\varrho^k,{\mathbf}{u}^k,{\mathbf}{N}^k)$ solves the above linearized problem with $(\varrho^k,{\mathbf}{u}^k,{\mathbf}{N}^k)$ in place of $(\varrho,{\mathbf}{u},{\mathbf}{N})$ and ${\mathbf}{u}^{k-1}$ in place of ${\mathbf}{v}$, and show the uniform boundedness of the solution sequence in some function space. Finally, we further prove that the sequence of solutions is a Cauchy sequence, and thus the limit function is the unique solution of the original problem. Throughout the rest of this article we shall repeatedly use the abbreviations: $$\begin{aligned} & H^1_\sigma:=\{{\mathbf}{u}\in H^1~\|~{\mathrm}{div}{\mathbf}{u}=0\},\ Q_T:=I_T\times \Omega,\\ & {V}_T:=\{{\mathbf}{v}\in C^0(\bar{I}_T,H^4)~\|~ {\mathbf}{v}_t\in C^0(\bar{I}_T,H^2),\ {\mathbf}{v}_{tt}\in C^0({I}_T,L^2),\\ &\qquad \qquad \qquad \qquad \qquad \quad {\mathbf}{v}_{t}\in L^2({I}_T,H^3),\ {\mathbf}{v}_{tt}\in L^2({I}_T,H^1),\ {\mathrm}{div}{\mathbf}{v}=0\},\\ & {H}_T:=\{(\varrho,{\mathbf}{N})\in C^0(\bar{I}_T,H^3)~\|~(\varrho_t,{\mathbf}{N}_t)\in C^0(\bar{I}_T,H^2),\ (\varrho_{tt},{\mathbf}{N}_{tt})\in C^0(\bar{I}_T,H^1)\},\\ & {F}_T:=\{{\mathbf}{f}\in C^0(\bar{I}_T,H^2)~\|~{\mathbf}{f}_{t}\in C^0(\bar{I}_T,L^2)\cap L^2({I}_T,H^1),\ {\mathbf}{f}_{tt}\in L^2({I}_T,L^2) \}. \end{aligned}$$ Unique solvability of linearized problems ----------------------------------------- To show the solvability of the linearized problem –, it suffices to solve the following two problems $$\label{hypobolic}\left\{\begin{array}{l} \varrho_t+{\mathbf}{v}\cdot\nabla \varrho=-{\mathbf}{v}\cdot\nabla \bar{\rho}, \\[1mm] {\mathbf}{N}_t+{\mathbf}{v}\cdot\nabla {\mathbf}{N}-\nabla {\mathbf}{v}{\mathbf}{N}=\bar{{\mathbf}{M}}\cdot\nabla {\mathbf}{v},\\ \mathrm{div}\mathbf{N}=0\end{array}\right.$$ with initial and boundary conditions $$\label{intbouadnry51} \begin{aligned} & (\varrho,\mathbf{N} )\|_{t=0}=(\varrho_0, \mathbf{N}_0)\in H^3\times H^3 ,\\ & \lim_{\|{x}_3\|\rightarrow +\infty}(\varrho,{\mathbf}{N})(t,\mathbf{x}',x_3)={\bf 0}\;\;\mbox{ for any }t>0, \end{aligned}$$ and $$\label{loc0503jjw} \left\{ \begin{array}{ll} (\varrho+\bar{\rho}){\bf u}_t+\nabla \tilde{q}-\mu \Delta \mathbf{u}={\mathbf}{f}, \\ \mathrm{div}\mathbf{u}=0 \end{array} \right.$$ with initial and boundary conditions $$\label{intabioudr502} \mathbf{u}\|_{t=0}={\mathbf u}_0\in H^4,\ \lim_{\|{x}_3\|\rightarrow +\infty}{\mathbf}{u}(t,\mathbf{x}',x_3)={\bf 0}\ \mbox{ for any }t>0,$$ where ${\mathbf}{f}:=({\mathbf}{N}+\bar{{\mathbf}{M}})\cdot\nabla {\mathbf}{N}-g \varrho {\mathbf}{e}_3- (\varrho+\bar{\rho}){\bf v}\cdot\nabla{\mathbf}{v}$. By the standard hyperbolic theory, we have the following result on the existence of a unique solution to the first problem –: \[lem:0501\]Let ${\mathbf}{v}\in {V}_T$, $(\varrho_0, {\mathbf}{N}_0)\in {\mathbf}{H}^3\times {\mathbf}{H}^3$ and ${\mathrm}{div}{\mathbf}{N}_0=0$, then, for any $T>0$, the problem – possesses a unique classical solution $(\varrho,{\mathbf}{N})\in {H}_T$ emanating from the initial data $(\varrho_0, {\mathbf}{N}_0)$. Moreover, $(\varrho,{\mathbf}{N})$ satisfies the following identities: $$\begin{aligned} &\label{fracrho}\frac{1}{2}\frac{d}{dt}\\|\partial^{\alpha}\varrho\\|_{L^2}^2 =-\int_\Omega[\partial^\alpha({\mathbf}{v}\cdot\nabla \bar{\rho})+\partial^{\alpha}( {\mathbf}{v}\cdot\nabla \varrho)]\partial^\alpha\varrho{\mathrm}{d}{\mathbf}{x},\\ &\label{fracrho12}\frac{1}{2}\frac{d}{dt}\\|\partial^{\alpha}{\mathbf}{N}\\|_{L^2}^2=\int_\Omega[\partial^\alpha(\bar{{\mathbf}{M}}\cdot \nabla{\mathbf}{v}+\nabla{\mathbf}{v}{\mathbf}{N})-\partial^{\alpha}({\mathbf}{v}\cdot \nabla{\mathbf}{N}) ]\partial^\alpha{\mathbf}{N}{\mathrm}{d}{\mathbf}{x},\end{aligned}$$ where we have defined that $${\mathbf}{v}\cdot \partial^{\alpha}\nabla\varrho=0\mbox{ and }{\mathbf}{v}\cdot \partial^{\alpha}\nabla{\mathbf}{N}=0\;\;\mbox{ for any }\|\alpha\|=3.$$ Following the proof of [@NASII04 Theorem 2.16], we can easily check that [@NASII04 Theorem 2.16] still holds with the horizontally periodic domain $\Omega$ in place of $\mathbb{R}^3$. Thus, we immediately get the unique solvability of the problem –. The strong continuity and can be shown by adapting the idea in the proof of [@NASII04 Lemm 6.9]. Here we give the proof for the reader’s convenience. Without a generality, we assume that $\alpha=(3,0,0)$. Similarly in [@NASII04 Lemm 6.9], we can deduce from that $$\begin{aligned}&\partial_t S_\varepsilon^2(\partial^3_{x_1}\varrho)+{\mathrm}{div}(S_\varepsilon^2(\partial^3_{x_1}\varrho){\mathbf}{v})-2r_\varepsilon S_\varepsilon(\partial^3_{x_1}\varrho)\\ &= -2 S_\varepsilon(\partial^3_{x_1}({\mathbf}{v}\cdot\nabla \bar{\rho})+\partial_{x_1}^2(\partial_{x_1}{\mathbf}{v}\cdot\nabla \varrho)+\partial_{x_1}^2{\mathbf}{v}\cdot \partial_{x_1}\nabla \varrho+2\partial_{x_1}{\mathbf}{v}\cdot \partial_{x_1}^2\nabla\varrho )S_\varepsilon(\partial^3_{x_1}\varrho) \end{aligned}$$ a.e. in $Q_T$, where $S_\varepsilon$ is a standard mollifier, and $r_\varepsilon ={\mathrm}{div}(S_\varepsilon(\partial^3_{x_1}\varrho){\mathbf}{v})-{\mathrm}{div}(S_\varepsilon(\partial^3_{x_1}\varrho{\mathbf}{v}))\to 0$ in $L^2(I_T,K)$ for any bounded domain $K\subset \Omega$. Note that $S_\varepsilon(f)\in H^1\cap C^\infty$ if $f\in H^1$. Let $\psi\in C_0^\infty(I_T)$ and $$\label{phconstctin}\begin{aligned} &\phi_n(x_3)=1\mbox{ for }\|x_3\|<n,\ \phi_n(x_3)=0\mbox{ for }\|x_3\|>2n,\\ &0\leq \phi_n(x_3)\leq 1,\ \|\phi_n'(x_3)\|\leq c/n\mbox{ for any }x_3\in (-2n,-n)\cup(n,2n), \end{aligned}$$ where the constant $c$ is independent of $n$. Multiplying $_1$ by $\phi_n\psi\in C_0^\infty(I_T)$, and integrating the resulting equality, we get $$\begin{aligned}&\int_0^T\psi_t\int_\Omega S_\varepsilon^2(\partial^3_{x_1}\varrho)\phi_n{\mathrm}{d}{\mathbf}{x}{\mathrm}{d}t +\int_0^T\psi\int_\Omega S_\varepsilon^2(\partial^3_{x_1}\varrho)v_3\phi_n'{\mathrm}{d}{\mathbf}{x}{\mathrm}{d}t +2\int_0^T\psi\int_\Omega r_\varepsilon S_\varepsilon(\partial^3_{x_1}\varrho)\phi_n{\mathrm}{d}{\mathbf}{x}{\mathrm}{d}t\\ &=2\int_0^T\psi\int_\Omega S_\varepsilon(\partial^3_{x_1}({\mathbf}{v}\cdot\nabla \bar{\rho})+\partial_{x_1}^2(\partial_{x_1}{\mathbf}{v}\cdot\nabla \varrho) +\partial_{x_1}^2{\mathbf}{v}\cdot \partial_{x_1}\nabla \varrho+2\partial_{x_1}{\mathbf}{v}\cdot \partial_{x_1}^2\nabla\varrho )S_\varepsilon(\partial^3_{x_1}\varrho)\phi_n{\mathrm}{d}{\mathbf}{x}{\mathrm}{d}t. \end{aligned}$$ Thus, taking $\varepsilon\to 0$, and then $n\rightarrow +\infty$, we immediately get . $\Box$ We turn to the existence proof of a unique solution to the second problem – by employing the Galerkin method, the domain expansion technique and a technique of improving regularity. \[lem:0502\] Let $T>0$, ${\mathbf}{u}_0\in H^4$, ${\mathbf}{v}\in {V}_T$, ${\mathbf}{f}\in {F}_T$, $\inf_{x\in \Omega}\{(\varrho_0+\bar{\rho})({\mathbf}{x})\}>0$ and $\varrho$ be constructed in Lemma \[lem:0501\], then there exists a unique classical solution $\mathbf{u} \in {V}_T$ to the problem – with an associated pressure $\tilde{q}$ satisfying $$\label{pressureesimta} \tilde{q}\in L^\infty(I_T,L^6),\;\;\nabla\tilde{q}\in C^0(\bar{I}_T,H^2),\;\; (\nabla \tilde{q})_t\in C^0(\bar{I}_T,L^2).$$ We divide the proof of Lemma \[lem:0502\] into three steps. \(1) First we solve – without pressure by the Galerkin method, and the existence of solutions to the Galerkin approximate problem is established by adapting the basic idea from the proof of [@GYTIPDE Theorem 4.3]. Since $H^2_{\sigma}:=\{{\mathbf}{v}\in H^2~\|~{\mathrm}{div}{\mathbf}{v}=0\}$ is separable, it possesses a countable basis $\{{\mathbf}{w}^j\}_{j=1}^\infty$. For each $m\geq 1$ we define a finite dimensional space $\mathcal{U}_m:={\mathrm}{span}\{{\mathbf}{w}^1,\ldots,{\mathbf}{w}^m\}\subset H^2_\sigma$ and an approximate solution $${\mathbf}{u}^m(t)=a_j^m(t){\mathbf}{w}_j\;\mbox{ with }\; a_j^m:\ I_T\rightarrow\mathbb{R}\;\mbox{ for }\; j=1,\ldots, m,$$ where as usual we have used the Einstein convention of summation of the repeated index $j$. Clearly, for each ${\mathbf}{v}\in H^2_\sigma$, we have $\mathcal{P}_m{\mathbf}{v}\rightarrow {\mathbf}{v}$ as $m\rightarrow \infty$ and $\\|\mathcal{P}_m{\mathbf}{u}\\|_{H^2}\leq \\|{\mathbf}{u}\\|_{H^2}$, where $\mathcal{P}_m:~H_\sigma^2\rightarrow \mathcal{U}_m$ is the $H^2_\sigma$ the orthogonal projection onto $\mathcal{U}_m$. In addition, by the method of characteristics, we can deduce from the mass equation $\eqref{hypobolic}_1$ that $$\tilde{\rho}:=\sup_{{\mathbf}{x}\in \Omega}\{(\varrho_0+\bar{\rho})({\mathbf}{x})\}\geq (\varrho+\bar{\rho})(t,{\mathbf}{x})\geq \underline{\rho}:= \inf_{{\mathbf}{x}\in \Omega}\{(\varrho_0+\bar{\rho})({\mathbf}{x})\}>0\;\; \mbox{ for any }(t,{\mathbf}{x})\in Q_T.$$ By the theory of ODEs, we can find the coefficients $a_j^m\in C^{2}(\bar{I}_T)$, such that $$\label{appweaksolu} \int_{\Omega}(\varrho+\bar{\rho}){\mathbf}{u}^m_t\cdot{\mathbf}{w}{\mathrm}{d}{\mathbf}{x}+\mu\int_{\Omega} \nabla {\mathbf}{u}^m:\nabla {\mathbf}{w}{\mathrm}{d}{\mathbf}{x}=\int_\Omega {\mathbf}{f}\cdot{\mathbf}{w}{\mathrm}{d}{\mathbf}{x}$$ with initial data $${\mathbf}{u}^m(0)=\mathcal{P}^m{\mathbf}{u}_0\in \mathcal{U}_m$$ for each ${\mathbf}{w}\in \mathcal{U}_m$. Next, we derive uniform estimates for ${\mathbf}{u}_m$. In what follows, we denote by $C(\cdots)$ a generic positive constant depending only on its variables. The notation $a\lesssim b$ means that $a\leq \tilde{C}b$ for a universal constant $\tilde{C}>0$, which may depend on $T$, $\bar{{\mathbf}{M}}$, $\mu$, $g$, $\underline{\rho}$, $\tilde{\rho}$, $\bar{\rho}$, and the norms $\\|(\varrho_0,{\mathbf}{N}_0)\\|_{H^3}$, $\\|{\mathbf}{u}_0\\|_{H^4}$, $\\|\varrho_t(0)\\|_{L^\infty}$, $\\|{\mathbf}{f}(0)\\|_{H^2}$, and $\\|{\mathbf}{f}_t(0)\\|_{L^2}$. Moreover, $\tilde{C}$ is increasing with $T$, and the norms $\\|\varrho_t(0)\\|_{L^\infty}$, $\\|{\mathbf}{f}(0)\\|_{H^2}$ and $\\|{\mathbf}{f}_t(0)\\|_{L^2}$, if $\tilde{C}$ depends on them. Taking ${\mathbf}{w}={\mathbf}{u}^m$ in , we see that $$\label{baseestimate} \frac{1}{2}\frac{{\mathrm}{d}}{{\mathrm}{d}t}\int_{\Omega}(\varrho+\bar{\rho})\|{\mathbf}{u}^m\|^2{\mathrm}{d}{\mathbf}{x} +\mu\int_{\Omega} \|\nabla {\mathbf}{u}^m\|^2{\mathrm}{d}{\mathbf}{x}=\int_\Omega {\mathbf}{f}\cdot{\mathbf}{u}^m{\mathrm}{d}{\mathbf}{x} +\frac{1}{2}\int_{\Omega}\varrho_t\|{\mathbf}{u}^m\|^2{\mathrm}{d}{\mathbf}{x},$$ which yields $$\frac{{\mathrm}{d}}{{\mathrm}{d}t}\\|\sqrt{\varrho+\bar{\rho}}{\mathbf}{u}^m\\|_{L^2}\leq2 \left\\|\frac{{\mathbf}{f}}{\sqrt{\varrho+\bar{\rho}}}\right\\|_{L^2} +\left\\|\frac{\varrho_t}{\varrho+\bar{\rho}}\right\\|_{L^\infty}\\|\sqrt{\varrho+\bar{\rho}}{\mathbf}{u}^m\\|_{L^2}.$$ Applying Gronwall’s inequality, we get $$\\|{\mathbf}{u}^m \\|_{L^2}^2\lesssim e^{\frac{2T}{\underline{\rho}}\\|\varrho_t\\|_{L^\infty(Q_T)}} \left(1 + {T}\\|{\mathbf}{f}\\|_{L^2(Q_T)}^2\right),$$ which, together with , implies that $$\label{lowestimesv} \\|{\mathbf}{u}^m(t)\\|_{L^2}^2+\int_0^t\\|\nabla {\mathbf}{u}^m(s)\\|_{L^2}^2{\mathrm}{d}s \lesssim e^{\frac{3T}{\underline{\rho}}\\|\varrho_t\\|_{L^\infty(Q_T)}} \left( 1 + T\\|{\mathbf}{f}\\|_{L^2(Q_T)}^2\right).$$ We differentiate in time and take ${\mathbf}{w}={\mathbf}{u}^m_t$ to deduce $$\begin{aligned} \frac{1}{2}\frac{{\mathrm}{d}}{{\mathrm}{d}t}\int_{\Omega}(\varrho+\bar{\rho})\|{\mathbf}{u}^m_t\|^2{\mathrm}{d}{\mathbf}{x} +\mu\int_{\Omega} \|\nabla {\mathbf}{u}^m_t\|^2{\mathrm}{d}{\mathbf}{x}=\int_\Omega {\mathbf}{f}_t\cdot{\mathbf}{u}^m_t{\mathrm}{d}{\mathbf}{x} +\frac{1}{2}\int_{\Omega}\varrho_t\|{\mathbf}{u}^m_t\|^2{\mathrm}{d}{\mathbf}{x}, \end{aligned}$$ which yields $$\begin{aligned} \\|{\mathbf}{u}^m_t(t)\\|^2_{L^2} +\int_{0}^t \\|\nabla {\mathbf}{u}^m_t(s)\\|^2_{L^2}{\mathrm}{d}s \lesssim e^{\frac{3T}{\underline{\rho}}\\|\varrho_t\\|_{L^\infty(Q_T)}}\left(1+T\\|{\mathbf}{f}_t \\|_{L^2(Q_T)}^2\right). \end{aligned}$$ Finally, taking ${\mathbf}{w}={\mathbf}{u}_t^m$ in , we find that $$\label{lowesti2} \begin{aligned} \\| \nabla {\mathbf}{u}^m(t)\\|_{L^2}^2 +\int_0^t\\|{\mathbf}{u}_t^m(s)\\|_{L^2}^2{\mathrm}{d}s\lesssim 1+\\|{\mathbf}{f}\\|_{L^2(Q_T)}^2. \end{aligned}$$ Summing up the previous estimates and making use of the estimate $$\\|{\mathbf}{f}(t)\\|_{L^2}=\left\\|{\mathbf}{f}(0)+\int_0^t{\mathbf}{f}_{s}(s){\mathrm}{d}t\right\\|_{L^2}\lesssim 1+\sqrt{T}\\|{\mathbf}{f}_t\\|_{L^2(Q_T)},$$ we arrive at $$\label{sumestimeat} \begin{aligned} &\\| {\mathbf}{u}^m(t)\\|_{H^1}^2+\\| {\mathbf}{u}^m_t(t)\\|_{L^2}^2+\int_0^t (\\|\nabla {\mathbf}{u}^m(s)\\|_{L^2}^2+\\|{\mathbf}{u}_t^m(s)\\|_{H^1}^2){\mathrm}{d}s\\ &\lesssim e^{\frac{3T}{\underline{\rho}}\\|\varrho_t\\|_{L^\infty(Q_T)}}\left(1+ T\\|{\mathbf}{f}_t\\|_{L^2(Q_T)}^2\right). \end{aligned}$$ \(2) Now we can obtain a strong solution from the above uniform estimates for the approximate solutions. In view of , up to the extraction of a subsequence, one has $$\begin{aligned} &{\mathbf}{u}^m\rightarrow {\mathbf}{u}\mbox{ weakly-* in }L^\infty(I_T,H^1_\sigma),\ {\mathbf}{u}^m_t\rightarrow {\mathbf}{u}_t\mbox{ weakly-* in }L^\infty(I_T,L^2),\\ &\nabla {\mathbf}{u}^m\rightarrow \nabla{\mathbf}{u}\mbox{ weakly in }L^2(I_T,L^2),\ {\mathbf}{u}^m_t\rightarrow {\mathbf}{u}_t\mbox{ weakly in }L^2(I_T,H^1),\\ &{\mathbf}{u}^m\rightarrow {\mathbf}{u}\mbox{ strongly in }C^0(\bar{I_T},L^2(K))\mbox{ for any bounded domain }K\subset \Omega,\ {\mathbf}{u}(0)={\mathbf}{u}_0, \end{aligned}$$ and ${\mathrm}{div}{\mathbf}{u}_t=0$. Thus, if we take limit in , we get $$\label{0507n1} \int_{\Omega}(\varrho+\bar{\rho}){\mathbf}{u}_t\cdot{\mathbf}{w}{\mathrm}{d}{\mathbf}{x}+\mu\int_{\Omega} \nabla{\mathbf}{u}:\nabla{\mathbf}{w}{\mathrm}{d}{\mathbf}{x} =\int_\Omega{\mathbf}{f}\cdot{\mathbf}{w}{\mathrm}{d}{\mathbf}{x}\;\;\mbox{ for any }{\mathbf}{w}\in H^2_\sigma,$$ which can be regarded as a weak solution of the following Stokes equations $$-\mu \Delta \mathbf{u}+ \nabla \tilde{q}={\mathbf}{F}:={\mathbf}{f}-(\varrho+\bar{\rho}){\bf u}_t\mbox{ in }\Omega,$$ where ${\mathbf}{F}\in L^\infty(I_T,L^2)\cap L^2(I_T,H^1)$. Next we proceed to derive more estimates for ${\mathbf}{u}$ in the Stokes equations by the domain expansion technique (i.e., the classical regularity theory on the Stokes equations). Let $L_n=(-n,n)$ and $\Omega_n=(2\pi L\mathbb{T})^2\times L_n$. Similarly to [@GYTIPDE Propositions 2.9 and 3.7], we can show that the Stokes problem $$\label{appstokes} \left\{ \begin{array}{ll} -\mu \Delta \mathbf{u}_n+\nabla \tilde{q}_n={\mathbf}{F} \mbox{ in }\Omega_n, \\ {\mathrm}{div}{\mathbf}{u}_n=0, \\ {\mathbf}{u}\|_{ x_3=-2\pi nL}={\mathbf}{u}\|_{x_3=2\pi nL}=0 \end{array} \right.$$ admits a unique strong solution ${\mathbf}{u}_n $ with a unique associated pressure $\tilde{q}_n$, such that $$\begin{aligned} & \\|\nabla^k {\mathbf}{u}_n\\|_{H^2(\Omega_n)}^2 + \\|\nabla^{k+1}\tilde{q}_n\\|_{L^2(\Omega_n)}^2 \leq C(\Omega_n,\mu) \\|\nabla^k {\mathbf}{F}\\|_{L^2(\Omega_n)}^2\;\;\mbox{ for }k=0,1,\nonumber\\ & \\|\tilde{q}_n\\|_{L^2(\Omega_n)}^2 \leq C(\Omega_n) \\|{\mathbf}{F}\\|_{L^2(\Omega_n)}^2\;\mbox{ and }\; \int_{\Omega_n}\tilde{q}_n{\mathrm}{d}{\mathbf}{x}=0. \end{aligned}$$ By scaling the spatial variables on a cuboid domain $(0,2\pi nL)^2\times L_n$ and using the horizontally periodic property (i.e., ${\mathbf}{u}_n(x_1,x_2,x_3)={\mathbf}{u}_n(x_1+2\pi L,x_2+2\pi L,x_3)$), we can deduce the following uniform estimates $$\begin{aligned} &\\| \nabla^{2+k} {\mathbf}{u}_n\\|_{L^2(\Omega_n)}^2+\\| \nabla^{1+k} \tilde{q}_n\\|_{L^2(\Omega_{n_0})}^2\leq C(\mu) \\|\nabla^k {\mathbf}{F}\\|_{L^{2}}^2 \quad\mbox{ for }k=0,1,\nonumber\\ & \\|\tilde{q}_n\\|_{L^2(\Omega_{n})}^2\leq C(\mu)n^2\\|{\mathbf}{F}\\|_{L^2}^2\;\mbox{ and }\;\\|\tilde{q}_n\\|_{L^6(\Omega_{n})}^2\leq C(\mu)\\|{\mathbf}{F}\\|_{L^2}^2.\end{aligned}$$ Moreover, using Hölder’s inequality, we get $$\\|\tilde{q}_n\\|_{L^2(\Omega_{n_0})}^2\leq C(\mu)n_0^{2/3}\\|{\mathbf}{F}\\|_{L^2}^2\quad\mbox{ for any }n\geq n_0.$$ We make use of and to deduce $$\mu\int_{\Omega_n} \|\nabla {\mathbf}{u}_n\|^2{\mathrm}{d}{\mathbf}{x}=\int_{\Omega_n} {\mathbf}{F}\cdot{\mathbf}{u}_n{\mathrm}{d}{\mathbf}{x} =\mu\int_{\Omega_n} \nabla {\mathbf}{u}:\nabla {\mathbf}{u}_n{\mathrm}{d}{\mathbf}{x},$$ whence, $$\\|\nabla {\mathbf}{u}_n\\|_{L^2(\Omega_n)}\leq \\|\nabla {\mathbf}{u}\\|_{L^2}.$$ Noting that $$\\| {\mathbf}{u}_n\\|_{L^{6}(\Omega_n)}\leq C(\Omega_n)\\| \nabla {\mathbf}{u}_n\\|_{L^{2}(\Omega_n)},$$ we may scale the spatial variables on a cuboid domain $(0,2\pi nL)^2\times L_n$ and utilize the horizontally periodic property to get $$\\| {\mathbf}{u}_n\\|_{L^{6}(\Omega_n)}\leq c\\|\nabla {\mathbf}{u}\\|_{L^2}\quad\mbox{ for some constant }c.$$ Now, recalling ${\mathbf}{F}\in L^\infty(I_T,L^2)\cap L^2(I_T,H^1)$ and using the established uniform estimates for $({\mathbf}{u}_n,\tilde{q}_n)$, up to the extraction of a subsequence, one has $$\begin{aligned}&\nabla^k{\mathbf}{u}_n\rightarrow\nabla^k \bar{{\mathbf}{u}}\mbox{ weakly-* in }L^\infty(I_T,L^2)\mbox{ for }k=1,2,\;\; \nabla^3{\mathbf}{u}_n\rightarrow \nabla^3 \bar{{\mathbf}{u}}\mbox{ weakly in }L^2(I_T,L^2),\\ & \nabla\tilde{q}_n\rightarrow\nabla\tilde{q}\mbox{ weakly-* in }L^\infty(I_T,L^2),\;\; \nabla^2 \tilde{q}_n\rightarrow \nabla^2 \tilde{q} \mbox{ weakly in }L^2(I_T,L^2),\\ &{\mathbf}{u}_n\rightarrow\bar{{\mathbf}{u}}\mbox{ weakly-* in }L^\infty(I_T, L^6),\;\;\tilde{q}_n\rightarrow \tilde{q}\mbox{ weakly-* in } L^\infty(I_T,L^6), \end{aligned}$$ where we have extended $\nabla^k {\mathbf}{u}_n$ and $\nabla^m \tilde{q}_n$ by zero extension outside $\Omega_n$. Thus, we have the following classical regularity estimates on the Stokes problem: $$\begin{aligned} &\label{stokestimates} \\| \nabla^{2} {\mathbf}{u}\\|_{L^\infty(I_T,L^2)}^2+\\| \nabla \tilde{q}\\|_{L^\infty(I_T,L^2)}^2+ \\|\tilde{q}\\|_{L^\infty(I_T,L^6)}^2\leq C(\mu) \\|{\mathbf}{F}\\|_{L^\infty(I_T,L^{2})}^2,\\ & \\| \nabla^{3} {\mathbf}{u}\\|_{L^2(I_T,L^2)}^2+\\| \nabla^2 \tilde{q}\\|_{L^2(I_T,L^2)}^2\leq C(\mu) \\|\nabla {\mathbf}{F}\\|_{L^2(Q_T)}^2,\nonumber\\ &\label{qomesgq}\\| \tilde{q}\\|_{L^\infty(I_T,L^2(\Omega_{n_0}))} \leq C(\mu) n_0^{\frac{1}{3}} \\|{\mathbf}{F}\\|_{L^\infty(I_T,L^2)}\quad\mbox{ for any }n_0>0.\end{aligned}$$ Moreover,$$\begin{aligned} &\label{stoequeiona} -\mu \Delta \bar{\mathbf{u}}+ \nabla \tilde{q}={\mathbf}{F}, \quad {\mathrm}{div}\bar{{\mathbf}{u}}=0, \\ &\label{0514n}\mu\int_\Omega \nabla \bar{\mathbf{u}}:\nabla {\mathbf}{w}{\mathrm}{d}{\mathbf}{x} =\int_\Omega{\mathbf}{F}\cdot {\mathbf}{w}{\mathrm}{d}{\mathbf}{x} \end{aligned}$$ for any $$\begin{aligned} {\mathbf}{w}\in C_{0,\sigma}^\infty:=\{{\mathbf}{w}\in C^\infty~\|~&{\mathrm}{div}{\mathbf}{w}=0\mbox{ and }\exists~R>0,\mbox{ such that }\\ &{\mathbf}{w}(x',x_3)=0\mbox{ for any }\|x_3\|>R\}. \end{aligned}$$ It should be noted that $C_{0,\sigma}^\infty$ is dense in $H_\sigma^1$, which will be proved in Lemma \[lem:0503\]. By a density argument, holds for any ${\mathbf}{w}\in H^1_\sigma$, which, together with , implies that $ {\mathbf}{u}=\bar{{\mathbf}{u}}$. Summing up the above estimates, we obtain the following inequality: $$\begin{aligned} &\\| {\mathbf}{u}(t)\\|_{H^2}^2+\\| {\mathbf}{u}_t(t)\\|_{L^2}^2+\\| \nabla \tilde{q}(t)\\|_{L^2}^2+ \int_0^t \left(\\| {\mathbf}{u}(s)\\|_{H^3}^2+\\|{\mathbf}{u}_t(s)\\|_{H^1}^2+\\|\nabla^2\tilde{q}\\|_{L^2}^2\right){\mathrm}{d}s\\ &\lesssim e^{\frac{3T}{\underline{\rho}}\\|\varrho_t\\|_{L^\infty(Q_T)}}\left(1+ \\|\varrho\\|_{H^2}^2\right) \left(1+T\\|{\mathbf}{f}_t\\|_{L^2(Q_T)}^2+\\|\nabla {\mathbf}{f} \\|_{L^2(Q_T)}^2\right). \end{aligned}$$ \(3) Before improving the regularity of ${\mathbf}{u}$, we show the continuity of $({\mathbf}{u},\nabla\tilde{q})$. First, similarly to [@YCHKimUnique Remark 6], we can prove that for a.e. $t\in I_T$, $$\label{eq0510} \begin{aligned} \frac{1}{2}\frac{\mathrm{d}}{\mathrm{d}t}\int_{\Omega}(\varrho+\bar{\rho})\|{\mathbf}{u}_t\|^2 {\mathrm}{d}{\mathbf}{x}=&\int_{\Omega}{\mathbf}{f}_t\cdot{{\mathbf}{u}}_t{\mathrm}{d}{\mathbf}{x}+\frac{1}{2}\int_\Omega \varrho_t\|{\mathbf}{u}_t\|^2{\mathrm}{d}{\mathbf}{x}-\mu\int_{\Omega}\|\nabla \mathbf{u}_t \|^2{\mathrm}{d}{\mathbf}{x}.\end{aligned}$$ It follows from that for any $\varphi\in H_\sigma^1$, $$\label{absolution} \begin{aligned} \frac{\mathrm{d}}{\mathrm{d}t}\int_{\Omega }(\varrho+\bar{\rho}){\mathbf}{u}_t\cdot\varphi{\mathrm}{d}{\mathbf}{x} =\int_{\Omega}{\mathbf}{f}_t\cdot\varphi{\mathrm}{d}{\mathbf}{x}-\mu \int_{\Omega} \nabla \mathbf{u}_t:\nabla \varphi{\mathrm}{d}{\mathbf}{x}\mbox{ holds for a.e. } t\in I_T. \end{aligned}$$ From the regularity of $(\varrho,{\mathbf}{u})$ we get $$\label{fracg-1}((\varrho+\bar{\rho}){\mathbf}{u}_t)_t\in L^2((0,T),H_\sigma^{-1})\mbox{ and } \frac{{\mathrm}{d}}{{\mathrm}{d}t}\int_{\Omega }(\varrho+\bar{\rho}){\mathbf}{u}_t\cdot\varphi {\mathrm}{d}{\mathbf}{x}=<((\varrho+\bar{\rho}){\mathbf}{u}_t)_t,\varphi>,$$ where $H^{-1}_\sigma$ denotes the dual space of $H_\sigma^1$ and $<\cdot,\cdot>$ the corresponding dual product. Therefore, the identity follows immediately from , and the following identity $$\frac{{\mathrm}{d}}{{\mathrm}{d}t}\int_{\Omega }(\varrho+\bar{\rho})\|{\mathbf}{u}_t\|^2 {\mathrm}{d}{\mathbf}{x}=2<((\varrho+\bar{\rho}){\mathbf}{u}_t)_t,{\mathbf}{u}_t>-\int_\Omega \varrho_t\|{\mathbf}{u}_t\|^2{\mathrm}{d}{\mathbf}{x},$$ which can be easily established by means of a classical regularization method. With in hand, we can infer that $$\label{continutiofu}{\mathbf}{u}_t\in C^0(\bar{I}_T,L^2).$$ In fact, it follows from and that $(\varrho+\bar{\rho}){\mathbf}{u}_t\in C^0(\bar{I}_T, H^{-1}_\sigma)$ and $\\|\sqrt{\varrho+\bar{\rho}}{\mathbf}{u}_t\\|_{L^2}^2\in C^0(\bar{I}_T)$. Since $(\varrho+\bar{\rho}){\mathbf}{u}_t\in L^\infty(I_T, L^2)$ and $L^2\hookrightarrow H^{-1}_\sigma$, it follows from the embedding theorem that $(\varrho+\bar{\rho}){\mathbf}{u}_t\in C(\bar{I}_T, L^2_{{\mathrm}{weak}})$. Thus, recalling $\varrho+\bar{\rho}\in C^0(\bar{I}_T,H^{2})$, we have $\sqrt{(\varrho+\bar{\rho})}{\mathbf}{u}_t\in C^0(\bar{I}_T,L^2)$ and . Noting that ${\mathbf}{u}\in C^0(\bar{I}_T,H^2)$ by virtue of the embedding theorem, we see that $\nabla \tilde{q}\in C^0(\bar{I}_T,L^2)$. In view of , we find that $$\Delta\tilde{q}=\frac{\nabla (\varrho+\bar{\rho})\cdot\nabla \tilde{q}}{\varrho+\bar{\rho}}-\frac{\mu\nabla (\varrho+\bar{\rho})\cdot\Delta{\mathbf}{u}}{\varrho+\bar{\rho}}+(\varrho+\bar{\rho}){\mathrm}{div}\left(\frac{{\mathbf}{f}}{\varrho+\bar{\rho}}\right) \in C^0(\bar{I}_T,L^2).$$ Hence, by virtue of , the elliptic estimates, the periodic property and a cut-off technique, there exists a constant $c$, such that $$\begin{aligned} \\|\nabla^2\tilde{q}(t)\\|_{L^2}\leq c\left(\\|\Delta\tilde{q}(t)\\|_{L^2}+\\|\nabla\tilde{q}(t)\\|_{L^2}+ \\|{\mathbf}{F}(t)\\|_{L^2}\right)\quad \mbox{ for a.e. }t\in I_T, \end{aligned}$$ which implies that $$\label{eqiatonqegH} \nabla^2 \tilde{q}\in L^\infty({I},L^2),$$ whence, $\nabla\tilde{q}\|_{t=0}\in H^1$. Now, we improve the regularity of ${\mathbf}{u}$. Consider the following problem: $$\label{loc0503jjwtimes} \left\{ \begin{array}{ll} (\varrho+\bar{\rho}){\bf v}_t + \nabla \tilde{p}-\mu \Delta \mathbf{v}= {\mathbf}{g}\quad\mbox{ in }\Omega, \\ {\mathrm}{div}{\mathbf}{v}=0,\\ {\mathbf}{v}\|_{t=0}= {\mathbf}{v}_0:= [({\mathbf}{f}-\nabla \tilde{q}+\mu \Delta \mathbf{u})/(\varrho +\bar{\rho})]\|_{t=0}, \end{array} \right.$$ where ${\mathbf}{g}:={\mathbf}{f}_t-\varrho_t{\mathbf}{u}_t$. Keeping in mind that ${\mathbf}{g}\in C^0(\bar{I}_T,{L^2})$ and ${\mathbf}{v}_0\in H^1_\sigma$, referring to the construction of the solution ${\mathbf}{u}$ and the derivation of regularity for ${\mathbf}{u}$, we can easily obtain a strong solution ${\mathbf}{v}$ to the problem with an associated pressure $\tilde{p}$, where ${\mathbf}{v}\in C^0(\bar{I}_T,H^1)\cap L^2(I_T,H^2),\ {\mathbf}{v}_t\in L^2(I_T,L^2)$, $\nabla \tilde{p}\in L^2(I_T,L^2)$ and $$\begin{aligned} &\\| {\mathbf}{v}(t)\\|_{H^1}^2+\int_0^t (\\|\nabla {\mathbf}{v}(s)\\|_{H^1}^2+\\|{\mathbf}{v}_t(s)\\|_{L^2}^2){\mathrm}{d}s\\ &\lesssim e^{\frac{6T}{\underline{\rho}}\\|\varrho_t\\|_{L^\infty(Q_T)}} \left(1+T\\|\varrho_t\\|^2_{L^\infty(Q_T)}\right) \left(1+\\|{\mathbf}{f}_t \\|_{L^2(Q_T)}^2\right). \end{aligned}$$ Moreover, for any $\varphi\in H_\sigma^1$, $$\label{equan0520}\frac{d}{dt}\int_{\Omega}(\varrho+\bar{\rho}){\mathbf}{v}\cdot\varphi{\mathrm}{d}{\mathbf}{x}+\mu\int_{\Omega} \nabla {\mathbf}{v}:\nabla \varphi{\mathrm}{d}{\mathbf}{x}=\int_\Omega ({\mathbf}{g}+\varrho_t{\mathbf}{v})\cdot\varphi{\mathrm}{d}{\mathbf}{x}.$$ Denoting $\tilde{{\mathbf}{v}}:={\mathbf}{v}-{\mathbf}{u}_t$, and recalling $\tilde{{\mathbf}{v}}\|_{t=0}=0$ and $\sqrt{(\varrho+\bar{\rho})}\tilde{{\mathbf}{v}}\in C(\bar{I}_T,L^2)$, we subtract from to infer that $$\frac{d}{dt} \int_{\Omega}(\varrho+\bar{\rho})\tilde{{\mathbf}{v}}\cdot\varphi{\mathrm}{d}{\mathbf}{x} + \mu\int_{\Omega} \nabla \tilde{{\mathbf}{v}}:\nabla \varphi{\mathrm}{d}{\mathbf}{x}{\mathrm}{d}t=\int_\Omega \varrho_t\tilde{{\mathbf}{v}}\cdot\varphi{\mathrm}{d}{\mathbf}{x}.$$ Similarly to the derivation of , we can establish $$\frac{d}{dt}\int_{\Omega}(\varrho+\bar{\rho})\|\tilde{{\mathbf}{v}}\|^2{\mathrm}{d}{\mathbf}{x}{\mathrm}{d}t+\mu \int_{\Omega} \|\nabla \tilde{{\mathbf}{v}}\|^2 {\mathrm}{d}{\mathbf}{x} =\int_\Omega \varrho_t\|\tilde{{\mathbf}{v}}\|^2{\mathrm}{d}{\mathbf}{x},$$ which implies ${\mathbf}{v}={\mathbf}{u}_t$. Therefore, ${\mathbf}{u}_t\in C^0(\bar{I}_T,H^1)$, ${\mathbf}{u}_{tt}\in L^2(I_T,L^2)$ and $\nabla \tilde{p}=(\nabla \tilde{q})_t$. Using the regularity estimate for the Stokes equations, we further have ${\mathbf}{u}\in L^\infty(\bar{I}_T,H^3)$. Hence, $\Delta\tilde{q}\in L^\infty(\bar{I}_T,H^1)$. In view of the elliptic interior regularity, we see that $\nabla \tilde{q}\in L^\infty({I},H^2_{{\mathrm}{loc}}(\mathbb{R}^3))$. Thus, similarly to the derivation of , we have $\nabla^3 \tilde{q}\in L^\infty({I},L^2)$, which yields $\nabla\tilde{q}\|_{t=0}\in H^2$. Consequently, ${\mathbf}{v}_0\in H^2$. Recalling that ${\mathbf}{g}_t\in L^2(I_T,{L^2})$, we easily see that ${\mathbf}{u}_t\in C^0(\bar{I}_T, H^2)$ solves the problem and enjoys the following estimate $$\begin{aligned} &\\| {\mathbf}{u}_t(t)\\|_{H^2}^2+\\| {\mathbf}{u}_{tt}(t)\\|_{L^2}^2+\int_0^t \left(\\| {\mathbf}{u}_t(s)\\|_{H^3}^2+\\|{\mathbf}{u}_{tt}(s)\\|_{H^1}^2\right){\mathrm}{d}s\\ &\lesssim e^{\frac{9T}{\underline{\rho}}\\|\varrho_t\\|_{L^\infty(Q_T)}} (1+\\|\varrho\\|_{H^2}^2) \left(1+T\\|\varrho_{t}\\|_{L^\infty(I_T,H^2)}^2+T\\|\varrho_{tt}\\|^2_{L^\infty(I_T,H^1)}\right)\\ &\qquad\times \left(1+T\\|\varrho_t\\|^2_{L^\infty(Q_T)} \right) \left(1+T\\|{\mathbf}{f}_{tt}\\|_{L^2(Q_T)}^2+\\|\nabla {\mathbf}{f}_t\\|_{L^2(Q_T)}^2\right); \end{aligned}$$ and moreover, $\nabla\tilde{p}\in C^0(\bar{I}_T,L^2)\cap L^2(I_T, H^1)$. Employing the regularity estimate on the Stokes equations again, we arrive at $$\label{estimatesofu} \begin{aligned} &\\| {\mathbf}{u}(t)\\|_{H^4}^2+\\| {\mathbf}{u}_t(t)\\|_{H^2}^2+\\| {\mathbf}{u}_{tt}(t)\\|_{L^2}^2+\int_0^t \left(\\| {\mathbf}{u}_t(s)\\|_{H^3}^2+\\|{\mathbf}{u}_{tt}(s)\\|_{H^1}^2\right){\mathrm}{d}s\\ &\leq \tilde{C} e^{\frac{9T}{\underline{\rho}}\\|\varrho_t\\|_{L^\infty(Q_T)}} (1+\\|\varrho\\|_{H^2}^2)^2 \left(1+T\\|\varrho_{t}\\|_{L^\infty(I_T,H^2)}^2+T\\|\varrho_{tt}\\|^2_{L^\infty(I_T,H^1)}\right)\\ &\quad\times \left(1+T\\|\varrho_t\\|^2_{L^\infty(Q_T)} \right) \left(1+T\\|{\mathbf}{f}_{tt}\\|_{L^2(Q_T)}^2+\\|\nabla {\mathbf}{f}_t\\|_{L^2(Q_T)}^2+\\|{\mathbf}{f}\\|_{L^\infty(I_T,H^2)}^2\right); \end{aligned}$$ and moreover, ${\mathbf}{u}\in C^0(\bar{I}_T, H^4)$ and $\nabla \tilde{q}\in C^0(\bar{I}_T,H^2)$. Summing up the above estimates, we obtain the desired conclusions immediately.$\Box$ Finally, we briefly prove that $C_{0,\sigma}^\infty$ is dense in $H_\sigma^1$. \[lem:0503\] $C_{0,\sigma}^\infty$ is dense in $H_\sigma^1$. Let ${\mathbf}{u}\in H_\sigma^1$ and $n>0$. We define $\Omega_n:=(2\pi L\mathbb{T}^2 )\times (-n,n)$, $\Omega_{n,2n}:=\Omega_{2n}/\Omega_{n}$, $\tilde{\Omega}_n:=(0,2\pi L)^2\times (-2n,2n)$ and $\phi_n(x_3)\in C_0^\infty(\mathbb{R})$ be constructed as in . In view of [@GPGAI Theorem III.3.4], for each given $n$, there exists a function ${\mathbf}{w}_n\in H^{1}(\Omega_{2n})\cap H_0^1(\tilde{\Omega}_n)$, satisfying $$\begin{aligned}&{\mathrm}{div}{\mathbf}{w}_n=-u_3 \phi_n',\\ & \\|\nabla {\mathbf}{w}_n\\|_{L^2(\Omega_{2n})}\leq c\\|u_3 \phi_n'\\|_{L^2(\Omega_{2n})}, \end{aligned}$$ where and throughout this proof, $c$ denotes a constant independent of the domain. Thus, $$\\|{\mathbf}{w}_n\\|_{L^2(\Omega_{2n})}\leq cn^{\frac{1}{3}}\\|{\mathbf}{w}_n\\|_{L^6} \leq cn^{\frac{1}{3}}\\|\nabla {\mathbf}{w}_n\\|_{L^2(\Omega_{2n})}\leq c\\|{u}_3\\|_{L^2(\Omega_{n,2n})}.$$ We define ${\mathbf}{w}_n=0$ on $\mathbb{R}^3\setminus{\Omega}_{2n}$. Let ${\mathbf}{v}_n=\phi_n{\mathbf}{u}+{\mathbf}{w}_n$ and $S_\varepsilon$ be a standard mollifier, then $ S_\varepsilon({\mathbf}{v}_{n})\in C_{0,\sigma}^\infty$. Moreover, $$\begin{aligned}\\|{\mathbf}{u}-S_\varepsilon({\mathbf}{v}_n)\\|_{L^2}\leq & \\|{\mathbf}{u}-{\mathbf}{v}_n\\|_{L^2}+\\|{\mathbf}{v}_n-S_\varepsilon({\mathbf}{v}_n)\\|_{L^2} \\ \leq& \\|(1-\phi_n){\mathbf}{u}\\|_{L^2}+\\|{\mathbf}{w}_n\\|_{L^2}+\\|{\mathbf}{v}_n-S_\varepsilon({\mathbf}{v}_n)\\|_{L^2}, \end{aligned}$$ which implies that $$S_\varepsilon({\mathbf}{v}_n)\rightarrow {\mathbf}{u}\mbox{ as }\varepsilon\rightarrow 0\mbox{ and }n\rightarrow +\infty.$$ This shows that $C_{0,\sigma}^\infty$ is dense in $H_\sigma^1$. $\Box$ Uniform estimates ----------------- Let $(\varrho_0,{\mathbf}{u}_0,{\mathbf}{N}_0)$ satisfy the assumption in Proposition \[pro:0401new\], then, in view of Lemmas \[lem:0501\] and \[lem:0502\], for any given $T>0$ and any given ${\mathbf}{v}\in V_T$ satisfying ${\mathbf}{v}(0)={\mathbf}{u}_0$, the linearized problem – possesses a unique classical solution $((\varrho,{\mathbf}{N}),{\mathbf}{u})\in H_T\times V_T$. To emphasize such relation between ${\mathbf}{u}$ and ${\mathbf}{v}$, we can define ${\mathbf}{u}:=S({\mathbf}{v})$. In addition, we can deduce from $_2$ that $$\label{partialut} \\|{\bf u}_t(0)\\|_{H^1}\leq C_0,$$ where the positive constant $C_0$ depends on the initial data $(\varrho_0,{\mathbf}{u}_0,{\mathbf}{N}_0)$ and other physical parameters in the perturbed equations, and is independent of ${\mathbf}{v}$. Now, let us denote $ {D}_{\kappa}^T=\{{\mathbf}{v}\in V_T~\|~\\|{\mathbf}{v}\\|_{V_T}\leq \kappa,\ \\|{\mathbf}{v}_t(0)\\|_{H^1}\leq C_0$, ${\mathbf}{v}(0)={\mathbf}{u}_0\}$ and $$\\|{\mathbf}{v}\\|_{V_T}:=\sqrt{\\|{\mathbf}{v}\\|_{L^\infty( {I}_T,H^4)}^2+\\|{\mathbf}{v}_t\\|_{L^\infty( {I}_T,H^2)}^2+\\|{\mathbf}{v}_{tt}\\|_{L^\infty( {I}_T,L^2)}^2+\\|{\mathbf}{v}_{t}\\|_{L^2( {I}_T,H^3)}^2 +\\|{\mathbf}{v}_{tt}\\|_{L^2({I}_T,H^1)}^2}.$$ We show next that there exists two positive constants $T\in (0,1)$ and $\kappa\geq 1$ depending on $(\varrho_0,{\mathbf}{u}_0,{\mathbf}{N}_0)$ and other physical parameters in the perturbed equations, such that $$\label{uniforestimate} \kappa^8 T= 1\mbox{ and }\\|S({\mathbf}{v})\\|_{V_{T}}\leq \kappa \quad \mbox{ for any } {\mathbf}{v}\in D^T_{\kappa}.$$ Throughout this and next subsections, the notation $a\lesssim b$ means that $a\leq Cb$ for some constant $C>0$ which may depend on $(\varrho_0,{\mathbf}{u}_0,{\mathbf}{N}_0)$ and other physical parameters in the perturbed equations. First we temporarily suppose that $\kappa^8 T=1$, $\kappa\geq 1$ and $T\in (0, 1)$. Similarly to the derivation of , we can deduce from $_1$ and $_2$ that $$\label{} \begin{aligned} \frac{d}{dt}\\|(\varrho,{\mathbf}{N}) \\|^2_{H^3}\lesssim \\|{\mathbf}{v}\\|_{H^4}(1+\\|(\varrho ,{\mathbf}{N})\\|_{H^3})\\|(\varrho, {\mathbf}{N}) \\|_{H^3}, \end{aligned}$$ which yields $$\label{engyrofdensiy}\\|(\varrho ,{\mathbf}{N})\\|_{L^\infty(I_T,H^3)}\lesssim e^{\kappa T}(1+\kappa T)\leq C.$$ Furthermore, we can get from $_1$ and $_2$ that $$\begin{aligned} & \\|\varrho_t(0)\\|_{L^\infty}\leq C,\\ & \label{Nutestimae}\\|(\varrho,{\mathbf}{N})_t\\|_{L^\infty(Q_T)} \lesssim \\|(\varrho,{\mathbf}{N})_t\\|_{L^\infty(I_T,H^2)} \lesssim \kappa,\\ &\\|(\varrho,{\mathbf}{N})_{tt}\\|_{L^\infty(I_T,H^1)}\lesssim \kappa+\kappa^2.\end{aligned}$$ Letting ${\mathbf}{f}:=({\mathbf}{N}+\bar{{\mathbf}{M}})\cdot\nabla {\mathbf}{N}-g \varrho {\mathbf}{e}_3-(\varrho+\bar{\rho}){\bf v}\cdot\nabla{\mathbf}{v}$, we see that $$\begin{aligned} &\\|{\mathbf}{f}(0)\\|_{H^2}^2+\\|{\mathbf}{f}_t(0)\\|_{L^2}^2\leq C,\\ & \\|{\mathbf}{f}_{tt}\\|_{L^2(Q_T)} \lesssim \kappa+\kappa^2+\kappa^3+\kappa^4,\\ &\label{} \\|\nabla {\mathbf}{f}_{t}\\|_{L^\infty(I_T,L^2)} \lesssim \kappa+\kappa^2+\kappa^3,\\ &\label{sevestemsnew} \\|{\mathbf}{f}\\|_{L^\infty(I_T,H^2)} \lesssim 1+\kappa^2 T.\end{aligned}$$ Recalling that ${\mathbf}{u}$ satisfies and $\tilde{C}$ is increasing in $\\|\varrho_t(0)\\|_{L^\infty}$, $\\|{\mathbf}{f}(0)\\|_{H^2}$ and $\\|{\mathbf}{f}_t(0)\\|_{L^2}$, we substitute – into to arrive at $$\label{finestim524} \begin{aligned}\\| S({\mathbf}{v})\\|_{V_T}\leq C. \end{aligned}$$ Finally, if we take $\kappa=\max\{C,1\}$ and $T:=\kappa^{-6}$ where the positive constant $C$ is from , then we obtain . Taking the limit ---------------- Let $(\varrho_0,{\mathbf}{u}_0,{\mathbf}{N}_0)$ satisfy the assumption in Proposition \[pro:0401new\] and ${\mathbf}{u}^0\equiv{\mathbf}{u}_0$. In view of Lemmas \[lem:0501\] and \[lem:0502\], we can construct a function sequence $\{(\varrho^k,{\mathbf}{u}^k,{\mathbf}{N}^k,\tilde{q}^{k})\}_{k=1}^\infty$ satisfying $((\varrho^k,{\mathbf}{N}^k),{\mathbf}{u}^k)\in H_T\times V_T$ and $$\label{iteratiequat}\left\{\begin{array}{l} \varrho_t^k+{\mathbf}{u}^{k-1}\cdot\nabla \varrho^k=-{\mathbf}{u}^{k-1}\cdot\nabla \bar{\rho}, \\[1mm] (\varrho^{k}+\bar{\rho}){\bf u}_t^{k}+\nabla\tilde{q}^{k}-\mu \Delta \mathbf{u}^{k}= ({\mathbf}{N}^{k}+\bar{{\mathbf}{M}})\cdot\nabla {\mathbf}{N}^{k}-g \varrho^{k} {\mathbf}{e}_3-(\varrho^{k}+\bar{\rho}){{\mathbf}{ u}^{k-1}}\cdot\nabla {\mathbf}{u}^{k-1},\\[1mm] {\mathbf}{N}_t^{k}+{\mathbf}{u}^{k-1}\cdot\nabla {\mathbf}{N}^{k}-\nabla {\mathbf}{u}^{k-1}{\mathbf}{N}^{k}= \bar{{\mathbf}{M}}\cdot\nabla {\mathbf}{u}^{k-1},\\[1mm] \mathrm{div}\mathbf{u}^{k}=0,\ \mathrm{div}\mathbf{N}^{k}=0\end{array}\right.$$ with initial data: $$(\varrho^k,\mathbf{u}^k,\mathbf{N}^k )\|_{t=0}=(\varrho_0,{\mathbf u}_0,\mathbf{N}_0).$$ Moreover, by virtue of the uniform estimates – and , there exits $T\in (0,1)$ such that the solution sequence $\{(\varrho^k,{\mathbf}{u}^k,{\mathbf}{N}^k)\}_{k=1}^\infty$ satisfies the following uniform estimate $$\label{n0531} \\|(\varrho^k,{\mathbf}{N}^k)\\|_{L^\infty(I_T,H^3)}+ \\|(\varrho^k,{\mathbf}{N}^k)_t\\|_{L^\infty(I_T,H^2)}+\\|{\mathbf}{u}^k\\|_{{V}_T}\leq C\quad\mbox{for any }k\geq 1,$$ which, together with the regularity estimates on the Stokes equations, implies that $$\label{n0531presure} \\|\nabla{\tilde{q}}^k\\|_{L^\infty(I_T,H^2)}+\\|\tilde{q}^k\\|_{L^\infty(I_T,L^6)} \leq C\quad\mbox{for any }k\geq 1.$$ In addition, by the mass equation $_1$, $$\label{}\inf_{\mathbf{x}\in\Omega}\{(\varrho_0+\bar{\rho})(\mathbf{x})\} \leq \varrho^k(t,{\mathbf}{x})+\bar{\rho}({\mathbf}{x})\leq \sup_{\mathbf{x}\in\Omega}\{(\varrho_0+\bar{\rho})({\mathbf{x}})\}\quad\mbox{for any }k\geq 1.$$ In order to take the limit in , we shall further show $\{{\mathbf}{u}_k\}_{k=1}^\infty$ is a Cauchy sequence. To this end, we define $$(\bar{\varrho}^{k+1},\bar{{\mathbf}{u}}^{k+1},\bar{{\mathbf}{N}}^{k+1},\nabla \bar{q}^{k+1})= ({\varrho}^{k+1}-{\varrho}^{k},{\mathbf}{u}^{k+1}-{\mathbf}{u}^k,{{\mathbf}{N}}^{k+1}-{{\mathbf}{N}}^{k},\nabla ( \tilde{q}^{k+1}-\tilde{q}^{k})),$$ which satisfies $$\label{difeequion}\left\{\begin{array}{l} \bar{\varrho}_t^{k+1}+{\mathbf}{u}^{k}\cdot\nabla \bar{\varrho}^{k+1}=-\bar{{\mathbf}{u}}^{k}\cdot\nabla \bar{\rho}-\bar{{\mathbf}{u}}^k\cdot\nabla \varrho^k, \\[1mm] (\varrho^{k+1}+\bar{\rho})\bar{{\mathbf}{ u}}_t^{k+1}+\nabla\bar{q}^{k+1}-\mu \Delta \bar{\mathbf{u}}^{k+1}= ({\mathbf}{N}^{k+1} +\bar{{\mathbf}{M}})\cdot\nabla \bar{{\mathbf}{N}}^{k+1}-g \bar{\varrho}^{k+1} {\mathbf}{e}_3-\bar{\varrho}^{k+1}{{\mathbf}{ u}}_t^{k}\\ -(\varrho^{k+1}+\bar{\rho}){{\mathbf}{ u}^{k}}\cdot\nabla \bar{{\mathbf}{u}}^{k}-\bar{{\mathbf}{N}}^{k+1}\cdot\nabla {{\mathbf}{N}}^{k}-(\varrho^{k}+\bar{\rho}){\bar{{\mathbf}{ u}}^{k}}\cdot\nabla {\mathbf}{u}^{k-1}-\bar{\varrho}^{k+1}{{{\mathbf}{ u}}^{k-1}}\cdot\nabla {\mathbf}{u}^{k-1}:={\mathbf}{f}_k,\\[1mm] \bar{{\mathbf}{N}}_t^{k+1}+{\mathbf}{u}^{k}\cdot\nabla \bar{{\mathbf}{N}}^{k+1}-\nabla {\mathbf}{u}^{k}\bar{{\mathbf}{N}}^{k+1}= \bar{{\mathbf}{M}}\cdot\nabla \bar{{\mathbf}{u}}^{k}-\bar{{\mathbf}{u}}^{k}\cdot\nabla {{\mathbf}{N}}^{k}+\nabla \bar{{\mathbf}{u}}^{k}{{\mathbf}{N}}^{k},\\[1mm] \mathrm{div}\mathbf{u}^{k+1}=0,\ \mathrm{div}\mathbf{N}^{k+1}=0\end{array}\right.$$ with initial data $$\label{}(\bar{\varrho}^{k+1},\bar{{\mathbf}{u}}^{k+1}, \bar{{\mathbf}{N}}^{k+1})=(0,{\mathbf}{0},{\mathbf}{0})\ \mbox{in}\ \Omega.$$ Recalling that $\kappa^8 T = 1$ and $T\leq 1$, we easily find by $_1$ and $_2$ that $$\begin{aligned} &\label{rhoNestimes}\sup_{t\in I_T}\\|(\bar{\varrho}^{k+1},\bar{{\mathbf}{N}}^{k+1})\\|_{H^1}\lesssim T\sup_{t\in I_T}\\|\bar{{\mathbf}{u}}^k\\|_{H^2}, \\ &\sup_{t\in I_T}\\|(\bar{\varrho}^{k+1},\bar{{\mathbf}{N}}^{k+1})_t\\|_{L^2}\lesssim \sup_{t\in I_T}\\|\bar{{\mathbf}{u}}^k\\|_{H^2}.\nonumber \end{aligned}$$ Similarly to the derivation of and , we get from $_2$ that $$\\|\bar{{\mathbf}{u}}^{k+1}(t) \\|_{L^2}^2+\int_0^t\\|\nabla \bar{{\mathbf}{u}}^{k+1}(s)\\|_{L^2}^2{\mathrm}{d}s \lesssim\\|{\mathbf}{f}_k\\|_{L^2(Q_T)}^2$$ and $$\begin{aligned} \\| \nabla \bar{{\mathbf}{u}}^{k+1}(t)\\|_{L^2}^2+\int_0^t \\|\bar{{\mathbf}{u}}_s^{k+1}(s)\\|_{L^2}^2{\mathrm}{d}s\lesssim \\|{\mathbf}{f}_k\\|_{L^2(Q_T)}^2\mbox{ for any }t\in {I}_t. \end{aligned}$$ Next we continue to derive a bound for $\nabla^2 \bar{{\mathbf}{u}}^{k+1}$. To this end, we differentiate $_2$ in $t$, multiply the resulting equations by $\bar{{\mathbf}{u}}_{t}^{k+1}$ in $L^2$ to obtain $$\begin{aligned} &\frac{1}{2}\frac{{\mathrm}{d}}{{\mathrm}{d}t}\int_{\Omega}(\varrho^{k+1}+\bar{\rho})\|\bar{{\mathbf}{u}}^{k+1}_t\|^2{\mathrm}{d}{\mathbf}{x} +\mu\int_{\Omega} \|\nabla \bar{{\mathbf}{u}}^{k+1}_t\|^2{\mathrm}{d}{\mathbf}{x}\\ &=\frac{1}{2}\int_{\Omega}\varrho_t^{k+1}\|\bar{{\mathbf}{u}}^{k+1}_t\|^2{\mathrm}{d}{\mathbf}{x} +\int_\Omega [{\mathbf}{N}^{k+1}_t\cdot\nabla \bar{{\mathbf}{N}}^{k+1}-\varrho^{k+1}_t{{\mathbf}{ u}^{k}}\cdot\nabla \bar{{\mathbf}{u}}^{k}-(\varrho^{k+1}+\bar{\rho}){\mathbf}{ u}^{k}_t\cdot\nabla \bar{{\mathbf}{u}}^{k}]\cdot{\bar{{\mathbf}{u}}^{k+1}_t}{\mathrm}{d}{\mathbf}{x} \\ &\quad -\int_\Omega \left\{({\mathbf}{N}^{k+1}+\bar{{\mathbf}{M}})\cdot\nabla \bar{{\mathbf}{u}}^{k+1}_{t}\cdot\bar{{\mathbf}{N}}^{k+1}_t -[(\varrho^{k+1}+\bar{\rho}){{\mathbf}{ u}^{k}}\cdot\nabla\bar{{\mathbf}{u}}^{k+1}_{t}+{{\mathbf}{ u}^{k}}\cdot\nabla (\varrho^{k+1}+\bar{\rho}) \bar{{\mathbf}{u}}^{k+1}_{t}]\cdot\bar{{\mathbf}{u}}^{k}_t\right\}{\mathrm}{d}{\mathbf}{x} \\ &\quad -\int_\Omega [g \bar{\varrho}^{k+1} {\mathbf}{e}_3+\bar{\varrho}^{k+1}{{\mathbf}{ u}}_t^{k} +\bar{{\mathbf}{N}}^{k+1}\cdot\nabla {{\mathbf}{N}}^{k}+(\varrho^{k}+\bar{\rho}){\bar{{\mathbf}{ u}}^{k}}\cdot\nabla {\mathbf}{u}^{k-1} +\bar{\varrho}^{k+1}{{{\mathbf}{ u}}^{k-1}}\cdot\nabla {\mathbf}{u}^{k-1}_t]_t\cdot\bar{{\mathbf}{u}}^{k+1}_t{\mathrm}{d}{\mathbf}{x}, \end{aligned}$$ where the terms on the right hand side can be bounded from above by $$C[\\|\sqrt{(\varrho^{k+1}+\bar{\rho})}\bar{{\mathbf}{u}}_t^{k+1}\\|_{L^2}^2 +\sup_{t\in I_T}(\\|\bar{{\mathbf}{u}}^k\\|_{H^2}^2+\\|\bar{{\mathbf}{u}}^{k}_t\\|_{L^2}^2)]+\frac{\mu}{2}\\|\nabla \bar{{\mathbf}{u}}_t^{k+1}\\|_{L^2}^2.$$ Therefore, we have $$\frac{{\mathrm}{d}}{{\mathrm}{d}t}\\|\sqrt{(\varrho^{k+1}+\bar{\rho})}\bar{{\mathbf}{u}}^{k+1}_t\\|^2_{L^2}+\\|\nabla \bar{{\mathbf}{u}}_t^{k+1}\\|_{L^2}^2 \lesssim \\|\sqrt{(\varrho^{k+1}+\bar{\rho})}\bar{{\mathbf}{u}}_t^{k+1}\\|_{L^2}^2 +\sup_{t\in I_T}(\\|\bar{{\mathbf}{u}}^k\\|_{H^2}^2+\\|\bar{{\mathbf}{u}}^{k}_t\\|_{L^2}^2).$$ Applying Gronwall’s lemma to the above inequality and recalling $\bar{{\mathbf}{u}}_t^{k+1}(0)=0$, we conclude $$\\|\bar{{\mathbf}{u}}^{k+1}_t(t)\\|_{L^2}^2 + \int_0^t\\|\nabla \bar{{\mathbf}{u}}_s^{k+1}(s)\\|_{L^2}^2{\mathrm}{d}s \lesssim T\sup_{t\in I_T}(\\|\bar{{\mathbf}{u}}^k\\|_{H^2}^2 + \\|\bar{{\mathbf}{u}}^{k}_t\\|_{L^2}^2)\mbox{ for any }t\in I_T,$$ With the help of the classical regularity estimate on the Stokes problem, we obtain $$\begin{aligned} &\begin{aligned} \\|\nabla^2 \bar{{\mathbf}{u}}^{k+1}\\|_{L^2}^2 +\\|\nabla\bar{q}^{k+1}\\|_{L^2}^2\lesssim \\|\bar{{\mathbf}{u}}^{k+1}_t\\|_{L^2}^2+\\|{\mathbf}{f}_k\\|_{L^2}^2, \end{aligned}\end{aligned}$$ where $\\|{\mathbf}{f}_k\\|_{L^2}$ can be bounded as follows. $$\label{fkestimates}\\|{\mathbf}{f}_k\\|_{L^2}^2\lesssim T\sup_{t\in I_T}(\\|\bar{{\mathbf}{u}}^k\\|_{H^2}^2+ \\|\bar{{\mathbf}{u}}^{k}_t\\|_{L^2}^2)+T\\|\nabla \bar{{\mathbf}{u}}_t^{k}\\|_{L^2(Q_T)}^2.$$ Summing up the estimates –, we arrive at $$\label{} \begin{aligned} &\sup_{t\in I_T}(\\|(\bar{\varrho}^{k+1},\bar{{\mathbf}{N}}^{k+1})\\|_{H^1}^2+\\|\nabla\bar{q}^{k+1}\\|_{L^2}^2+\\|\bar{{\mathbf}{u}}^{k+1} \\|_{H^2}^2+\\|\bar{{\mathbf}{u}}^{k+1}_t \\|_{L^2}^2)+\\|\nabla \bar{{\mathbf}{u}}^{k+1}_t\\|_{L^2(Q_T)}^2 \\ & \quad \leq \bar{C}T\left[\sup_{t\in I_T}(\\|\bar{{\mathbf}{u}}^k\\|_{H^2}^2 + \\|\bar{{\mathbf}{u}}^{k}_t\\|_{L^2}^2)+\\|\nabla \bar{{\mathbf}{u}}_t^{k}\\|_{L^2(Q_T)}^2\right]\\ &\quad \leq \frac{1}{2}\left[\sup_{t\in I_T}(\\|\bar{{\mathbf}{u}}^k\\|_{H^2}^2+\\|\bar{{\mathbf}{u}}^k_t\\|_{L^2}^2)+\\|\nabla \bar{{\mathbf}{u}}_t^{k}\\|_{L^2(Q_T)}^2\right]\qquad\mbox{ for any }k\geq 1, \end{aligned}$$ provided $T$ is chosen so small that $T\leq ({2\bar{C}})^{-1}$, where the constant $\bar{C}$ may depend on $(\varrho_0,{\mathbf}{u}_0,{\mathbf}{N}_0)$ and other physical parameters in the perturbed equations. The above inequality implies that $$\sum_{k=1}^\infty\left(\\|(\bar{\varrho}^{k+1},\bar{{\mathbf}{N}}^{k+1})\\|_{L^\infty(I_T,H^1)}^2+ \\|\bar{{\mathbf}{u}}^{k+1}\\|_{L^\infty(I_T,H^2)}^2+\\|\bar{{\mathbf}{u}}^{k+1}_t\\|_{L^\infty(I_T,L^2)}^2\right)<\infty.$$ Hence, $({\varrho}^{k},{{\mathbf}{u}}^{k},{{\mathbf}{N}}^{k},{\mathbf}{u}_t^{k})$ is a Cauchy sequence in $C(\bar{I}_T,H^1)\times C(\bar{I}_T,H^2)\times C(\bar{I}_T,H^1)\times C(\bar{I}_T,L^2)$, and $$\label{strongconvegneuN} ({\rho}^{k},{{\mathbf}{u}}^{k},{{\mathbf}{N}}^{k},{\mathbf}{u}_t^{k})\rightarrow ( {\rho}, {{\mathbf}{u}},{{\mathbf}{N}},{\mathbf}{u}_t)$$ strongly in $C(\bar{I}_T,H^1)\times C(\bar{I}_T,H^2)\times C(\bar{I}_T,H^1)\times C(\bar{I}_T,L^2)$. Consequently, by , and , we easily verify that $((\varrho,{\mathbf}{N}),{\mathbf}{u})\in H_T\times V_T$ is a unique solution to the following problem with an associated pressure $\tilde{q}$ enjoying the regularity : $$\left\{\begin{array}{l} \varrho_t+{\bf u}\cdot\nabla \varrho=-{\bf u}\cdot\nabla\bar{\rho} , \\[1mm] (\varrho+\bar{\rho}){\bf u}_t+(\varrho+\bar{\rho}){\bf u}\cdot\nabla {\bf u}+\nabla\tilde{q}-\mu \Delta \mathbf{u}= ({\mathbf}{N}+\bar{{\mathbf}{M}})\cdot\nabla {\mathbf}{N}-g \varrho {\mathbf}{e}_3,\\[1mm] {\mathbf}{N}_t+{\mathbf}{u}\cdot\nabla {\mathbf}{N}= ({\mathbf}{N}+\bar{{\mathbf}{M}})\cdot\nabla {\mathbf}{u},\\[1mm] \mathrm{div}\mathbf{u}=0,\ \mathrm{div}\mathbf{N}=0 \end{array}\right.$$ with initial data $$(\varrho,\mathbf{u},\mathbf{N} )\|_{t=0}=(\varrho_0,{\mathbf u}_0,\mathbf{N}_0).$$ Finally, if we define $${q}:=\tilde{q}-\|{\mathbf}{N}+\bar{{\mathbf}{M}}\|^2/2,$$ and use the facts that $$(\nabla \times({\mathbf}{N}+\bar{{\mathbf}{M}}) )\times ({\mathbf}{N}+\bar{{\mathbf}{M}})=({\mathbf}{N}+\bar{{\mathbf}{M}})\cdot\nabla {\mathbf}{N}-\nabla \frac{\|{\mathbf}{N}+\bar{{\mathbf}{M}}\|^2}{2},$$ and $$\nabla \times ({\mathbf}{u}\times ({\mathbf}{N}+\bar{{\mathbf}{M}}))= ({\mathbf}{N}+\bar{{\mathbf}{M}})\cdot\nabla {\mathbf}{u}-{\mathbf}{u}\cdot\nabla {\mathbf}{N},$$ we obtain Proposition \[pro:0401new\]. Acknowledgements. The research of Fei Jiang was supported by the NSFC (Grant Nos. 11301083 and 11471134) and the NSF of Fujian Province of China (Grant No. 2014J01011), the research of Song Jiang by the National Basic Research Program under the Grant 2011CB309705, NSFC (Grant Nos. 11229101, 11371065) and the Beijing Center for Mathematics and Information Interdisciplinary Sciences. [99]{} R.A. Adams, John J.F. Fournier, Sobolev Space, 2005, Academic Press: New York. H. Cabannes, Theoretical Magnetofluiddynamics, Academic Press, New York, 1970. Y. Choa, H. Kim, Unique solvability for the density-dependent Navier-Stokes equations, Nonlinear Analysis 59(2004), 465-489. T.G. Cowling, Magnetohydrodynamics, Interscience Publishers, New York, 1957. R. Duan, F. Jiang, S. Jiang, On the Rayleigh Taylor instability for incompressible, inviscid magnetohydrodynamic flows, SIAM J. Appl. Math. [71]{} (2011), 1990-2013. D. Erban, The equations of motion of a perfect fluid with free boundary are not well posed, Comm. PDE [12]{} (1987), 1175-1201. C.L. Feffermana, D.S. McCormick, J.C. Robinsonb, J.L. Rodrigo, Higher order commutator estimates and local existence for the non-resistive MHD equations and related models, J. Funct. Anal., [267]{}(2014), 1035-1056. G.P. Galdi, An Introduction to the Mathematical Theory of the Navier-Stokes Equations: Steady-State Problems (Second Edition), Academic Press: Springer. Loukas Grafakos, Classical fourier analysis, second edition, 2008, Springer. Y. Guo, C. Hallstrom, D. Spirn, Dynamics near unstable, interfacial fluids, Comm. Math. Phys. [270]{} (2007), 635-689. Y. Guo, W. Strauss, Instability of periodic BGK equilibria, Comm. Pure Appl. Math. [48]{} (1995), 861-894. Y. Guo, I. Tice, Linear Rayleigh-Taylor instability for viscous, compressible fluids, SIAM J. Math. Anal. [42]{} (2011), 1688-1720. Y. Guo, I. Tice, Compressible, inviscid Rayleigh-Taylor instability, Indiana Univ. Math. J. [60]{} (2011), 677-712. Y. Guo, I. Tice, Local well-posedness of the viscous surface wave problem without surface tension, Analysis and PDE [6]{} (2013), 287-369. R. Hide, Waves in a heavy, viscous, incompressible, electrically conducting fluid of variable density, in the presence of a magnetic field, Proc. Roy. Soc. (London) A [233]{} (1955), 376-396. H. Hwang, Variational approach to nonlinear gravity-driven instability in a MHD setting, Quart. Appl. Math. [66]{} (2008), 303-324. H. Hwang, Y. Guo, On the dynamical Rayleigh-Taylor instability, Arch. Rational Mech. Anal. [167]{} (2003), 235-253. X.P. Hu, F.H. Lin, Global existence for two dimentional incompressible magnetohydrodynamic flow with zero magnetic diffusivity, arXiv:1405.0082v1 \[math.AP\] 1 May 2014. J. Jang, I. Tice, Instability theory of the Navier-Stokes-Poisson equations, [6]{} (2013), 1121–1181. F. Jiang, S. Jiang, G. Ni, Nonlinear instability for nonhomogeneous incompressible viscous fluids, Science China Math. [56]{} (2013), 665-686. F. Jiang, S. Jiang, W. Wang, On the Rayleigh-Taylor instability for the incompressible viscous magnetohydrodynamic equations, Comm. PDE [39]{} (2014), 399–438. F. Jiang, S. Jiang, W. Wang, On the Rayleigh-Taylor instability for two uniform viscous incompressible flows, Chinese Annals of Mathematics, Series B, [35]{} (2014), 907-940. S. Kawashima, Systems of a hyperbolic-parabolic composite type, with applications to the equations of magnetohydrodynamics, Ph. D. Thesis, Kyoto University, 1983. M. Kruskal, M. Schwarzschild, Some instabilities of a completely ionized plasma, Proc. Roy. Soc. (London) A [233]{} (1954), 348-360. A.G. Kulikovskiy, G.A. Lyubimov, Magnetohydrodynamics, Addison-Wesley, Reading, MA, 1965. L.D. Landau, E.M. Lifshitz, L.P. Pitaevskii, Electrodynamics of Continuous Media, Vol.8, 1984 (Translated from the Russian). X. Li, N. Su, D. Wang, Local strong solution to the compressible magnetohydrodynmic flow with large data, J. Hyper. Diff. Eqns. [8]{} (2011), 4315-4369. F. Lin, P. Zhang, Global small solutions to an MHD type system: the three-dimensional case, Comm. Pure. Appl. Math. [[67]{}]{} (2014), 531-580. A. Novotn[y]{}, I. Stra[š]{}kraba, Introduction to the Mathematical Theory of Compressible Flow, Oxford University Press, USA, 2004. J. Prüss, G. Simonett, On the Rayleigh-Taylor instability for the two-phase Navier-Stokes equations, Indiana Univ. Math. J. [59]{} (2010), 1853-1871. L. Rayleigh, Analytic solutions of the Rayleigh equations for linear density profiles, Proc. London. Math. Soc. [14]{} (1883), 170-177. Y. Wang, Critical magnetic number in the MHD Rayleigh-Taylor instability, J. Math. Phys. [53]{} (2012), 073701. Y. Wang, I. Tice, The viscous surface-internal wave problem: nonlinear Rayleigh-Taylor instability, Comm. PDE [37]{} (2012), 1967-2028.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Codalema is one of the experiments devoted to the detection of ultra high energy cosmic rays by the radio method. The main objective is to study the features of the radio signal induced by the development in the atmosphere of extensive air showers (EAS) generated by cosmic rays in the energy range of $10^{16}$-$10^{18}~eV$. After a brief presentation of the detector features, the main results obtained are reported (emission mechanism, lateral distribution of the electric field, energy calibration, etc.). The first studies of the radio wave front curvature are discussed as new preliminary results.' address: - 'SUBATECH IN2P3-CNRS/Université de Nantes/École des Mines de Nantes France' - 'LESIA, Observatoire de Paris-Meudon France - Station de Radioastronomie de Nançay France' author: - 'Rebai, A.' - the CODALEMA collaboration bibliography: - 'rebai.bib' title: Some recent results of the Codalema Experiment --- UHECR, radiodetection, antennas, radio emission mechanism, EAS Introduction ============ A century after the discovery of cosmic rays, several fundamental issues related to the nature and the origin of ultra high energy cosmic rays (UHECR) remain unanswered, despite intensive experimental efforts. The main difficulty opposing the progress is due to the extremely low flux of UHECR (1 particle/$km^2$/century at $10^{20}~eV$) and the present performances of particle detectors arrays and fluorescence telescopes [@AugerComp2010PRL; @HiresComp2010PRL]. In the recent decade, the measurement of the radio counterparty of EAS becomes a promising technique. Many scientific collaborations like CODALEMA in France [@GeoMagOrigCodAstroPh] and LOPES in Germany [@LDFRadioSigEasLOPESAstroPh] have demonstrated the feasibility of this method to deduce the EAS features. The potentialities of the radiodetection resides in several avantages, among them may be mentioned: the operating duty cycle close to $100\%$, the sensitivity to the shower longitudinal development in the atmosphere and the mechanical robustness, the simplicity and the low cost of the antennas. The CODALEMA experiment is installed at the radio observatory site (Nançay, 47.3$^{\circ}$N, 2.1$^{\circ}$E and 137 m above sea level). Its main goal is to improve the pioneer experimental results [@AllanRevue], taking avantage of ultrafast electronic devices and a quiet radio environment from anthropic transmitters in the detection bandwidth [@RadioElecFeaturEASCodaAstroPh]. From the phenomenological point of view, CODALEMA has made progress in the understanding of radio signal origin, showing that the geomagnetic field is the main actor in radio signal emission via the geomagnetic mechanism [@GeoMagOrigCodAstroPh], and showing recently the contribution of a second emission mechanism due to the shower negative charge excess [@ICRC2011MarinExcess]. This contribution reports on the last results of the experiment CODALEMA with updated data set, recalling the observations of the north-south asymmetry and of the energy correlation. Finally the reconstruction method of the radio wavefront radius of curvature is presented. Experimental situation ====================== The CODALEMA experiment is made of two main arrays of detectors. The first array is built with $24$ short active dipoles antennas distributed on a cross geometry with dimensions $400~m$ by $600~m$. This apparatus is used to study the EAS radio counterpart. The dipole antenna is made by two radiator arms each $60~cm$ long at a height of $1.2~m$. The antenna design was optimized to reach an almost isotropic pattern. Low noise amplifier (LNA) is used to amplify the electric signal. It is conceived to be sensitive to the radio galactic background and is linear over a wide frequency band from $0.1$ MHz up to $230$ MHz. The second apparutus is a ground-based particle detector array formed by $17$ plastic scintillators placed on a square with $340~m$ side. It measures the primary particle energy and provides the trigger signal to the other detector arrays. The entire acquisition system (DAQ) is trigged by the passage of secondary particles in coincidence through each of the five central scintillators. A North-South asymmetry: a geomagnetic effect signature in the production of the radio signal ============================================================================================= CODALEMA is the first experience which has reported a large and stable assymetry in counting rates between showers coming from north and south [@GeoMagOrigCodAstroPh; @OlivierArena2010]. This asymmetry has been interpreted as a signature of the geomagnetic field effect in the radio signal emission process. The local geomagnetic field $\vec{B}$ is the main cause of this assymmetry through the action of Lorentz force on the secondary charged particles via the $\vec{v} \wedge \vec{B}$ term. We find that among the $2030$ events detected in coincidences with the two arrays: $1708$ events coming from the north and $322$ coming from the south (see Fig. \[NS\_anisotropy\], Left). The respective ratio equal to $0.188$ is consistent with the value of $0.17$ obtained with a previous statistic used in [@GeoMagOrigCodAstroPh]. A pure statistical effect is excluded with $15~\sigma$. One can enhance here on the fact that this asymmetry is stable over time and with different data samples. ![[Left:]{} A plot sky showed the arrival directions distribution of detected EAS events. The red dot indicates the local geomagnetic field ($\theta=27^\circ$,$\phi=0^\circ$). [Right:]{} Sky map of the geomagnetic model theoretical prediction based on the projection of the product $\vec{v} \wedge \vec{B}$ on the East-West direction multiplied by the trigger coverage map. The colormap is normalized to 1.[]{data-label="NS_anisotropy"}](./rebai_fig1 "fig:"){width="40.00000%"} ![[Left:]{} A plot sky showed the arrival directions distribution of detected EAS events. The red dot indicates the local geomagnetic field ($\theta=27^\circ$,$\phi=0^\circ$). [Right:]{} Sky map of the geomagnetic model theoretical prediction based on the projection of the product $\vec{v} \wedge \vec{B}$ on the East-West direction multiplied by the trigger coverage map. The colormap is normalized to 1.[]{data-label="NS_anisotropy"}](./rebai_fig2 "fig:"){width="30.00000%"} Energy calibration of the Codalema antenna array ================================================ The energy of the primary particle $E_p$ is one of the most important parameters for studying EAS. To avoid the use of the particles detectors array, the radio technique must demonstrate its ability to estimate the primary particle energy with the information given by the radio signal alone. In this perspective, a study of the correlation between the shower energy and the electric field created at the shower axis $E_{0}$ is the natural way to determine the energy calibration response of the antenna array. In this goal, after sampling the particles density at ground a NKG lateral distribution is deduced to measure the total number of charged paticles $N_e$ (mostly electrons and positrons) in the shower front. The energy $E_p$ is deduced from the constant intensity cuts method (CIC). The procedure gives a relative error equal to $30\%$. The electric field $E_{0}$ is determined by the lateral distribution function (LDF). The Allan formula has been used to fit the radio LDF with an exponential law (Fig. \[LDFstudies\]) following this formula: $$E_i = E_0.exp(-\frac{\sqrt{( (x_i - x_c)^2 + (y_i - y_c)^2 - ((x_i - x_c).cos(\phi).sin(\theta) + (y_i - y_c).sin(\phi).sin(\theta))^2)}}{d_0})$$ ![Radio lateral distributions reconstructed from antenna signals. The full line show the result of an exponential law fit (Allan formula). [Left:]{} We show the lateral distribution of an event detected by 19 antennas. The primary particle energy is estimated to $E_{p}=7.97~10^{16}~eV$. Error bars correspond to the radio galactic background. In this example, the signal amplitudes are close to the noise level. [Right:]{} We show the lateral distribution of an other event ,detected by 16 antennas, with larger energy $E_p=1.64~10^{18}~eV$. In this case, error bars are small compared to signals.[]{data-label="LDFstudies"}](./rebai_fig3 "fig:"){width="45.00000%"} ![Radio lateral distributions reconstructed from antenna signals. The full line show the result of an exponential law fit (Allan formula). [Left:]{} We show the lateral distribution of an event detected by 19 antennas. The primary particle energy is estimated to $E_{p}=7.97~10^{16}~eV$. Error bars correspond to the radio galactic background. In this example, the signal amplitudes are close to the noise level. [Right:]{} We show the lateral distribution of an other event ,detected by 16 antennas, with larger energy $E_p=1.64~10^{18}~eV$. In this case, error bars are small compared to signals.[]{data-label="LDFstudies"}](./rebai_fig8 "fig:"){width="45.00000%"} where, $\theta$ and $\phi$ are respectivly zenithal and azimuthal angles reconstructed by a planar fit. This fit has four free parameters $E_{0}$, the LDF decay distance $d_{0}$ and the radio-shower core coordinates on the ground $(x_c,y_c)$, providing the antennas coordinates $(x_i,y_i)$ and the radio filtered pulse amplitude $E_i$ for each antennas. The error on $E_0$ has been estimated through a Monte Carlo analysis. It consist of repeating the LDF fit with $E_i$ values, randomly selected from gaussian probability density function centred on the measured values $E_i$ and a gaussian standard deviation $\sigma_{i}$ took as the RMS of measured radio noise on the antenna. This Monte Carlo method enables to explore the entire phase space and thus to estimate the $E_0$ error on a by event basis. At the end of the procedure a statistical error less than $22~\%$ is deduced. A linear regression is then used to deduce the correlation coefficients in the $(E_p, E_0)$ plan, assuming some hypothesis: (i) The two variables are represented on a linear scale. (ii) Gaussian errors are used on both observables. (iii) The two observables $E_p$ and $E_0$ are assumed to be independent, to cancel the nonlinear covariance term. Results of this procedure is presented on Fig. \[Correlationstudies\]. ![[Left:]{} $E_0$ statistical errors $\frac{\sigma E_0}{E_0}$ histogram of the LDF fit method. [Right:]{} Fit of the correlation between the primary particle energy $E_p$ and the electric field at the shower axis $E_0$. This study strongly depends on both $E_p$ and $E_0$ errors.[]{data-label="Correlationstudies"}](./rebai_fig4 "fig:"){width="45.00000%"} ![[Left:]{} $E_0$ statistical errors $\frac{\sigma E_0}{E_0}$ histogram of the LDF fit method. [Right:]{} Fit of the correlation between the primary particle energy $E_p$ and the electric field at the shower axis $E_0$. This study strongly depends on both $E_p$ and $E_0$ errors.[]{data-label="Correlationstudies"}](./rebai_fig5 "fig:"){width="45.00000%"} It shows that the electric field created at the shower axis $E_{0}$ is clearly correlated with the primary particle energy $E_p$ following the relationship $E_0=a.E_p+b$. This correlation shows that the radio signal emitted by EAS seems coherent. An energy calibration relationship can be deduced, with the following form $E_R=(1/a).E_0-(b/a)$ which allows to have a radio estimation $E_R$ of the energy through the radio method once we have measured $E_0$. It is conceivable that for a future autonomous antennas array, we can measure the primary particle energy only with a pure radio observable. Radius of curvature reconstruction with CODALEMA ================================================ Because theoretical developments indicate that the radio signal shape depends on the shower longitudinal development, it is waited that the wavefront shape provides information on the nature of the primary particle. In the first step of the CODALEMA’s analysis, the wavefront was assimilated to a plan determined by a simple planar fit using the arrival times and locations of each tagged antenna. More detailed studies indicates today (Fig. \[RadiusCurvature\]. Left) that the measured wavefront differs slightly from the plan in most cases and that it exhibits a curved geometry, favoring the idea of a privileged center for the radio emission during the shower development [@RebaiJJC2010; @Schroder]. To take into account of these experimental observations, one of the simplest hypothesis is to assume that the maximum of the filtered pulse is linked to a radio signal emission center located along the shower axis. This leads to define a curvature radius $R_c$. Several modelisations suggest that this observable could then be related to the shower maximum, $X_{max}$, which is directly correlated to the UHECR chemical composition. This possibility has been investigated using the present data. Our fitting method is based on the fit of the residue between the real wavefront determined by arrivals times distribution and the planar wavefront. A parabolic dependence is used to account for the difference. Results of the calculations are presented in Fig. \[RadiusCurvature\]. The distribution of the $R_c$ presents a maximum at $4~km$ in global aggrement with the waited characteristic altitude of the emission maximum. However, the tail of the distribution extends up to $20~km$. A present time, the physical interpretation of this long tail is not well understood. This may be due both to a poor estimate of arrival time and to a biased assessment of the estimated error on its time measurement, an arrival time error of $10~ns$ is used. Improvement of these points are underway. ![[Left:]{} Expected delay versus measured delay. The black line presents the plane wave fit, Despite the $10~ns$ error bars on both axes, many points are located far from the line, which shows that the wavefront differs from a plane. [Right:]{} The histogram of the radius of curvature $R_c$ distribution for 1010 selected events. The distribution maximum is located at $4~km$.[]{data-label="RadiusCurvature"}](./rebai_fig6 "fig:"){height="0.3\textheight" width="45.00000%"} ![[Left:]{} Expected delay versus measured delay. The black line presents the plane wave fit, Despite the $10~ns$ error bars on both axes, many points are located far from the line, which shows that the wavefront differs from a plane. [Right:]{} The histogram of the radius of curvature $R_c$ distribution for 1010 selected events. The distribution maximum is located at $4~km$.[]{data-label="RadiusCurvature"}](./rebai_fig7 "fig:"){height="0.3\textheight" width="45.00000%"} Conclusions =========== Codalema is installed in a radio quiet environment. This advantage has enabled very high accuracy radio signal measurements and several progress in the understanding of the radio emission mechanisms like the geomagnetic field effect or the energy calibration method. A method for the wavefront radius of curvature reconstruction has been presented. The first analysis of the radius of curvature presented shows extremely interesting perspectives with the aim of determining the cosmic rays nature. The deduced location of the emission point are in a satisfactory agreement (for radii less than $10~km$) with the waited values, but for the larger curvatures the values remain poorly understood. Unfortunately the low timing accuracy ($<10~ns$) may limit the analysis with current data. The current array extension which uses standalone radio stations to reach a larger array ($1.5~km^{2}$) will increased the available statistic and improve our interpretations. This work has been made a part under a grant from R' egion Pays de la Loire. The author wishes to thank the CNRS (Centre national de la recherche scientifique) for funding his work. The author likes to thank the SF2A week organizers for their hospitality.	{ "pile_set_name": "ArXiv" }
--- author: - Zhenghan Liao - 'William T. M. Irvine' - Suriyanarayanan Vaikuntanathan title: 'Supplemental Material: Rectification of energy and motion in non-equilibrium parity violating metamaterials' --- The formula for energy flux =========================== In this section, we derive the formula for energy flux, [Eq.]{} (4) in the main text. The force $F$ in this section is a general conservative force, which does not need to be linear. We use the following strategy to determine the energy flux. First we define the energy $E_i$ of particle $i$, and then write down an energy balance relation, that expresses the infinitesimal energy change $\dd E_i$ using stochastic calculus. Finally, we collect terms in $\dd E_i$ that couple neighboring particles together and identify this as the energy transfer among particles. The energy of particle $i$ is defined as $$\label{eqnS:energy_individual} E_i = \frac{1}{2}m_iv_i^Tv_i + U_{ii} + \frac{1}{2}\sum_{j\neq i}U_{ij},$$ where the first term is the kinetic energy, the second term denotes the on-site potential, and the last term is the shared spring energy between the particle and its neighbors. We use Ito’s formula to calculate $\dd E_i$. Ito calculus provides the advantage that the stochastic terms in $\dd E_i$ vanish under time-averaging. For a stochastic differential equation (SDE) of variable $X$(vector) with drift $\mu$(vector) and diffusion $\sigma$(matrix) $$\label{eqnS:SDE_general} \dd X = \mu \dd t + \sigma \dd W ,$$ where $\dd W$ is a vector consisting of standard Wiener processes, Ito’s formula gives the SDE of function $f(X)$ $$\label{eqnS:SDE_ito} \dd f(X) = ((\nabla_X^Tf)\mu + \frac{1}{2}\tr[\sigma \sigma^T \nabla_X\nabla_X^Tf])\dd t + (\nabla_X^Tf) \sigma \dd W,$$ where $\nabla_X$ denotes the gradient with respect to $X$, the superscript $T$ denotes the transpose, and $\tr$ denotes the trace. We begin by writing the equation of motion of our system [Eq.]{} (1) in the form of a stochastic differential equation [Eq.]{} . We represent $N$ particles’ position by a column vector $z = \sum_{i=1}^N \ket{i}\otimes z_i$, where $\ket{i}$ denotes the 2D subspace of particle $i$. Similar representations are also applied to $v$ and $\eta$. Then we get $$\begin{aligned} X &= \pmqty{ z & v & \eta }^T, \label{eqnS:SDE_X}\\ \mu &= \pmqty{ v \\ \frac{1}{m}(-\nabla_zU - BAv - \gamma v + \eta) \\ -\frac{1}{\tau}\eta }, \label{eqnS:SDE_mu} \\ \sigma &= \text{diag} \pmqty{0 & 0 & \frac{\sqrt{2\gamma T_a}}{\tau} I}, \label{eqnS:SDE_sigma}\end{aligned}$$ where $U$ is the total energy of the system, $A$ is an antisymmetric matrix $A=\sum_i \ket{i}\bra{i}\otimes \mqty(0 & 1 \\ -1 & 0)$, and $\text{diag}()$ means a block-diagonal matrix. Now we apply Ito’s formula [Eq.]{} to our system by associating with the function $f(X)$, the energy of particle $i$, $E_i(X)$. The nonzero terms in the gradient of $E_i$ are $$\begin{aligned} \nabla_{z_i}E_i &= -(F_{ii} + \frac{1}{2}\sum_jF_{ji}), \\ \nabla_{z_j}E_i &= -\frac{1}{2}F_{ij}, \\ \nabla_{v_i}E_i &= m_iv_i.\end{aligned}$$ The term $(\nabla_X^TE_i)\mu$ reads $$\begin{split} (\nabla_X^TE_i)\mu &= (\nabla_{z_i}^TE_i)v_i + \sum_j(\nabla_{z_j}^TE_i)v_j + (\nabla_{v_i}^TE_i)m_i^{-1}(-\nabla_{z_i}U - BAv_i - \gamma v_i + \eta_i) \\ &= -(F_{ii} + \frac{1}{2}\sum_jF_{ji})^T v_i - \sum_j\frac{1}{2}F_{ij}^T v_j + v_i^T (F_{ii} + \sum_jF_{ji}) - \gamma v_i^Tv_i + v_i^T\eta_i \\ &= -\sum_j\frac{1}{2}(v_i + v_j)^T F_{ij} - \gamma v_i^Tv_i + v_i^T\eta_i , \end{split}$$ where we used $F_{ji} = -F_{ij}$ and $v_i^TAv_i = 0$. The term $\frac{1}{2}\tr[\sigma \sigma^T \nabla_X\nabla_X^Tf]$ and $\nabla_X^Tf$ are zero. Finally, the energy change can be written as $$\begin{aligned} \dd E_i &= -\sum_jJ_{ij}\dd t + h_i \dd t, \label{eqnS:flux_dEi} \\ J_{ij} &= \frac{1}{2}(v_i + v_j)^T F_{ij}, \label{eqnS:flux_Jij} \\ h_i &= -\gamma_i v_i^Tv_i +v_i^T\eta_i. \label{eqnS:flux_hi}\end{aligned}$$ $J_{ij}$ is identified as the energy transferred per unit time from particle $i$ to $j$, and $h_i$ is identified as the energy transferred from the bath to particle $i$. As for the steady-state average of $J_{ij}$, we use $\expval{\dd U_{ij} / \dd t =0}$ and the chain rule to simplify [Eq.]{} $$0 = v_i^T F_{ji} + v_j^T F_{ij} = -v_i^T F_{ij} + v_j^T F_{ij},$$ and arrive at the expresion $$\expval{J_{ij}} = \expval{v_j^T F_{ij}}.$$ Numerical method for solving the energy flux ============================================ One straightforward numerical method to compute the flux $J$ is as follows. Our system is determined by the network geometry and parameters $m, k_g, k, B, \gamma, \tau, T_a$. Given the equation of motion [Eq.]{} ,-, one can numerically solve for the covariance $C=\expval{XX^T}$ from the matrix equation $\mu C + C \mu^T = \sigma\sigma^T$ [@Gardiner2009ItoCalculus; @Ceriotti2010ColoredNoiseThermostats]. Finally, the flux [Eq.]{} , which is bilinear in $x$ and $v$, can be extracted from the covariance $C$. Numerical calculations of $\expval{J}$ are performed using Mathematica with custom code [@WolframResearch2018MathematicaVersion] Energy flux from linear response theory ======================================= Following [@Kundu2011LargeDeviations], we derive the expression of the energy flux ([Eq.]{} (7) and (10) in the main text) using a spectral linear response theory. Fourier modes for energy flux ----------------------------- We define Fourier transform (FT) of a function $f(t)$ as $$\begin{aligned} \tilde{f}(\omega) &= \frac{1}{t} \int_0^t dt'\ f(t')e^{-i\omega t'},\quad \omega = \frac{2\pi n}{t} ,\\ f(t) &= \sum_{\omega=-\infty}^{\infty} \tilde{f}(\omega) e^{i\omega t} .\end{aligned}$$ The FT of the equation of motion [Eq.]{} ,- reads $$\begin{aligned} \tilde{v}(\omega) &= i\omega \tilde{z}(\omega) ,\label{eqnS:FT_v}\\ \tilde{z}(\omega) &= G^+(\omega) \tilde{\eta}(\omega) ,\label{eqnS:FT_z}\\ \tilde{\eta}(\omega) &= \frac{\sqrt{2\gamma T_a}}{1 + i\omega \tau} \tilde{\xi}(\omega) ,\label{eqnS:FT_eta}\end{aligned}$$ where $G^+$ is the response function $$\label{eqnS:response} G^{\pm}(\omega) = [K \pm i\omega(\gamma I + BA) - m\omega^2I]^{-1} .$$ The energy flux $J_{ij}$ from [Eq.]{} can be expressed as a bilinear function in $z$ and $v$, by writing the linearized force $F$ in terms of $z$, $$\begin{aligned} J_{ij} &= kv^TA^{J}z \\ \begin{split} A^{J} &\equiv \frac{1}{2} (\ket{i}\bra{i} \otimes e_{ij}e_{ij}^T + \ket{i}\bra{j} \otimes e_{ij}e_{ji}^T \\ &\quad + \ket{j}\bra{i} \otimes (-e_{ji}e_{ij}^T) + \ket{j}\bra{j} \otimes (-e_{ji}e_{ji}^T)) . \end{split}\end{aligned}$$ This bilinear form enables us to write the time integral of energy flux $Q = \int_0^t \dd{t'} J(t')$ as a sum of Fourier modes $\tilde{q}_\omega$ using Parseval’s theorem, $$\begin{aligned} Q &= t\sum_{\omega=-\infty}^{\infty} \tilde{q}_\omega, \\ q_\omega &= k\tilde{v}^T A^{J} \tilde{z}^* = i\omega k \tilde{\eta}^TG^{+T}A^{J}G^-\tilde{\eta}^* ,\end{aligned}$$ where the superscript $$ denotes the complex conjugate. Pairing $\tilde{q}_\omega$ and its conjugate $\tilde{q}_{-\omega}$ gives a real function, which would be beneficial for subsequent derivations. $$\begin{gathered} Q = t\sum_{\omega=2\pi/t}^{\infty} (\tilde{q}_\omega + \tilde{q}_{-\omega}), \label{eqnS:qmode_sum}\\ \tilde{q}_\omega + \tilde{q}_{-\omega} = \tilde{\eta}(\omega)^T A_\omega^q \tilde{\eta}(\omega)^ \\ A_\omega^q = -i\omega k G^{+T}(\omega) A^{as} G^-(\omega), \\ A^{as} = -(A^{J} - {A^{J}}^T) = -\ket{i}\bra{j} \otimes e_{ij}e_{ji}^T + \ket{j}\bra{i} \otimes e_{ji}e_{ij}^T.\end{gathered}$$ Averaging $\tilde{q}_\omega + \tilde{q}_{-\omega}$ over the noise $\tilde{\eta}(\omega)$ using the relationship between $\tilde{\eta}$ and $\tilde{\xi}$ [Eq.]{} , and $\expval{\tilde{\xi}(\omega) \tilde{\xi}^T(\omega')} = \frac{1}{t} I \delta(\omega + \omega')$, we get $$\label{eqnS:qmode_single} \begin{split} \expval{\tilde{q}_\omega + \tilde{q}_{-\omega}} &= \frac{2\gamma T_a}{1 + \omega^2 \tau^2} \tr[A_\omega^q \expval{\tilde{\xi}(-\omega)\tilde{\xi}(\omega)^T}] \\ &= \frac{1}{t} \frac{2\gamma T_a}{1 + \omega^2 \tau^2} \tr A_\omega^q . \end{split}$$ Integrating over the Fourier modes ---------------------------------- In long time limit, the sum can be approximated by an integral $$\frac{1}{t} \sum_{\omega=2\pi/t}^{\infty} = \frac{1}{2t} \sum_{\omega=-\infty}^{\infty} \frac{t \Delta\omega}{2\pi} \approx \frac{1}{4\pi}\int_{-\infty}^{\infty} \dd{\omega} .$$ [Eq.]{} and can then be turned to an integral expression of the flux $$\label{eqnS:flux_integral_raw} \expval{J} = \lim_{t\rightarrow \infty} \frac{\expval{Q}}{t} = \frac{\gamma T_a}{2\pi} \int_{-\infty}^{\infty} \dd{\omega} \frac{\tr A_\omega^q}{1+\omega^2\tau^2} .$$ In the next steps, we simplify this integral with the help of the property [@Kundu2011LargeDeviations] $$\label{eqnS:response_property} G^-(\omega) - G^{+T}(\omega) = 2i\omega\gamma G^-(\omega)G^{+T}(\omega) .$$ Using this property, the trace of $A_\omega^q$ becomes $$\begin{split} \tr A_\omega^q &= -i\omega k \tr G^{+T} A^{as} G^- \\ &= -i\omega k \frac{1}{2i\omega \gamma} \tr (G^- - G^{+T})A^{as} \\ &= -\frac{k}{\gamma} \Re\tr G^+A^{as} . \end{split}$$ Plugging this trace into [Eq.]{} , we get the integral form for the flux [Eq.]{} (7) $$\expval{J} = -\frac{T_ak}{2\pi} \int_{-\infty}^{\infty} \dd{\omega} \frac{\Re \tr G^+A^{as}}{1+\omega^2\tau^2} .$$ This integral can be calculated using the residue theorem. Since $\Im G^+(-\omega) = -\Im G^+(\omega)$, $\frac{\Im \tr G^+A^{as}}{1+\omega^t\tau^2}$ is an odd function of $\omega$, and its line integral vanishes. $$\expval{J} = -\frac{T_ak}{2\pi} \int_{-\infty}^{\infty} \dd{\omega} \frac{\tr G^+A^{as}}{1+\omega^2\tau^2} .$$ The integrand vanishes at $\omega \rightarrow \infty$, so the line integral can be converted to a contour integral along the counter-clockwise semicircle $R$ in the lower-half plane $$\expval{J} = \frac{T_ak}{2\pi} \oint_R \dd{\omega} \frac{\tr G^+A^{as}}{1+\omega^2\tau^2} .$$ The noise correlation $\tau$ introduces a pole of the integrand at $\omega=-i/\tau$, thus the contour integral can be evaluated as $$\label{eqnS:flux_residue} \expval{J} = -\frac{T_ak}{2\tau} \tr G^+(-\frac{i}{\tau})A^{J,as} ,$$ and the response function at $\omega=-i/\tau$ reads $$\label{eqnS:response_at_pole} G^+(-\frac{i}{\tau}) = [K + (\frac{\gamma}{\tau} + \frac{m}{\tau^2})I + \frac{B}{\tau}A]^{-1} .$$ In theory, the equation [Eq.]{} provides the analytical solution of the flux, because the inverse matrix [Eq.]{} can be expressed analytically. In practice, analytical solutions can be easily calculated for small networks, but are hard for large networks. Nevertheless, some general properties of the flux can be obtained from [Eq.]{} after some algebra. For network with only horizontal and vertical bonds ( 1b), all fluxes are zero. For two networks whose slanted bonds have opposite angles ( 1b), their fluxes are opposite. Changing $B$ to $-B$ would change the flux $J$ to $-J$. Kirchoff’s law ============== The derivation of the Kirchoff’s law is similar to the derivation of the energy flux, except that we use the energy from the bath to the particle $h_i$ in [Eq.]{} instead of $J_{ij}$ in [Eq.]{} . Following the procedure in the last section from [Eq.]{} to , we arrive at an integral expression for the $\expval{h_i}$ with a different $A_\omega^q$ $$\begin{gathered} \expval{h_i} = \frac{\gamma T_a}{2\pi} \int_{-\infty}^{\infty} \dd{\omega} \frac{\tr A_\omega^q}{1+\omega^2\tau^2}, \label{eqnS:flux_hi_fourier}\\ A_\omega^q = i\omega (G^{+T} \rho_i - \rho_iG^-) - 2\gamma\omega^2 G^{+T} \rho_i G^- ,\\ \rho_i = \ket{i}\bra{i} .\end{gathered}$$ Using the property of $G^\pm$ [Eq.]{} , we get $$\begin{split} \tr (G^{+T} \rho_i - \rho_iG^-) &= \tr \rho_i(G^{+T} - G^-) \\ &= -\tr \rho_i 2i\omega \gamma G^- G^{+T} \\ &= -2i\omega\gamma \tr G^{+T}\rho_i G^- , \end{split}$$ so the trace of $A_\omega^q$ vanishes $$\tr A_\omega^q = i\omega \tr (G^{+T} \rho_i - \rho_iG^-) - \tr 2\gamma\omega^2 G^{+T} \rho_i G^- = 0 .$$ From [Eq.]{} , $\expval{h_i}$ is also zero, so on average there is no energy exchange between the particle and the bath. Because the average change of $E_i$ is zero, and $\expval{\dot{E_i}} = -\sum_j\expval{J_{ij}} + \expval{h_i}$, we obtain the Kirchoff’s law $$-\sum_j\expval{J_{ij}} = \sum_j\expval{J_{ji}} = 0 .$$ Connection to isolated gyroscopic networks ========================================== ![ Comparison between a boundary-localized eigenmode of the undamped isolated network and the Fourier modes of the damped network at the same frequency. First order dynamics (by setting $m=0$) are used. Numerical calculations are performed with all other parameters set to $1$. (a) Eigenmode of undamped gyroscopic system. For the frequency chosen, the eigenmode is localized on the boundary. Blue disks represent the orbit of particles. (b) The Fourier mode of damped variant of our model at small $\gamma$ ($\gamma=0.001$) resembles the eigenmode. (c) The Fourier mode at larger $\gamma$ ($\gamma=1$) is no longer close to the eigenmode. []{data-label="fig:Fourier_modes"}](Fourier_modes.pdf){width="80.00000%"} Since our model is built upon the well-studied isolated system in Refs [@Nash2015TopologicalMechanics; @Susstrunk2016ClassificationTopological; @Mitchell2018AmorphousTopological; @Lee2018TopologicalDynamics], we would like to build a connection between our energy flux in the active system and eigenmodes in those studies. In this section, we show that the flux formula [Eq.]{} (10) can be decomposed to a weighted sum over eigenmodes [Eq.]{} . Then we apply this result to a honeycomb network as an example. The Fourier analysis from [Sec.]{} IV in the main text is not suitable for this connection, because Fourier modes and eigenmodes are related only at small $\gamma$’s ( \[fig:Fourier\_modes\]a and b), but they become dissimilar at larger $\gamma$’s ( \[fig:Fourier\_modes\]a and c). The underlying discrepancy between Fourier modes and eigenmodes comes from the fact that eigenmodes are a natural basis for the isolated network, whereas Fourier modes have an extra factor of friction or damping. In addition to this extra factor $\gamma$, the active system also has extra factors of $m$ and $\tau$. The factor $m$ comes from the order of dynamics: the active system is second order in time, while the gyroscopic dynamics in [@Nash2015TopologicalMechanics] is first order, which corresponds to the $m\rightarrow 0$ limit. Our starting point is [Eq.]{} (10). The key point is that, in the function $G^+(-i/\tau)$ from the equation, $\gamma,m,\tau$ are not independent factors. Rather, they act collectively through $$k_{g,\tau} \equiv k_g+\frac{\gamma}{\tau}+\frac{m}{\tau^2}.$$ In effect, the extra factors $m,\gamma, \tau$ only add a modification to $k_g$. Following these ideas, we imagine a reference isolated system with modified on-site spring constant $k_{g,\tau}$. Then after some algebra, the flux $\expval{J}$ in active system can be written as a weighted sum of the flux of each eigenmode $J^{\text{eig}}_{\omega_e}$ in the reference system (see the next section for the derivation), $$\label{eqnS:flux_eigen} \expval{J} = \sum_{\omega_e} \frac{1}{1+\omega_e^2\tau^2} J^{\text{eig}}_{\omega_e}.$$ Here $\omega_e$ is the discrete eigen-frequency of the reference system, not to be confused with the continuous Fourier frequency $\omega$. The amplitude of eigenmode is set such that its energy is $T_a$, and $J^{\text{eig}}_{\omega_e}$ is the time-averaged energy flux. A related result is a so called “sum rule", namely, the unweighted sum of all modes is zero, $$\begin{aligned} \sum_{\omega_e} J^{\text{eig}}_{\omega_e} = 0. \label{eqnS:eigen_sum_rule}\end{aligned}$$ This “sum rule" can be derived from direct calculations (see the next section). From this eigenmode decomposition, the discussion of time-reversal symmetry in the isolated system [@Nash2015TopologicalMechanics] immediately carries over to the active system. For network geometries that satisfy time-reversal symmetries, the energy flux of eigenmodes are zero. Thus through [Eq.]{} , the flux in active system is also zero. This result can alternatively be obtained from [Eq.]{} (10) through some linear algebra. ![ Using the eigenmode decomposition, we explain how the flux in honeycomb network is CCW, even though its edgemodes contribute to CW fluxes. (a) Network used for calculation, which consists of one row of hexagons (51 unit cells) and has periodic boundary in $x$ direction. Parameters: $k_{g,\tau}=1, k=10$, others are $1$. (b) Band structure of the network (marked with different colors). The yellow/green band contains CW flux localized on the top/bottom edge (an example mode is shown in (d)/(e)). The blue band contains bulk modes with CCW flux (also see (f)). (c) Weighted flux $J_{\omega_e}^{eig}$ from $1$ to $2$ (marked in (a)) of the four bands. Total flux in the green band with CW edge modes and the blue band with CCW bulk modes are $-0.106$ and $0.115$, respectively. As a result, the net flux is CCW. []{data-label="fig:eigen_modes"}](eigen_modes.pdf){width="80.00000%"} As an application, we will analyze the flux in the honeycomb network using the eigenmode decomposition [Eq.]{} and the “sum rule" [Eq.]{} . The flux pattern in the active honeycomb network displays CCW flux localized on the boundary ( 1b). This localization is reminiscent of the edgemode in [@Nash2015TopologicalMechanics] ( \[fig:Fourier\_modes\]b), however, their directions are opposite. From the decomposition [Eq.]{} , the edgemodes should contribute a large CW flux in the active system, but somewhat surprisingly, the net flux is CCW. To better analyze the contribution from each eigenmode, we look at a simple honeycomb lattice with only one layer ( \[fig:eigen\_modes\]a). This lattice has four bands ( \[fig:eigen\_modes\]b), two bulk bands (blue, red) and two edge bands (green, yellow). The weighted flux of each band is plotted in \[fig:eigen\_modes\]c. We see that the CW edge band does contribute a large CW flux (green curve in \[fig:eigen\_modes\]c), however, due to the “sum rule", the unweighted sum of other bands has to be CCW. In the honeycomb lattice, it happens that many of this CCW fluxes are contained in the lower bulk band (blue curve in \[fig:eigen\_modes\]c and example mode in \[fig:eigen\_modes\]f). When the flux gets weighted, the CCW flux from lower bulk band outweighs CW flux from the edgemodes, the other two bands (yellow and red curve in \[fig:eigen\_modes\]c) also contribute to CCW flux, although relatively small. As a result, the net flux is CCW, which is opposite to the flux of the edgemode. Derivation of eigenmode decomposition of energy flux ==================================================== To derive the eigenmode decomposition [Eq.]{} , we first look at the reference isolated system, write down its eigenmodes [Eq.]{} and time-averaged energy flux [Eq.]{} . Then we turn to the active system and decompose the flux [Eq.]{} using the eigenmodes to get [Eq.]{} . Finally we show that the flux from these two sides are actually related in [Eq.]{} . Lastly we also derive the “sum rule" [Eq.]{} . Reference isolated system ------------------------- The reference isolated system has first-order gyroscopic dynamics as in [@Nash2015TopologicalMechanics]. In the setup with Lorentz force, the dynamical equation can be obtained by setting the mass to zero, and replacing the force matrix $K$ by ${K^\tau \equiv K + (\frac{\gamma}{\tau} + \frac{m}{\tau^2})I}$ $$\dot{z} = \frac{1}{B} A K^\tau z.$$ Following [@Nash2015TopologicalMechanics], we convert to complex representation with $z^c \equiv \pmqty{x + iy & x - iy}^T$ $$i \dot{z}^c = \Omega z^c,\quad K^\tau = i B A O^{-1} \Omega O$$ where $O,O^{-1}$ are the transformations between $z$ and $z^c$ $z^c = Oz, z = O^{-1}z^c$. Writing the eigenvalue problem as $$\Omega u_{\omega_e} = \omega_e u_{\omega_e} ,$$ the eigenmode with eigen-frequency $\omega_e$ reads $$\label{eqnS:mode_zct} z^c_{\omega_e}(t) = (u_{\omega_e} e^{-i\omega_e t} + u_{-\omega_e} e^{i\omega_e t})z_0 ,$$ where $z_0$ is the amplitude, and it will be specified shortly. The eigenmode needs a combination of $\omega_e$ and $-\omega_e$ to ensure that the motion of $x$ and $y$ is real-valued. Mathematically, this combination is possible because of a symmetry in this eigenvalue problem, when there is $\omega_e$, there is also solution $-\omega_e$ with $u_{-\omega_e} = \pmqty{ 0 & I \\ I & 0 } u_{\omega_e}^$. A related property we need later is that, the left eigenvector $v_{\omega_e}$ can be expressed as $v_{\omega_e} = c_{\omega_e} \pmqty{ -I & 0 \\ 0 & I } u_{\omega_e}$, where $c_{\omega_e}$ is a real prefactor to ensure normalization $v_{\omega_e}^T u_{\omega_e} = 1$. If there are degenerate eigenvectors (like $v_{\omega_e}^{1},v_{\omega_e}^{2},\dots$), we choose an orthonormal basis set, i.e. $v_{\omega_e}^{i,T} u_{\omega_e}^{j} = 0$ for $i \neq j$. With the introduction of $c_{\omega_e}$, we now set the amplitude $z_0$ to $z_0^2 = -2 c_{\omega_e} T_a / \omega_e B$, such that the energy of the eigenmode is $T_a$. The instantaneous energy flux $J_{\omega_e}$ of mode $z^c_{\omega_e}$ writes $$\begin{split} J_{\omega_e} &= (O^{-1} v^{c}_{\omega_e})^T A^J O^{-1}z^c_{\omega_e} \\ &= \tr O^{-1,T} A^J O^{-1}z^c_{\omega_e} v^{cT}_{\omega_e} . \end{split}$$ From the expression of mode [Eq.]{} , $$z^c_{\omega_e} v^{cT}_{\omega_e} = -i\omega_e(u_{\omega_e} e^{-i\omega_e t} + u_{-\omega_e}e^{i\omega_e t})(u_{\omega_e}^T e^{-i\omega_e t} - u_{-\omega_e}^Te^{i\omega_e t})z_0^2 .$$ When averaging over time, terms like $e^{\pm 2i\omega_e t}$ vanish, so we get $$\overline{z^c_{\omega_e} v^{cT}_{\omega_e}} = i\omega_e(u_{\omega_e} u_{-\omega_e}^T - u_{-\omega_e} u_{\omega_e}^T)z_0^2 .$$ Plugging in $z_0^2 = -2 c_{\omega_e} T_a / \omega_e B$, the time-averaged flux of the eigenmode $J_{\omega_e}^\text{eig}$ reads $$\label{eqnS:mode_J_isolated} J_{\omega_e}^\text{eig} = -\frac{2T_ak}{B} ic_{\omega_e} \tr O^{-1,T}A^JO^{-1} (u_{\omega_e} u_{-\omega_e}^T - u_{-\omega_e} u_{\omega_e}^T) .$$ Active system ------------- Now we turn to the active system, and the starting point is [Eq.]{} . We need to decompose $G^\tau \equiv G^+(-i/\tau)$ into modes as below, $$\begin{aligned} G^\tau &= \frac{i}{B} O^{-1} (\Omega - \frac{i}{\tau}I)^{-1} OA ,\\ (\Omega - \frac{i}{\tau}I)^{-1} &= \sum_{\omega_e > 0}\frac{i\tau}{1 + \omega_e^2 \tau^2} (u_{\omega_e} v_{\omega_e}^T + u_{-\omega_e} v_{-\omega_e}^T) + \\ &\quad \sum_{\omega_e > 0}\frac{\omega_e\tau^2}{1 + \omega_e^2 \tau^2} (u_{\omega_e} v_{\omega_e}^T - u_{-\omega_e} v_{-\omega_e}^T) ,\\ G^\tau &= \sum_{\omega_e > 0}\frac{-\tau/B}{1 + \omega_e^2 \tau^2} O^{-1} (u_{\omega_e} v_{\omega_e}^T + u_{-\omega_e} v_{-\omega_e}^T) OA + \\ &\quad \sum_{\omega_e > 0}\frac{i\omega_e\tau^2/B}{1 + \omega_e^2 \tau^2} O^{-1} (u_{\omega_e} v_{\omega_e}^T - u_{-\omega_e} v_{-\omega_e}^T) OA .\end{aligned}$$ The averaged flux $\expval{J}$ reads $$\begin{split} \expval{J} &= \sum_{\omega_e > 0}\frac{T_ak/(2B)}{1 + \omega_e^2 \tau^2} \tr O^{-1} (u_{\omega_e} v_{\omega_e}^T + u_{-\omega_e} v_{-\omega_e}^T) OAA^{as} + \\ &\quad \sum_{\omega_e > 0}\frac{-i\omega_e T_ak\tau/(2B)}{1 + \omega_e^2 \tau^2} \tr O^{-1} (u_{\omega_e} v_{\omega_e}^T - u_{-\omega_e} v_{-\omega_e}^T) OAA^{as}. \end{split}$$ The second term can be shown to be zero, $\tr OAA^{as}O^{-1} (u_{\omega_e} v_{\omega_e}^T - u_{-\omega_e} v_{-\omega_e}^T) = 0$. So the mode decomposition in its preliminary form reads $$\begin{gathered} \expval{J} = \sum_{\omega_e} \expval{J}_{\omega_e} ,\\ \expval{J}_{\omega_e} \equiv \frac{T_ak}{2B}\frac{1}{1 + \omega_e^2 \tau^2} \tr OAA^{as}O^{-1} (u_{\omega_e} v_{\omega_e}^T + u_{-\omega_e} v_{-\omega_e}^T). \label{eqnS:mode_J_active}\end{gathered}$$ The relationship between isolated system and active system ---------------------------------------------------------- Now we need to find relationship between these two fluxes $\expval{J}_{\omega_e}$ [Eq.]{} and $J_{\omega_e}^\text{eig}$ [Eq.]{} . We will write $J_{\omega_e}^\text{eig}$ in a form that looks similar to $\langle J\rangle_{\omega_e}$. Converting $A^{J}$ to $A^{as}$ using $A^{as}=-(A_J-A_J^T)$, and $u_{\omega_e}$ to $v_{\omega_e}$ using $v_{\omega_e} = c_{\omega_e} \pmqty{ -I & 0 \\ 0 & I } u_{\omega_e}$, we get $$J_{\omega_e}^\text{eig} = -\frac{iT_ak}{B} \tr AO^{-1,T}A^{as}O^{-1} (u_{\omega_e} v_{\omega_e}^T + u_{-{\omega_e}}v_{-{\omega_e}}^T) .$$ From direct calculation, $AO^{-1,T} = \frac{i}{2} OA$, and $J_{\omega_e}^\text{eig}$ becomes the same as $\expval{J}_{\omega_e}$ apart from a factor $$\label{eqnS:mode_J_isolated_mod} J_{\omega_e}^\text{eig} = \frac{T_ak}{2B} \tr OAA^{as}O^{-1} (u_{\omega_e} v_{\omega_e}^T + u_{-{\omega_e}}v_{-{\omega_e}}^T) .$$ Comparing [Eq.]{} with , we derived the relationship between flux from active system and isolated system $$\label{eqnS:mode_J_relation} \expval{J}_{\omega_e} = \frac{1}{1 + {\omega_e}^2 \tau^2} J_{\omega_e}^\text{eig} .$$ Lastly, we show that the unweighted sum of $J_{\omega_e}^\text{eig}$ is zero. This unweighted sum reads $$\sum_{\omega_e} J_{\omega_e}^\text{eig} = \frac{T_ak}{2B} \tr [O A A^{as} O^{-1} UV^T],$$ where $U$ is the collection of all right eigenvectors $U = \pmqty{u_{\omega_e,1} & u_{\omega_e,2} & \cdots}$, and likewise for $V$. Since $UV^T = I$ from orthonormality, this sum vanishes $$\label{eqnS:mode_J_sum} \sum_{\omega_e} J_{\omega_e}^\text{eig} = \frac{T_ak}{2B} \tr A A^{as} = 0.$$ Path summation of energy flux ============================= To derive the path summation formula [Eq.]{} (11) and the path rules, we start from [Eq.]{} , expand around a noninteracting reference system with $k=0$ to get [Eq.]{} , discuss the convergence radius in [Eq.]{} , then insert resolution of identity to make each term representable by a path as in [Eq.]{} , and arrive at the path summation formula in . We also provide a convenient way to calculate $S_{-l}$ in [Eq.]{} , and a heuristic interpretation of $S_l$ in [Eq.]{} . Derivation of path summation formula ------------------------------------ Similar to the last section, the central object is $G^\tau$. In the noninteracting case ($k=0$), $G^\tau$ is analytically solvable. We denote $G^\tau(k=0) = G^\tau_0$. The inverse $(G^\tau_0)^{-1}$ has a block diagonal form, $$(G^\tau_0)^{-1} = k_{g,\tau} I + \frac{B}{\tau}A = \sum_i \ketbra{i}{i} \otimes (k_{g,\tau}I + \frac{B}{\tau} \pmqty{0 & 1 \\ -1 & 0}).$$ where $k_{g,\tau} \equiv k_g + \frac{\gamma}{\tau} + \frac{m}{\tau^2}$. Then $G^\tau_0$ is also block diagonal, with each block the inverse of the blocks above, $$\label{eqnS:smallk_Gtau0} G^\tau_0 = \sum_i \ketbra{i}{i} \otimes \frac{1}{(k_{g,\tau})^2 + (B/\tau)^2}(k_{g,\tau} I - \frac{B}{\tau} \pmqty{0 & 1 \\ -1 & 0}) = \sum_i \ketbra{i}{i} \otimes \frac{1}{k_0}R_\alpha,$$ where $k_0 \equiv \sqrt{(k_{g,\tau})^2 + (B/\tau)^2}$, and $R_\alpha$ is the rotation matrix with angle $\alpha \equiv \arcsin{\frac{B/\tau}{k_0}}$, $R_\alpha = \pmqty{\cos\alpha & -\sin\alpha \\ \sin\alpha & \cos\alpha}$. We now turn on $k$. We denote the inter-particle part of the force matrix $K$ as $k K_s$, where the factor $k$ is extracted so that the matrix $K_s$ is dimensionless. The blocks of $K_s$ read $$\label{eqnS:smallk_matKs} \mel{i}{K_s}{i} = \sum_{i'} e_{ii'}e_{ii'}^T, \quad \mel{i}{K_s}{j} = e_{ij}e_{ji}^T.$$ Then $G^\tau$ reads $$G^\tau = \frac{1}{(G^\tau_0)^{-1} + k K_s} = \frac{1}{k_0} [(k_0 G^\tau_0)^{-1} + \frac{k}{k_0}K_s]^{-1}$$ In small $k/k_0$ regime, this matrix inversion can be expanded as $$\label{eqnS:smallk_expand} \begin{split} G^\tau &= \frac{1}{k_0} [(k_0 G^\tau_0) + \frac{k}{k_0} (k_0 G^\tau_0) (-K_s) (k_0 G^\tau_0) + (\frac{k}{k_0})^2 (k_0 G^\tau_0) (-K_s) (k_0 G^\tau_0) (-K_s) (k_0 G^\tau_0) + \dots ]\\ &= \frac{1}{k_0} (k_0 G^\tau_0) \sum_{n=0}^{\infty}[\frac{k}{k_0}(-K_s)(k_0 G^\tau_0)]^n. \end{split}$$ To find the convergence radius, we can write the eigen-decomposition of the matrix $(-K_s)(k_0 G^\tau_0)$ as $(-K_s)(k_0 G^\tau_0) = W\Lambda W^{-1}$, where $\Lambda$ is the diagonal matrix that contains all eigenvalues $\lambda_i$’s, then the flux becomes $$\begin{split} \expval{J} &\propto \tr G^\tau A^{as} \propto \sum_{n=0}^{\infty}\tr (k_0 G^\tau_0) [\frac{k}{k_0}W\Lambda W^{-1}]^n A^{as} \\ &= \sum_i [W^{-1} A^{as} (k_0 G^\tau_0) W]_{ii} \sum_n (\frac{k}{k_0} \lambda_i)^n. \end{split}$$ For terms in the series to be convergent, $\frac{k}{k_0}$ should satisfy $$\label{eqnS:smallk_convergence} \frac{k}{k_0} < \frac{1}{\max_i\abs{\lambda_i}}.$$ Before inserting resolution of identity to make paths, we note that the matrix $A^{as}$ and $K_s$ have common blocks, $A^{as} = -\ket{i}\bra{j} \otimes e_{ij}e_{ji}^T + \ket{j}\bra{i} \otimes e_{ji}e_{ij}^T$ and $\mel{i}{K_s}{j} = e_{ij}e_{ji}^T$, so that $A^{as}$ can merge with the series of $G^\tau$. $$\label{eqnS:smallk_J_block} \begin{split} \frac{\expval{J}}{T_a/\tau} = -\frac{k}{2} (\tr G^\tau A^{as}) &= -\frac{k}{2} (\tr \mel{i}{G^\tau}{j} e_{ji}e_{ij}^T - \tr \mel{j}{G^\tau}{i} e_{ij}e_{ji}^T) \\ &= \frac{k}{2} (\tr \mel{i}{G^\tau}{j} \mel{j}{-K_s}{i} - \tr \mel{j}{G^\tau}{i} \mel{i}{-K_s}{j}). \end{split}$$ Now we use the expansion [Eq.]{} , and look at the contribution of its $(n-1)$’th-order term to the first term of the flux, $k \tr \mel{i}{\frac{1}{k_0} (k_0 G^\tau_0) [\frac{k}{k_0}(-K_s)(k_0 G^\tau_0)]^{n-1}}{j} \mel{j}{-K_s}{i}$. If $n-1=0$, this term vanishes, so we only need to consider $n-1 \ge 1$ case. Insert $n-1$ resolution of identity $I = \sum_{l_a=1}^N \ketbra{l_a}{l_a}$, and plug in $k_0 G^\tau_0$ [Eq.]{} and $K_s$ [Eq.]{} , we get $$\begin{split} &\frac{k}{k_0} \tr \mel{i}{(k_0 G^\tau_0) [\frac{k}{k_0}(-K_s)(k_0 G^\tau_0)]^{n-1}}{j} \mel{j}{-K_s}{i} \\ =& (\frac{k}{k_0})^n \sum_{l_1,l_2,\dots,l_{n-1}} \tr \bra{i} (k_0G^\tau_0) \ket{l_{n-1}}\bra{l_{n-1}} (-K_s) (k_0G^\tau_0) \cdots \ket{l_1}\bra{l_1} (-K_s) (k_0G^\tau_0) \ket{j} \mel{j}{-K_s}{i}\\ =& (\frac{k}{k_0})^n \sum_{l_1,l_2,\dots,l_{n-2}} \tr R_\alpha (-K_s)_{il_{n-2}} R_\alpha \cdots (-K_s)_{l_1j} R_\alpha (-K_s)_{ji}, \end{split}$$ where $(-K_s)_{l_bl_a} \equiv \mel{l_b}{-K_s}{l_a}$. We will denote path $l = i\rightarrow j\rightarrow l_1\rightarrow l_2\rightarrow \dots \rightarrow l_{n-2}\rightarrow i$, and its corresponding term in the above summation as $S_l$ $$\label{eqnS:smallk_Sl} S_l = (\frac{k}{k_0})^n \tr R_\alpha (-K_s)_{il_{n-2}} R_\alpha \cdots (-K_s)_{l_1j} R_\alpha (-K_s)_{ji}.$$ The second term of the flux in can be treated similarly, and it results in $S_{-l}$, where $-l$ means path $l$ in its reversed order. Combining [Eq.]{} and , we get the path summation formula of the flux $$\label{eqnS:smallk_path_sum} \frac{\expval{J}}{T_a/\tau} = \sum_l J^\text{path}_l = \sum_l \frac{1}{2}(S_l - S_{-l}).$$ Path rules and discussions -------------------------- The path rules can be extracted from the expression of $S_l$ and $J^\text{path}$. From the element $(-K)_{l_bl_a}$ in $S_l$, we see that either $l_a,l_b$ are bonded, or $l_a=l_b$, otherwise $(-K)_{l_bl_a}=0$. So the path has to be a closed walk along the edges of the network. From $J^\text{path}_l$ for flux from $i$ to $j$, we see that if the path contains equal numbers of $i\rightarrow j$ and $j\rightarrow i$, the net contribution is zero. Because, either $l=-l$, so $J^\text{path}_l \propto S_l - S_{-l} = 0$, or $l'\equiv -l$ is another path, and $J^\text{path}_l + J^\text{path}_{l'} = 0$. To calculate $S_{-l}$, there is a convenient way given that $S_l$ is known. Based on the transformation below, $S_{-l}$ can be obtained by taking the result of $S_l$, then replacing $\alpha$ by $-\alpha$. $$\label{eqnS:smallk_S-l} \begin{split} S_{-l} / (\frac{k}{k_0})^n &= \tr (R_\alpha (-K_s)_{ij} R_\alpha (-K_s)_{jl_1} \cdots R_\alpha (-K_s)_{l_{n-2}i})^T \\ &= \tr (-K_s)_{l_{n-2}i}^T R_\alpha^T \cdots (-K_s)_{jl_1}^T R_\alpha^T (-K_s)_{ij}^T R_\alpha^T \\ &= \tr R_{-\alpha} (-K_s)_{il_{n-2}} R_{-\alpha} \cdots (-K_s)_{l_1j} R_{-\alpha} (-K_s)_{ji}. \end{split}$$ To interpret $S_l$ in a more heuristic way, we insert $I = e_{ij}e_{ij}^T + e_{ij,\perp}e_{ij,\perp}^T$ to the trace in [Eq.]{} , where $e_{ij,\perp}$ denotes the unit direction perpendicular to $e_{ij}$. Because $(-K_s)_{ji}e_{ij,\perp} = 0$, the trace reduces to a matrix product $$\label{eqnS:smallk_path_vector} S_l/(\frac{k}{k_0})^n = e_{ij}^T R_\alpha (-K_s)_{i l_{n-2}} R_\alpha \cdots (-K_s)_{l_1j} R_\alpha (-K_s)_{ji} e_{ij}.$$ This expression means the following operations: starting from a unit displacement of $i$ along $e_{ij}$, $j$ would be displaced according to the force $(-K_s)_{ji} e_{ij}$, after which $j$ is rotated by angle $\alpha$; then start from $j$ and perform similar operations for $(-K_s)_{l_1j}$ and $R_\alpha$; finally, the transmission goes back to $i$; we project the displacement onto $e_{ij}$, and this value is $S_l$ (apart from the prefactor $(\frac{k}{k_0})^n$). Flux of polygon paths --------------------- Here we write down the flux formula for a polygon path without loops. It is easier to work in local coordinates, where each node has its own coordinate system. For node $i$ in the path, let the outer angle from $i$ to $i-1$ be $\pi$, and the angle from $i$ to $i+1$ be $\theta_i$. Then the matrix $(-K_s)_{i+1,i}$ reads $$(-K_s)_{i+1,i} = -e_{i+1,i} e_{i,i+1}^T = -\pmqty{-1 \\ 0} \pmqty{\cos\theta_i & \sin\theta_i} = \pmqty{1 \\ 0} \pmqty{\cos\theta_i & \sin\theta_i}.$$ The trace in $S_l$ becomes $$\begin{split} S_l / (\frac{k}{k_0})^n &= \tr \prod_i (-K_s)_{i+1,i} R_\alpha = \tr \prod_i \pmqty{1 \\ 0} \pmqty{\cos\theta_i & \sin\theta_i} R_\alpha \\ &= \prod_i \pmqty{\cos\theta_i & \sin\theta_i} R_\alpha \pmqty{1 \\ 0} = \prod_i \cos(\theta_i - \alpha) \end{split}$$ So the flux for this path without loops writes $$\label{eqnS:smallk_path_polygon} J^\text{path}_\text{polygon} = \frac{1}{2} (\frac{k}{k_0})^n (\prod_i \cos(\theta_i - \alpha) - \prod_i \cos(\theta_i + \alpha)).$$ Contribution from higher-order paths ------------------------------------ ![ Higher-order paths in a tailed-square network. (a) Direction of flux can be controlled by the orientation of the side-chain. (b) From diagrammatic approach, the flux of the lowest-order path (square) vanishes, and the first non-vanishing path is affected by the side-chain. []{data-label="fig:path_square_tail"}](path_square_tail.pdf){width="50.00000%"} ![ Decay of fluxes away from the boundary of honeycomb networks, and explanation using the diagrammatic technique. (a) Schematic of a honeycomb network which is periodic in the $x$ direction. Layers from the boundary are indexed as $n_l$. (b) Semi-log plot of flux $\expval{J}$ at layer $n_l$. The flux starting from layer $n_l=1$ shows exponential decay, with decay length $d_l$. Parameters used for numerical calculations are $\theta = \pi/6, k/k_0 = 0.01, \alpha = \pi/4$. (c) Decay length $d_l$ changes with the network angle $\theta$ non-monotonically, and the curve has a cusp at $\theta = \alpha = \pi/4$. At small $k/k_0$, perturbation theory results agree with numerical calculations. (d) The first non-vanishing path pair for $n_l=1$ has length $7$. The two paths do not cancel, because the loop in the bulk and at the boundary have different values. []{data-label="fig:path_decay"}](path_decay.pdf){width="60.00000%"} In some situations, the contribution of polygonal paths vanish, and higher order paths with loops become dominant. Unlike in polygonal paths, paths with loops are affected by side-chains. One situation is when the polygon path itself vanishes. In \[fig:path\_square\_tail\]a, the flux of lowest order path, square ( \[fig:path\_square\_tail\]b), is zero, so the main contribution comes from the path with length $5$ ( \[fig:path\_square\_tail\]b). Through the loop in this path, the orientation of the side-chain controls the flux direction in the main square, without changing the geometry of the main cycle (as seen in 1b). Another situation is that two polygon paths cancel each other, which happens in honeycomb-like networks away from the boundary ( \[fig:path\_decay\]a). With careful calculations, the fluxes for $n_l\ge 1$ are not zero, they rather appear as an exponential decay ( \[fig:path\_decay\]b). By changing the geometric angle $\theta$, the decay length varies non-monotonically, and has a cusp at $\theta=\alpha$ ( \[fig:path\_decay\]c). This decay and its relationship with $\theta$ can be explained by considering the paths. While the hexagon path constitutes the lowest-order path at the boundary, it vanishes for $n_l\ge 1$ due to cancellations. The first non-vanishing pair of paths for $n_l=1$ is shown in \[fig:path\_decay\]d, in which the loop exploits the asymmetry between the bulk side (with a vertical bond at the blue loop) and the boundary side (with no vertical bonds at the red loop). For every increment of one layer, the length of paths increases by $4$. So the flux at layer $n_l$ is on the order of $k^{4n_l+3}$, which exhibits an exponential decay $e^{-(n_l-1)/d_l}$. Through the calculation of these paths, we get the decay length $d_l = -1/\log[4(k/k_0)^4(\sin(\theta+\alpha)\sin(\theta-\alpha))^2]$. From this result, we see that the cusp at $\theta=\alpha$ in \[fig:path\_decay\]c is due to the term $\sin(\theta-\alpha)$. In fact, at the special point $\theta=\alpha$, paths like \[fig:path\_decay\]d vanish, and we need to consider even higher-order paths. Simulation of active gyroscopic network coupled with a passive segment ====================================================================== A simulation is shown in the Supplemental Video, which presents both the motion of particles and the energy flux through the color-labelled bonds. The energy fluxes are in general random. During the period when $J$ is large, $J$ shows successive peaks, indicating a large energy flow from left to right. The spacing between the peaks matches the sound speed of the elastic chain ($\sqrt{k/m}$). Although the averaged direction of energy flux is from left to right, the instantaneous flux can also transport from right to left, shown as negative peaks. The simulation is performed using LAMMPS [@Plimpton1995FastParallel] with Moltemplate toolkit [@Jewett2013MoltemplateCoarseGrained] and custom code. We used a Trotter splitting method [@Tuckerman1992ReversibleMultiple; @Bussi2007AccurateSampling] to simulate the underdamped Langevin dynamics. The integrator combines the integrator for colored noise [@Ceriotti2010ColoredNoiseThermostats] and that for Lorentz force [@Chin2008SymplecticEnergyconserving]. We did not simulate the commonly-used overdamped Langevin dynamics, because some intricacy arises when the system also experiences a Lorentz force [@Chun2018EmergenceNonwhite]. Below, we first define each step in the integrator, then present the combined result. The velocity-Verlet step $U_{vv}$ is the integrator when both Lorentz force and the colored noise are absent. It is defined as $$\begin{aligned} U_{vv}(\Delta t):\quad &v \leftarrow v + F(x) \Delta t / (2m) \\ &x \leftarrow x + v \Delta t \\ &v \leftarrow v + F(x) \Delta t / (2m),\end{aligned}$$ where $F(x)$ is the conservative force, including on-site and inter-particle potentials. Writing the Lorentz force part as $$\dot{v} = -\pmqty{ 0 & B/m \\ -B/m & 0 } \pmqty{ v_x \\ v_y } \equiv -a_p v ,$$ then its integrator $U_L$ is a rotation of the velocity $$U_{L}(\Delta t):\quad v \leftarrow e^{-\Delta t a_p} v .$$ Writing the colored noise part as $$\begin{gathered} \frac{d}{dt} \pmqty{ v \\ \eta } = -A_p \pmqty{ v \\ \eta } + B_p \pmqty{ \xi_w \\ \xi_a }, \\ A_p = \pmqty{ \frac{\gamma}{m} & -\frac{1}{m} \\ 0 & \frac{1}{\tau} },\quad B_p = \pmqty{ 0 & 0 \\ 0 & \frac{\sqrt{2\gamma T_a}}{\tau} },\end{gathered}$$ then its integrator $U_{OUP}$ reads $$U_{OUP}(\Delta t):\quad \pmqty{ v \\ \eta } \leftarrow T(\Delta t) \pmqty{ v \\ \eta } + S(\Delta t) \pmqty{ 0 \\ N_a },$$ where $N_a$ is the standard Gaussian random variable, and $$\begin{aligned} T(\Delta t) &= e^{-\Delta t A_p} ,\\ S(\Delta t)S(\Delta t)^T &= C_p - T(\Delta t) C_p T(\Delta t) ^T .\end{aligned}$$ $C_p$ is the solution of $A_p C_p + C_p A_p^T = B_pB_p^T$. $S(\Delta t)$ can be solved as an upper-triangle matrix. Combining these steps together, the integrator for one time step $\Delta t$ reads $$U(\Delta t) = U_{OUP}(\frac{\Delta t}{2})U_{L}(\frac{\Delta t}{2})U_{vv}(\Delta t)U_{L}(\frac{\Delta t}{2})U_{OUP}(\frac{\Delta t}{2}),$$ where the order of operations is right-to-left. Relationship between swimmer’s speed and energy flux ==================================================== To understand the proportionality between $V_s$ and $\expval{J}$, we turn to the diagrammatic technique. Different from previous cases, this path sum can be computed exactly, so the result holds beyond small $k$ regime. First we rewrite $V_s$ as the following $$\frac{V_s}{7a/24L^2} = \expval{J_{12}^s} + \expval{J_{23}^s} + \expval{J_{31}^s}, \label{eqnS:swimmer_Vs}$$ where we have defined $\expval{J_{ij}^s} \equiv \expval{(x_i-x_j)(v_i+v_j)}$. $\expval{J_{ij}^s}$ is proportional to the energy flux via $\expval{J_{ij}} = \frac{k_{ij}}{2}\expval{J_{ij}^s}$, where $k_{12}=k_{23}=k$, and $k_{31}=0$ (because $\expval{J_{31}} = 0$, there is no energy flux from $3$ to $1$). We see that both $\expval{J_{12}^s}$ and $\expval{J_{23}^s}$ are proportional to the flux $\expval{J}$ apart from a factor $k$, so the remaining task is to find the relationship between $\expval{J_{31}^s}$ and $\expval{J}$ or $\expval{J_{12}^s}$. ![ One example of $J_{31}^s$ path (red) and $J_{12}^s$ path (blue). Passive particles are boxed in gray, and the active ones are boxed in red. []{data-label="fig:swimmer_path"}](swimmer_path.pdf){width="40.00000%"} We use a diagrammatic technique with the modification that the paths should contain only one $3\rightarrow 1$ segment. This modification is a consequence of the fact that particle $3$ and $1$ are not bonded. We now illustrate the correspondence between the paths for $\expval{J_{31}^s}$ and for $\expval{J_{12}^s}$. For each path $l$ for $\expval{J_{31}^s}$, we can construct $n$ paths for $\expval{J_{12}^s}$ by reversing $l$ then replacing $1\rightarrow 3$ by $1\rightarrow 2(\rightarrow 2)^n \rightarrow 3$, where $n=0,1,\dots$. An example construction of paths is shown in \[fig:swimmer\_path\]. For $\expval{J_{12}^s}$, all its paths can be constructed in this way. As a result, there is a $1$ to $n$ correspondence between the paths for $\expval{J_{31}^s}$ and for $\expval{J_{12}^s}$, which leads to the relationship $$\expval{J_{12}^s} = \frac{k}{k_0}\sum_{n=0}^\infty (-2\frac{k}{k_0})^n (-\expval{J_{31}^s}) = \frac{k/k_0}{1-(-2k/k_0)} (-\expval{J_{31}^s}), \label{eqnS:swimmer_path_mapping}$$ where $k_0 = k_g + m/\tau^2$ ($B,\gamma=0$ for the passive part), and the factor $-2\frac{k}{k_0}$ comes from the loop $2\rightarrow 2$. Plugging [Eq.]{} to the expression of $V_s$ [Eq.]{} , we obtain the proportionality $$\frac{V_s}{7a/24L^2} = -\frac{k_0}{k} \frac{\expval{J}}{k/2},$$ which is [Eq.]{} (15) in the main text. Since we have considered all the paths, this result can be analytically continued to arbitrarily large $k$. From this diagrammatic technique we also see that, the proportionality constant is independent of the geometry of the active part of the network. This is because the paths through the active part for $\expval{J_{31}^s}$ and for $\expval{J_{12}^s}$ are identical. [14]{}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 in @noop []{} (, ) ed., Chap. , p. [*, ()](\doibase 10.1021/ct900563s) @noop [“,” ]{} () [ (), 10.1088/1742-5468/2011/03/P03007](\doibase 10.1088/1742-5468/2011/03/P03007), [, ()](\doibase 10.1073/pnas.1507413112), [, ()](\doibase 10.1073/pnas.1605462113), [ (), 10.1038/s41567-017-0024-5](\doibase 10.1038/s41567-017-0024-5) [, ()](\doibase 10.1103/PhysRevB.97.085110), [, ()](\doibase 10.1006/JCPH.1995.1039) [, ()](\doibase 10.1016/j.bpj.2012.11.953) [ (), 10.1063/1.463137](\doibase 10.1063/1.463137) [, ()](\doibase 10.1103/PhysRevE.75.056707), [, ()](\doibase 10.1103/PhysRevE.77.066401) [**, ()](\doibase 10.1103/PhysRevE.97.032117),	{ "pile_set_name": "ArXiv" }
--- abstract: \| Scaling relations, such as the IPAT equation and the Kaya identity, are useful for quickly gauging the scale of economic, technological, and demographic changes required to reduce environmental impacts and pressures; in the case of the Kaya identity, the environmental pressure is greenhouse gas emissions. However, when considering large-scale economic transformation, as with a shift to a low-carbon economy, the IPAT and Kaya identities and their cousins fail to capture the legacy of existing capital, on the one hand, and the need for new investment, on the other. While detailed models can capture these factors, they do not allow for rapid exploration of widely different alternatives, which is the appeal of the IPAT and Kaya identities. In this paper we present an extended Kaya identity that includes investment and capital stocks. The identity we propose is a sum of terms, rather than a simple scaling relation. Nevertheless, it allows for quick analysis and rapid exploration of a variety of different possible paths toward a low-carbon economy. Keywords: IPAT, Erlich identity, Kaya identity, capital intensity, low-carbon economy author: - \| Eric Kemp-Benedict\ Stockholm Environment Institute\ [email protected] bibliography: - 'ipat\_with\_capital.bib' title: Confronting the Kaya Identity with Investment and Capital Stocks --- Introduction ============ In 1972 Ehrlich proposed an identity that captured the main drivers of impacts on the environment [@ehrlich_bulletin_1972]. Called the “IPAT equation” or “Ehrlich identity” it proposes that impact, $I$, is equal to the product of population, $P$, affluence, $A$, and technology, $T$. The IPAT equation is a popular method for decomposing environmental impacts, and has spawned several variants [@chertow_ipat_2000; @waggoner_framework_2002; @york_stirpat_2003]. More recently, structural decomposition of changes in input-output matrices provide a detailed basis for the IPAT identity [@baiocchi_understanding_2010]. Within the climate community, the closely-related Kaya identity decomposes carbon dioxide (CO$_2$) emissions into a product of factors: population, GDP per capita, energy intensity per GDP, and emissions intensity per unit energy [@nakienovi_special_2000]. Combining some of these factors, the Kaya identity is equivalent to expressing total CO$_2$ emissions, $E$, as a product of GDP, conventionally denoted $Y$, and emissions intensity $\epsilon$, $$\label{kaya_compressed} E = \epsilon Y.$$ A typical analysis would examine the factors that enter the Ehrlich or Kaya identities in order to recommend policy or individual action (e.g., [@ekins_step_2004; @fan_analyzing_2006; @jackson_prosperity_2009; @raupach_global_2007]). In this paper we argue that the IPAT and Kaya identities, while useful, are deficient when they are used to evaluate the need for or potential of substantial economic transformations. This is because they do not take into account the legacy of existing capital and the investments required for such a transformation. Popular movements in high-income countries in support of zero economic growth [@jackson_prosperity_2009] or de-growth [@martinez-alier_sustainable_2010] (or, perhaps more appropriately, “a-growth”, meaning “indifferent to growth”, as proposed in [@van_den_bergh_environment_2011]), sometimes assume a return to low-technology modes of consumption and production, but many are premised on the development and dissemination of new technologies, and the retirement of older technologies [@chertow_ipat_2000; @iea_energy_2010; @iea_world_2010; @vollebergh_role_2005]. The Kaya identity hides the changes in investment flows and capital stocks required for such a transformation. To illustrate the difficulties, suppose that a high-income country introduces policies designed to reduce emissions by 80 per cent relative to 2010 levels by 2050. As part of the policy package, the country adopts a “Factor 10 by 2050” strategy whereby emissions intensities in 2050 are one-tenth those in 2010.[^1] The Kaya identity suggests that to meet these goals the economy must grow at or below 1.7 per cent per year. While this is a mathematical necessity given the assumptions, it is not clear whether such substantial emissions intensity reductions are consistent with that rate of growth. In this paper we propose an alternative to the Kaya identity that includes consumption, capital stocks, and investment. It must of necessity include a sum of more than one term. Thus, unlike the IPAT and Kaya identities, which are simple scaling relationships, the formulation that we propose involves relative changes in different terms, and is therefore more difficult to use than IPAT-like expressions. We argue that the additional complexity is unavoidable, because, as we demonstrate, it captures an interrelated set of structural changes in the economy. CO$_2$ emissions from economic activities ========================================= We express total CO$_2$ emissions $E(t)$ as a sum of three terms: one for emissions from operating capital, $E_K(t)$, such as industrial machinery, power stations, and some transport infrastructure; one for immediate consumption, $E_C(t)$; and one for investment, $E_I(t)$, \[total\_emissions\] $$\begin{aligned} E(t) &= E_K(t) + E_C(t) + E_I(t)\label{total_emissions_a}\\ &= E_K(t) + \epsilon_C(t) C(t) + \epsilon_I(t) I(t)\label{total_emissions_b}\\ &= Y(t) \left[\bar{\epsilon}_K(t) + \left(1-s(t)\right)\epsilon_C(t) + s(t)\epsilon_I(t) \right]\label{total_emissions_c}.\end{aligned}$$ In passing from Equation (\[total\_emissions\_a\]) to Equation (\[total\_emissions\_b\]) we assume that emissions from consumption and investment activities can be expressed in terms of the volume of current consumption and investment, with time-varying emissions coefficients $\epsilon_C(t)$ and $\epsilon_I(t)$; that is, we treat consumption and investment as flows, with instantaneous emissions associated with those flows. In contrast, as we explain below, emissions from capital are best thought of as a stock that depends on the quality of the capital put in place with prior investment. In passing from Equation (\[total\_emissions\_b\]) to Equation (\[total\_emissions\_c\]), we assume that total economic output $Y(t)$ is either saved or consumed, and that all savings result in investment. The savings rate $s(t)$ is therefore given by $I(t)/Y(t)$ and, as indicated, it may change over time. Carbon dioxide emissions from energy combustion by economic sector are available for most countries from the International Energy Agency (IEA) [@iea_co2_2007], while emissions from cement manufacture are available from the Climate Dioxide Information Analysis Center (CDIAC) [@boden_global_2010]. As a proxy for emissions from operating capital we calculate total energy combustion emissions from the energy sector, from rail and pipeline transport, and from manufacturing (excluding iron and steel and machinery manufacture). For emissions from investment activities we add emissions from cement manufacture to energy combustion emissions from the iron and steel and machinery manufacture sectors. Finally, for consumption emissions we include energy combustion emissions from air and road transport, households, the service sector, and agriculture. With these proxies, we estimate that of total 2005 emissions in the United States, 58.2 per cent was due to the operation of capital, 40.4 per cent due to consumption, and 1.4 per cent due to investment. The trend in emissions over the 30 years from 1975 to 2005 is shown in Figure \[hist\_emiss\]. We note that while emissions due to investment activities are small, if a low-carbon economy requires significant new investment, and other emissions drop strongly, they may not remain small as a share of the total, as it is very difficult to reduce emissions from iron and steel and cement manufacture. ![Historical emissions from fuel combustion and cement manufacture in the US[]{data-label="hist_emiss"}](hist_emiss) Capital Investment ------------------ Before formulating an expression for emissions from operating capital as a stock, we first consider (physical) capital itself, $K(t)$, which is a standard exemplar of a stock, fed by a flow of investments and drained through depreciation, $$\label{capital_balance} \frac{dK(t)}{dt} = - \delta K(t) + I(t).$$ As indicated in the equation, we assume the the depreciation rate $\delta$ is constant in time. A key measure of the structure of an economy is its capital intensity, $k(t)$, which is given by the ratio $K(t)/Y(t)$. The capital intensity, which has units of time, captures the extent to which production relies on physical capital, rather than other inputs, notably labour. We therefore divide Equation (\[capital\_balance\]) by the size of the economy, $Y(t)$. This gives $$\label{capital_balance_div_by_Y} \frac{1}{Y(t)}\frac{dK(t)}{dt} = - \delta k(t) + s(t).$$ This equation can be recast as a dynamic equation for the capital intensity $k(t)$. Consider \[time\_change\_k\_prelim\] $$\begin{aligned} \frac{dk(t)}{dt} &= \frac{d\,}{dt} \frac{K(t)}{Y(t)}\label{time_change_k_prelim_a}\\ &= \frac{1}{Y(t)}\frac{dK(t)}{dt} - k(t) \frac{1}{Y(t)}\frac{dY(t)}{dt}.\label{time_change_k_prelim_b}\end{aligned}$$ Note that $(1/Y(t))dY(t)/dt = d/dt \ln Y(t)$ is the instantaneous growth rate of the economy, which we denote by $r(t)$. Substituting into Equation (\[capital\_balance\_div\_by\_Y\]), we find $$\label{time_change_capital_intens} \frac{dk(t)}{dt} + r(t)k(t) = - \delta k(t) + s(t).$$ This equation can be integrated—in the case of constant $r(t)$ and $s(t)$ it can be integrated exactly—but more interesting is the asymptotic value $k_\infty$ of the capital intensity of the economy at constant growth and savings rates, which can be found by setting the derivative of $k(t)$ in Equation (\[time\_change\_capital\_intens\]) equal to zero and solving for $k$, $$\label{k_infty} k_\infty = \frac{s}{r+\delta}.$$ Because the growth rate and savings rate are assumed constant, we have dropped the time argument in this equation. This equation, a standard result in economic growth theory, is interesting because it relates the long-term capital intensity of the economy—a key indicator of the type of economy—to the savings, growth, and depreciation rates. Emissions from Capital as a Stock --------------------------------- As new capital is put in place through investment, it has the emissions profile corresponding to the technology at the time the investment is made, aside from the changes in emissions that arise from deterioration and upgrading. For simplicity, we ignore deterioration, and consider upgrades to old equipment as new investment. We can therefore treat the emissions from operating capital as something that has been put in place through historical flows of investment and removal—that is, the stock of capital corresponds to a stock of emissions from operating that capital. If new capital has an emissions intensity $\epsilon_K(t)$, then $$\label{capital_emiss_balance} \frac{dE_K(t)}{dt} = - \delta E_K(t) + \epsilon_K(t) I(t).$$ We note that the units of the emissions intensity of new capital $\epsilon_K(t)$ that appears in this equation are the same as those of average emissions from operating capital $\bar{\epsilon}_K(t)$ in Equation (\[total\_emissions\_c\]); that is, tonnes of CO$_2$ per USD. We also note that in this equation we assume that the depreciation rate $\delta$ reflects the rate of removal of old capital from the economy. In fact, economic depreciation also captures the effect of wear and tear. While a more refined approach could distinguish these two phenomena, we equate them for simplicity. From this point the analysis for average emissions from capital, $\bar{\epsilon}_K(t)$, is the same as for capital intensity. Dividing Equation (\[capital\_emiss\_balance\]) by the size of the economy $Y(t)$ gives $$\label{capital_emiss_balance_per_Y} \frac{1}{Y(t)}\frac{dE_K(t)}{dt} = - \delta\bar{\epsilon}_K(t) + \epsilon_K(t) s(t),$$ where $\bar{\epsilon}_K(t)$ is emissions from capital divided by the size of the economy, as in Equation (\[total\_emissions\_c\]). By analogy with Equations (\[time\_change\_k\_prelim\]), this can be shown to be equivalent to $$\label{time_change_epsilon_K} \frac{d\bar{\epsilon}_K(t)}{dt} + r(t) \bar{\epsilon}_K(t) = - \delta\bar{\epsilon}_K(t) + \epsilon_K(t) s(t).$$ This equation can be integrated if the time rate of change of $\epsilon_K(t)$, $r(t)$, and $s(t)$ are known. We assume that the emissions intensity of new capital declines exponentially at a rate $\alpha$; that is, $$\label{exp_decline_alpha} \epsilon_K(t) = \epsilon_{K0}e^{-\alpha t}.$$ We take historical data for the growth rate $r(t)$ and savings rate $s(t)$ when fitting the model, and assume they are constant in time when exploring possible futures. We also assume that $\alpha$ is constant in time; while we do not calculate formal statistics when fitting our model, the historical data we have compiled for the US appears consistent with an assumption of stable parameters. Using standard techniques, the solution to Equation (\[time\_change\_epsilon\_K\]) is $$\label{epsilon_k_full_equation} \bar{\epsilon}_K(t) = e^{-\delta t - \int_0^t dt'\,r(t')}\left[ \bar{\epsilon}_{K0} + \epsilon_{K0} \int\limits_{0}^{t}dt'\,e^{-(\alpha + \delta) t' - \int_0^{t'} dt''\,r(t'')} s(t') \right].$$ When applied to historical data we evaluate the integrals numerically (using Newton’s method). In the case of constant $r$ and $s$, this expression can be evaluated exactly, to give $$\label{epsilon_K} \bar{\epsilon}_K(t) = \left(\bar{\epsilon}_{K0} + s\epsilon_{K0}f(t)\right)e^{-(r+\delta)t},$$ where $$\label{def_of_ft} f(t) = \left\{ \begin{matrix} t &\mbox{ if $r+\delta-\alpha=0$} \\ \frac{1}{r+\delta-\alpha}\left[e^{(r+\delta-\alpha)t}-1\right] &\mbox{ otherwise} \end{matrix} \right. .$$ We fit Equation (\[epsilon\_k\_full\_equation\]) using historical data from the US. As for Figure \[hist\_emiss\], we used data on emissions from combustion from IEA [@iea_co2_2007] and from cement manufacture from CDIAC [@boden_global_2010]. For the savings rate and GDP growth rate we used data from the World Bank [@world_bank_world_2011]. We constructed a time series from 1971, when IEA began reporting disaggregated emissions by manufacturing sector, through 2005. For later convenience when generating future scenarios, we set the reference time $t=0$ to the final year of the time series, 2005, rather than the initial year, 1971. Minimizing the mean squared deviation using Microsoft Excel 2007’s Solver utility, we estimated $\alpha = 0.9$ per cent per year, $\delta = 3.7$ per cent per year (corresponding to a time constant $1/\delta = 27$ years), $\bar{\epsilon}_{K0} = 3059$ MtCO$_2$, and $\epsilon_{K0} = 1029$ MtCO$_2$ per year. The fitted model and historical data are shown in Figure \[model\_vs\_observed\_capital\]. ![Modelled and estimated emissions from operating capital in the US[]{data-label="model_vs_observed_capital"}](model_vs_observed_capital) Total Emissions --------------- Referring to Equation (\[total\_emissions\_c\]), and substituting for $\bar{\epsilon}_K(t)$ from Equation (\[epsilon\_K\]), gives $$\label{total_emissions_ver1} E(t) = Y(t) \left[\left(\bar{\epsilon}_{K0} + s\epsilon_{K0}f(t)\right)e^{-(r+\delta)t} + \left(1-s\right)\epsilon_C(t) + s\epsilon_I(t) \right],$$ where we assume a constant savings rate $s$, as in Equation (\[epsilon\_K\]). Assuming that the emissions intensity of consumption declines exponentially at a rate $\beta$ and emissions from investment at a rate $\gamma$, this can be written $$\label{total_emissions_ver2} E = Y(t) \left[\left(\bar{\epsilon}_{K0} + s\epsilon_{K0}f(t)\right)e^{-(r+\delta)t} + (1-s)\epsilon_{C0}e^{-\beta t} + s\epsilon_{I0}e^{-\gamma t} \right].$$ Fitting historical data for log consumption emissions from 1971 to 2005 against time using ordinary least squares yields an estimated value for $\beta$ of 2.8 per cent per year. The historical coefficient of investment emissions, $\epsilon_{I}(t)$, shows two distinct periods: a rapid decline from 1971 to 1981, and a much slower decline since. Fitting log investment emissions against time using data from the period 1982 to 2005 gives an estimated value for $\gamma$ of 2.0 per cent per year. Alternative Futures =================== We now apply the model developed in this paper to the future of the US economy to 2050 under different assumptions for rates of decline in emissions from consumption and new capital ($\alpha$, $\beta$), the depreciation rate $\delta$, growth rate $r$, and long-term capital intensity $k_\infty$ relative to the present. From Equation (\[k\_infty\]) this allows us to calculate an implied savings rate $s$. We assume that the rate of reduction in emissions intensity from investment activities continues as in the past, reflecting the challenges to reducing emissions in the steel and cement sectors. We present different combinations of parameters descriptively in Table \[experiments\_text\], and quantitatively in Table \[experiments\]. $r$ $\delta$ $k_\infty$ $\alpha$, $\beta$ $\gamma$ ------------------------- ------------ ------------- ------------ ------------------- ------------ -- -- Extrapolation historical historical present historical historical Accelerated reductions - - - Factor 10 by 2050 - Aggressive reductions - - - Factor 10 by 2030 - Accelerated retirement - accelerated - Factor 10 by 2030 - Steady state, high tech zero accelerated 50% higher Factor 10 by 2030 - : Descriptions of parameter assumptions for alternative future CO$_2$ emissions[]{data-label="experiments_text"} ------------------------- -------- ---------- ------------ ---------- --------- ---------- ------------ $r$ $\delta$ $k_\infty$ $\alpha$ $\beta$ $\gamma$ $s$ (%/yr) (%/yr) (%/yr) (%/yr) (%/yr) (%, calc.) Extrapolation 3.1 3.7 1.0 0.9 2.8 2.0 14 Accelerated reductions 3.1 3.7 1.0 5.6 5.6 2.0 14 Aggressive reductions 3.1 3.7 1.0 10.9 10.9 2.0 14 Accelerated retirement 3.1 10.0 1.0 10.9 10.9 2.0 27 Steady state, high tech 0.0 10.0 1.5 10.9 10.9 2.0 31 ------------------------- -------- ---------- ------------ ---------- --------- ---------- ------------ : Parameter sets for alternative future CO$_2$ emissions[]{data-label="experiments"} First, assuming no change in parameter values from 2010 onward (the extrapolation set from Table \[experiments\]) yields continuously growing emissions as shown in Figure \[future\_exrapolated\]. By 2050 in this scenarios emissions for the US are just under 10 Gt CO$_2$ per year. To put this value in context, it is close to the estimated global emissions in 2050 consistent with a low (6 to 32 per cent) probability of exceeding 2$^{\circ}$C [@meinshausen_greenhouse-gas_2009], so it is inconsistent with the goal of avoiding dangerous climate change. ![Extrapolation: historical and future emissions assuming no change in parameters[]{data-label="future_exrapolated"}](future_extrapolated) Next, suppose that rates of emissions intensity reductions from consumption and new capital follow a “Factor 10 by 2050” strategy, as in the parameter set accelerated reductions. As explained above, we assume that emissions improvements for investment are difficult to change, and leave the rate at its historical value. This gives the trajectory shown in Figure \[future\_accel\_emiss\_reduction\]. As seen in the figure, total emissions decline steadily, reaching 71 per cent of their 1990 value in 2050. However, they also decline slowly, suggesting a need for more aggressive action. ![Accelerated reductions: historical and future emissions with a “Factor 10 by 2050” strategy for new capital and consumption[]{data-label="future_accel_emiss_reduction"}](future_accel_emiss_reduction) We consider more rapid reductions in emissions intensity of new capital and consumption in the third parameter set, aggressive reductions. In this case we suppose that both $\epsilon_K(t)$ and $\epsilon_C(t)$ decline follow a “Factor 10 by 2030” strategy in which they drop to 10 per cent of their 2010 value by 2030 and then a similar drop between 2030 and 2050. The resulting trajectory is shown in Figure \[future\_aggressive\_emiss\_reduction\]. In this figure emissions from capital and consumption both decline through the period, bringing total emissions in 2050 to 30 per cent of 1990 emissions. This is a considerable reduction, but given a global cap of 10 GtCO$_2$ per year by 2050, it is still not sufficient. ![Aggressive reductions: historical and future emissions with a “Factor 10 by 2030” strategy for new capital and consumption[]{data-label="future_aggressive_emiss_reduction"}](future_aggressive_emiss_reduction) With the measures taken in the aggressive reductions parameter set the impact of legacy capital becomes clear. Capital emissions do not fall as quickly as consumption emissions because the emissions profile has already been determined by previous investment. Therefore, in the accelerated retirement parameter set we assume that the depreciation rate is increased to 10 per cent per year. Keeping the economic growth rate and long-term capital intensity at their 2010 levels, this implies a substantial—but not unprecedented—increase in the savings rate, from 14 per cent per year to 27 per cent per year. While this is high, rates in rapidly-growing Asian economies have been higher [@world_bank_world_2011]. The results are shown in Figure \[future\_accel\_capital\_retirement\]. As seen in the figure, emissions from capital and consumption both fall more quickly than in the previous cases. Emissions from capital fall more rapidly because older, “dirtier” capital is removed from production; emissions from consumption fall more quickly because people reduce their consumption in favour of saving to replace the retired capital. Emissions from investment activities increase, but total emissions in 2050 are only 14 per cent of 1990 levels. ![Accelerated retirement: historical and future emissions with a “Factor 10 by 2030” strategy for new capital and consumption and faster-than-historical capital depreciation[]{data-label="future_accel_capital_retirement"}](future_accel_capital_retirement) Going beyond the emissions reductions shown in Figure \[future\_accel\_capital\_retirement\] can be accomplished in more than one way. We present one option in the steady state, high tech parameter set. In this case emissions from new capital, consumption, and investment all decline at higher than historical rates. The economy becomes “steady-state” in that the growth rate is zero, but it also becomes more capital intensive, the long-term capital intensity being 50 per cent higher than in the past. The implied savings rate is 31 per cent—high, but still not unprecedented. The results are shown in Figure \[future\_steadystate\_hightech\]. As seen in the figure, emissions drop sharply from 2011 to 2050, reaching 6 per cent of 1990 levels by the end of the period. The trajectory in Figure \[future\_steadystate\_hightech\] is not very different from that in Figure \[future\_accel\_capital\_retirement\], suggesting that strong emissions reductions are not (in principle) incompatible with a growing economy, but as the principle has never been active in the past, there is reason to be sceptical that it will be followed in the future [@jackson_prosperity_2009]. ![Steady state, high tech: historical and future emissions with high rates of emissions intensity reduction, zero growth, and increased capital intensity[]{data-label="future_steadystate_hightech"}](future_steadystate_hightech) Discussion and Conclusion ========================= We have extended the Kaya identity to take into account the effects of investment activities and legacy capital on carbon emissions. Unlike the Kaya identity, the identity in this paper is a sum of three terms, representing emissions from the operation of capital, consumption, and investment. It is therefore more difficult to manipulate than the Kaya identity, and less appropriate for “back-of-the-envelope” calculations. Nevertheless, we have tried to keep the formulation as close to the spirit and form of the Kaya identity as possible, and the resulting formula is, compared to a full economic model, relatively easy to use. Also, it captures an important reality of the climate challenge, that highly industrialized economies must retire capital early and turn it over rapidly in order to capture the benefits of improved capital equipment. We applied our identity to the US economy and generated possible future emissions trajectories. Based on that analysis, we argue that the large emissions reductions required to avoid dangerous climate change will require accelerated turnover of capital and rapid improvements in the emissions profiles of consumption activities and capital equipment. To a lesser degree, it may require efforts to reduce emissions from investment activities. The identity we propose, which is an accounting identity rather than a model, can represent alternatives being discussed today in high-income countries, including no growth and de-growth scenarios, and both low-technology and high-technology routes to reducing emissions. It should also be appropriate in rapidly-growing low-income and middle-income countries that are currently investing in new capital equipment. An accounting identity, whether ours or the Kaya identity, cannot replace a full technical and economic analysis, but can provide a rough estimate of the magnitude of change required to meet a target emissions level. While not conclusive, the reasonable fit shown in Figure \[model\_vs\_observed\_capital\] suggests enough stability in the estimated parameters (for the US economy, at least) that our proposed identity can be useful for broadly delimiting the requirements for a low-carbon economy. [^1]: Reductions in material intensity per unit service (MIPS) to one-tenth of present values are promoted by the Factor 10 Institute (http://www.factor10-institute.org). The concept is explained in [@hinterberger_dematerialization_1999] and put in the context of other approaches to strategic sustainable development in [@robert_strategic_2002]. In this paper we are concerned with carbon emissions per unit economic activity, which are not the same as MIPS. Most important, carbon emissions face no fundamental physical constraints and net emissions can even be negative. We use it here because it is a prominent, and physically achievable, benchmark value for thinking broadly about alternative futures.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Let $\operatorname{G}$ be a complex affine algebraic group and $\operatorname{H}, \operatorname{F}\subset \operatorname{G}$ be closed subgroups. The homogeneous space $\operatorname{G}/ \operatorname{H}$ can be equipped with structure of a smooth quasiprojective variety. The situation is different for double coset varieties $\dcosets{\operatorname{F}}{\operatorname{G}}{\operatorname{H}}$. In this paper we give examples showing that the variety $\dcosets{\operatorname{F}}{\operatorname{G}}{\operatorname{H}}$ does not necessarily exist. We also address the question of existence of $\dcosets{\operatorname{F}}{\operatorname{G}}{\operatorname{H}}$ in the category of constructible spaces and show that under sufficiently general assumptions $\dcosets{\operatorname{F}}{\operatorname{G}}{\operatorname{H}}$ does exist as a constructible space.' address: 'Department of Higher Algebra, Faculty of Mechanics and Mathematics, Lomonosov Moscow State University Leninskie Gory 1, GSP-1, Moscow 119991, Russia' author: - Artem Anisimov title: On existence of double coset varieties --- [^1] [^2] Introduction ============ Let $\operatorname{G}$ be a complex affine algebraic group and $\operatorname{H}\subseteq \operatorname{G}$ be a closed subgroup. By Chevalley Theorem the set of left $\operatorname{H}$-cosets can be equipped with a uniquely defined structure of a smooth quasiprojective variety such that $\operatorname{G}$ act morphically on $\operatorname{G}/ \operatorname{H}$. Moreover, the projection $\operatorname{G}\rightarrow \operatorname{G}/ \operatorname{H}$ is a geometric quotient for the action of $\operatorname{H}$ on $\operatorname{G}$ by right multiplication. The construction of the homogeneous space $\operatorname{G}/ \operatorname{H}$ has a natural generalisation: one can take another subgroup $\operatorname{F}\subset \operatorname{G}$ and consider double cosets, i. e. the sets $\operatorname{F}g \operatorname{H}$, $g \in \operatorname{G}$. These cosets are orbits of the action of $\operatorname{F}\times \operatorname{H}$ on $\operatorname{G}$ given by the formula $(f,h) \circ g = fgh^{-1}$. It is clear that such action, unlike the action of $\operatorname{H}$ on $\operatorname{G}$ by multiplication, can have orbits of different dimensions, thus it does not necessarily admit a geometric quotient. Because of this we consider a weaker quotient, namely, a categorical one. The double coset variety $\dcosets{\operatorname{F}}{\operatorname{G}}{\operatorname{H}}$ is defined to be the underlying space of the categorical quotient $\operatorname{G}\rightarrow \dcosets{\operatorname{F}}{\operatorname{G}}{\operatorname{H}}$ with respect to the described action of $\operatorname{F}\times \operatorname{H}$, if this quotient exists. If the subgroups $\operatorname{F}$ and $\operatorname{H}$ are reductive then this variety exists and coincides with the spectrum $\operatorname{Spec}({}^{\operatorname{F}} \kk[\operatorname{G}]^{\operatorname{H}})$ of the algebra of regular functions on $\operatorname{G}$ invariant under the action of $\operatorname{F}\times \operatorname{H}$. Moreover, if $\operatorname{G}$ is also reductive then by a result of Luna [@LunaClosedOrbits] the action $\operatorname{F}\times \operatorname{H}: \operatorname{G}$ is stable[^3], hence $\dcosets{\operatorname{F}}{\operatorname{G}}{\operatorname{H}}$ parametrises generic (closed) double cosets. In this paper we consider the case when the subgroups $\operatorname{F}$ and $\operatorname{H}$ are not reductive. In this setting one can not guarantee that $\dcosets{\operatorname{F}}{\operatorname{G}}{\operatorname{H}} = \operatorname{Spec}({}^{\operatorname{F}}\kk[\operatorname{G}]^{\operatorname{H}})$; moreover, $\dcosets{\operatorname{F}}{\operatorname{G}}{\operatorname{H}}$ does not necessarily exist. To illustrate this we give the following examples: I. \[example0\] A unipotent group $\operatorname{G}$ and a subgroup $\operatorname{U}$ of $\operatorname{G}$ such that the variety $\dcosets{\operatorname{U}}{\operatorname{G}}{\operatorname{U}}$ does not exist. II. \[example1\] A reductive group $\operatorname{G}$ and two subgroups $\operatorname{F}, \operatorname{H}$ such that the variety $\dcosets{\operatorname{F}}{\operatorname{G}}{\operatorname{H}}$ does not exist. III. \[example2\] A semisimple group $\operatorname{G}$ and two subgroups $\operatorname{F}, \operatorname{H}$ such that the algebra of $\operatorname{F}\times \operatorname{H}$-invariant regular functions $R = {}^{\operatorname{F}} \kk[ \operatorname{G}]^{\operatorname{H}}$ is finitely generated and the natural morphism $\pi : \operatorname{G}\rightarrow \operatorname{Spec}R$ is surjective, but $\pi$ is not a categorical quotient. It is interesting to remark that though $\dcosets{\operatorname{U}}{\operatorname{G}}{\operatorname{U}}$ considered in Example \[example0\] does not exist as an algebraic variety, it does exist as a constructible space. Thus, here we observe the same phenomenon as in [@ArzhCelik], [@CampoNeuen] and [@Celik], namely, an action that admits no quotient in the category of algebraic varieties does admit one in the category of constructible spaces. In Example \[example2\] the categorical quotient $\dcosets{\operatorname{F}}{\operatorname{G}}{\operatorname{H}}$ exists in the category of algebraic varieties; its underlying space is the blow-up of $\operatorname{Spec}({}^{\operatorname{F}} \kk[ \operatorname{G}]^{\operatorname{H}})$ at one point; moreover, in this example the categorical quotient separates generic double cosets. The author would like to thank I. V. Arzhantsev for stating the problem and helpful discussions. Preliminaries on categorical quotients ====================================== Let an algebraic group $\operatorname{G}$ act on an algebraic variety $X$. Recall that the categorical quotient of this action is a $\operatorname{G}$-invariant (i. e., constant on ) morphism such that every $\operatorname{G}$-invariant morphism $\varphi : X \rightarrow Z$ factors uniquely through $\pi_{\operatorname{G}}$, that is, there is a unique morphism $\tilde{\varphi}$ making the following diagram commutative: $$\xymatrix{ X \ar[rr]^{\varphi} \ar[rd]_{\pi_{\operatorname{G}}} && Z. \\ & Y \ar[ru]_{\tilde{\varphi}} }$$ The universal property of $\pi_{\operatorname{G}}$ implies that $Y$ is defined uniquely up to isomorphism. Remark that $\pi_{\operatorname{G}}$ is necessarily surjective. By abuse of language we will sometimes call the variety $Y = X \catquot \operatorname{G}$ the categorical quotient. If $\operatorname{G}$ is reductive and $X$ is affine then the categorical quotient for the action $\operatorname{G}: X$ is $\pi_{\operatorname{G}} : X \rightarrow Y = \operatorname{Spec}\kk[X]^{\operatorname{G}}$ with morphism $\pi_{\operatorname{G}}$ corresponding to inclusion $\kk[X]^{\operatorname{G}} \subset \kk[X]$; in this case $\pi_{\operatorname{G}}$ has an important additional property: it separates closed orbits. If $\operatorname{G}$ is not reductive then the quotient $X \catquot \operatorname{G}$ does not necessarily exist. Examples of actions not admitting a categorical quotient are given in [@VinbergPopov 4.3], [@CampoNeuen], [@ArzhCelik]. Let us point out one example that we will make use of. [@VinbergPopov 4.3]\[simple-example\] There is no categorical quotient for the action of a one-dimensional unipotent group $\operatorname{U}$ on space $\operatorname{Mat}_{2 \times 2}$ of $2 \times 2$-matrices given by the formula $$\lambda \circ \begin{pmatrix}a_{11} & a_{12} \\ a_{21} & a_{22}\end{pmatrix} = \begin{pmatrix}1 & \lambda \\ 0 & 1\end{pmatrix} \begin{pmatrix}a_{11} & a_{12} \\ a_{21} & a_{22}\end{pmatrix}.$$ Remark that we have $\kk [\operatorname{Mat}_{2 \times 2}]^{\operatorname{U}} = \kk [ a_{21}, a_{22}, \det ]$ and the canonical morphism $\pi : \operatorname{Mat}_{2 \times 2} \rightarrow \operatorname{Spec}\kk [\operatorname{Mat}_{2 \times 2}]^{\operatorname{U}} \isom \AA^3$ separates $\operatorname{U}$-orbits of generic points having $a_{21} \neq 0$ or $a_{22} \neq 0$. The image of this morphism is $\AA^3$ without the punctured line $\{ a_{21} = a_{22} = 0, \det \neq 0 \}$. Since the image of $\pi$ is not open, by [@ArzhCelik Corollary 1.4] the action $\operatorname{U}: \operatorname{Mat}_{2 \times 2}$ has no categorical quotient in the category of algebraic varieties. The morphism $\pi$ considered in Example \[simple-example\] can be regarded as a quotient morphism after an appropriate modification to the definition of categorical quotient. It turns out that admitting only morphisms into varieties as categorical quotients is overly restrictive for certain actions $\operatorname{G}: X$. To work around this Bialynicki-Birula introduced in [@BirulaDC] the category of dense constructible subsets. This approach has been further developed in [@ArzhCelik] to permit maps into constructible spaces as candidates for quotient morphisms. Recall that a constructible space is a topological space with a sheaf of functions admitting a finite cover by subsets that are isomorphic (as spaces with functions) to constructibe subsets of affine varieties. A morphism of constructible spaces is a morphism of spaces with functions. We say that a constructible quotient is a categorical quotient in the category of constructible spaces. It is possible for an action $\operatorname{G}: X$ to have no quotient in the category of algebraic varieties, but to have a constructible quotient. \[simple-example-constructible\] Let a unipotent group $\operatorname{G}$ act on a vector space $V$. It follows from [@ArzhCelik Corollary 1.2] that the action $\operatorname{G}: V$ admits a constructible quotient, provided that $\kk [V]^{\operatorname{G}}$ is finitely generated. If $\rho : V \rightarrow \operatorname{Spec}\kk [V]^{\operatorname{G}}$ is the morphism corresponding to the inclusion $\kk[V]^{\operatorname{G}} \subset \kk[V]$ then the constructible quotient is $\rho : V \rightarrow \rho(V)$. In particular, the map $\pi$ in Example \[simple-example\] is a constructible quotient for the action $\operatorname{U}: \operatorname{Mat}_{2 \times 2}$. Let us point out a fact concerning quotients under two commuting actions; it will be used to identify double coset varieties with quotients of homogeneous spaces. Let $\operatorname{F}\times \operatorname{H}$ act on a variety $X$ and let there be a categorical quotient $\pi_{\operatorname{F}} : X \rightarrow Y = X \catquot \operatorname{F}$. The group $\operatorname{H}$ acts on $Y$ as an abstract group: if $y = \pi_{\operatorname{F}}(x)$ then $h \circ y = \pi_{\operatorname{F}}(h \circ x)$. By [@Birula Theorem. 7.1.4] this action is regular. Moreover, existence of $Y \catquot \operatorname{H}$ is equivalent to existence of $X \catquot (\operatorname{F}\times \operatorname{H})$ and these two quotients coincide: $$\xymatrix{ X \ar[rr]^{\pi_{\operatorname{F}\times \operatorname{H}}} \ar[rd]_{\pi_{\operatorname{F}}} && X \catquot (\operatorname{F}\times \operatorname{H}) = Y \catquot \operatorname{H}. \\ & Y = X \catquot \operatorname{F}\ar[ru]_{\pi_{\operatorname{H}}} }$$ The following statement will be used in Proposition \[example1-proof\]. \[uzhosnah\] Let an algebraic group $\operatorname{G}$ act on an algebraic variety $Y$. Suppose that there is $y_0 \in Y$ that belongs to closure of every $\operatorname{G}$-orbit. Consider the action $\operatorname{G}: X \times Y$, where $X$ is a normal variety and $\operatorname{G}$ acts trivially on the first factor. Let $W \subseteq X \times Y$ be a open subset. Suppose that $W$ contains $X_0 \times Y$, where $X_0 \subseteq X$ is a dense subset, and , where $\operatorname{pr}$ is the projection onto the first factor. Then the action $\operatorname{G}: W$ has $\operatorname{pr}: W \rightarrow X$ as a categorical quotient both in the category of algebraic varieties and in the category of constructible spaces. Let us fix a $\operatorname{G}$-invariant morphism $\varphi : W \rightarrow Z$ into an algebraic variety $Z$ (resp., into a constructible space) and show that it factors uniquely through $\operatorname{pr}$. Step 1. We claim that $\varphi$ extends to a continuous map on $W \cup X \times \{y_0\}$. Let us fix a point $(x^\prime, y_0) \not\in X \times \{y_0\}$ and an arbitrary sequence $\{x_n\} \subset X_0$ such that $x_n \rightarrow x^\prime$. Now we show that the sequence $\varphi(x_n, y_0)$ converges. Since $\operatorname{pr}(W) = X$, there is a point $(x^\prime, y)$ in $W$ for some $y \in Y$. The points $(x_n, y_0)$ and $(x_n, y)$ belong to $W$, hence by $\operatorname{G}$-invariance of $\varphi$ we have $\varphi(x_n, y) = \varphi(x_n, y_0)$, thus $\lim\limits_{n \rightarrow +\infty} \varphi(x_n, y_0) = \lim\limits_{n \rightarrow +\infty} \varphi(x_n, y) = \varphi(x^\prime,y)$. Since a converging sequence can have only one limit, $\lim\limits_{n \rightarrow +\infty} \varphi(x_n, y_0)$ does not depend on choice of $(x^\prime, y) \in W$. For the extended map $\varphi$ we have $\varphi(x,y) = \varphi(x,y_0)$, so continuity of $\varphi\|_{X \times \{y_0\}}$ implies continuity of $\varphi$ on $W \cup X \times \{y_0\}$. Step 2. Now we show that $X \times \{y_0\}$ can be covered by open affine sets $X_i \times \{y_0\}$ such that the image of $\varphi : X_i \times \{y_0\} \rightarrow Z$ is contained in some affine subset of $Z$. Let $\{Z_i\}$ be an affine covering of $Z$ and $\{U_i\}$ be an affine covering of $X$. The set $V_{ij} = \varphi^{-1}(\varphi(U_i) \cap Z_j)$ is open in $U_i$. Every set $V_{ij}$ is a union of principal open subsets $V_{ij} = \cup_k V_{ijk}$. The sets $V_{ijk}$ make up the required covering of $X \times \{y_0\}$. Step 3. Since $\varphi(x,y) = \varphi(x, y_0)$, we have $\varphi = \tilde{\varphi} \circ \rho$, where $\tilde{\varphi} = \varphi\|_{X \times \{y_0\}}$ and $\rho$ is the map $W \rightarrow X \times \{y_0\}$, $\rho(x,y) = (x,y_0)$. Denote $\imath$ the identification of $X$ and $X \times \{y_0\}$: $\imath(x) = (x,y_0)$. We have $\varphi = \tilde{\varphi} \circ \imath \circ \pi$, so $\varphi$ factors through $\pi$. It remains to verify that $\varphi\|_{X \times \{y_0\}}$ is a morphism. The variety $X$ is normal, hence the affine opens $X_i \times \{y_0\}$ constructed at step 2 are normal varieties, too. Restrictions of $\varphi$ to these opens are morphisms of affine varieties; if $Z$ is a constructible space then $\varphi\|_{X_i \times \{y_0\}}$ is a morphism into a constructible set, but it can be regarded as a morphism into an affine variety containing $\varphi(X_i \times \{y_0\})$. By theorem on removable singularities the continuous extensions of $\varphi\|_{X_i \times \{y_0\}}$ are morphisms. Existence and non-existence of double coset varieties {#main-section} ===================================================== \[main-section\].1. Consider a unipotent group $\operatorname{G}$ and a subgroup $\operatorname{U}$: $$\begin{matrix} \operatorname{G}= \begin{pmatrix} 1 & * & * & * \\ & 1 & * & * \\ & & 1 & 0 \\ & & & 1 \end{pmatrix}, & \operatorname{U}= \begin{pmatrix} 1 & * & 0 & 0 \\ & 1 & 0 & 0 \\ & & 1 & 0 \\ & & & 1 \end{pmatrix}, \end{matrix}$$ here $$ denotes an arbitrary number. We claim that if we take $\operatorname{F}= \operatorname{H}= \operatorname{U}$ then the double coset variety $\dcosets{\operatorname{F}}{\operatorname{G}}{\operatorname{H}}$ does not exist. Remark that the group $\operatorname{F}\times \operatorname{H}= \operatorname{U}\times \operatorname{U}$ is unipotent, hence every double $(\operatorname{U}, \operatorname{U})$-coset is closed [@VinbergPopov 1.3]; had $\operatorname{F}$ and $\operatorname{H}$ been reductive, this would have implied existence of the geometric* quotient $\operatorname{G}\rightarrow \operatorname{G}/ (\operatorname{F}\times \operatorname{H})$. The action $\operatorname{U}\times \operatorname{U}: \operatorname{G}$ has no categorical quotient in the category of algebraic varieties. It admits a constructible quotient and the constructible quotient parametrises generic double cosets. Consider the action of $\operatorname{G}$ on space of $4 \times 2$-matrices by left multiplication. The subgroup $\operatorname{U}$ is the stabiliser of the matrix $$\label{definition-of-M} M = \begin{pmatrix} 0 & 0 \\ 0 & 0 \\ 1 & 0 \\ 0 & 1 \end{pmatrix}.$$ Therefore, the homogeneous space $\operatorname{G}/ \operatorname{U}$ is isomorphic to $\AA^4$ and can be identified with the variety of matrices $$\begin{pmatrix} * & * \\ * & * \\ 1 & 0 \\ 0 & 1 \end{pmatrix}.$$ After this identification the action of $\operatorname{U}$ on $\operatorname{G}/ \operatorname{U}$ becomes the matrix multiplication; it is therefore isomorphic to the action of $\operatorname{U}$ on space $\operatorname{Mat}_{2 \times 2}$ of $2 \times 2$-matrices by left multiplication. Example \[simple-example\] shows that this action does not admit a categorical quotient. Therefore, $\dcosets{\operatorname{U}}{\operatorname{G}}{\operatorname{U}} = (\operatorname{G}/ \operatorname{U}) \catquot \operatorname{U}$ does not exist. From Example \[simple-example-constructible\] it follows that the action $\operatorname{U}\times \operatorname{U}: \operatorname{G}$ has a constructible quotient which separates generic double cosets. The constructible quotient $\pi : \operatorname{G}\rightarrow \dcosets{\operatorname{U}}{\operatorname{G}}{\operatorname{U}}$ does not separate all closed double cosets. Indeed, all $2 \times 2$-matrices with $a_{21} = a_{22} = 0$ (we use notation of Example \[simple-example\]) are fixed under the action of $\operatorname{U}$ and have $\det = 0$, hence their preimages in $\operatorname{G}$ are closed $(\operatorname{U}, \operatorname{U})$-cosets, which are mapped by $\pi$ to $0 \in \AA^3$. \[main-section\].2. Take $\operatorname{G}= \operatorname{GL}_4$ and consider the action of $\operatorname{G}$ on $4 \times 2$-matrices. Let $\operatorname{H}$ be the stabiliser of the matrix $M$, $M$ being the same as in \[definition-of-M\]. The homogeneous space $W = \operatorname{G}/ \operatorname{H}$ is identified with the variety of $4 \times 2$-matrices with non-zero columns. Let $\operatorname{F}$ be the subgroup of $\operatorname{G}$ consisting of the following matrices: $$\begin{pmatrix} 1 & a & 0 & 0 \\ & 1 & 0 & 0 \\ & & s & 0 \\ & & & s \end{pmatrix},~a \in \kk,~s \in \kk^\times.$$ The subgroup $\operatorname{F}$ acts on $W$ via matrix multiplication. \[example1-proof\] The action $\operatorname{F}\times \operatorname{H}: \operatorname{G}$ does not admit a categorical quotient in the category of algebraic varieties, but has a constructible quotient. The group $\operatorname{F}$ is a direct product $\operatorname{F}= \operatorname{U}\times \operatorname{S}$ of one-dimensional unipotent group $\operatorname{U}$ and one-dimensional torus $\operatorname{S}$. The categorical quotient for the action $\operatorname{S}: W$ is $\operatorname{pr}: W \rightarrow \operatorname{Mat}_{2 \times 2}$, where the projection $\operatorname{pr}$ erases the bottom half of matrices of $W$. Indeed, one can apply Lemma \[uzhosnah\] with the acting group $\operatorname{S}$ and $X = Y = \operatorname{Mat}_{2 \times 2}$ representing top and bottom halves of matrices respectively, and with $X_0$ consisting of matrices with non-zero columns. Thus, had the quotient $W \catquot \operatorname{F}= \dcosets{\operatorname{F}}{\operatorname{GL}_4}{\operatorname{H}}$ existed, it would have been also $(W \catquot \operatorname{S}) \catquot \operatorname{U}= \operatorname{Mat}_{2 \times 2} \catquot \operatorname{U}$, but, according to Example \[simple-example\], the latter quotient does not exist. By Lemma \[uzhosnah\] and Example \[simple-example-constructible\], the actions $\operatorname{S}: W$ and $\operatorname{U}: \operatorname{Mat}_{2 \times 2}$ both have a constructible quotient, thus $\dcosets{\operatorname{F}}{\operatorname{GL}_4}{\operatorname{H}} = (W \catquot \operatorname{S}) \catquot \operatorname{U}$ exists as a constructible space. \[main-section\].3. This example is based on [@ArzhCelik 4.5]. Consider the following symmetric bilinear form on $\kk^4$: $(e_1, e_4) = (e_2, e_3) = 1$ and the other pairings of basis vectors are zero. The cone $$X = \{ {\bf x} \in \kk^4\ \|\ x_1x_4 + x_2x_3 = 0 \} \setminus \{(0,0,0,0)\}$$ is the collection of non-zero isotropic vectors, therefore $X = \operatorname{SO}_4 / \operatorname{H}$, where $\operatorname{H}$ is the stabiliser of a non-zero isotropic vector. As $\operatorname{F}$ we take the following unipotent subgroup of $\operatorname{SO}_4$: $$\begin{pmatrix} 1 & a & 0 & 0 \\ & 1 & 0 & 0 \\ & & 1 & -a \\ & & & 1 \end{pmatrix},~a \in \kk.$$ The algebra $\kk [X]^{\operatorname{F}}$ is freely generated by $x_2$ and $x_4$; indeed, these two functions are $\operatorname{F}$-invariant and generic orbits meet the plane $\{x_1 = x_3 = 0\}$, so there are no other generators. It is clear that the canonical morphism $\pi : X \rightarrow \operatorname{Spec}\kk [X]^{\operatorname{F}} = \AA^2$ is surjective. Nevertheless, $\operatorname{Spec}\kk [X]^{\operatorname{F}}$ is not the quotient for $\operatorname{F}: X$ because the following morphism $\varphi : X \rightarrow \PP^1$ does not factor through $\pi$: $$\varphi(x_1, x_2, x_3, x_4) = (x_2 : x_4) = (x_1 : -x_3).$$ Indeed, $\pi(x_1, x_2, x_3, x_4) = (x_2, x_4)$, and from $\varphi = \tilde{\varphi} \circ \pi$ it follows that $\tilde{\varphi}(x_2,x_4) = (x_2 : x_4)$ when $x_2 \neq 0$ or $x_4 \neq 0$, hence $\tilde{\varphi}$ is not continuous in $(0,0)$, which is not possible. Let us show that the considered action has a categorical quotient, though it does not coincide with $\operatorname{Spec}\kk[X]^{\operatorname{F}}$. The action $\operatorname{F}: X$ has a categorical quotient in the category of algebraic varieties. The quotient is the blow-up of the origin in $\AA^2$ $$\hat{\AA}^2 = \{ ((x, y), (u:v)) \in \AA^2 \times \PP^1\ \|\ xv - yu = 0 \}$$ with the morphism $\rho : X \rightarrow \hat{\AA}^2$, . Let us check that every $\operatorname{F}$-invariant morphism $\varphi : X \rightarrow Z$ factors through $\rho$. By $\operatorname{F}$-invarience of $\varphi$ we have $\varphi(cx_1, x_2, cx_3, x_4) = \varphi(x_1, x_2, x_3, x_4)$. Indeed, if one of $x_2, x_4$ is not zero then the point $(x_1, x_2, x_3, x_4)$ and the points $(cx_1, x_2, cx_3, x_4)$ belong to one orbit of $\operatorname{F}$; if $x_2 = x_4 = 0$ then we have $$\begin{aligned} \varphi(x_1, 0, x_3, 0) = \lim\limits_{t \rightarrow 0} \varphi(x_1, tx_1, &x_3, -tx_3) = \\ & = \lim\limits_{t \rightarrow 0} \varphi(cx_1, tx_1, cx_3, -tx_3) = \varphi(cx_1, 0, cx_3, 0).\end{aligned}$$ Define $\tilde{\varphi} : \hat{\AA}^2 \rightarrow Z$ as the morphism taking $((x, y), (u:v))$ to $\varphi(u, x, -v, y)$. The reasoning above shows that $\tilde{\varphi}$ is well defined. Thus, $\varphi = \tilde{\varphi} \circ \rho$, i. e., $\varphi$ factors through $\rho$. Since $\rho(X) = \hat{\AA}^2$, the morphism $\tilde{\varphi}$ can be chosen uniquely. It is clear that $\rho : X \rightarrow \hat{\AA}^2$ separates orbits of points having $x_2 \neq 0$ or . However, $\rho$ does not separate all closed orbits: the points $z = (x_1, 0, x_3, 0)$ and $z^\prime = (cx_1, 0, cx_3, 0)$ are $\operatorname{F}$-fixed, but $\rho(z) = \rho(z^\prime)$. Thus, the quotient $q : \operatorname{SO}_4 \rightarrow \dcosets{\operatorname{F}}{\operatorname{SO}_4}{\operatorname{H}} = \hat{\AA}^2$ separates generic double cosets, but fails to separate all closed double cosets. \[main-section\].4. Remark that in Examples \[example0\] and \[example1\] the actions $\operatorname{F}\times \operatorname{H}: \operatorname{G}$ have no categorical quotient in the category of algebraic varieties but do admit one in the category of constructible spaces. Let $\operatorname{G}$ be a connected affine algebraic group and $\operatorname{F}, \operatorname{H}$ be closed subgroups in $\operatorname{G}$. Is it true that $\dcosets{\operatorname{F}}{\operatorname{G}}{\operatorname{H}}$ exists as a constructible space? The following proposition gives a partial answer to this question. Let $\operatorname{G}$ be a connected affine algebraic group and $\operatorname{F}, \operatorname{H}\subset \operatorname{G}$ be closed connected subgroups with trivial character groups. Suppose that the algebra ${}^{\operatorname{F}} \kk[\operatorname{G}]^{\operatorname{H}}$ is finitely generated and let $\pi : \operatorname{G}\rightarrow \operatorname{Spec}({}^{\operatorname{F}} \kk[\operatorname{G}]^{\operatorname{H}})$ be the canonical morphism. Then $\dcosets{\operatorname{F}}{\operatorname{G}}{\operatorname{H}}$ exists as a constructible space and the map is the constructible quotient for the action of $\operatorname{F}\times \operatorname{H}$ on $\operatorname{G}$. By [@PopovPicardGroups Theorem 6], the underlying variety of $\operatorname{G}$ has a finite divisor class group. Additionally, $\operatorname{F}$ and $\operatorname{H}$ have trivial character groups, therefore every $\operatorname{F}\times \operatorname{H}$-invariant hypersurface $D \subset \operatorname{G}$ is the zero set of an invariant function $f_D \in {}^{\operatorname{F}} \kk[\operatorname{G}]^{\operatorname{H}}$. It follows from [@ArzhCelik Corollary 1.2] that the action $\operatorname{F}\times \operatorname{H}: \operatorname{G}$ has $\pi : \operatorname{G}\rightarrow \pi(\operatorname{G})$ as a constructible quotient. One can often give a positive answer to the question on finite generation of ${}^{\operatorname{F}} \kk[\operatorname{G}]^{\operatorname{H}}$. Recall that if $\operatorname{R}$ is a reductive group, $Z$ is an affine $\operatorname{R}$-variety and $\operatorname{U}\subset \operatorname{R}$ is a maximal unipotent subgroup then the algebra $\kk[Z]^{\operatorname{U}}$ is finitely generated [@Kraft Chapter 3.2]. Thus, the constructible space is guaranteed to exist if both groups $\operatorname{F}$ and $\operatorname{H}$ are maximal unipotent subgroups in bigger reductive subgroups $\operatorname{F}^\prime, \operatorname{H}^\prime \subseteq \operatorname{G}$ or if one of them is semisimple and the other one is a maximal unipotent subgroup in a bigger reductive subgroup. Other results on finite generation of algebras of invariants can be found in [@Grosshans]. [HD]{} A. A’Campo-Neuen, J. Hausen: Examples and Counterexamples of Existence of Categorical Quotients. J. Algebra 231 (1999), 67–85. I. V. Arzhantsev, D. Celik, J. Hausen: Factorial algebraic group actions and categorical quotients. arXiv:0908.0443v2 \[math.AG\], 11 pages. A. Bialynicki-Birula: Algebraic quotients. Invariant Theory and Algebraic Transformation Groups, II, in: Encyclopaedia Math. Sci., vol. 131, Springer-Verlag, Berlin, 2002. A. Bialynicki-Birula: Categorical quotients. J. Algebra 239 (2001), 35–55. D. Celik: A categorical quotient in the category of dense constructivle subsets. Coll. Math. 116 (2009), no. 2, 147–151. F. Grosshans: Algebraic Homogeneous Spaces and Invariant Theory Lecture Notes in Mathematics 1673, Springer-Verlag, Berlin, Heidelberg, 1997. J. Humphreys: Linear algebraic groups. Graduate Texts in Mathematics 21, Springer-Verlag, New-York, 1998. H. Kraft: Geometrische methoden in der Invariantentheorie. Aspects of Math., Vieweg, 1984. D. Luna: Sur les orbites fermées des groupes algébriques réductifs. Invent. Math. 16 (1972), 1–5. V.L. Popov: Picard groups of homogeneous spaces of linear algebraic groups and one-dimensional homogeneous vector bundles. Izv. Akad. Nauk SSSR Ser. Mat. 8 (1974), no. 2, 301–327. V. L. Popov, E. B. Vinberg: Invariant theory. Algebraic geometry IV, Encyclopaedia Math. Sci., vol. 55, pp. 123–284, Springer-Verlag, Berlin, 1994. [^1]: 2010 Mathematics Subject Classification: Primary 14L30; Secondary 14M17, 14R20. [^2]: Key words and phrases: Algebraic group, double coset variety, categorical quotient. [^3]: Reductivity of $\operatorname{G}$ is essential: consider the group $\operatorname{B}$ of the upper-triangular matrices and its subgroup $\operatorname{T}$ of the diagonal matrices; the action $\operatorname{T}\times \operatorname{T}: \operatorname{B}$ is not stable.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The Pierre Auger Observatory is the currently largest experiment dedicated to unveil the nature and origin of the highest energetic cosmic rays. The software framework has been developed by the Pierre Auger Collaboration for joint analysis of data from different detector systems used in the observatory. While reconstruction modules are specific to the Pierre Auger Observatory components of the framework are also used by other experiments. The software framework has recently been extended to analyze also data from the Auger Engineering Radio Array (AERA), the radio extension of the Pierre Auger Observatory. The reconstruction of the data from such radio detectors requires the repeated evaluation of complex antenna gain patterns which significantly increases the required computing resources in the joint analysis. In this contribution we explore the usability of massive parallelization of parts of the code on the GPU. We present the result of a systematic profiling of the joint analysis of the software framework aiming for the identification of code areas suitable for parallelization on GPUs. Possible strategies and obstacles for the usage of GPGPU in an existing experiment framework are discussed.' author: - \| [ Tobias Winchen$^1$, Marvin Gottowik$^1$, and Julian Rautenberg$^1$ for the Pierre Auger Collaboration$^{2,3}$]{}\ $^1$ Bergische Universit[ä]{}t Wuppertal, Gau[ß]{}str. 20, 42119 Wuppertal, Germany\ $^2$ Pierre Auger Observatory, Av. San Martín Norte 304, 5613 Malarg[ü]{}e, Argentina.\ $^3$ Full author list: <http://www.auger.org/archive/authors_2014_09.html> bibliography: - 'winchen\_tobias.bib' title: Prospects of GPGPU in the Auger Offline Software Framework --- Introduction ============ Although cosmic rays have been intensively studied for more than 100 years, fundamental questions about the phenomenon remain unclear. In particular, the origin, the acceleration mechanism, and the chemical composition of the highest energetic cosmic rays with energies up to several hundred EeV (1 EeV = $10^{18}$ eV) remain open questions. As the flux of the highest energy cosmic rays is of the order of one particle per square kilometer per century, direct observation is impracticable. Instead, the Earth’s atmosphere is used as a calorimeter in which the cosmic rays are detected indirectly by the particle cascades, or ‘air showers’, they induce. For detailed reviews on cosmic rays see e.g. references [@Kotera2011; @Letessier-Selvon2011]. The currently largest detector for cosmic rays at the highest energies is the Pierre Auger Observatory in Argentina. The Pierre Auger Observatory [@PAO2004] is designed as a hybrid of two complementary detector systems. The ‘surface detector’ [@Allekotte2008] consists of 1660 water-Cherenkov stations that sample the secondary particles at the ground level. The stations are arranged in a hexagonal grid that covers an area of 3000 km$^2$. The surface detector array is surrounded by 27 telescopes stationed at four sites to detect the fluorescence light emitted by air molecules that have been excited by the particle cascade. This ‘fluorescence detector’ [@Abraham2010] provides a direct calorimetric measurement independent of hadronic interaction models and thus a better energy resolution than the surface detector. Furthermore, as it measures the development of the shower in the atmosphere, it is sensitive to the mass of the primary cosmic ray. However, the fluorescence detector can operate only during clear and moonless nights, whereas the surface detector has no principal constraint on the uptime. As complementary detector without principal time constraints, the Auger Engineering Radio Array (AERA) [@Neuser2014] detects the radio pulses emitted by the particle cascades, due to geomagnetic and charge-excess effects [@Aab2014a]. Currently, AERA consists of 124 stations in a hexagonal grid covering an area of about 6 km$^2$. Each station is equipped with an antenna designed for polarized measurements from 30 to 80 MHz. The Auger Software Framework [@Argiro2007] provides the tools and infrastructure for the reconstruction of events detected with the Pierre Auger Observatory. Components of the framework are also used by other experiments [@Sipos2012]. It is designed to support and incorporate the ideas of physicists for the projected lifetime of the experiment of more than 20 years. This is achieved by a strict separation of configuration, reconstruction modules, event and detector data, and utilities to be used in the algorithms. The reconstruction of radio events detected with AERA is fully integrated in the Auger Software Framework which allows joint analyses of events from all detector systems [@PAO2011g]. The main reconstruction algorithm for the radio reconstruction and its performance profile is described in the next section. Radio Reconstruction in the Auger Offline Framework =================================================== ![Schematic of the individual steps in the radio reconstruction.[]{data-label="fig:OfflineRadioReconstruction"}](winchen_tobias_fig1.pdf){width="\textwidth"} The main algorithm for the reconstruction of radio events implemented in several modules is depicted in Fig. \[fig:OfflineRadioReconstruction\]. After a trigger, the voltage traces of the two channels at each station are read out, and, after noise filtering and signal enhancing, processed as follows. First, as an initial estimate of the event timing in each station, the maxima of the voltage traces is used. Second, from the timing information of the individual stations, the direction of the shower is obtained by triangulation. Third, the antenna pattern for this direction is evaluated, and the E-field trace at each station reconstructed from the voltage traces. Finally, the maximum of the envelope of the E-field trace is used as updated timing information for a new iteration of the procedure. On convergence, the timing information yields the incident direction of the primary particle whereas the E-field distribution on the ground allows a derivation of the energy and particle type of the cosmic ray.\ The execution of this algorithm in requires an uncomfortable amount of time. Using the Linux kernel profiler ‘perf’ [@linux_perf] we identified two main performance bottlenecks in this algorithm. First, about 15% of the computation time is spent in calculating Fourier transformations with the FFTW library [@fftw]. Second, about 25% of the time is used for the interpolation of the antenna patterns. All other parts of the algorithm use less than 5% of the time. The same bottlenecks are identified using ‘google-perftools’ [@google_perf] or ‘Intel VTune amplifier’ [@intel_vtune] as alternative profilers. In the next sections we discuss the elimination of these bottlenecks by implementing the relevant parts on the GPU using the Cuda framework. In both cases we followed a minimum invasive approach that leaves the general interfaces in intact. To select between the CPU and GPU implementation, a preprocessor directive is used. Interpolations of Antenna Patterns ================================== To reconstruct the electric field vector from the measured voltage traces the antenna pattern in the direction of the electromagnetic wave must be known. The antenna pattern can be conveniently expressed as a two dimensional complex vector, the ‘vector effective length (VEL)’. For each antenna, the VEL is known for discrete frequencies and zenith and azimuth angles from measurements or simulations. Between theses nodes the VEL has to be interpolated for arbitrary directions. The interpolation of textures is a core task of graphics cards, which have dedicated circuits for the interpolation. The usage of these promises a great speedup, but is available for single precision data of limited size only. In the baseline implementation, the antenna pattern is evaluated in double precision. As a linear interpolation requires six elementary operations, the maximum relative uncertainty from the limited floating-point precision can be estimated as $2.5{\times}10^{-4}$% in single precision. This is smaller than other uncertainties in the reconstruction and thus negligible here. The largest antenna pattern considered here consists of complex data for the vector effective length at 98 frequencies, 31 zenith angles, and 49 azimuth angles. In single precision this corresponds to approximately 2.4 MB of data which is small compared to the total available memory size of a few GB on modern GPUs and much below the maximal size of a 3D texture of typically at least 2048$\times$2048$\times$2048 elements. The patterns for all antennas used in AERA can be stored simultaneously on the device. To speed up the interpolation on the CPU, the pattern has already been buffered in the baseline implementation for look up on repeated access. In the GPU implementation this is unnecessary. Here, the patterns are copied and converted to allow binding to texture memory only once on first access. Fourier Transformations and Hilbert Envelopes ============================================= In the baseline implementation in , calls to the FFTW library are wrapped in an object-oriented interface. The interface provides several distinct classes for the operation on real or complex data of given dimensionality. Shared functionality is implemented in a base class and propagated by inheritance. In the radio reconstruction, the wrapper operates within a container that stores a trace simultaneously in the time and frequency domains. After modification of either, the other is updated in a lazy evaluation scheme. To calculate the FFT on the GPU, FFTW calls are replaced by calls to the CUDA FFT library (CuFFT). In contrast to the CPU implementation, all instances of the GPU implementation share static memory on the GPU to avoid time consuming allocations. This is safe here as memory copies are blocking and performed immediately before and after the FFT calculation. ![Calculation of the Hilbert envelope (a) using the CuFFT wrapper only and (b) using a dedicated kernel.[]{data-label="fig:HilbertImplementations"}](winchen_tobias_fig2.pdf){width="\textwidth"} However, in the reconstruction several FFTs are calculated in context of obtaining the Hilbert envelope of the radio traces. The envelope $E(t)$ of a trace $x(t)$ is $E(t) = \sqrt{x(t)^2 + H(x(t)^2)}$ with the Hilbert transform of the signal $H(x(t))$. The Hilbert transformation shifts the phase of negative frequencies, i.e. for band-limited signal frequencies below the mid-frequency, by $-90^\circ$ and positive frequencies by $+90^\circ$. The GPU implementation as described above thus results in a non-optimal memory access pattern for the envelope calculation as shown in Fig. \[fig:HilbertImplementations\] (a). However, with a dedicated computing kernel not only can two memory copies be avoided, but also the phase shift and summation are calculated in parallel (cf. Fig \[fig:HilbertImplementations\] (b)). Discussion ========== ![Summary of the speedup achieved in the individual improvements.[]{data-label="fig:SpeedupSummary"}](winchen_tobias_fig3.pdf){width="\textwidth"} The results obtained from the new GPU implementations are consistent with the results from the baseline implementation. While the FFTs yield identical results on CPU and GPU, the interpolation is not only a different implementation but also only in single precision. This amounts to a relative difference in the directional antenna patterns between the individual implementations of typically below $\pm 0.8\%$ and thus small compared to other uncertainties. The performance improvements obtained by the modifications are summarized in Fig. \[fig:SpeedupSummary\]. As test systems we used here a typical recent desktop PC and a combined CPU/GPU cluster. The desktop is equipped with an AMD A8-6600K processor and an NVIDIA GeForce 750 Ti graphics card. The cluster contains 4 Intel Xeon X5650 CPUs and 4 NVIDIA Tesla M2090 GPUs. On the desktop system the GPU implementation of FFT and the Hilbert transformation yield a speedup of 1.3, doing also the interpolations on the GPU increased this speedup to approximately 2. On the cluster system the total achieved speedup is 1.5. The lower speedup on the cluster system is due to the higher relative performance of the cluster CPU and GPU compared to the desktop system. As only selected isolated parts of the code are moved to the GPU, the time used for computing on the GPU is low compared to the time needed for memory copy, and also only 7% of the copy-time is overlapped by computing time. However, increasing the GPU utilization would require the traces to be kept permanently on the GPU so that more analysis steps can benefit from porting to the GPU. This, however, would require non-trivial changes in the framework, in particular, modifications of the internal structure and interfaces. Conclusion ========== The calculation of Fourier transformations and the interpolation of antenna response patterns have been identified as bottlenecks in the AERA event reconstruction using a performance profiler. Eliminating both by re-implementating the calculation in CUDA while keeping the structure of intact yields a speedup of 1.49 to 2.04 depending on the test system. The largest speedup is obtained here on a typical desktop PC equipped with an entry level graphics card. Considering the relative costs of about €[500]{} for a desktop PC and €[100]{} for an entry level GPU, even such selected applications of GPGPU in existing frameworks are a possibility to be considered in planning future computing strategies.	{ "pile_set_name": "ArXiv" }
--- abstract: 'How do large-scale participants in parimutuel wagering events affect the house and ordinary bettors? A standard narrative suggests that they may temporarily benefit the former at the expense of the latter. To approach this problem, we begin by developing a model based on the theory of large generalized games. Constrained only by their budgets, a continuum of diffuse (ordinary) players and a single atomic (large-scale) player simultaneously wager to maximize their expected profits according to their individual beliefs. Our main theoretical result gives necessary and sufficient conditions for the existence and uniqueness of a pure-strategy Nash equilibrium. Using this framework, we analyze our question in concrete scenarios. First, we study a situation in which both predicted effects are observed. Neither is always observed in our remaining examples, suggesting the need for a more nuanced view of large-scale participants.' address: - 'Department of Mathematics, University of Michigan, Ann Arbor, Michigan 48109' - 'Department of Mathematics, University of Michigan, Ann Arbor, Michigan 48109' author: - Erhan Bayraktar - Alexander Munk bibliography: - 'MFGBib.bib' title: 'High-Roller Impact: A Large Generalized Game Model of Parimutuel Wagering' --- [^1] Introduction {#intro sect} ============ Suppose that a collection of bettors are wagering on an upcoming event. The payoffs are determined via a [(frictionless) parimutuel system]{} if whenever Outcome $i$ occurs, Bettor $A$ receives $$\begin{aligned} \left(\text{Total Amount Wagered} \right) \left( \frac{ \text{Bettor $A$'s Wager on Outcome $i$}}{\text{Total Amount Wagered on Outcome $i$}} \right) . \end{aligned}$$ The idea is that players with correct predictions will proportionally share the final betting pool. Prizes are reduced in practice by transaction costs such as the [house take]{}, a percentage fee collected by the [betting organizer]{} (or [house]{}). For example, Bettor $A$ might only win $$\begin{aligned} \label{friction pmg payoff words} \kappa \left(\text{Total Amount Wagered} \right) \left( \frac{ \text{Bettor $A$'s Wager on Outcome $i$}}{\text{Total Amount Wagered on Outcome $i$}} \right) \end{aligned}$$ when Outcome $i$ occurs, if the house take is $\left( 1- \kappa \right) \%$ for some $0 < \kappa <1$. This mechanism was invented in the context of horse race gambling by Oller in the late 1800’s ([@oller+bio]) and remains widely employed in that setting: In 2014, worldwide parimutuel betting on horse races totaled around seventy-five billion euros ([@horse+racing+fed]). It also typically determines wagering payoffs for other sports such as jai alai and races involving bicycles, motorcycles, motorboats, and greyhounds ([@baron2007parimutuel]). Certain prizes for major lotteries such as Mega Millions, Powerball, and “EuroMillions” are computed in a parimutuel fashion ([@euromil]). Parimutuel systems are increasingly popular methods for distributing payoffs in online prediction markets as well ([@Peters2007]). Goldman Sachs Group, Inc., Deutsche Bank AG, CME Group Inc., Deutsche Börse AG, and ICAP have even facilitated the development of parimutuel derivatives on economic indicators ([@baron2007parimutuel]). The first scholarly publication on parimutuel wagering was written by Borel in 1938 ([@borel1938pari]), and more recent surveys and anthologies ([@thaler1988parimutuel]; [@hausch2011handbook]; [@hausch2008efficiency]) attest to the substantial academic interest garnered by this topic since. A vast range of issues from optimal betting ([@isaacs]; [@rosner]; [@bolt+chap]) to market efficiency ([@hzr]; [@ASCH1982187]; [@hurley1995note]) to market microstructure ([@lange+econ]; [@Peters2007]; [@Pennock]) has been extensively studied. Significant attention has been paid to strategic interactions among bettors ([@quandt+bet+eq]; [@ott+sor+surprise]; [@Koessler2008733]; [@terr+farmer]; [@plott]). Similar to the rise of high-frequency and algorithmic traders in financial markets, a growing number of parimutuel wagering event participants are organizations employing large-scale strategies based upon advanced mathematical, statistical, and computational techniques ([@wired]). There are fundamental differences between these bettors and more traditional wagerers. The new firms typically have access to vast budgets, making their betting totals orders of magnitude beyond the amounts wagered by regular players. Often, they can place their wagers at speeds impossible for ordinary bettors to match. Presumably, their use of complex methods also makes their forecasts and corresponding wagering strategies generally superior. The house collects a percentage of the total amount wagered and, therefore, may initially benefit from the presence of large-scale wagering firms. After all, their activities should increase the size of the pool, at first anyway. The factors just described are thought to put ordinary bettors at an extreme disadvantage, though. Since payouts are calculated according to (\[friction pmg payoff words\]), ordinary bettors’ profits may even directly decline as a result of the large-scale firms’ wagers. If this discourages enough regular players from betting, then pool sizes may eventually dwindle, hurting the house’s revenue. In fact, many betting organizers have publicly expressed strong concerns about the new breed of wagerers. Betting organizers have even occasionally banned these participants from parimutuel wagering events ([@wired]). How reasonable is this narrative? Our goal is to quantify the impact of large-scale participants in parimutuel wagering events on the house and ordinary bettors. First, using the theory of [large generalized games]{}, i.e., games with a continuum of diffuse (or non-atomic/minor) players and finitely many atomic (or major) players, we develop a model of parimutuel betting. The bets made by individual atomic players affect all others because they change the final payoff per unit bet on Outcome $i$: $$\begin{aligned} \label{payouts per unit bet words} \kappa \left( \frac{ \text{Total Amount Wagered}}{\text{Total Amount Wagered on Outcome $i$}} \right) . \end{aligned}$$ Aggregate decisions made by the diffuse players also affect every player for the same reason. A key feature is that an individual diffuse player cannot change (\[payouts per unit bet words\]) by revising her wagers. In fact, her specific choices have no effect whatsoever on the rest of the game’s participants. We view diffuse and atomic players as stand-ins for ordinary bettors and large-scale wagering firms, respectively. The approximation is motivated by the observation that the total amount wagered by a single traditional bettor is generally negligible compared to the total amount wagered by an entire betting firm. Our main theoretical result, Theorem \[main thm\], provides necessary and sufficient conditions for the existence and uniqueness of a pure-strategy Nash equilibrium. Other scholars have shown the existence of equilibria in a broad class of large generalized games ([@Balder1999207]; [@Balder2002437]; [@Carmona2014130]; [@riascos2013pure]). Such results often rely upon sophisticated technology including variants of the Kakutani fixed-point theorem. We choose to employ a more elementary fixed-point argument instead. Advantages of our approach include its simplicity and the possibility of proving the equilibrium’s uniqueness. A simple algorithm for computing relevant equilibrium quantities immediately presents itself as well. Having such an algorithm allows us to analyze our problem in specific scenarios. In accordance with the prevailing narrative, we observe that because of the atomic player, the house is temporarily better off and the diffuse players are worse off in Example \[exmp 1\]. For varying reasons, at least one of these effects is not observed in each of the remaining situations. In Example \[exmp 2\], the diffuse players are better off in the presence of the atomic player. Intuitively, when the event is [too close to call]{}, the atomic player can bet on the [wrong]{} outcome even if her prediction is assumed to be quite accurate. Such an error is to the advantage of the diffuse players. In Example \[exmp 3\], the diffuse players [believe]{} that they are better off when there is an atomic player. Roughly, if the diffuse players’ beliefs are too homogeneous but the atomic player disagrees with them, the diffuse players’ expected profits per unit bet rise when the atomic player takes the other side of their wagers. In Example \[exmp 4\], we argue that the diffuse players are better off but the house is (immediately) worse off because of the atomic player, exactly the opposite of the prevailing narrative.[^2] To make this point, we recast our model as a two-stage game taking into account the house’s strategic decisions. Effectively, because the atomic player considers her impact on (\[payouts per unit bet words\]), she has a lower tolerance for unfavorable betting conditions than the diffuse player. When she is absent, this means that the house can more easily prey upon the diffuse players. Before offering further details, we more thoroughly discuss related literature in Section \[lit review sect\]. We carefully present our model in Section \[model sect\] and our main theoretical result in Section \[main result sect\]. We numerically investigate our concrete examples in Section \[leb meas sect\]. Appendix \[corr app\] highlights a few technical aspects of Section \[main result sect\]’s work. We give our longer formal proofs in Appendices \[prop 2 proof app\] and \[main thm proof app\]. Related Literature {#lit review sect} ================== The individual states of our diffuse players are only coupled via the empirical distribution of controls. Since each diffuse player is too small to influence this distribution, she treats it as fixed when determining her own strategy. Assumptions like these have appeared in the literature on continuum games ([@aumann]; [@schmeidler]; [@MASCOLELL]; [@rath]) and mean-field games ([@Lasry2007]; [@huang2006]).[^3] Our paper bears a stronger resemblance to work in the former category. For instance, we model parimutuel wagering as a static game. Such a choice is quite common in a continuum game study; however, stochastic differential games are more often the focus in mean-field game theory. Also, we restrict ourselves to an intuitive argument for viewing ordinary bettors as diffuse players. Papers on mean-field games often rigorously present their continuum model as a limit of finite population models, while those on continuum games typically emphasize other issues. General mean-field interactions among players can be described by complex functions of the empirical distributions of states and/or controls. On the other hand, parimutuel wagerers affect one another through (\[payouts per unit bet words\]) alone, a comparatively simple scenario. This makes parimutuel wagering an especially strong candidate for modeling by either theory. Continuum games have already been applied in this way ([@ottaviani2006timing]; [@Takahiro199785]). Watanabe considered a two-stage game with a betting organizer and a continuum of risk-neutral diffuse players with heterogeneous beliefs ([@Takahiro199785]). First, the betting organizer selects a value for the house take. The diffuse players then determine whether to place a unit bet on one of two outcomes or bet nothing at all. Using techniques from set-valued analysis, Watanabe showed that an equilibrium always exists, provided the house take is not too large. As long as each player can only bet negligible amounts, Watanabe also found that equilibria in parimutuel wagering games are [regular]{}. That is, if a player wagering on Outcome $i$ believes that Outcome $i$ will occur with probability $p$, all players who believe that Outcome $i$ will occur with probability $p^{\prime} > p$ also wager on Outcome $i$. The paper’s results on the betting organizer’s optimal strategy were in the context of specific examples. Ottaviani and Sørenson developed their continuum game to explain two phenomena frequently observed in the context of parimutuel wagering on horse races: [late informed betting]{} and the [favorite-longshot bias]{} ([@ottaviani2006timing]). The first states that more accurate information about a race’s outcome can be gleaned from late bets than early bets. The second says that the public tends to excessively bet on unlikely outcomes and wager too little on likely outcomes. A continuum of privately informed risk-neutral players decide when to place their individual bets in a discrete-time setting. They can wager a unit amount on one of two outcomes or abstain from betting. The paper gave conditions under which all of these players simultaneously wager at the terminal time, and the corresponding equilibrium is shown to always feature the favorite-longshot bias. Others have more implicitly created infinite-player parimutuel wagering models by assuming that there are [many]{} bettors ([@hurley1995note]; [@Ottaviani08thefavorite-longshot]). These references have sought to identify other possible causes of the favorite-longshot bias. The most important new feature of our setup is that, in addition to the diffuse bettors, we introduce an atomic bettor. The parimutuel wagering studies we just discussed only incorporate diffuse players. We would not be able to understand the effects of large-scale wagering organizations on ordinary bettors, if we made a similar assumption. Because of this addition, our model belongs to the class of continuum game models known as [large generalized games]{}. Games that include both diffuse and atomic players can be found in mean-field game theory as well under the heading [major-minor player models]{} ([@huang]; [@nguyen+huang]; [@nourian+caines]; [@jaimungal2015mean]; [@wang+tang+huang]). Applications of large generalized games are known to be diverse and already include a collection of problems from politics ([@Correa2014]) to oligopolistic markets ([@wiszniewska2008dynamic]). General results on the existence of equilibria in large generalized games have also been obtained ([@Balder1999207]; [@Balder2002437]; [@Carmona2014130]; [@riascos2013pure]). We choose not to rely upon these, as the simple structural aspects of parimutuel wagering just discussed, combined with a convenient modeling assumption (see Section \[model sect\]), allow us to use elementary arguments. Ultimately, we employ a mean-field approximation for the standard reason: By doing so, we make our model tractable, hopefully while preserving the critical macroscopic properties of our original problem. Issues other than the impact of large-scale wagering organizations have been resolved in finite-player settings, though such studies have usually invoked other strong assumptions. Weber gave sufficient conditions for the existence of an equilibrium in a simultaneous parimutuel betting game with $N$ atomic players, each of whom is risk-neutral and must bet a specific total amount ([@MR637486]). Watanabe, Nonoyama, and Mori’s setup and goals are similar to those in Watanabe’s continuum game paper, except that the diffuse players in the latter are replaced by finitely many atomic players ([@wat+non+mori]; [@Takahiro199785]). Explanations of the favorite-longshot bias have been offered using equilibrium results for $N$-player parimutuel wagering games ([@chadha1996betting]; [@Koessler2008733]; [@ott+sor+surprise]; [@potters+wit]). Games in which finitely many bettors wager sequentially have also been investigated for various purposes ([@Feeney2001165]; [@cheung]; [@koess+zieg+broi]; [@Koessler2008733]; [@thrall]). For example, Thrall showed that if risk-neutral atomic bettors with homogeneous beliefs wager sequentially, their profits tend to zero as the number of bettors increases ([@thrall]). Note that some of these works do consider limiting cases in which the population of wagerers grows arbitrarily large to complement their other insights ([@thrall]; [@ott+sor+surprise]). Model Details {#model sect} ============= Our players have the opportunity to wager on an event that can unfold in two mutually exclusive ways: Outcome 1 might occur. If not, Outcome 2 will occur. For now, we view $\kappa \in \left( 0 , 1 \right)$ to be exogenously given. Inspired by Watanabe et al. ([@wat+non+mori]; [@Takahiro199785]), we later informally consider what happens when we allow the house take to be optimally selected by the betting organizer in the first stage of a two-stage game (see Example \[exmp 4\]). Our results in Section \[main result sect\] are unaffected by such a shift. The unit interval describes the diffuse bettors’ views on the likelihood of Outcome 1: the bettors whose views are indexed by $p \in \left[ 0 , 1 \right]$ believes that Outcome 1 will occur with probability $p$. Initially, each diffuse bettor has some (negligible) unit wealth. A finite Borel measure $\mu$ with a continuous everywhere positive density characterizes the distribution of the diffuse bettors. More precisely, the total initial wealth of all diffuse bettors whose views are contained in a Borel set $A$ is $\mu \left( A\right)$. That $\mu$ has a continuous everywhere positive density is our [convenient modeling assumption]{} from Section \[lit review sect\]. It is similar to a key hypothesis in Ottaviani and Sørenson’s work, although the posterior beliefs of their diffuse bettors are obtained after updating a common prior belief using a private signal and Bayes’ rule ([@ottaviani2006timing]). Effectively, Watanabe assumed that $\mu$ need not have a density, and even when it does, the density need not be positive everywhere ([@Takahiro199785]). These choices necessitated a set-valued approach, which we are able to avoid. Continuity merely simplifies a few of our arguments, e.g., see Step \[step 6 alt varphi dec\] of Theorem \[main thm\]’s proof. The rest of the assumption plays a more critical role. Results on the existence of equilibria in parimutuel wagering games often include a hypothesis such as the following: for any given outcome, at least two bettors believe that the outcome will occur with positive probability.[^4] By supposing that the density is positive, we assume this as well. Watanabe has shown that an equilibrium may not be unique, if $\mu \left( \left\{ p \right\} \right) > 0$ for some fixed $p$ ([@Takahiro199785]). Obviously, this situation is prevented by the density’s existence. Our atomic player believes that Outcome 1 will occur with probability $q \in \left[ 0 , 1 \right]$. She has (non-negligible) finite initial wealth $w > 0$. Throughout, we treat all players’ beliefs as exogenously determined. We do not address how the players generate their estimates; however, this process is of great interest both practically and academically ([@hausch2011handbook]; [@hausch2008efficiency]). For our theoretical results in Section \[main result sect\], we also do not specify how the players’ estimates compare to the [actual]{} probability that Outcome 1 will occur. Since large-scale betting organization allegedly produce highly accurate forecasts, we could choose $q$ to be some small perturbation of the actual probability of Outcome 1. We informally experiment with this extra assumption in Examples \[exmp 1\] and \[exmp 2\]. All players decide how much to wager on each outcome. Their choices are constrained only by their initial wealth: a betting strategy is [feasible]{} (or [admissible]{}) for an individual bettor as long as the sum of her wagers is no more than her wealth. For example, a bettor could choose to wager 100% of her wealth on Outcome 1, 55% of her wealth on Outcome 1 and 30% of her wealth on Outcome 2, or not wager at all. We formalize this as follows.[^5] \[strat profile defn\] A [feasible strategy profile for the diffuse players]{} is a measurable function $$f = \left( f_1 , f_2 \right) : \left( 0 , 1 \right) \longrightarrow \left\{ \left(x_1 , x_2 \right) \in \mathbb{R}^2_{\geq 0} \, : \, x_1 + x_2 \leq 1 \right\}.$$ A [feasible strategy profile for the atomic player]{} is a vector $a = \left( a_1 , a_2 \right) \in \mathbb{R}^2_{\geq 0}$ such that $$a_1 + a_2 \leq w.$$ We call the pair $\left( f , a \right)$ a [feasible strategy profile]{}. Under $\left( f , a \right)$, the atomic player wagers $a_i$ on Outcome $i$. Each diffuse player who believes that Outcome 1 will occur with probability $p$ wagers $f_i \left( p \right) \times 100 \%$ of her (negligible) unit initial wealth on Outcome $i$. Our space of feasible strategy profiles is slightly atypical. Previously, diffuse players in parimutuel wagering games have only been able to place unit bets, if they bet at all ([@ottaviani2006timing]; [@Takahiro199785]). We could have made this restriction as well without loss of generality due to Proposition \[diff equilib bets prop\]. Watanabe allowed groups of diffuse players to wager differently, even if they held identical beliefs ([@Takahiro199785]). In equilibrium, such a discrepancy could only arise among the diffuse players who believed that their expected profits would be zero. We encounter a related ambiguity in our framework (see our discussion of Proposition \[diff equilib bets prop\]). For us, the $\mu$-measure of this set of bettors is zero, and we anticipate that all of our main results would remain the same, if we were to relax the diffuse bettors’ [same beliefs-same bets]{} restriction. More significantly, atomic players have been frequently constrained to wager a fixed amount in total or a unit amount on a single outcome when they bet ([@MR637486]; [@wat+non+mori]; [@chadha1996betting]; [@ott+sor+surprise]).[^6] Proposition \[atom equilib bets prop\] suggests that imposing these restrictions would have a severe effect. The total amount that the diffuse players wager on Outcome $i$, denoted $d_i$, is given by $$d_i = \displaystyle\int_{0}^1 f_i\left( p \right) \mu \left( d p \right) .$$ Since $\mu$ has a density, we immediately confirm that the bets placed by any given diffuse player are too small to affect the amount wagered on any specific outcome. Of course, if she revises her strategy, then a particular diffuse player affects neither the total amount wagered nor (\[payouts per unit bet words\]). All of these quantities could change when aggregations of diffuse players, that is, collections of diffuse players whose beliefs are contained in a Borel set $A$ with positive $\mu$-measure, revise their wagers. Payoffs are determined according to (\[friction pmg payoff words\]). Each player selects her wagering strategy simultaneously in order to maximize her expected profit according to her beliefs. We implicitly assume that every bettor knows $\kappa$, $\mu$, $q$, and $w$, so that she can select the best response to her opponents’ collective actions. Similar assumptions can be found in many other equilibrium studies on parimutuel wagering ([@MR637486]; [@wat+non+mori]; [@Takahiro199785]; [@chadha1996betting]; [@ott+sor+surprise]).[^7] Since each diffuse player starts out with negligible unit wealth, technically, we should only discuss the expected profits of diffuse bettors whose views lie in some Borel set $A$. We nevertheless compute and refer to the expected profits of an individual diffuse bettor in an obvious, but admittedly informal, way. Doing so helps motivate our definition of a [pure-strategy Nash equilibrium]{} (see Definition \[maj min NE def\]) and highlight the intuition underlying our results. A seemingly more formidable concern is how to handle (\[friction pmg payoff words\]) in pathological cases. It is trivial to produce a feasible strategy profile $\left( f, a \right)$ such that for some $p$, we have $f_j \left( p \right) > 0$ but $$d_j = a_j = 0 .$$ Na[ï]{}vely translating (\[friction pmg payoff words\]), we conclude that a diffuse player whose views are indexed by $p$ receives $$\begin{aligned} \label{diff pay eqn bef impl prob} \kappa \left( \displaystyle\sum_{i=1 }^2 \left( d_i + a_i \right) \right) \left( \displaystyle\frac{ f_j \left( p \right) }{ d_j + a_j } \right)\end{aligned}$$ whenever Outcome $j$ occurs. If $$d_1 = d_2 = a_1= a_2 = 0 ,$$ then the total amount wagered is zero. Practically, the betting organizer would probably cancel such an event, which suggests that it is natural to set all players’ payoffs to zero in this scenario. Alternatively, we might have $$d_i+ a_i >0$$ for $i \ne j$. There are now two appealing options for the diffuse player’s payoff. First, we might choose to set the payoff to zero. Practically, no bettor would receive a payout, if Outcome $j$ occurred but no one wagered on it. We also might set the payoff to $+\infty$, as in Watanabe’s work ([@Takahiro199785]). In practice, if the amount wagered on Outcome $j$ were zero, each player who believed that Outcome $j$ would occur with positive probability would want to place an arbitrarily small bet on Outcome $j$. Setting the payoff to $+\infty$ captures this intuition. We choose the first option, but selecting the second instead would not change our results. Only equilibrium payoffs need to be computed, and in an equilibrium, positive amounts are always wagered on both outcomes. Essentially, the scenario we describe never arises. One reason is that we do not allow the [trivial]{} (or [null]{}) equilibrium in which no player wagers. We could,[^8] but as Watanabe observed, that case is comparatively uninteresting and practically unimportant ([@Takahiro199785]). Roughly, the other reason is the same as our justification for possibly setting the payoff to $+\infty$. Before making this discussion precise, we introduce some notation. \[impl prob defn\] Given a feasible strategy profile $\left( f , a \right)$ such that at least one of the $d_i$’s or $a_i$’s is positive, the [implied (or subjective) probability]{} that Outcome 1 will occur, denoted $P^{f,a}$, is defined by $$P^{f,a} = \displaystyle\frac{ d_1 + a_1 }{ \sum_{i=1 }^2 \left( d_i + a_i \right) } .$$ We refer to $$1- P^{f,a} = \displaystyle\frac{ d_2 + a_2 }{ \sum_{i=1 }^2 \left( d_i + a_i \right) }$$ as the [implied (or subjective) probability]{} that Outcome 2 will occur. $P^{f,a}$ is the ratio of the amount wagered on Outcome 1 to the total amount wagered, assuming the latter is positive. Our previous discussion implies that $P^{f,a} \in \left( 0 , 1 \right)$ in equilibrium.[^9] Since the argument was informal, we do not yet take this as fact. In particular, the amount received by a diffuse player who believes that Outcome 1 will occur with probability $p$ is[^10] $$\begin{aligned} \kappa \left( \displaystyle\sum_{i=1 }^2 \left( d_i + a_i \right) \right) \left( \displaystyle\frac{ f_1 \left( p \right) \mathds{1}_{ \left\{ d_1 + a_1 \ne 0 \right\} } }{ d_1 + a_1 } \right) = \displaystyle\frac{ \kappa f_1 \left( p \right) \mathds{1}_{ \left\{ P^{f,a} \ne 0 \right\} } }{ P^{f,a} } \end{aligned}$$ and $$\begin{aligned} \kappa \left( \displaystyle\sum_{i=1 }^2 \left( d_i + a_i \right) \right) \left( \displaystyle\frac{ f_2 \left( p \right) \mathds{1}_{\left\{ d_2 + a_2 \ne 0 \right\} } }{ d_2 + a_2 } \right) = \displaystyle\frac{ \kappa f_2 \left( p \right) \mathds{1}_{ \left\{ P^{f,a} \ne 1 \right\} } }{ 1- P^{f,a} } \end{aligned}$$ when Outcomes 1 and 2 occur, respectively. Hence, this diffuse player believes that her expected profit is $$\begin{aligned} \label{exp prof diff player intuit} f_1 \left( p \right) \left( \displaystyle\frac{ \kappa \, \mathds{1}_{\left\{ P^{f, a}\ne 0 \right\}} \, p }{ P^{f, a}} - 1 \right) + f_2 \left( p \right) \left( \displaystyle\frac{ \kappa \, \mathds{1}_{ \left\{ P^{f, a} \ne 1 \right\} } \, \left( 1 - p \right) }{ 1 - P^{f, a} } - 1 \right) . \end{aligned}$$ Similarly, the atomic player thinks that her expected profit is $$\begin{aligned} \label{exp prof atom player intuit} a_1 \left( \displaystyle\frac{ \kappa \, \mathds{1}_{\left\{ P^{f, a}\ne 0 \right\}} \, q }{ P^{f, a}} - 1 \right) + a_2 \left( \displaystyle\frac{ \kappa \, \mathds{1}_{ \left\{ P^{f, a} \ne 1 \right\} } \, \left( 1 - q \right) }{ 1 - P^{f, a} } - 1 \right) . \end{aligned}$$ \[maj min NE def\] A [pure-strategy Nash equilibrium]{} is a feasible strategy profile $\left( f^\star , a^\star \right)$ such that 1. \[NE def orig i\] at least one of the $d_i^{\star}$’s or $a_i^{\star}$’s is positive; 2. \[NE def orig ii\] for any $p \in \left[ 0 , 1 \right]$, $$\begin{aligned} &f_1^{\star} \left( p \right) \left( \displaystyle\frac{ \kappa \, \mathds{1}_{\left\{ P^{f^{\star}, a^{\star}}\ne 0 \right\}} \, p }{ P^{f^{\star}, a^{\star}}} - 1 \right) + f_2^{\star} \left( p \right) \left( \displaystyle\frac{ \kappa \, \mathds{1}_{ \left\{ P^{f^{\star}, a^{\star}} \ne 1 \right\} } \, \left( 1 - p \right) }{ 1 - P^{f^{\star}, a^{\star}} } - 1 \right) \notag \\ &\quad = \displaystyle\sup_{\substack{ b_1 , b_2 \geq 0 \\ b_1 + b_2 \leq 1 }} \left\{ b_1 \left( \displaystyle\frac{ \kappa \, \mathds{1}_{\left\{ P^{f^{\star}, a^{\star}}\ne 0 \right\}} \, p }{ P^{f^{\star}, a^{\star}}} - 1 \right) + b_2 \left( \displaystyle\frac{ \kappa \, \mathds{1}_{ \left\{ P^{f^{\star}, a^{\star}} \ne 1 \right\} } \, \left( 1 - p \right) }{1 - P^{f^{\star}, a^{\star}} } - 1 \right) \right\} ;\end{aligned}$$ 3. \[NE def orig iii\] and $$\begin{aligned} & a_1^{\star} \left( \displaystyle\frac{ \kappa \, \mathds{1}_{\left\{ P^{f^{\star}, a^{\star}}\ne 0 \right\}} \, q }{ P^{f^{\star}, a^{\star}}} - 1 \right) + a_2^{\star} \left( \displaystyle\frac{ \kappa \, \mathds{1}_{ \left\{ P^{f^{\star}, a^{\star}} \ne 1 \right\} } \, \left( 1 - q\right) }{ 1 - P^{f^{\star}, a^{\star}} } - 1 \right) \notag \\ &\quad = \displaystyle\sup_{\substack{ b_1 , b_2 \geq 0 \\ b_1 + b_2 \leq w \\ d_1^{\star}, d_2^{\star}, b_1 \text{ or } b_2 > 0}} \left\{ b_1 \left( \displaystyle\frac{ \kappa \, \mathds{1}_{\left\{ P^{f^{\star}, b }\ne 0 \right\}} \, q }{ P^{f^{\star}, b }} - 1 \right) + b_2 \left( \displaystyle\frac{ \kappa \, \mathds{1}_{ \left\{ P^{f^{\star}, b } \ne 1 \right\} } \, \left( 1 -q \right) }{1 - P^{f^{\star}, b} } - 1 \right) \right\} . \end{aligned}$$ \[NE def orig i\] formally excludes the case in which the total amount wagered is zero. \[NE def orig ii\] and \[NE def orig iii\] ensure that each player maximizes her expected profit according to her beliefs, given her opponents’ wagers.[^11] First, observe that each player requires very little information about her opponents’ strategies. For a given diffuse bettor, knowing $P^{f^{\star}, a^{\star}}$ alone is enough. The atomic player must be able to compute $P^{f^{\star}, b}$ for all of her feasible strategy profiles $b$, so it is sufficient for her to know $d_1^{\star}$ and $d_2^{\star}$. The difference for the two kinds of players reflects that an individual diffuse player cannot affect the implied probability of Outcome 1, while the atomic player can. These remarks explain why we informally claim that only the atomic player and aggregations of diffuse players affect the other participants. Notice that each player’s strategy depends [anonymously]{} on her opponents’ bets: how, specifically, her opponents wagers produced $P^{f^{\star}, a^{\star}}$, $d_1^{\star}$, and $d_2^{\star}$ is irrelevant. Although we only optimize over $b$ in \[NE def orig iii\], the apparent possibility that $d_1^{\star} = d_2^{\star} =0$ motivates our use of the extra constraint $$d_1^{\star}, d_2^{\star}, b_1 \text{ or } b_2 > 0 .$$ One might be concerned that we do not consider the feasible strategy $b_1 = b_2 = 0$ for the atomic player, if $d_1^{\star} = d_2^{\star} =0$. Recall that all players receive a payoff of zero in such a situation. A simple calculation shows that the supremum is then also zero, so no issue is caused by our omission. The last important concept for our modeling framework is uniqueness. \[uniq equilib def\] A pure-strategy Nash equilibrium $\left( f^\star , a^\star \right)$ is [unique]{} if for any other pure-strategy Nash equilibrium $\left( f^\diamond , a^\diamond \right)$, we have $f^{\star} = f^{\diamond}$ $\mu$-a.s. and $a^{\star}=a^{\diamond}$. We allow $f^{\star}$ and $f^{\diamond}$ to disagree on a set of $\mu$-measure zero since, ultimately, all relevant equilibrium quantities such as the implied probabilities are unaffected by such a difference. Soon, we see that $f^{\star} \left( p \right)$ and $f^{\diamond} \left( p \right)$ must be equal for all but two points, at most. There is only uncertainty about the behavior of the diffuse bettors who believe that their expected profits are zero (cf. our discussion about our space of feasible strategy profiles). Theoretical Results {#main result sect} =================== We now state and prove[^12] our theoretical results, beginning with Propositions \[diff equilib bets prop\] and \[atom equilib bets prop\]. The former describes how the diffuse players should wager in response to the atomic player’s strategy. The latter tells us how the atomic player should bet, given the diffuse players’ wagers. We use these observations to prove Theorem \[main thm\], our main result.[^13] Recall that it offers necessary and sufficient conditions for the existence and uniqueness of a pure-strategy Nash equilibrium. We conclude Section \[main result sect\] with Corollaries \[atom pl exp prof pos cor\], \[reg equilib cor\], and \[impl prob kappa 0.5\]. Corollary \[atom pl exp prof pos cor\] says that the atomic player wagers on a particular outcome if and only if the final expected profit per unit bet on that outcome is positive. The next corollary states that our equilibria are regular in a particular sense, while Corollary \[impl prob kappa 0.5\] says that the implied probability of Outcome 1 tends to $ 0.5$ uniformly as the house take approaches $50 \%$. \[diff equilib bets prop\] Let $\left( f , a \right)$ be a feasible strategy profile such that $d_1$, $d_2 > 0$. $f$ satisfies $$\begin{aligned} \label{max problem prop 1} &f_1 \left( p \right) \left( \displaystyle\frac{ \kappa \, p }{ P^{f , a }} - 1 \right) + f_2 \left( p \right) \left( \displaystyle\frac{ \kappa \, \left( 1 - p \right) }{ 1 - P^{f , a } } - 1 \right) \notag \\ &\quad = \displaystyle\sup_{\substack{ b_1 , b_2 \geq 0 \\ b_1 + b_2 \leq 1 }} \left\{ b_1 \left( \displaystyle\frac{ \kappa \, p }{ P^{f , a }} - 1 \right) + b_2 \left( \displaystyle\frac{ \kappa \, \left( 1 - p \right) }{1 - P^{f , a } } - 1 \right) \right\} \end{aligned}$$ for all $p \in \left[ 0 , 1 \right]$ if and only if $$\begin{aligned} f_1 \left( p \right) &= \left\{ \begin{array}{cc} 1 & \quad \text{if } \, \, p > P^{f,a} / \kappa \\ & \\ 0 & \quad \text{if } \, \, p < P^{f,a} / \kappa \\ \end{array} \right. \qquad \qquad f_2 \left( p \right) = \left\{ \begin{array}{cc} 1 & \quad \text{if } \, \, 1 - p > \left( 1 - P^{f,a} \right) / \kappa \\ &\\ 0 & \quad \text{if } \, \, 1 - p < \left( 1 - P^{f,a} \right) / \kappa \\ \end{array} \right. .\end{aligned}$$ First, notice that $ P^{f,a} \in \left( 0 ,1 \right)$ since both $d_1$ and $d_2$ are positive. Comparing (\[max problem prop 1\]) and \[NE def orig ii\] of Definition \[maj min NE def\], we find that Proposition \[diff equilib bets prop\] provides a simple characterization of the diffuse players’ equilibrium strategies in this case, given the bets of the atomic player. Our assumption is not too restrictive, as Step \[step 1 orig\] of Theorem \[main thm\]’s proof says that $d_1$ and $d_2$ are always both positive in equilibrium. Three distinct groups of diffuse bettors emerge: The first group, containing the diffuse bettors who believe that Outcome 1 will occur with probability greater than $P^{f, a}/ \kappa$, wager all of their initial wealth on Outcome 1. Diffuse bettors who think that Outcome 2 will occur with probability greater than $\left( 1 - P^{f, a}\right) / \kappa$ make up the second group. These players bet their entire fortunes on Outcome 2. The remaining diffuse players, except those whose beliefs are indexed by $p = P^{f, a}/ \kappa$ or $\left( 1 - P^{f, a}\right) / \kappa$, do not wager at all. None of the groups overlap, since $0 < \kappa < 1$ implies that $$\label{fi star nonoverlap eqn} 1 - \left( \displaystyle\frac{ 1 - P^{f, a}}{ \kappa} \right) < \displaystyle\frac{P^{f, a}}{\kappa} .$$ The proof’s underlying intuition is easy to explain. It is equivalent to show that $f$ satisfies (\[max problem prop 1\]) on $\left[ 0 , 1 \right]$ if and only if $$\begin{aligned} f_1 \left( p \right) &= \left\{ \begin{array}{cc} 1 & \quad \text{if } \, \, \frac{ \kappa p }{ P^{f, a}} - 1 > 0\\ & \\ 0 & \quad \text{if } \, \, \frac{ \kappa p }{ P^{f, a}} - 1 < 0\\ \end{array} \right. \qquad \qquad f_2 \left( p \right) = \left\{ \begin{array}{cc} 1 & \quad \text{if } \, \, \frac{ \kappa \left( 1 - p \right) }{ 1 - P^{f, a} } - 1 > 0 \\ &\\ 0 & \quad \text{if } \, \, \frac{ \kappa \left( 1 - p \right) }{ 1 - P^{f, a} } - 1 < 0 \\ \end{array} \right. .\end{aligned}$$ The terms $$\displaystyle\frac{ \kappa p }{ P^{f, a}} - 1 \qquad \text{and} \qquad \displaystyle\frac{ \kappa \left( 1 - p \right) }{ 1 - P^{f, a} } -1$$ describe the expected profit per unit bet on Outcomes 1 and 2, respectively, from the perspective of the diffuse player who believes that Outcome 1 will occur with probability $p$. Individual diffuse players are risk-neutral and do not affect these quantities, so they wager on an outcome only if the corresponding term is positive. For a particular diffuse player, this is true of at most one outcome as $\kappa \in \left( 0 , 1 \right)$. Consequently, if a diffuse player has identified a profitable wagering opportunity, she bets her entire fortune on it. Despite their large space of feasible strategies, the diffuse players, aside from those whose beliefs are indexed by $p = P^{f, a}/ \kappa$ or $\left( 1 - P^{f, a}\right) / \kappa$, wager either $100\%$ or $0\%$ of their wealth on each outcome. The value of $f_i$ at $P^{f, a} / \kappa$ and $\left( 1 - P^{f,a} \right) / \kappa$ is ambiguous because, if a given diffuse player’s expected profit per unit bet on Outcome $i$ is zero, then she is indifferent to the size of her bet on Outcome $i$.[^14] Though their setups differed from our own (see Sections \[lit review sect\] - \[model sect\]), Ottaviani, Sørenson, and Watanabe found similar groupings of diffuse players in equilibrium ([@ottaviani2006timing]; [@Takahiro199785]). We return to this observation during our discussion of Corollary \[reg equilib cor\], which roughly says that these groupings persist even when we take into account the atomic player’s wagers. There is little to formalize beyond our heuristic discussion above. We only comment that rearranging (\[fi star nonoverlap eqn\]) shows that we can never have both $$\displaystyle\frac{ \kappa p }{ P^{f, a}} - 1 > 0 \qquad \text{and} \qquad \displaystyle\frac{ \kappa \left( 1 - p \right) }{ 1 - P^{f, a} } - 1 > 0 .$$ \[atom equilib bets prop\] Let $\left( f , a \right)$ be a feasible strategy profile such that $d_1$, $d_2 > 0$. Consider the equation $$\begin{aligned} \label{atom max prob eqn prop} & a_1 \bigg( \displaystyle\frac{ \kappa \, q }{ P^{f , a }} - 1 \bigg) + a_2 \left( \displaystyle\frac{ \kappa \, \left( 1 - q\right) }{ 1 - P^{f , a } } - 1 \right) \notag \\ &\quad = \displaystyle\sup_{\substack{ b_1 , b_2 \geq 0 \\ b_1 + b_2 \leq w }} \left\{ b_1 \bigg( \displaystyle\frac{ \kappa \, q }{ P^{f , b }} - 1 \bigg) + b_2 \left( \displaystyle\frac{ \kappa \, \left( 1 -q \right) }{1 - P^{f , b} } - 1 \right) \right\} . \end{aligned}$$ 1. \[atom bet prop orig i\] When $$\label{atom bets out1 ineq alone} q > \displaystyle\frac{ d_1 }{\kappa \left( d_1 + d_2 \right)} ,$$ $a$ satisfies (\[atom max prob eqn prop\]) if and only if $a_2 = 0$ and $$\begin{aligned} \label{atom bets out1 soln alone} a_1 &= \displaystyle\min \left\{ w , \, \displaystyle\sqrt{ \displaystyle\frac{ \kappa q d_1 d_2}{1 - \kappa q } } - d_1 \right\}. \end{aligned}$$ 2. \[atom bet prop orig ii\] When $$\label{atom bets out2 ineq alone} 1 - q > \displaystyle\frac{ d_2 }{\kappa \left( d_1 + d_2 \right)} ,$$ $a$ satisfies (\[atom max prob eqn prop\]) if and only if $a_1 = 0$ and $$\begin{aligned} \label{atom bets out2 soln alone} a_2 &= \displaystyle\min \left\{ w , \, \sqrt{ \displaystyle\frac{ \kappa \left( 1 - q \right) d_1 d_2 }{1 - \kappa \left( 1 - q \right) } } - d_2 \right\}. \end{aligned}$$ 3. \[atom bet prop orig iii\] When $$\label{atom bets noth ineq alone} q \leq \displaystyle\frac{ d_1 }{\kappa \left( d_1 + d_2 \right)} \qquad \text{and} \qquad 1 - q \leq \displaystyle\frac{ d_2 }{\kappa \left( d_1 + d_2 \right)},$$ $a$ satisfies (\[atom max prob eqn prop\]) if and only if $a_1 = a_2 = 0$. As in our discussion of the last result, $P^{f,a} \in \left( 0 ,1 \right)$ and that we only study the case in which $d_1$, $d_2 > 0$ does not matter. Proposition \[atom equilib bets prop\] can be interpreted as the complement of Proposition \[diff equilib bets prop\]: It characterizes the atomic player’s equilibrium strategy, given the diffuse players’ bets. A calculation similar to that in (\[fi star nonoverlap eqn\]) shows that (\[atom bets out1 ineq alone\]) and (\[atom bets out2 ineq alone\]) never hold simultaneously. Notice that $a_1 > 0$ under \[atom bet prop orig i\], while $a_2 > 0$ under \[atom bet prop orig ii\]. For example, (\[atom bets out1 ineq alone\]) implies that $$\label{1 bc impli} \displaystyle\frac{ d_1 }{d_1 + d_2 } < \kappa q \qquad \text{and} \qquad \displaystyle\frac{ d_2 }{d_1 + d_2 } > 1- \kappa q.$$ In this case, $$a_1 = \displaystyle\min \left\{ w , \, \displaystyle\sqrt{ \displaystyle\frac{ \kappa q d_1 d_2}{1 - \kappa q } } - d_1 \right\} \geq \displaystyle\min \left\{ w , \, \displaystyle\sqrt{ \left( \displaystyle\frac{d_1}{d_2} \right) d_1 d_2 } - d_1 \right\} > 0.$$ Hence, under (i), the atomic player wagers on Outcome 1 alone. The atomic player only bets on Outcome 2 in (ii), while she does not wager at all in (iii). Despite having the opportunity to do so, she never simultaneously wagers on both possibilities. Proposition \[diff equilib bets prop\] revealed similar behavior for diffuse bettors. Important ideas in the proofs of Propositions \[diff equilib bets prop\] and \[atom equilib bets prop\] are closely related. By rearranging (\[atom bets out1 ineq alone\]) and (\[atom bets out2 ineq alone\]), we get $$\displaystyle\frac{ \kappa \left( d_1 + d_2 \right) q }{ d_1} - 1 > 0 \qquad \text{and} \qquad \displaystyle\frac{ \kappa \left( d_1 + d_2 \right) \left( 1 - q \right) }{ d_2} - 1 > 0 ,$$ respectively. Given the wagers of the diffuse players, the first term describes the expected profit per unit bet on Outcome 1 according to the atomic player. The second term has the analogous interpretation for Outcome 2. As in our analysis for the diffuse players, the atomic player is risk-neutral and bets only when one of these inequalities holds,[^15] leading directly to \[atom bet prop orig iii\]. Isaacs first proved this while modeling parimutuel wagering as the control problem faced by a single risk-neutral atomic player unconstrained by a budget ([@isaacs]). We include \[atom bet prop orig iii\] in Proposition \[atom equilib bets prop\] merely to assist with our presentation. Unlike our previous analysis, we cannot conclude that the atomic player bets her entire fortune on an initially profitable wagering opportunity. The reason is simple: the atomic player’s choices affect (\[payouts per unit bet words\]). In fact, all else being equal, the payoffs per unit bet on an outcome decrease as the atomic player raises her wager on that outcome. Balancing the desires to increase her expected profit by betting more and keep her expected profit per unit bet high by betting less leads to (\[atom bets out1 soln alone\]) and (\[atom bets out2 soln alone\]), not the all-or-nothing wagers of Proposition \[diff equilib bets prop\]. More precisely, if (\[atom bets out1 ineq alone\]) holds and we relax our wealth constraint, this trade-off makes it optimal for the atomic player to wager $$\label{isaacs wager outcome 1} \displaystyle\sqrt{ \displaystyle\frac{ \kappa q d_1 d_2}{1 - \kappa q } } - d_1$$ on Outcome 1. This solution was also first discovered by Isaacs ([@isaacs]).[^16] The idea behind (\[atom bets out1 soln alone\]) is then clear: If the atomic player cannot afford to bet (\[isaacs wager outcome 1\]) on Outcome 1, she instead wagers as much as she can. This seems reasonable, intuitively, since up to (\[isaacs wager outcome 1\]), the positive impact of raising her Outcome 1 bet on her expected profit should outweigh the negative impact. The interpretation of \[atom bet prop orig ii\] is similar. We present the formal proof of Proposition \[atom equilib bets prop\] in Appendix \[prop 2 proof app\]. \[main thm\] A pure-strategy Nash equilibrium exists if and only if $\kappa > 0.5$.[^17] When an equilibrium exists, it is unique. The connection between low transaction costs (or large $\kappa$) and the existence of non-trivial equilibria has been observed in other parimutuel wagering studies. For example, a non-trivial equilibrium exists in Watanabe’s model only if the house take is sufficiently small ([@Takahiro199785]). Ottaviani and Sørenson make a similar observation ([@ottaviani2006timing]). Intuitively, $\kappa > 0.5$ is necessary for the existence of an equilibrium in our framework because there are only two possible outcomes. Suppose that $\left( f^{\star} , a^{\star} \right)$ is a pure-strategy Nash equilibrium. Regardless of her type, if a player believes that Outcome 1 will occur with probability $p$ and wagers on Outcome 1, it should be true that her final expected profit per unit bet on Outcome 1 is positive: $$\displaystyle\frac{ \kappa \, p }{ P^{f^{\star}, a^{\star}}} - 1 > 0 .$$ Similarly, if she wagers on Outcome 2, then $$\displaystyle\frac{ \kappa \, \left( 1 - p \right) }{ 1 - P^{f^{\star} , a^{\star} } } - 1 > 0.$$ Positive amounts are wagered on both outcomes (see Section \[model sect\]), implying that $$\label{both profitable ineq} \displaystyle\frac{ \kappa }{ P^{f^{\star}, a^{\star}}} - 1 > 0 \qquad \text{and} \qquad \displaystyle\frac{ \kappa }{ 1 - P^{f^{\star} , a^{\star} } } - 1 > 0 .$$ Rearranging (\[both profitable ineq\]) shows that $\kappa > 0.5$. More generally, it is easy to see that in a parimutuel betting game with $n$ outcomes and risk-neutral players who can elect not to bet, $\kappa > 1/n$ is necessary for the existence of an equilibrium. Parimutuel wagering games in the literature occasionally possess multiple equilibria. For instance, Watanabe et al. adapt the work of Harsanyi and Selten to select one equilibrium out of several that arise in their atomic player model ([@wat+non+mori]). As mentioned in Section \[model sect\], Watanabe’s continuum game model can feature multiple equilibria when $\mu \left( \left\{ p \right\} \right) > 0$ for some fixed $p$. We suspect that our equilibrium would no longer be unique, if we incorporated additional atomic players or relaxed our assumption that $\mu$ had a density, but we leave this issue to a future study. We break our proof into \[step 5 orig\] steps. Step \[step 1 orig\] allows us to use Propositions \[diff equilib bets prop\] and \[atom equilib bets prop\]. It says that in an equilibrium, the total amount wagered by the diffuse players on each outcome is always positive. Step \[step 2 orig\] formalizes our discussion above and shows that $\kappa > 0.5$ is necessary for the existence of an equilibrium. To finish, we find an equivalent formulation of our original problem. More precisely, after presenting some preliminary notation in Steps \[step 3 alt 2\] - \[step 4 alt 2\], we define our so-called [implied probability map]{} $\varphi$ in Step \[step 5 alt 2\]. Our work in Steps \[step 6 alt varphi dec\] - \[step 6 uniq fp\] shows that this map has a unique fixed-point. Due to its construction, its fixed-point corresponds to a pure-strategy Nash equilibrium and vice versa. We can then conclude that our game has a unique equilibrium when $\kappa > 0.5$ in Steps \[step 4 orig\] - \[step 5 orig\]. Our approach is motivated by the following observation: an equilibrium is essentially determined by the implied probability of Outcome 1. Given this quantity, we immediately recover the diffuse players’ wagers from Proposition \[diff equilib bets prop\]. Technically, we do not know how the diffuse players whose beliefs are indexed by $P^{f^{\star}, a^{\star}}/ \kappa$ or $\left( 1 - P^{f^{\star}, a^{\star}}\right) / \kappa$ behave, but this does not matter. By Proposition \[atom equilib bets prop\], we then identify the atomic player’s wagers. This observation leads to the definition of $\varphi$ in Step \[step 5 alt 2\]. In a certain sense, our recipe is only meaningful at a fixed-point of $\varphi$, which underlies the correspondence just discussed. Our complete proof of Theorem \[main thm\] can be found in Appendix \[main thm proof app\]. We close Section \[main result sect\] with Corollaries \[atom pl exp prof pos cor\], \[reg equilib cor\], and \[impl prob kappa 0.5\]. The first result says that the atomic player wagers on a particular outcome if and only if her final expected profit per unit bet on that outcome is positive. This is fairly obvious from \[NE def orig iii\] of Definition \[maj min NE def\], and we basically assume this to be true during our discussion of Theorem \[main thm\]. Still, recall that Proposition \[atom equilib bets prop\] identifies the outcome, if any, on which the atomic player wagers based upon only the diffuse players’ bets. For instance, according to that result, the atomic player wagers on Outcome 1 in an equilibrium $\left( f^{\star} , a^{\star} \right)$ if and only if $$\label{at bet o1 in equlib star} q > \displaystyle\frac{ d_1^{\star} }{\kappa \left( d_1^{\star} + d_2^{\star} \right)} .$$ Since $a_1^{\star} > 0$, we might be concerned about the possibility that $$\frac{P^{f^{\star},a^{\star}}}{\kappa} \geq q > \displaystyle\frac{ d_1^{\star} }{\kappa \left( d_1^{\star} + d_2^{\star} \right)} .$$ The specific form of $a_1^{\star}$ in (\[atom bets out1 soln alone\]) ultimately prevents this. Notice that Proposition \[diff equilib bets prop\] already [almost]{} implies the corresponding result for diffuse players. We say [almost]{} because of the undetermined behavior of the diffuse bettors whose final expected profit per unit bet on some outcome is 0. From that perspective, the atomic and diffuse players identify profitable wagering opportunities using the same criteria. Strategically, they just differ in how they size their equilibrium wagers. \[atom pl exp prof pos cor\] Suppose that $\left( f^{\star} , a^{\star} \right)$ is an equilibrium. Then $a^{\star}_1 > 0$ if and only if $$\displaystyle\frac{ \kappa q }{ P^{f^{\star}, a^{\star}}} - 1 > 0 ,$$ while $a^{\star}_2 > 0$ if and only if $$\displaystyle\frac{ \kappa \left( 1 - q \right) }{ 1 - P^{f^{\star}, a^{\star}} } -1 >0 .$$ We present the argument for the first case. The other is similar. Recall that $d_1^{\star}$, $d_2^{\star} >0$ (see Step \[step 1 orig\] of Theorem \[main thm\]’s proof). Assume that $a_1^{\star} > 0$. According to Proposition \[atom equilib bets prop\], (\[atom bets out1 ineq alone\]) holds and $a_2^{\star} = 0$. Using the notation from Step \[step 4 alt 2\] of Theorem \[main thm\]’s proof, $$P^{f^{\star}, a^{\star}} \leq \displaystyle\frac{ \zeta_1 \left( P^{f^{\star}, a^{\star}}\right) + d_1^{\star} }{ \zeta_1 \left( P^{f^{\star}, a^{\star}}\right) + d_1^{\star} + d_2^{\star} } .$$ That $$\label{atom exp prof pub pos ineq} \displaystyle\frac{ \kappa q }{ P^{f^{\star}, a^{\star}}} - 1 > 0$$ follows from Isaacs’ work ([@isaacs]). To prove the remaining direction, assume that (\[atom exp prof pub pos ineq\]) is satisfied. Exactly one of (\[atom bets out1 ineq alone\]), (\[atom bets out2 ineq alone\]), and (\[atom bets noth ineq alone\]) holds. We cannot have (\[atom bets out2 ineq alone\]), since it would follow that $a_2^{\star} > 0$. Arguing as we just did, we would get $$\displaystyle\frac{ \kappa \left( 1 - q \right) }{ 1 - P^{f^{\star}, a^{\star}} } -1 >0 ,$$ which would lead to $$\displaystyle\frac{ \kappa q }{ P^{f^{\star}, a^{\star}}} - 1 < 0.$$ (\[atom bets noth ineq alone\]) cannot hold either, as this would give the contradiction $$q \leq \displaystyle\frac{ d_1^{\star} }{ \kappa \left( d_1^{\star} + d_2^{\star} \right) } = \frac{P^{f^{\star},a^{\star}} }{\kappa}.$$ Hence, (\[atom bets out1 ineq alone\]) holds and $a_1^{\star} > 0$. To explain our next result, suppose that some player is wagering on a particular outcome. If another player believes that this outcome will occur with higher probability than the original player, Corollary \[reg equilib cor\] says that the new player also wagers on the outcome. Recall from Section \[lit review sect\] that Watanabe called an equilibrium with this property [regular]{} ([@Takahiro199785]). All equilibria in Watanabe’s framework and Ottaviani and Sørenson’s framework are regular ([@ottaviani2006timing]; [@Takahiro199785]). One might suspect that such a result generally holds, but this is not the case ([@wat+non+mori]). To the best of our knowledge, regularity has only been consistently observed in the literature when each player’s initial wealth is negligible ([@ottaviani2006timing]; [@Takahiro199785]). Corollary \[reg equilib cor\] shows that our model is an example of a parimutuel wagering game in which the equilibrium is regular, even when an atomic player is active. It is an immediate consequence of our observation that both atomic and diffuse players decide to wager on some outcome by determining whether or not the final expected profit per unit bet on that outcome is positive. \[reg equilib cor\] An equilibrium $\left( f^{\star} , a^{\star} \right)$ is always regular. Since $d_1^{\star}$, $d_2^{\star} >0$ by Step \[step 1 orig\] of Theorem \[main thm\]’s proof, the result directly follows from Proposition \[diff equilib bets prop\] and Corollary \[atom pl exp prof pos cor\]. Section \[main result sect\]’s last result says that the implied probability of Outcome 1 tends to $ 0.5$ as the house take approaches $ 0.5$, regardless of the other parameters that we choose for our model. In fact, the convergence is uniform. Initially, this finding may appear rather odd. For example, it is easy to ensure that $P^{f^{\star},a^{\star}}$ lies between $49.9\%$ and $50.1\%$ when $q= 0$ and the $\mu$-mass of $\left[ 0 , 1 \right]$ is [almost]{} entirely concentrated near $p = 0$. If essentially the whole population believes that Outcome 2 is guaranteed to occur, how can the total amounts wagered on each outcome be roughly equal? Our discussion of Theorem \[main thm\] outlines the key intuition. Simply notice that instead of rearranging (\[both profitable ineq\]) to show that $\kappa > 0.5$, we can show that $P^{f^{\star}, a^{\star}} \in \left( 1 - \kappa , \kappa \right)$. This holds even in the extreme scenario where virtually all of the initial wealth is held by those who believe that Outcome 2 is a sure bet. Still, our first instinct has some merit: Here, $P^{f^{\star}, a^{\star} } \approx 1 - \kappa$ for all $\kappa$ (see Appendix \[corr app\]). \[impl prob kappa 0.5\] Fix $\mu$, $q$, and $w$ and consider the map defined on $\left( 0.5 , 1 \right)$ by $$\kappa \mapsto P^{f^{\star}, a^{\star}} .$$ As $\kappa \downarrow 0.5$, the values of the map approach $ 0.5$. Simply notice that the map is well-defined by Theorem \[main thm\] and that $P^{f^{\star}, a^{\star}} \in \left( 1 - \kappa , \kappa \right)$ by Step \[step 5 orig\] of its proof. Numerical Results {#leb meas sect} ================= The theoretical results from Section \[main result sect\] allow us to return to our central question: How do large-scale participants in parimutuel wagering events affect the house and ordinary bettors? We explore this issue by analyzing several concrete scenarios (see Examples \[exmp 1\] - \[exmp 4\]). We use the house’s revenue to quantify the atomic player’s impact on the house. Given an equilibrium $\left( f^{\star} , a^{\star} \right)$, the house’s revenue is simply the product of the house take and the total amount wagered: $$\label{house rev eqn} \left( 1 - \kappa \right) \left( d_1^{\star} + d_2^{\star} + a_1^{\star} + a_2^{\star} \right) .$$ Notice that this quantity is deterministic and does not depend on the [actual]{} probability of Outcome 1, as the house collects (\[house rev eqn\]) regardless of which outcome occurs. To quantify the atomic player’s effect on diffuse bettors, we use one of two quantities. In Examples \[exmp 1\] - \[exmp 2\], we select values for the [actual]{} probability of Outcome 1. Making this choice lets us compute the [actual]{} total expected profit of the diffuse players. If $\left( f^{\star} , a^{\star} \right)$ is an equilibrium and the [actual]{} probability of Outcome 1 is $\bar{p}$, then it is given by $$\label{act exp prof db eqn} d_1^{\star} \bigg( \displaystyle\frac{ \kappa \, \bar{p} }{ P^{f^{\star}, a^{\star}}} - 1 \bigg) + d_2^{\star} \left( \displaystyle\frac{ \kappa \, \left( 1 - \bar{p} \right) }{ 1 - P^{f^{\star}, a^{\star}} } - 1 \right) .$$ While we use (\[act exp prof db eqn\]) to describe how the atomic player affects the diffuse players in Examples \[exmp 1\] - \[exmp 2\], we cannot do so in Examples \[exmp 3\] - \[exmp 4\]. The reason is that we make no assumption about the [actual]{} probability of Outcome 1 in the latter situations. Instead, we quantify the impact on diffuse bettors using their total [subjective]{} expected profit, which is given by $$\label{subj exp prof db eqn} \displaystyle\int_{0}^{1} \left[ f_1^{\star} \left( p \right) \bigg( \displaystyle\frac{ \kappa \, p }{ P^{f^{\star}, a^{\star}}} - 1 \bigg) + f_2^{\star} \left( p \right) \left( \displaystyle\frac{ \kappa \, \left( 1 - p \right) }{ 1 - P^{f^{\star}, a^{\star}} } - 1 \right) \right] d \mu \left( p \right)$$ in an equilibrium $\left( f^{\star} , a^{\star} \right)$. Here, we merely compute each diffuse player’s expected profit according to her beliefs and aggregate the results over all diffuse players. From Section \[intro sect\], recall that the standard narrative says that the presence of the atomic player should increase the house’s revenue but decrease the diffuse players’ total expected profit ([actual]{} or [subjective]{}, as applicable). The eventual decline in the house’s revenue should only be seen over time, not in our static game model. Technically, since we specified that $w > 0$ in Section \[model sect\], the atomic player can never be absent in our framework. Still, we can model the atomic player’s absence by choosing an extremely low value for $w$, say $w = 10^{-10}$. Her wagers are then too small to materially affect any equilibrium quantities. We do this for all of the Case 1’s in Examiples \[exmp 1\] - \[exmp 4\]. The atomic player is present, that is, $w > > 0$, in all of our upcoming Case 2’s. The following collection of measures is convenient for our purposes. \[dens fxn exmp defn\] For $n \geq 1$, let $\mu_n$ be the Borel measure on $\left[ 0 , 1 \right]$ whose density $g_n$ is defined by $$\begin{aligned} g_n \left( p \right) &= \left\{ \begin{array}{cc} -2 n \left( n - 1 \right) p + 2 \left( n -1 \right) + 1/n & \quad \text{if } \, \, p < 1/n \\ 1/n & \quad \text{if } \, \, p \geq 1/n \\ \end{array} \right. .\end{aligned}$$ In Figure \[fig: gn varying n plot\], we give the plots of $g_n$ for $n = 1$, 3, and 9. Here are the key observations: 1. \[gn cont pos\] $g_n$ is continuous and positive on $\left[ 0 , 1 \right]$. 2. \[un meas 1\] $\mu_n \left( \left[ 0 , 1 \right] \right) = 1$ for all $n$. 3. \[mu1 is Leb\] $\mu_1$ is the Lebesgue measure on $\left[ 0 , 1 \right]$. 4. \[gn conv Dirac\] $g_n$ converges in distribution to the Dirac delta function as $n \uparrow \infty$. \[gn cont pos\] and \[un meas 1\] ensure that $\mu_n$ is a suitable candidate for the measure describing the initial wealth of the diffuse bettors, that is, all of our results apply when $\mu = \mu_n$. In this case, \[mu1 is Leb\] says that the initial wealth of the diffuse bettors is uniformly distributed when $n =1$. \[gn conv Dirac\] says that their initial wealth is increasingly concentrated among those who believe that Outcome 2 will occur with high probability as $n$ increases. Another key feature is that the diffuse bettors’ total initial wealth is always equal to 1 (see \[un meas 1\]). ![[]{data-label="fig: gn varying n plot"}](nWedgeO2FxnPlotVaryN_FINAL) Before we proceed, we remark that all of our figures are generated with the help of the ideas in Theorem \[main thm\]’s proof. More precisely, assume that $\kappa \in \left( 0.5 , 1 \right)$. The function $\varphi$ defined in Step \[step 5 alt 2\] has a unique fixed-point. Since $\varphi$ is also continuous and decreasing, we can efficiently approximate this value with arbitrary precision using binary search. Step \[step 4 orig\] shows how to reconstruct the pure-strategy Nash equilibrium $\left( f^{\star} , a^{\star} \right)$, given such an estimate. \[exmp 1\] We now show that both effects predicted by the usual narrative can be observed. In Cases 1 and 2, we set $\mu = \mu_1$ and $q = 0.9$. We assume that the atomic player’s beliefs are exactly correct, i.e., the [actual]{} probability that Outcome 1 will occur is also 0.9. The only difference between Cases 1 and 2 is that $w = 10^{-10}$ in the former but $w = 1$ in the latter. Recall that choosing $\mu = \mu_1$ means that the diffuse bettors’ initial wealth is uniformly distributed. Since $q = 0.9$, the atomic player (correctly) believes that Outcome 1 is quite likely. In Figure \[fig: Diff Pl EP Exmp 1\], we plot the diffuse players’ [actual]{} total expected profit. Collectively, the diffuse players beliefs are rather inaccurate, so it is not surprising that their expect profit is negative. Still, as $\kappa \uparrow 1$, the diffuse players become increasingly worse off in Case 2. Intuitively, the atomic player quickly raises her bet on Outcome 1 as $\kappa \uparrow 1$. The implied probability of Outcome 1 rises,[^18] causing more diffuse players to bet on Outcome 2 and less to bet on Outcome 1. Since the [actual]{} probability of Outcome 1 is 0.9, this transition negatively affects the diffuse players. We plot the house’s revenue in Figure \[fig: House Rev Exmp 1\]. The house’s revenue is higher in Case 2 than in Case 1 for all $\kappa \in \left( 0.5, 1 \right)$, as a result of the higher wagering totals in Case 2. ![[]{data-label="fig: House Rev Exmp 1"}](DiffPlay_ACTUALTotExpProf_FxnKappa_DiffAtom_EXMP1_FINAL) ![[]{data-label="fig: House Rev Exmp 1"}](HouseRev_FxnKappa_DiffAtom_EXMP1_FINAL) \[exmp 2\] Still, there are cases in which diffuse players are positively affected by the activities of the atomic player. We now choose $\mu = \mu_1$ and $q = 0.57$ for Cases 1 and 2. The [actual]{} probability of Outcome 1 is 0.47. The only difference between the two scenarios is that $w = 10^{-10}$ in Case 1, while $w = 1$ in Case 2. Compared to Example \[exmp 1\], Outcome 1 is slightly less likely here. Also, we still assume that the atomic player’s forecast is quite accurate, but her prediction is no longer perfect. We plot the diffuse players’ [actual]{} expected profit in Figure \[fig: Diff Pl EP Exmp 2\]. The graphs for Cases 1 and 2 are the same for most values of $\kappa$; however, we see that the diffuse players’ expected profit is [higher]{} in Case 2 when $\kappa$ is large enough. Roughly, the atomic player does not start betting (on Outcome 1) until the house take is low, since she is nearly ambivalent. This explains why the curves are initially identical. When the atomic player begins to wager on Outcome 1, the diffuse players reshuffle their bets as in Example \[exmp 1\].[^19] The shift benefits them, essentially because the atomic player wagers on the [wrong]{} outcome. In Figure \[fig: House Rev Exmp 2\], we plot the house’s revenue. We see an extremely small improvement in Case 2 when $\kappa$ is large, but the graphs are almost indistinguishable, visually. As in Example \[exmp 1\], the increase corresponds to the increase in the total amount wagered. It is slight, as the atomic player’s uncertainty about what will occur causes her to bet very little in Case 2, even for $\kappa \approx 1$. ![[]{data-label="fig: House Rev Exmp 2"}](DiffPlay_ACTUALTotExpProf_FxnKappa_DiffAtom_EXMP2_FINAL) ![[]{data-label="fig: House Rev Exmp 2"}](HouseRev_FxnKappa_DiffAtom_EXMP2_FINAL) \[exmp 3\] We can argue that the diffuse players are better off in the presence of the atomic player, even without making assumptions about the [actual]{} probability of Outcome 1. In Cases 1 and 2, we now choose $\mu = \mu_{10}$ and $q = 0.95$. As in Examples \[exmp 1\] - \[exmp 2\], we set $w = 10^{-10}$ in Case 1 and $w = 1$ in Case 2. The atomic player believes that Outcome 1 is highly likely. Collectively, the diffuse players believe that Outcome 2 will probably happen: over $90\%$ of their initial wealth is held by those who believe that the probability of Outcome 2 is at least 0.9. We make no judgment about the accuracy of the players’ beliefs. We plot the diffuse players’ [subjective]{} expected profit in Figure \[fig: Diff Pl EP Exmp 3\]. The graphs are the same for low $\kappa$, but eventually, the diffuse players’ [subjective]{} expected profit is much higher in Case 2. Intuitively, the atomic player raises her wager on Outcome 1 as $\kappa \uparrow 1$, since she believes that Outcome 1 will occur. The implied probability of Outcome 1 then rises, a boon to those who believe that Outcome 2 will occur. This includes [most]{} of the diffuse players. In Figure \[fig: House Rev Exmp 3\], we plot the house’s revenue. The house’s revenue in Case 2 is at least as large as its revenue in Case 1 for all values of $\kappa$. Often, the improvement is significant. Roughly, there is too much agreement among the diffuse bettors. They wager more in the presence of the atomic player, [believing]{} that they can profit from her [supposed]{} wagering mistakes. The pool size increases, leading to greater revenue for the house in Case 2. ![[]{data-label="fig: House Rev Exmp 3"}](DiffPlay_TotExpProf_FxnKappa_DiffAtom_EXMP3_FINAL) ![[]{data-label="fig: House Rev Exmp 3"}](HouseRev_FxnKappa_DiffAtom_EXMP3_FINAL) \[exmp 4\] In keeping with the standard narrative, it seems like the house is always immediately better off because of the atomic player. We can cast some doubt on this too. We used the same $\mu$ for Cases 1 and 2 in Examples \[exmp 1\] - \[exmp 3\]. This choice captures the idea that the presence of the atomic player should not affect the diffuse players’ initial wealth in a static parimutuel wagering game.[^20] Alternatively, it might be reasonable to fix the distribution of initial wealth across the [entire]{} population, not just the diffuse population. For instance, we could specify that the amount held by those who believe that Outcome 1 will occur is roughly the same as the amount held by those who believe that Outcome 2 will occur. We could then compare the situations in which the wealth is held by only diffuse players and in which some wealth is held by the atomic player. This is the approach we now take. In Case 1, $w = 10^{-10}$ and the density of $\mu$ is defined by $$p \mapsto \frac{ g_{100} \left(p \right) + g_{100} \left(1 -p \right) }{2} .$$ For Case 2, we set $w = 1$ and $\mu = \mu_{100}$. We choose $q = 1$ for both scenarios and, again, make no assumption about Outcome 1’s [actual]{} probability. Intuitively, [half]{} of the diffuse players in Case 1 believe that Outcome 1 is going to occur, while the other [half]{} believes that Outcome 2 will occur. For Case 2, the diffuse players [all]{} believe that Outcome 2 is going to occur. The atomic player, whose wealth is equal to the collective wealth of the diffuse players, believes that Outcome 1 will occur. Notice that the diffuse players’ total initial wealth is equal in Cases 1 and 2; however, the wealth of the entire population in Case 2 is twice what it is in Case 1. This is consistent with our parameter selections in Examples \[exmp 1\] - \[exmp 3\]. When studying the house’s revenue here, we might have also chosen $\mu = 0.5 \times \mu_{100} $ and $w = 0.5$ in Case 2, ensuring that the wealth of the entire population is identical in both scenarios. We comment on this shortly. One might suspect that Cases 1 and 2 are quite similar, but this is not true. In Figure \[fig: Diff Pl EP Exmp 4\], we plot the diffuse players’ [subjective]{} expected profit. For $\kappa \approx 1$, their [subjective]{} expected profit is higher in Case 2 than in Case 1. Otherwise, it is lower for all $\kappa$, and often, the drop is significant. To explain this observation, we plot the implied probability of Outcome 1 in Figure \[fig: Impl Prob Exmp 4\]. The implied probability of Outcome 1 is about 0.5 for all $\kappa$ in Case 1. For Case 2, it is almost $1 - \kappa$ for low $\kappa$ but approaches $0.5$ as $\kappa \uparrow 1$. Since the diffuse players believe that Outcome 2 will occur but the atomic player believes that Outcome 1 will occur, the intuition appears to be as follows. Roughly, the diffuse players are willing to tolerate unfavorable betting conditions more than the atomic player. Despite the fact that each type of player is [basically]{} sure that the outcome they are betting on will occur, the diffuse players in Case 2 raise the size of their wagers much faster than the atomic player when $\kappa$ is low. The diffuse players do this at the expense of their own [subjective]{} expected profit, which is nearly zero in Case 2 until $\kappa \approx 0.84$. The atomic player does not substantially raise her wager on Outcome 1 until the values of the implied probability of Outcome 1 and $\kappa$ make her [subjective]{} expected profit per unit bet very high. One could argue that the atomic player’s strategy is superior to the strategies employed by the diffuse players, although we have made no assumption about the accuracy of her prediction. Perhaps the reason is that unlike the diffuse players, she considers her individual impact on the other wagerers because of her substantial wealth. In Figure \[fig: House Rev Exmp 4\], we plot the house’s revenue. We see that the house’s revenue is higher in Case 1 for small $\kappa$ but higher in Case 2 for large $\kappa$. After selecting $\mu = 0.5 \times \mu_{100} $ and $w = 0.5$ in Case 2 (see our explanation above), we re-plot the house’s revenue in Figure \[fig: Norm House Rev Exmp 4\]. Now the house’s revenue is higher in Case 1 for all $\kappa$. Regardless of our normalization, the key point is that the house’s maximum revenue is always much higher in Case 1. Intuitively, the diffuse players appear to have a greater tolerance for poor betting conditions than the atomic player, as discussed previously. They are willing to bet even when $\kappa \approx 0.5$, and consequently, about $99\%$ of the diffuse players are wagering in Case 1 when $\kappa \approx 0.51$. The house loses revenue by raising $\kappa$, since the entire population has already wagered [almost]{} everything that it can. In Case 2, the atomic player’s reluctance to significantly raise her wager on Outcome 1 until betting conditions improve means that the total amount wagered is low for most $\kappa$. The pool is large only if $\kappa$ is quite high, so the house does not collect much. In Examples \[exmp 1\] - \[exmp 3\], the house’s revenue in Case 2 is at least as great as the house’s revenue in Case 1 for all $\kappa$. Of course, the house’s maximum revenue is then higher in Case 2 for these studies. Here, the pointwise analysis is not as clean, leading us to more directly consider the house’s strategic behavior. Thus far, we have understood the house take to be exogenously determined. One way to relax this assumption is to break our game into two stages. In the first stage, the betting organizer chooses a value for the house take in order to maximize her revenue. The second stage is identical to our current setup. Under the new framework, the house needs to account for more than the distribution of initial wealth across the entire population. The house must also consider how the initial wealth is distributed across the population [for each type of player]{}. Independent of our normalization of the entire population’s wealth, the house’s revenue is maximized at $\kappa \approx 0.506$ in Case 1 and at $\kappa \approx 0.839$ in Case 2. Hence, the house selects these values of $\kappa$ in Cases 1 and 2, respectively. In Figure \[fig: Diff Pl EP Exmp 4\], the diffuse players’ [subjective]{} expected profit is about 0.0085 when $\kappa \approx 0.506$ in Case 2, while it is about 0.023 when $\kappa \approx 0.839$ in Case 2. These numbers arise from our original parameter choices, ensuring that the total initial wealth of the diffuse players is 1 in both scenarios. The diffuse players are not well off in either Case 1 or Case 2. We could still argue that the setup of Case 2 is to their advantage, in contrast to our earlier pointwise analysis. Roughly, because the atomic player is less willing to bet under poor wagering conditions, the house more easily preys upon the diffuse players when she is absent. ![[]{data-label="fig: Impl Prob Exmp 4"}](DiffPlay_TotExpProf_FxnKappa_DiffAtom_EXMP4_FINAL) ![[]{data-label="fig: Impl Prob Exmp 4"}](P1star_FxnKappa_DiffAtom_EXMP4_FINAL) ![[]{data-label="fig: Norm House Rev Exmp 4"}](HouseRev_FxnKappa_DiffAtom_EXMP4_FINAL) ![[]{data-label="fig: Norm House Rev Exmp 4"}](RESCALED_HouseRev_FxnKappa_DiffAtom_EXMP4_FINAL) Properties of $P^{f^{\star}, a^{\star}}$ {#corr app} ======================================== We briefly revisit our discussion of Corollary \[impl prob kappa 0.5\] by studying the properties of $P^{f^{\star}, a^{\star}}$ as a function of $\kappa$. We hope this highlights a few theoretical features of our model, though we do not relate our findings in Appendix \[corr app\] to our central question. To generate these figures, we use the approach and notation of Section \[leb meas sect\]. Recall our claim that in an extreme case where nearly all of the initial wealth is held by those who believe that Outcome 2 is highly likely, $P^{f^{\star}, a^{\star} } \approx 1- \kappa$ for all $\kappa$. We illustrate this in Figure \[fig: P1 star kappa All O2\], which depicts the map $\kappa \mapsto P^{f^{\star}, a^{\star} }$ when $\mu = \mu_{100}$, $q = 0$, and $w = 1$. Here, over 99% of the diffuse players’ initial wealth belongs to those who believe that the probability of Outcome 2 is at least 0.99. The atomic player believes that Outcome 2 is guaranteed to occur. As expected, the plot is visually indistinguishable from a plot of the map $\kappa \mapsto 1- \kappa$. Often, it is more difficult to broadly describe attributes of the function $\kappa \mapsto P^{f^{\star}, a^{\star}}$. We know that $P^{f^{\star}, a^{\star}} \in \left( 1 - \kappa, \kappa \right)$, but our players’ heterogeneity allows for a wide range of possibilities within these bounds. There is a rich interplay between their differing beliefs, effects on (\[payouts per unit bet words\]), and wealth constraints. By plotting the map $\kappa \mapsto P^{f^{\star}, a^{\star}}$ under varying assumptions on $q$, $w$, and $\mu$, Figure \[fig: Wild P1star graph\] displays a few of the myriad possibilities. For instance, the density $g$ of the $\mu$ used to generate Line $A$ is a linear combination of Gaussian densities with different means. Line $A$’s oscillations arise because of $g$’s distinct peaks. ![[]{data-label="fig: Wild P1star graph"}](P1star_FxnKappa_AllO2) ![[]{data-label="fig: Wild P1star graph"}](P1star_FxnKappa_WildP1star_FINAL) Proof of Proposition \[atom equilib bets prop\] {#prop 2 proof app} =============================================== We only prove (i). The argument for (ii) is similar, and as observed above, \[atom bet prop orig iii\] is due to Isaacs ([@isaacs]). Suppose that (\[atom bets out1 ineq alone\]) holds. Then $$\label{key isaacs ineq for prop} \displaystyle\frac{1-q}{d_2} < \displaystyle\frac{1- \kappa q}{d_2} < \displaystyle\frac{1}{d_1 + d_2} < \displaystyle\frac{1}{\kappa \left( d_1 + d_2 \right) } < \displaystyle\frac{q}{d_1}.$$ We get the first and third inequalities because $\kappa \in \left( 0 , 1\right)$. The second inequality is due to (\[1 bc impli\]), while the last inequality is a rearrangement of (\[atom bets out1 ineq alone\]). For now, the critical observation is that the leftmost quantity is less than the rightmost, allowing us to use the work of Isaacs ([@isaacs]). Define the map $$\Phi : \mathbb{R}^2_{\geq 0} \longrightarrow \mathbb{R}$$ by $$\begin{aligned} \Phi \left( b_1 , b_2 \right) &= b_1 \left( \displaystyle\frac{\kappa \left( b_1 + d_1 + b_2 + d_2 \right) q}{b_1 + d_1} - 1 \right) + b_2 \left( \displaystyle\frac{\kappa \left( b_1 + d_1 + b_2 + d_2 \right) \left( 1 - q \right) }{b_2 + d_2} - 1 \right) . \end{aligned}$$ $\Phi \left( b_1 , b_2 \right)$ is the atomic player’s expected profit, given that she wagers $b_i$ on Outcome $i$, and it allows us to rewrite (\[atom max prob eqn prop\]) as $$a_1 \bigg( \displaystyle\frac{ \kappa \, q }{ P^{f , a }} - 1 \bigg) + a_2 \left( \displaystyle\frac{ \kappa \, \left( 1 - q\right) }{ 1 - P^{f , a } } - 1 \right) = \displaystyle\sup_{\substack{ b_1 , b_2 \geq 0 \\ b_1 + b_2 \leq w }} \Phi \left( b_1 , b_2 \right) .$$ Technically, strategy profiles $\left( b_1 , b_2 \right)$ with $$b_1 + b_2 > w$$ are not feasible, but this is unimportant. Isaacs showed that $\Phi$ has a unique global maximum on $\mathbb{R}^2_{\geq 0}$, denoted $\left( b_1^{\star} , b_2^{\star} \right)$, given by $$\left( b_1^{\star} , b_2^{\star} \right) = \left( \displaystyle\sqrt{ \displaystyle\frac{ \kappa q d_1 d_2}{1 - \kappa q } } - d_1 , 0 \right) .$$ This proves that (i) holds when $w \geq b_1^{\star}$. The case where $w < b_1^{\star}$ is handled with elementary calculus. First, observe that (\[atom bets out1 ineq alone\]) implies that $q > 0$ and $$\begin{aligned} \partial_{b_1 } \Phi \left( 0 , 0 \right) = \displaystyle\frac{ \kappa \left( d_1 + d_2 \right) q}{ d_1 } - 1 > 0 .\end{aligned}$$ Since $\partial_{b_1 } \Phi \left( b_1^{\star} , 0 \right) = 0$ and $$\partial_{b_1 b_1} \Phi \left( b_1 , b_2 \right) = - \displaystyle\frac{ 2 \kappa d_1 \left( b_2 + d_2 \right) q}{\left( b_1 + d_1 \right)^3}$$ is always negative, it follows that $\partial_{b_1} \Phi \left( b_1 , 0 \right) > 0$ for $0 \leq b_1 < b_1^{\star}$. The interpretation is that as long as the atomic player has not yet wagered on Outcome 2, she should wager as much as she can up to $b_1^{\star}$ on Outcome 1. We only need to argue that she should never wager on Outcome 2, that is, $\partial_{b_2} \left( b_1 , b_2 \right)$ is negative whenever $b_1 + b_2 \leq w$. Rearranging (\[key isaacs ineq for prop\]) implies that $$\partial_{b_2 } \Phi \left( 0, 0 \right) = \displaystyle\frac{ \kappa \left( d_1 + d_2 \right) \left( 1 - q \right)}{ d_2 } - 1 < 0 .$$ Now $\partial_{b_2 } \Phi \left( b_1^\star , 0 \right) \leq 0$[^21] and $$\partial_{b_2 b_1 } \Phi \left( b_1 , b_2 \right) = \displaystyle\frac{ \kappa d_1 q }{ \left( b_1 + d_1 \right)^2} + \displaystyle\frac{ \kappa d_2 \left( 1 - q \right) }{ \left( b_2 + d_2 \right)^2}$$ is positive everywhere, so $\partial_{b_2} \Phi \left( b_1 , 0 \right)< 0$ for $0 \leq b_1 < b_1^\star$. Since $$\partial_{b_2 b_2} \Phi \left( b_1 , b_2 \right) = - \displaystyle\frac{ 2 \kappa d_2 \left( b_1 + d_1 \right) \left( 1 - q \right)}{\left( b_2 + d_2 \right)^3}$$ is always non-positive, we conclude that $\partial_{b_2} \left( b_1 , b_2 \right)$ must be negative whenever $0 \leq b_1 < b_1^\star$. In particular, $\partial_{b_2} \left( b_1 , b_2 \right)$ is negative, if $ b_1 + b_2 \leq w$. Proof of Theorem \[main thm\] {#main thm proof app} ============================= [Step : If $\left(f^{\star}, a^{\star} \right)$ is an equilibrium, then $d_1^{\star}$, $d_2^{\star} > 0$. ]{} We show that the total amount wagered on each outcome by the diffuse players is positive in equilibrium: $d_1^{\star}$, $d_2^{\star} > 0$. We use this result to complete the proof of the [only if]{} direction of Theorem \[main thm\] in Step \[step 2 orig\]. It also allows us to use Propositions \[diff equilib bets prop\] and \[atom equilib bets prop\] in the proof of the [if]{} direction. Suppose instead that we can find a pure-strategy Nash equilibrium $\left( f^{\star} , a^{\star} \right)$ with $d_1^{\star} = 0$. It follows immediately that $d_2^{\star} > 0$: otherwise, \[NE def orig iii\] of Definition \[maj min NE def\] implies that $a_1^{\star}= a_2^{\star}= 0$, which contradicts \[NE def orig i\] of Definition \[maj min NE def\]. A consequence of Isaacs’ work is that $q = 0$ ([@isaacs]).[^22] Then \[NE def orig iii\] of Definition \[maj min NE def\] implies that $a_1^{\star}= a_2^{\star}= 0$. In particular, $P^{f^{\star}, a^{\star}} = 0$. By \[NE def orig i\] of Definition \[maj min NE def\], $f_2^{\star} \equiv 0$, which is impossible since $d_2^{\star} > 0$. Hence, $d_1^{\star} > 0$. It follows similarly that $d_2^{\star} > 0$. Step : If an equilibrium exists, then $\kappa > 0.5$. We finish the proof of the [only if]{} direction of Theorem \[main thm\] by formalizing the heuristics given previously. Suppose that $\left( f^{\star} , a^{\star} \right)$ is an equilibrium. Step \[step 1 orig\] implies that $d_1^{\star}$, $d_2^{\star} > 0$. Then both $$\label{Pstar in 1 min al al} \displaystyle\frac{P^{f^{\star}, a^{\star} }}{ \kappa } < 1 \qquad \text{and} \qquad \displaystyle\frac{1 - P^{f^{\star}, a^{\star} } }{ \kappa } < 1$$ by Proposition \[diff equilib bets prop\]. Rearranging (\[Pstar in 1 min al al\]) finishes the argument. Step : Definition and discussion of $\bar{p}_i$ (when $\kappa > 0.5$). Due to Step \[step 2 orig\], we assume that $\kappa > 0.5$ for the remainder of the proof (Steps \[step 3 alt 2\] - \[step 5 orig\]). This assumption ensures that our discussions are not vacuous, as we implicitly rely on the positive length of the interval $\left[ 1- \kappa, \kappa \right]$. We now define and discuss the quantities $\bar{p}_1$, $\bar{p}_2 \in \left[ 1- \kappa , \kappa \right]$. We use this notation when we define our $\zeta_i$ maps in Step \[step 4 alt 2\] and $\varphi$ in Step \[step 5 alt 2\]. Roughly, $\bar{p}_1$ is a naïve lower bound for the implied probability of Outcome 1 when the atomic player wagers on Outcome 1. Similarly, $\bar{p}_2$ is an upper bound for the implied probability of Outcome 1 when the atomic player wagers on Outcome 2. Both are derived from Propositions \[diff equilib bets prop\] and \[atom equilib bets prop\]. Since $\mu$’s density is positive, the map $$\label{out1 bc map} p \mapsto \displaystyle\frac{ \mu \left( \frac{ p }{\kappa} , 1 \right] }{ \kappa \left( \mu \left( \frac{ p }{\kappa} , 1 \right] + \mu \left[0, 1 - \frac{1 -p }{\kappa} \right) \right) }$$ is decreasing and continuous on $\left[ 1 - \kappa , \kappa \right]$. Its value is $1/ \kappa$ at $p = \left( 1- \kappa \right)$ and 0 at $p = \kappa$. Hence, there is a unique $\bar{p}_1 \in \left( 1 - \kappa , \kappa \right]$ such that $$\label{pbar1 defn eqn} q = \displaystyle\frac{ \mu \left( \frac{ \bar{p}_1 }{\kappa} , 1 \right] }{ \kappa \left( \mu \left( \frac{ \bar{p}_1 }{\kappa} , 1 \right] + \mu \left[0, 1 - \frac{1 - \bar{p}_1 }{\kappa} \right) \right) } .$$ Clearly, $\bar{p}_1 = \kappa$, or equivalently, $\left( \bar{p}_1 , \kappa \right]$ is empty, if and only if $q = 0$. Regardless of $q$’s value, we know $$\begin{aligned} \label{q ineq pbar1 ints} \left\{ \begin{array}{cc} q \leq \displaystyle\frac{ \mu \left( \frac{ p }{\kappa} , 1 \right] }{ \kappa \left( \mu \left( \frac{ p }{\kappa} , 1 \right] + \mu \left[0, 1 - \frac{1 -p }{\kappa} \right) \right) } \qquad & \qquad \text{if} \quad p \in \left[ 1- \kappa, \bar{p}_1 \right] \\ & \quad \\ q > \displaystyle\frac{ \mu \left( \frac{ p }{\kappa} , 1 \right] }{ \kappa \left( \mu \left( \frac{ p }{\kappa} , 1 \right] + \mu \left[0, 1 - \frac{1 -p }{\kappa} \right) \right) } \qquad & \qquad \text{if} \quad p \in \left( \bar{p}_1 , \kappa \right] \end{array} \right. .\end{aligned}$$ We later connect this observation to (\[atom bets out1 ineq alone\]) and (\[atom bets noth ineq alone\]) when defining $\zeta_i$ and $\varphi$. Similarly, the map $$\label{out2 bc map} p \mapsto \displaystyle\frac{ \mu \left[0, 1 - \frac{1 -p }{\kappa} \right) }{ \kappa \left( \mu \left( \frac{ p }{\kappa} , 1 \right] + \mu \left[0, 1 - \frac{1 -p }{\kappa} \right) \right) }$$ is increasing and continuous on $\left[ 1 - \kappa , \kappa \right]$. Its value is 0 at $p = \left( 1- \kappa \right)$ and $1/ \kappa$ at $p = \kappa$, so there is a unique $\bar{p}_2 \in \left[ 1 - \kappa , \kappa \right)$ such that $$1 - q = \displaystyle\frac{ \mu \left[0, 1 - \frac{1 - \bar{p}_2 }{\kappa} \right) }{ \kappa \left( \mu \left( \frac{ \bar{p}_2 }{\kappa} , 1 \right] + \mu \left[0, 1 - \frac{1 - \bar{p}_2 }{\kappa} \right) \right) } .$$ Now $\bar{p}_2 = \left( 1-\kappa \right)$, i.e., $\left[ 1- \kappa , \bar{p}_2 \right)$ is empty, if and only if $q = 1$. In any case, $$\begin{aligned} \label{q ineq pbar2 ints} \left\{ \begin{array}{cc} 1 - q > \displaystyle\frac{ \mu \left[0, 1 - \frac{1 -p }{\kappa} \right) }{ \kappa \left( \mu \left( \frac{ p }{\kappa} , 1 \right] + \mu \left[0, 1 - \frac{1 -p }{\kappa} \right) \right) } \qquad & \qquad \text{if} \quad p \in \left[ 1 - \kappa, \bar{p}_2 \right) \\ &\\ 1 - q \leq \displaystyle\frac{ \mu \left[0, 1 - \frac{1 -p }{\kappa} \right) }{ \kappa \left( \mu \left( \frac{ p }{\kappa} , 1 \right] + \mu \left[0, 1 - \frac{1 -p }{\kappa} \right) \right) } \qquad & \qquad \text{if} \quad p \in \left[ \bar{p}_2 , \kappa \right] \end{array} \right. .\end{aligned}$$ This comment relates to (\[atom bets out2 ineq alone\]) and (\[atom bets noth ineq alone\]), as reflected in our definitions of $\zeta_i$ and $\varphi$. We conclude by observing that we have $\bar{p}_2 < \bar{p}_1$ since $\kappa \in \left( 0.5 , 1 \right)$. This is another key remark for our future definition of $\varphi$. Step : Definition and discussion of $\zeta_i$ (when $\kappa > 0.5$). We define and discuss the functions $\zeta_1$ and $\zeta_2$. We use this notation in our definition of $\varphi$ in Step \[step 5 alt 2\]. Intuitively, $\zeta_i \left( p \right)$ represents the amount that the atomic player wagers on Outcome $i$ when she makes a positive wager on Outcome $i$, does not face a budget constraint, and the implied probability of Outcome 1 is $p$ (cf. (\[atom bets out1 soln alone\]) and (\[atom bets out2 soln alone\])). For $p \in \left[ \bar{p}_1 , \kappa \right]$, define $\zeta_1$ by $$\zeta_1 \left( p \right) = \sqrt{\frac{\kappa q}{1 - \kappa q} \, \mu \left( \frac{p}{\kappa} , 1 \right] \mu \left[0, 1 - \frac{1 -p}{\kappa} \right) } - \mu \left( \frac{p}{\kappa} , 1 \right] .$$ Since $\mu$ has a positive density, (\[q ineq pbar1 ints\]) implies that $$\mu \left( \frac{p}{\kappa} , 1 \right] > 0 , \qquad \mu \left[0, 1 - \frac{1 -p}{\kappa} \right) > 0 , \qquad \text{and} \qquad q > \displaystyle\frac{ \mu \left( \frac{ p }{\kappa} , 1 \right] }{ \kappa \left( \mu \left( \frac{ p }{\kappa} , 1 \right] + \mu \left[0, 1 - \frac{1 -p }{\kappa} \right) \right) }$$ for $p \in \left( \bar{p}_1 , \kappa \right)$. It follows as in (\[1 bc impli\]) that $\zeta_1$ is positive on $\left( \bar{p}_1 , \kappa \right)$. We also have $$\zeta_1 \left( \bar{p}_1 \right) = \zeta_1 \left( \kappa \right) = 0$$ from (\[pbar1 defn eqn\]). We define $\zeta_2$ for $p \in \left[ 1 - \kappa, \bar{p}_2 \right]$ by $$\zeta_2 \left( p \right) = \sqrt{\frac{\kappa \left( 1 - q\right) }{1 - \kappa \left( 1 - q \right)} \, \mu \left( \frac{p}{\kappa} , 1 \right] \mu \left[0, 1 - \frac{1 -p}{\kappa} \right) } - \mu \left[0, 1 - \frac{1 -p}{\kappa} \right) .$$ Using the techniques from our discussion of $\zeta_1$, we see that $\zeta_2$ is positive on $\left( 1 - \kappa, \bar{p}_2 \right)$ and $$\zeta_2 \left( 1- \kappa \right) = \zeta_2 \left( \bar{p}_2 \right) = 0 .$$ Step : Definition and discussion of $\varphi$ (when $\kappa > 0.5$). We introduce the [implied probability map]{} $\varphi$. We see in Steps \[step 4 orig\] and \[step 5 orig\] that a fixed-point of $\varphi$ corresponds to a pure-strategy Nash equilibrium in an obvious way and vice versa, which ultimately allows us to complete Theorem \[main thm\]’s proof. $\varphi$’s domain is the set of candidates $p$ for the implied probability of Outcome 1. We need only consider $p \in \left[ 1- \kappa, \kappa \right]$, as we observe in Step \[step 5 orig\]. Proposition \[diff equilib bets prop\] says that $$\begin{aligned} \label{diff bet examps step 3} \mu \left( \frac{p}{\kappa} , 1 \right] \qquad \text{and} \qquad \mu \left[0, 1 - \frac{1 -p}{\kappa} \right) \end{aligned}$$ are the total amounts wagered by the diffuse players on Outcomes 1 and 2, respectively. The atomic player’s wagers are described by Proposition \[atom equilib bets prop\]. For example, if $$q > \displaystyle\frac{ \mu \left( \frac{ p }{\kappa} , 1 \right] }{ \kappa \left( \mu \left( \frac{ p }{\kappa} , 1 \right] + \mu \left[0, 1 - \frac{1 -p }{\kappa} \right) \right) } ,$$ that is, $p \in \left( \bar{p}_1 , \kappa \right]$, then the atomic player wagers nothing on Outcome 2 and $$\displaystyle\min \left\{ w , \zeta_1 \left( p \right) \right\}$$ on Outcome 1. Recalculating the implied probability of Outcome 1 using these bets, we get $$\displaystyle\frac{ \displaystyle\min \left\{ w , \zeta_1 \left( p \right) \right\} + \mu \left( \frac{p}{\kappa} , 1 \right] }{ \displaystyle\min \left\{ w , \zeta_1 \left( p \right) \right\} + \mu \left( \frac{p}{\kappa} , 1 \right] + \mu \left[0, 1 - \frac{1 -p}{\kappa} \right) }$$ from Definition \[impl prob defn\]. We set $\varphi \left( p \right)$ to this value. In some sense, this procedure is only potentially meaningful when $p$ is equal to $\varphi \left( p \right)$, leading to our focus on fixed-points. Here is the complete definition of $\varphi$ suggested by this explanation: $$\begin{aligned} \varphi \left( p \right) = \left\{ \begin{array}{cc} \displaystyle\frac{\mu \left( \frac{p}{\kappa} , 1 \right] }{ \displaystyle\min \left\{ w , \zeta_2 \left( p \right) \right\} + \mu \left( \frac{p}{\kappa} , 1 \right] + \mu \left[0, 1 - \frac{1 -p}{\kappa} \right) } & \quad \text{if} \quad p \in \left[ 1 - \kappa, \bar{p}_2 \right) \\ &\\ \displaystyle\frac{ \mu \left( \frac{p}{\kappa} , 1 \right] }{ \mu \left( \frac{p}{\kappa} , 1 \right] + \mu \left[0, 1 - \frac{1 -p}{\kappa} \right) } & \quad \text{if} \quad p \in \left[ \bar{p}_2 , \bar{p}_1 \right] \\ &\\ \displaystyle\frac{ \displaystyle\min \left\{ w , \zeta_1 \left( p \right) \right\} + \mu \left( \frac{p}{\kappa} , 1 \right] }{ \displaystyle\min \left\{ w , \zeta_1 \left( p \right) \right\} + \mu \left( \frac{p}{\kappa} , 1 \right] + \mu \left[0, 1 - \frac{1 -p}{\kappa} \right) } & \quad \text{if} \quad p \in \left( \bar{p}_1 , \kappa \right] \end{array} \right. .\end{aligned}$$ Observe that $\varphi$ is continuous on $\left[ 1- \kappa , \kappa \right]$ since $\mu$ has a positive density and $\zeta_i \left( \bar{p}_i \right) = 0$ (see Step \[step 4 alt 2\]). This helps us prove that $\varphi$ has a unique fixed-point in Step \[step 6 uniq fp\]. Step : $\varphi$ is decreasing (when $\kappa > 0.5$). We show that $\varphi$ is decreasing. We use this property in Step \[step 6 uniq fp\] to argue that $\varphi$ has a unique fixed-point. Since $\mu$’s density is positive everywhere, $\varphi$ is decreasing on $\left[ \bar{p}_2 , \bar{p}_1 \right]$. It suffices to show that $\varphi$ is decreasing on both $\left[ 1 - \kappa, \bar{p}_2 \right)$ and $\left( \bar{p}_1 , \kappa \right]$ because $\varphi$ is continuous. The proofs are similar, so we consider the former case. We are done, if $\left[ 1 - \kappa, \bar{p}_2 \right)$ is empty. Suppose that it is not. Recall from Step \[step 3 alt 2\] that this is equivalent to assuming that $q < 1$. Define two functions $\varphi_2$ and $\varphi_2^w$ on $\left[ 1 - \kappa, \bar{p}_2 \right)$ by $$\begin{aligned} \varphi_2 \left( p \right) &= \displaystyle\frac{\mu \left( \frac{p}{\kappa} , 1 \right] }{ \zeta_2 \left( p \right) + \mu \left( \frac{p}{\kappa} , 1 \right] + \mu \left[0, 1 - \frac{1 -p}{\kappa} \right) } \\ \varphi_2^w \left( p \right) &= \displaystyle\frac{\mu \left( \frac{p}{\kappa} , 1 \right] }{ w+ \mu \left( \frac{p}{\kappa} , 1 \right] + \mu \left[0, 1 - \frac{1 -p}{\kappa} \right) } .\end{aligned}$$ The point is that $$\varphi \left( p \right) = \displaystyle\max \left\{ \varphi_2 \left( p \right) , \varphi_2^w \left( p \right) \right\}$$ on $\left[ 1 - \kappa, \bar{p}_2 \right)$, so it is enough to show that $\varphi_2$ and $\varphi_2^w$ are both decreasing. Clearly, $\varphi_2^w$ is decreasing. Denoting the positive and continuous density of $\mu$ by $g$, we see that $\varphi_2$ is decreasing because $$\begin{aligned} \varphi_{2}^{\prime} \left( p \right) &= - \displaystyle\frac{ g \left( \frac{p}{\kappa} \right) \sqrt{\frac{\kappa \left( 1 -q \right)}{1 - \kappa \left( 1 -q \right)} \mu \left( \frac{p }{\kappa} , 1 \right] \mu \left[0, 1 - \frac{1 -p}{\kappa} \right) } }{ 2\kappa \left( \sqrt{\frac{\kappa \left( 1 -q \right)}{1 - \kappa \left( 1 -q \right)} \mu \left( \frac{p }{\kappa} , 1 \right] \mu \left[0, 1 - \frac{1 -p}{\kappa} \right) }+ \mu \left( \frac{p }{\kappa} , 1 \right] \right)^2 } \\ &\qquad - \displaystyle\frac{ \left( \frac{\kappa \left( 1 -q \right)}{1 - \kappa \left( 1 -q \right)} \right)^{ 0.5} \mu \left( \frac{p}{\kappa} , 1 \right]^{3/2} \mu \left[0, 1 - \frac{1 -p}{\kappa} \right)^{- 0.5} g \left( 1- \frac{1- p}{\kappa} \right) }{ 2\kappa \left( \sqrt{\frac{\kappa \left( 1 -q \right)}{1 - \kappa \left( 1 -q \right)} \mu \left( \frac{p }{\kappa} , 1 \right] \mu \left[0, 1 - \frac{1 -p}{\kappa} \right) }+ \mu \left( \frac{p }{\kappa} , 1 \right] \right)^2 } . \end{aligned}$$ Step : $\varphi$ has a unique fixed-point (when $\kappa > 0.5$). We show that $\varphi$ has a unique fixed-point. We use the existence of $\varphi$’s fixed-point to prove the existence of a pure-strategy Nash equilibrium in Step \[step 4 orig\], while we use the uniqueness of $\varphi$’s fixed-point to demonstrate that the equilibrium is unique in Step \[step 5 orig\]. The proof is simple: $\varphi \left( 1 - \kappa \right) = 1$ and $\varphi \left( \kappa \right) = 0$. Since $\varphi$ is continuous and decreasing (see Steps \[step 5 alt 2\] - \[step 6 alt varphi dec\]), it must have a unique fixed-point. Step : An equilibrium exists (when $\kappa > 0.5$). We show that a pure-strategy Nash equilibrium exists. Based on Step \[step 6 uniq fp\], we need only describe how to construct an equilibrium from a fixed-point of $\varphi$. Suppose that $\hat{P}$ is a fixed-point of $\varphi$. The proofs for each case are similar, so we only present the argument when $\hat{P} \in \left[ 1 - \kappa , \bar{p}_2 \right)$. Define a feasible strategy profile $\left( f , a \right)$ by $$\begin{aligned} f_1 \left( p \right) &= \left\{ \begin{array}{cc} 1 & \quad \text{if } \, \, p \geq \hat{P} / \kappa \\ &\\ 0 & \quad \text{if } \, \, p < \hat{P} / \kappa \\ \end{array} \right. , \qquad \qquad f_2 \left( p \right) = \left\{ \begin{array}{cc} 1 & \quad \text{if } \, \, 1 - p \geq \left( 1 - \hat{P} \right) / \kappa \\ &\\ 0 & \quad \text{if } \, \, 1 - p < \left( 1 - \hat{P} \right) / \kappa \\ \end{array} \right. ,\end{aligned}$$ $a_1 = 0$, and $$a_2 = \displaystyle\min \left\{ w , \zeta_2 \left( \hat{P} \right) \right\} .$$ In particular, $$d_1 = \mu \left( \frac{\hat{P} }{\kappa} , 1 \right] \qquad \text{and} \qquad d_2 = \mu \left[0, 1 - \frac{1 - \hat{P}}{\kappa} \right) .$$ (\[q ineq pbar2 ints\]) implies that (\[atom bets out2 ineq alone\]) is satisfied. Since $\varphi \left( 1 - \kappa \right) = 1$, we know that $\hat{P} \ne \left( 1 - \kappa \right)$. Hence, $d_1$, $d_2 > 0$ because the density of $\mu$ is positive everywhere. We then have \[NE def orig iii\] of Definition \[maj min NE def\] by Proposition \[atom equilib bets prop\]. Since $\hat{P}$ is a fixed-point of $\varphi$, $$\hat{P} = \frac{ \mu \left( \frac{\hat{P}}{\kappa} , 1 \right] }{ \displaystyle\min \left\{ w , \zeta_2 \left( \hat{P} \right) \right\} + \mu \left( \frac{\hat{P}}{\kappa} , 1 \right] + \mu \left[0, 1 - \frac{1 -\hat{P}}{\kappa} \right) } = P^{f,a}.$$ From Proposition \[diff equilib bets prop\], we see that \[NE def orig ii\] of Definition \[maj min NE def\] holds. This completes the proof, as (i) of Definition \[maj min NE def\] is obviously satisfied. Step : The equilibrium in Step \[step 4 orig\] is unique (when $\kappa > 0.5$). We conclude Theorem \[main thm\]’s proof by showing that the equilibrium in Step \[step 4 orig\] is unique. The key observation is that $\varphi$’s fixed-point is also unique (see Step \[step 6 uniq fp\]). First, we describe how to construct a fixed-point of $\varphi$, given an equilibrium $\left( f^{\star} , a^{\star} \right)$. From Proposition \[diff equilib bets prop\] and Step \[step 1 orig\], we know that $$d_1^{\star} = \mu \left( \frac{P^{f^{\star}, a^{\star}} }{\kappa} , 1 \right] > 0 \qquad \text{and} \qquad d_2^{\star} = \mu \left[0, 1 - \frac{1 - P^{f^{\star}, a^{\star}} }{\kappa} \right) > 0 .$$ In particular, $P^{f^{\star}, a^{\star}} \in \left( 1 - \kappa , \kappa \right)$. By Proposition \[atom equilib bets prop\], there are three possibilities for $a^{\star}$. Assume that $a_1^{\star} > 0$ and $a_2^{\star} = 0$. The other cases can be handled similarly. By Proposition \[atom equilib bets prop\], (\[atom bets out1 ineq alone\]) holds. Hence, $P^{f^{\star}, a^{\star}} \in \left( \bar{p}_1 , \kappa \right)$ due to (\[q ineq pbar1 ints\]) and $$\begin{aligned} a_1^{\star} = \displaystyle\min \left\{ w , \sqrt{ \displaystyle\frac{ \kappa q d_1^{\star} d_2^{\star} }{1 - \kappa q } } - d_1^{\star} \right\} = \displaystyle\min \left\{ w , \zeta_1 \left( P^{f^{\star}, a^{\star}} \right) \right\} . \end{aligned}$$ By Definition \[impl prob defn\], $$\begin{aligned} P^{f^{\star}, a^{\star}} &= \displaystyle\frac{ \displaystyle\min \left\{ w , \zeta_1 \left( P^{f^{\star}, a^{\star}} \right) \right\} + \mu \left( \frac{ P^{f^{\star}, a^{\star}}}{\kappa} , 1 \right] }{ \displaystyle\min \left\{ w , \zeta_1 \left( P^{f^{\star}, a^{\star}} \right) \right\} + \mu \left( \frac{ P^{f^{\star}, a^{\star}}}{\kappa} , 1 \right] + \mu \left[0, 1 - \frac{1 - P^{f^{\star}, a^{\star}}}{\kappa} \right) } \\ &= \varphi \left( P^{f^{\star}, a^{\star}} \right), \end{aligned}$$ that is, $ P^{f^{\star}, a^{\star}}$ is a fixed-point of $\varphi$. Now suppose that we have another equilibrium $\left( f^{\diamond} , a^{\diamond} \right)$. Using the method just described, we see that $P^{f^{\diamond}, a^{\diamond}}$ is a fixed-point of $\varphi$. Since there is exactly one fixed-point of $\varphi$ by Step \[step 6 uniq fp\], $P^{f^{\diamond}, a^{\diamond}} = P^{f^{\star}, a^{\star}}$. By Step \[step 1 orig\] and Proposition \[diff equilib bets prop\], $f^{\star}$ and $f^{\diamond}$ necessarily agree everywhere, except when $$\begin{aligned} p = P^{f^{\star}, a^{\star}} / \kappa \qquad \text{or} \qquad 1- p = \left( 1 - P^{f^{\star}, a^{\star}} \right) / \kappa .\end{aligned}$$ Since $\mu$ has a density, it follows that $f^{\star} = f^{\diamond}$ $\mu$-a.s. Clearly, $d_1^{\star} = d_1^{\diamond}$ and $d_2^{\star} = d_2^{\diamond}$. Proposition \[atom equilib bets prop\] implies that $a^{\star} = a^{\diamond}$. [^1]: This work is supported by the National Science Foundation under DMS-1613170. We gratefully acknowledge the anonymous referees for their valuable advice on improving the paper. [^2]: Recall that the alleged decline in the house’s revenue is thought to occur over time. [^3]: Originally, players were coupled via the empirical distribution of states, not controls, in the mean-field game literature. Recent advances have shown that models with additional interactions through the controls also have promising applications, say in the contexts of price impact, optimal execution, high-frequency trading, and oligopolistic energy market problems ([@2013arXiv1305.2600G]; [@carmona2015]; [@2015arXiv150807914G]; [@chan+sirc+frack]; [@chan+sirc+courn]). [^4]: This is true of Weber’s work, for instance ([@MR637486]). Roughly, if only one player believes that a particular outcome will occur with positive probability, she should wager an arbitrarily small amount on that outcome. In the absence of a positive minimum bet size constraint, it follows that an equilibrium does not exist. [^5]: In practice, concerns about large-scale wagering firms are also driven by their alleged ability to bet [faster]{} than ordinary players. Our static game only has two outcomes, so it does not make sense to model this feature here. We hope to revisit the issue in a future work. [^6]: An exception is Cheung’s thesis, though that work’s focus is quite different from our own ([@cheung]). [^7]: Even so, our players might seem unrealistically knowledgeable and confident in their beliefs. Real bettors would presumably have a more complex prior for the outcomes’ likelihoods. They probably would not have access to such comprehensive information about their opponents. If they somehow had a sense of these details, they might also wish to update their own forecasts. An ad-hoc but simple way to address these objections could be to consider $\kappa$ as some (publicly known) negative perturbation of the true $\kappa$ (abusing notation). The idea is that each player would model transaction costs as being higher than their actual value, artificially lowering their perceived edge and encouraging them to bet more cautiously than they would otherwise. A more thorough study could begin by determining whether or not approximating parimutuel wagering using the Nash equilibrium solution concept is, indeed, reasonable. If it is, one might endow the bettors with more sophisticated priors and enable them to update their beliefs using the equilibrium implied probabilities (see Definition \[impl prob defn\]). This would lead to an extra condition in Definition \[maj min NE def\]. If the solution concept is unreasonable, one could devise a new scenario in which players independently determine their betting strategies according to individual reference models and appropriately penalized alternative models for outcome likelihoods, as well as their opponents’ parameters. Broadly speaking, this treatment of a single player’s optimization problem has seen widespread use across macroeconomics and finance ([@herr+jmk+hedg]). One might then investigate what unfolds when all players simultaneously participate in the same parimutuel wagering event. We leave further consideration of these topics for a future work. [^8]: This situation could be considered an equilibrium because if any single player unilaterally revised her wagers, intuitively, she should incur a loss of at least $\left( 1 - \kappa \right)\%$. [^9]: We later observe that $P^{f,a} \in \left( 1- \kappa , \kappa \right)$ in equilibrium (see Step \[step 5 orig\] of Theorem \[main thm\]’s proof). [^10]: Of course, we still abuse notation here. When one of our indicator functions is equal to zero, the corresponding fraction is actually of the form $0/0$, not zero as we suppose. [^11]: In \[NE def orig ii\], we determine the equilibrium wagers for all diffuse bettors whose beliefs are indexed by $p$; however, a single diffuse bettor with these beliefs solves the same maximization problem. [^12]: We describe the ideas underlying all of our proofs in Section \[main result sect\]; however, we delay our formal arguments for Proposition \[atom equilib bets prop\] and Theorem \[main thm\] until Appendices \[prop 2 proof app\] and \[main thm proof app\], respectively. [^13]: Roughly, this suggests that our large generalized game can almost be viewed as a game with two players: the mean-field of diffuse players and the atomic player. [^14]: Since $\mu$ has a density, the $\mu$-measure of a set with two points is zero. It follows that this ambiguity has no bearing on an equilibrium’s uniqueness. A more serious concern is that we could have feasible strategy profiles satisfying Definition \[maj min NE def\] with different implied probabilities. Precluding this possibility is a key part of Theorem \[main thm\]’s proof. [^15]: The outcome, if any, on which the atomic player wagers can be identified based upon her opponents’ wagers alone. In particular, this identification can be made without knowledge of the implied probability of Outcome 1. Still, from (\[atom max prob eqn prop\]), it is clear that the atomic player bets on Outcome $i$ in equilibrium if and only if the expected profit per unit bet on Outcome $i$ is positive. We rigorously prove and discuss this further in Corollary \[atom pl exp prof pos cor\]. [^16]: Related expressions are also seen in equilibrium studies with $N$ risk-neutral atomic players constrained to wager a specific total amount ([@chadha1996betting]). In an obvious way, Proposition \[atom equilib bets prop\] fills the small gap between these two settings. Recall that no explicit solutions are available for Cheung’s atomic player, who faces a budget constraint like ours but is risk-averse ([@cheung]). [^17]: In practice, this inequality almost always holds. [^18]: We remark that $P^{f^{\star}, a^{\star}} \in \left[ 0.5, 0.7 \right]$ for all $\kappa \in \left( 0.5 , 1 \right)$ in Example \[exmp 1\]. In particular, the [favorite-longshot bias]{} results, regardless of the house take. [^19]: Like Example \[exmp 1\], we see the [favorite-longshot bias]{} here because $P^{f^{\star}, a^{\star}} \in \left[ 0.5, 0.53 \right]$ for all $\kappa \in \left( 0.5 , 1 \right)$. [^20]: Over the course of many events, diffuse players may stop participating because of the atomic player, making this intuition questionable in another context (see Section \[intro sect\]). [^21]: We know that $\partial_{b_1 } \Phi \left( b_1^{\star} , 0 \right) = 0$ and $\partial_{b_2 } \Phi \left( b_1^\star , 0 \right) \leq 0$ since $\Phi$ has a unique global maximum on $\mathbb{R}^2_{\geq 0}$ at $\left( b_1^{\star} , b_2^{\star} \right)$. [^22]: If $q > 0$, then $a_1^{\star}$ is undefined. Roughly, the atomic player needs to make an arbitrarily small bet on Outcome 1 but is not allowed to do so.	{ "pile_set_name": "ArXiv" }
=8,8mm On finite groups whose Sylow subgroups are submodular **V.A.Vasilyev [Abstract]{} [m[8mm]{}m[138mm]{}]{} & A subgroup $H$ of a finite group $G$ is called submodular in $G$, if we can connect $H$ with $G$ by a chain of subgroups, each of which is modular (in the sense of Kurosh) in the next. If a group $G$ is supersoluble and every Sylow subgroup of $G$ is submodular in $G$, then $G$ is called strongly supersoluble. The properties of groups with submodular Sylow subgroups are obtained. In particular, we proved that in a group every Sylow subgroup is submodular if and only if the group is Ore dispersive and every its biprimary subgroup is strongly supersoluble. [Keywords:]{} finite group, modular subgroup, submodular subgroup, strongly supersoluble group, Ore dispersive group. Mathematics Subject Classification (2010): 20D10, 20D20, 20D40. [ [Introduction]{} ]{} Throughout this paper, all groups are finite. The notion of a normal subgroup takes a central place in the theory of groups. One of its generalizations is the notion of a modular subgroup, i.e. a modular element (in the sense of Kurosh [@sch2 Chapter 2, p. 43]) of a lattice of all subgroups of a group. Recall that a subgroup $M$ of a group $G$ is called modular in $G$, if the following hold: 1\) $\langle X,M \cap Z \rangle=\langle X,M \rangle \cap Z$ for all $X \leq G, Z \leq G$ such that $X \leq Z$, and 2\) $\langle M,Y \cap Z \rangle=\langle M,Y \rangle \cap Z$ for all $Y \leq G, Z \leq G$ such that $M \leq Z$. Properties of modular subgroups were studied in the book [@sch2]. Groups with all subgroups are modular were studied by R. Schmidt [@sch2], [@sch11] and I. Zimmermann [@zim]. By parity of reasoning with subnormal subgroup, in [@zim] the notion of a submodular subgroup was introduced. [Definition [@zim].*]{} [A subgroup $H$ of a group $G$ is called submodular in $G$, if there exists a chain of subgroups $H=H_0 \leq H_1 \leq \ldots \leq \ H_{s-1} \leq H_{s} = G$ such that $H_{i-1}$ is a modular subgroup in $H_{i}$ for $i=1,\ldots,s$.]{} If $H \neq G$, then the chain can be compacted to maximal modular subgroups. It’s well known that in a nilpotent group every Sylow subgroup is normal (subnormal). In the paper [@zim] groups with submodular subgroups were studied. In particular, it was proved that in a supersoluble group $G$ every Sylow subgroup is submodular if and only if $G/F(G)$ is abelian of squarefree exponent. A criterion of the submodularity of Sylow subgroups in an arbitrary group was found. This paper is devoted to the further study of groups with submodular Sylow subgroups. A group we will call strongly supersoluble, if it is supersoluble and every its Sylow subgroup is submodular in it. The class of all strongly supersoluble groups we will denote $s\mathfrak{U}$. We proved that the class of groups $s\mathfrak{U}$ is a hereditary local formation. We obtained that a group is strongly supersoluble if and only if it is metanilpotent and every its Sylow subgroup is submodular. The class of all groups with submodular Sylow subgroups we denote $sm\mathfrak{U}$. We proved that $sm\mathfrak{U}$ forms a hereditary local formation and its local screen was found. We established that in a group every Sylow subgroup is submodular if and only if the group is Ore dispersive and every its biprimary subgroup is strongly supersoluble. [ *[1. Preliminaries]{}** ]{} We use the standard notation and terminology (see [@shem] and [@DH]). Recall some of them. Let $G$ be a group. ${\rm {Syl}}_p(G)$ is a set of all Sylow $p$-subgroups of $G$ for some prime $p$; ${\rm {Syl}}(G)$ is a set of all Sylow subgroups of $G$; $M_G$ is the core of subgroup $M$ of $G$, i.e. the intersection of all subgroups conjugated with $M$ in $G$; $F(G)$ is the Fitting subgroup of $G$, i.e. the product of all normal nilpotent subgroups of $G$; $F_p(G)$ is a $p$-nilpotent radical of $G$, i.e. the product of all normal $p$-nilpotent subgroups of $G$, $p$ is some prime. A group $G$ of order $p_1^{n_1}p_2^{n_2}\dots p_n^{n_k}$ is called Ore dispersive [@shem p. 251], if $p_1>p_2>\dots >p_n$ and $G$ has a normal subgroup of order $p_1^{n_1}p_2^{n_2}\dots p_n^{n_i}$ for every $i=1, 2,\dots, k$. We use the following notation for concrete classes of group: $\frak{S}$ is the class of all soluble groups; $\frak{U}$ is the class of all supersoluble groups; $\frak{N}$ is the class of all nilpotent groups; $\frak{A}(p-1)$ is the class of all abelian groups of exponent dividing $p-1$. By ${\cal M}(\frak{X})$ is denoted the class of all minimal non $\frak{X}$-groups, i.e. such groups $G$ for which all proper subgroups of $G$ are contained in $\frak{X}$, but $G\not\in \frak{X}$. A class of groups $\frak F$ is called a [formation]{} if the following conditions hold: (a) every quotient group of a group lying in $\frak F$ also lies in $\frak F$; (b) if $H/A\in \frak F$ and $H/B\in \frak F$ then $H/A\cap B\in \frak F$. A formation $\frak{F}$ is called [hereditary]{} whenever $\frak{F}$ together with every group contains all its subgroups, and [saturated]{}, if $G/\Phi(G)\in\frak{F}$ implies that $G\in\frak{F}$. The $\frak F$-residual of a group $G$ for nonempty formation is denoted by $G^{\frak {F}}$, i.e. the smallest normal subgroup of $G$ with $G/G^{\frak {F}}\in \frak {F}$. A function $f:\Bbb{P}\rightarrow \{$formations$\}$ is called a [local screen]{}. A formation $\frak{F}$ is called [local]{}, if there exists a local screen $f$ such that $\frak{F}$ coincides with the class of groups $(G \| G/C_G(H/K)\in f(p)$ for every chief factor $H/K$ of $G$ and $p\in\pi(H/K))$. It’s denoted by $\frak{F}=LF(f)$. Recall that a subgroup $H$ of $G$ is called maximal modular in $G$, if $H$ is modular in $G$ and from $H \leq M < G$ it always follows $H = M$ for every modular subgroup $M$ in $G$. [Lemma 1.1 [@zim Lemma 1].]{} Let $G$ be a group and $T \leq G$. Then the following hold: $1)$ if $T$ is submodular in $G$ and $U \leq G$, then $U \cap T$ is submodular in $U$; $2)$ if $T$ is submodular in $G$, $N$ is normal in $G$ and $N \leq T$, then $T/N$ is submodular in $G/N$; $3)$ if $T/N$ is submodular in $G/N$, then $T$ is submodular in $G$; $4)$ if $T$ is submodular in $G$, then $T^x$ is submodular in $G$ for every $x \in G$; $5)$ if $T_1$ and $T_2$ are submodular subgroups in $G$, then $T_1 \cap T_2$ is a submodular subgroup in $G$; $6)$ if $T$ is submodular in $G$, then $TN$ is submodular in $G$ for every normal in $G$ subgroup $N$. [Lemma 1.2 [@sch11 Lemma 1].]{} [A subgroup $M$ of a group $G$ is maximal modular in $G$ if and only if either $M$ is maximal normal subgroup in $G$, or $G/M_G$ is nonabelian of order $pq$, where $p$ and $q$ are primes. ]{} [Lemma 1.3 [@hein Lemma 2].]{} Let $G=AB$ be a product of nilpotent subgroups $A$ and $B$, and $G$ has a minimal normal subgroup $N$ such that $N=C_G(N) \neq G$. Then $1)$ $A \cap B = 1$; $2)$ $N \subseteq A \cup B$; $3)$ if $N \leq A$, then $A$ is a $p$-group for some prime $p$ and $B$ is $p'$-group. [Lemma 1.4 [@shem Lemma 3.9 (1)].]{} [If $H/K$ is a chief factor of a group $G$ and $p \in \pi(H/K)$, then $G/C_G(H/K)$ is not contain nonidentity normal $p$-subgroups, and besides $F_p(G) \leq C_G(H/K)$.]{} [Theorem 1.5 [@wei Theorem 1.4].]{} [Let $H/K$ be a $p$-chief factor of a group $G$. Then $\|H/K\|=p$ if and only if $Aut_G(H/K)$ is abelian of exponent dividing $p-1$.]{} [Theorem 1.6 [@DH Chapter IV, Theorem 4.6].]{} [A formation is saturated if and only if it is local.]{} [Lemma 1.7 [@shem Lemma 4.5].]{} [Let $f$ be a local screen of a formation $\frak{F}$. A group $G$ belongs to $\frak{F}$ if and only if $G/F_p(G)\in f(p)$ for every $p\in\pi(G)$.]{} We need the following property of the class of all supersoluble groups (see, for example, [@shem p. 35] or [@DH p. 358]). [Lemma 1.8.]{} [The class of all supersoluble groups has a local screen $f$ such that $f(p)=\frak{A}(p-1)$ for every prime $p$.]{} [ **[2. Strongly supersoluble groups]{} ]{} [Lemma 2.1.*]{} [Let $p$ be the largest prime divisor of $\|G\|$ and $G_p \in \text{\rm Syl}_p(G)$. If $G_p$ is submodular subgroup in $G$, then $G_p$ is normal in $G$.]{} [Proof.]{} We will use an induction by $\|G\|$. We can consider that $G_p \neq G$ and there exists a chain of subgroups $G_p=H_0 < H_1 < \ldots < H_{s-1} < H_{s} = G$ such that $H_{i-1}$ is maximal modular subgroup in $H_{i}$ for $i=1,\ldots,s$. By induction, $G_p$ is normal in $H_{s-1} = M$. By Lemma 1.2, either $M$ is normal in $G$, or $G/M_G$ is nonabelian of order $rq$, where $r$ and $q$ are primes. In the first case, $G_p$ is normal in $G$. So let $\|G/M_G\|=rq$ and $G/M_G$ be a nonabelian group. It follows $\|G:M\|$ is a prime different from $p$. So we can assume that $\|G:M\|=q \neq p$. If $N_G(G_p) \neq G$, then by the Theorem of Sylow $\|G:M\| = \|G:N_G(G_p)\| = q \equiv 1\ ({\rm{mod}}\ p)$. We got a contradiction with $q < p$. So $N_G(G_p) = G$. Lemma is proved. [Corollary 2.1.1 [@zim Proposition 9].]{} [If every Sylow subgroup of the group $G$ is submodular in $G$, then $G$ is Ore dispersive.]{} [Definition 2.2.]{} [A group $G$ we will call strongly supersoluble if $G$ is supersoluble and every Sylow subgroup of $G$ is submodular in $G$.]{} Denote $s\mathfrak{U}$ the class of all strongly supersoluble groups. [Proposition 2.3 [@zim Proposition 10].]{} [A group $G$ is strongly supersoluble if and only if $G$ is supersoluble and $G/F(G)$ is abelian of squarefree exponent.]{} In the paper we denote $\mathfrak{B}$ the class of all abelian groups of exponent free from squares of primes. [Lemma 2.4.]{} [The class of groups $\mathfrak{B}$ is a hereditary formation.]{} [Proof.]{} It’s clear, if $G \in \mathfrak{B}$, then $H \in \mathfrak{B}$ and $G/N \in \mathfrak{B}$ for any subgroup $H$ and any normal subgroup $N$ of $G$. Let $G$ be a group of the smallest order such that $G/N_i \in \mathfrak{B}, N_i \unlhd G, i=1,2$, but $G/N_1 \cap N_2 \not \in \mathfrak{B}$. If $N_1 \cap N_2 \neq 1$, then in $N_1 \cap N_2$ there is a normal subgroup $K$ in $G$ . From $\|G/K\| < \|G\|$ and $G/K/N_i/K \simeq G/N_i \in \mathfrak{B}$, $i=1,2$, it follows $G/K/(N_1/K \cap N_2/K) \simeq G/N_1 \cap N_2 \in \mathfrak{B}$. This contradicts the choice of $G$. Let $N_1 \cap N_2 = 1$. Since $ \mathfrak{A}$ is a formation and $G/N_i \in \mathfrak{B} \subseteq \mathfrak{A}$, then $G \in \mathfrak{A}$. Let’s show that the exponent of $G$ is free from squares of primes. Let $z$ be an element of order $q^n$ from $G$, where $q$ is a prime, and $Z = \langle z \rangle$. Assume that $Z \cap N_1 \neq 1$ and $Z \cap N_2 \neq 1$. Since $Z$ is cyclic $q$-group, then there exists $i \in \{1,2\}$ such that $Z \cap N_i \leq Z \cap N_{3-i}$. Hence we get a contradiction $1 \ne Z \cap N_i \leq (Z \cap N_1) \cap (Z \cap N_2) =1$. Hence $Z \cap N_j = 1$ for some $j \in \{1,2\}$. Then from $G/N_j \in \mathfrak{B}$ and $ZN_j/N_j \simeq Z$ it follows $n < 2$ and $G \in \mathfrak{B}$. Lemma is proved. Note that the class of groups $\mathfrak{B}$ is not saturated. For example, a cyclic group $G=\langle z \| z^4=1\rangle \not \in \mathfrak{B}$, but $G/\Phi(G) \in \mathfrak{B}$. [Theorem 2.5.]{} Let $G$ be a group. Then the following hold:* $1)$ if $G$ is strongly supersoluble, then every subgroup of $G$ is strongly supersoluble; $2)$ if $G$ is strongly supersoluble and $N \unlhd G$, then $G/N$ is strongly supersoluble; $3)$ if $N_i \unlhd G$ and $G/N_i$ is strongly supersoluble for $i=1,2$, then $G/N_1 \cap N_2$ is strongly supersoluble; $4)$ if $H_i \unlhd G$, $H_i$ is strongly supersoluble, $i=1,2$ and $H_1 \cap H_2 =1$, then $H_1 \times H_2$ is strongly supersoluble; $5)$ if $G/\Phi(G)$ is strongly supersoluble, then $G$ is strongly supersoluble; $6)$ the class of groups $s\mathfrak{U}$ is a hereditary saturated formation. [Proof.]{} Show the validity of 1). Let $G\in s\mathfrak{U}$ and $H\leq G$. In view of Propositon 2.3 and Lemma 2.4, from $G/F(G)\in \mathfrak{B}$ it follows that $H/H\cap F(G)\simeq HF(G)/F(G) \in \mathfrak{B}$. Since $\mathfrak{B}$ is a homomorph and $H\cap F(G)\leq F(H)$, we get $H/F(H)\simeq H/H\cap F(G)/F(H)/H\cap F(G)\in \mathfrak{B}$. By Proposition 2.3, $H\in s\mathfrak{U}$. Prove Statement 2). Let $G\in s\mathfrak{U}$ and $N \unlhd G$. By Proposition 2.3, $G/F(G)\in \mathfrak{B}$. Since $\mathfrak{B}$ is a homomorph, from $F(G)N/N\leq F(G/N)=F/N$ we conclude $G/N/F(G/N)\simeq G/F\simeq G/F(G)/F/F(G)\in \mathfrak{B}$. By Proposition 2.3, $G/N\in s\mathfrak{U}$. Prove Statement 3). Let $G$ be a group of the smallest order such that $N_i \unlhd G$ and $G/N_i\in s\mathfrak{U}$ for $i=1,2$, but $G/N_1 \cap N_2\not\in s\mathfrak{U}$. Since $\mathfrak{U}$ is a formation, $G/N_1 \cap N_2\in \mathfrak{U}$. If $N_1\cap N_2\not=1$, then take from $N_1\cap N_2$ a subgroup $K\unlhd G$. From the choice of $G$ and $G/K/N_i/K\simeq G/N_i\in s\mathfrak{U}$ for $i=1,2$ it follows $G/K/(N_1/K\cap N_2/K)\simeq G/N_1 \cap N_2\in s\mathfrak{U}$. This contradicts the choice of $G$. Let $N_1\cap N_2=1$. For every Sylow $p$-subgroup $P$ of $G$ a quotient group $PN_i/N_i\in \text{\rm Syl}_p(G/N_i), i=1, 2$. From the strongly supersolubility of $G/N_i$ it follows $PN_i/N_i$ is submodular in $G/N_i, i=1, 2$. By 3) of Lemma 1.1, $PN_i$ is submodular in $G, i=1, 2$. From properties of Sylow subgoups and 5) of Lemma 1.1, it follows $PN_1\cap PN_2=P(N_1\cap N_2)=P$ is submodular in $G$. So $G\in s\mathfrak{U}$. This contradicts the choice of $G$. Statement 3) is proved. Statement 4) follows from 3). Prove Statement 5). Let $G/\Phi(G)\in s\mathfrak{U}$. From $s\mathfrak{U}\subseteq \mathfrak{U}$ and the saturation of the class of groups $\mathfrak{U}$ it follows $G\in \mathfrak{U}$. So $F(G/\Phi(G))=F(G)/\Phi(G)$. Then $G/F(G)\simeq G/\Phi(G)/F(G/\Phi(G)) \in \mathfrak{B}$, i.e. $G\in s\mathfrak{U}$. Statement 6) follows from 1)–3) and 5). Theorem is proved. [Theorem 2.6.]{} [The class of all strongly supersolubility groups is a local formation and has a local screen $f$ such that $f(p)=\mathfrak{A}(p-1)\cap\mathfrak{B}$ for any prime $p$. ]{} [Proof.]{} Since $f(p)=\mathfrak{A}(p-1)\cap\mathfrak{B}$ is a formation, $f$ is a local screen. Let a local formation $LF(f)$ be defined by a screen $f$. Let’s show that $s\mathfrak{U} = LF(f)$. Let $G\in s\mathfrak{U}$ and $H/K$ be any chief factor of $G$. From $G/F(G)\in \mathfrak{B}$ and $F(G)\leq C_G(H/K)$ it follows $G/C_G(H/K)\in \mathfrak{B}$. Since $G$ is supersoluble, $\|H/K\|=p$ is some prime. By Lemma 1.5, $G/C_G(H/K)\in \mathfrak{A}(p-1)$. So $G/C_G(H/K)\in f(p)$. Then $G\in LF(f)$ and $s\mathfrak{U}\subseteq LF(f)$. Let now $G\in LF(f)$. Then $G/C_G(H/K) \in f(p)\subseteq \mathfrak{B}$ for any chief factor $H/K$ of $G$ and $p\in \pi(H/K)$. Since $\mathfrak{B}$ is a formation, we conclude $G/F(G) \in \mathfrak{B}$. So $LF(f) \subseteq s\mathfrak{U}$. Thus $LF(f) = s\mathfrak{U}$. Theorem is proved. [Proof.]{} Let $G$ be a counterexample of minimal order to Theorem. Then, by Theorem of Wielandt-Kegel [@wiel; @keg], $G$ is soluble. Let $N$ be a minimal normal subgroup of $G$. Then $AN/N \simeq A/A \cap N \in \mathfrak{N},\ BN/N \simeq B/B \cap N \in \mathfrak{N}$. By 6) and 2) of Lemma 1.1, $AN/N$ and $BN/N$ are submodular in $G/N$. By the choice of $G$, it follows $G/N \in s\mathfrak{U}$. By 3)–5) of Theorem 2.5, we conclude that $N$ is the only minimal normal subgroup in $G$ and $\Phi(G) = 1$. Then $G=MN$, where $M$ is a maximal subgroup in $G$, $M \cap N = 1$, $N=C_G(N)$ and $\|N\|=p^n$ for some prime $p$. By 1) of Lemma 1.3, $A \cap B = 1$. From $N \subseteq A \cup B$ it follows either $N \leq A$, or $N \leq B$. Without loss of generality, we may suppose $N \leq A$. Then, by 3) of Lemma 1.3, $A$ is a $p$-subgroup, $B$ is a $p'$-subgroup. Let $q$ be the largest prime divisor of $\|G\|$. If $q \neq p$, then $B$ has some Sylow $q$-subgroup $Q$ of $G$. From $Q \unlhd B$ and the submodularity of $B$ in $G$ it follows that $Q$ is submodular in $G$. By Lemma 2.1, $Q \unlhd G$. Then $N \leq Q$. We get a contradiction with $q \neq p$. So $q=p$. In view of Lemma 2.1, $A \unlhd G$. By Lemma 1.4, $O_p(M)=1$. Then $M \cap A = 1$ and $A=N$, $B$ is a maximal subgroup of $G$ and $B_G=1$. Hence, $B$ is a maximal modular subgroup in $G$. In view of $B_G=1$ by Lemma 1.2, we conclude $\|G\|=pr$, where $r$ is a prime and $p\not= r$. So $G \in s\mathfrak{U}$. This contradicts the choice of $G$. Theorem is proved. In Theorem 2.7 we can’t discard the submodularity of one of subgroups. [Example 2.8.]{} In group $G=AB$, where $A\simeq Z_{17}$ and $B\simeq Aut(Z_{17}) \simeq Z_{16}$, the subgroup $A$ is submodular, but the subgroup $B$ is not submodular in $G$. The group $G$ is supersoluble, but not strongly supersoluble. The example also shows that $s\mathfrak{U}\not=\mathfrak{U}$. [Theorem 2.9.]{} [A group $G$ is strongly supersoluble if and only if $G$ is metanilpotent and any Sylow subgroup of $G$ is submodular in $G$.]{} [Proof.]{} Necessity follows from that the strongly supersoluble group is supersoluble, and so it has a nilpotent commutator subgroup, i.e. it is metanilpotent. Sufficiency. Let there exists metanilpotent groups which have all Sylow subgroups are submodular in a group, but the group is not strongly supersoluble. Let’s choose from them a group $G$ of the smallest order. Let $N$ be a minimal normal subgroup of $G$. Then $G/N \in s\mathfrak{U}$ in view of the choice of $G$. Since, by Theorem 2.5, the class $s\mathfrak{U}$ is a saturated formation, then $N$ is the only minimal normal subgroup in $G$ and $\Phi(G)=1$. So $N = C_G(N)$ and $G=NM$, where $M$ is a maximal subgroups of $G$, $M \cap N =1$. From the metanilpotency of $G$ and $N=F(G)$ it follows $G/N \simeq M \in \mathfrak{N}$. Let $p$ be the largest prime divisor of $\|G\|$. Since $G$ is Ore dispersive, it follows that $N$ is contained in some Sylow $p$-subgroup of $G$. In view of $O_p(M)=1$, we conclude that $N \in \text{Syl}_p(G)$ and $M$ is a $p'$-group. Let $S \in \text{Syl}_q(M)$. If $G=SN$, then, by Theorem 2.7, $G$ is strongly supersoluble. This contradicts the choice of $G$. Let $G\not=SN$ for every $S \in \text{Syl}_q(M)$. Denote $L=SN$. Then $L$ is strongly supersoluble by the choice of $G$. From $C_G(N)=N$ it follows $O_{p'}(L)=1$. Then $N=F_p(L)$. By Lemma 1.7 and Theorem 2.6, we get $S\simeq L/F_p(L)\in \mathfrak{A}(p-1)\cap\mathfrak{B}$. Hence and from the nilpotency of $M$ it follows that $M \in \mathfrak{A}(p-1)$. Since $N=F_p(G)$, we conclude that $M\simeq G/F_p(G)\in \mathfrak{A}(p-1)$. By Lemmas 1.7 and 1.8, $G$ is supersoluble. By Definition 2.2, $G$ is strongly supersoluble. This contradicts the choice of $G$. Theorem is proved. **[3. Groups with submodular Sylow subgroups]{} Denote $sm\mathfrak{U}=(\ G\ \|$ every Sylow subgroup of the group $G$ is submodular in $G$ ). [Theorem 3.1.*]{} Let $G$ be a group. Then the following hold:* $1)$ if $G \in sm\mathfrak{U}$ and $H \leq G$, then $H \in sm\mathfrak{U}$; $2)$ if $G \in sm\mathfrak{U}$ and $N \unlhd G$, then $G/N \in sm\mathfrak{U}$; $3)$ if $N_i \unlhd G$ and $G/N_i \in sm\mathfrak{U}$, $i=1,2$, then $G/N_1 \cap N_2 \in sm\mathfrak{U}$; $4)$ if $H_i \in sm\mathfrak{U}$, $H_i \unlhd G$, $i=1,2$ and $H_1 \cap H_2 =1$, then $H_1 \times H_2 \in sm\mathfrak{U}$; $5)$ if $G/\Phi(G) \in sm\mathfrak{U}$, then $G \in sm\mathfrak{U}$; $6)$ the class of groups $sm\mathfrak{U}$ is a hereditary saturated formation. [Proof.]{} The validity of Statements 1) and 2) of Theorem follows from Statements 1), 2) and 6) of Lemma 1.1, in view of $H \cap G_p \in \text{Syl}_p(H)$ for some $G_p \in \text{Syl}_p(G)$ and $R/N = G_qN/N \in \text{Syl}_q(G/N)$ for some $G_q \in \text{Syl}_q(G)$. Prove Statement 3). Let $G$ be a group of the smallest order such that $G/N_i \in sm\mathfrak{U},\ N_i \unlhd G,\ i=1,2$, but $G/N_1 \cap N_2 \not \in sm\mathfrak{U}$. We can suppose that $N_1 \cap N_2 = 1$. Let $P \in \text{Syl}_p(G)$. Then $PN_i/N_i \in \text{Syl}_p(G/N_i),\ i=1,2$. So $PN_i/N$ is submodular in $G/N_i$. By 3) of Lemma 1.1, $PN_i$ is submodular in $G$. By the property of Sylow subgroups and Statement 5) of Lemma 1.1, $PN_1 \cap PN_2 = P(N_1 \cap N_2)=P$ is submodular in $G$, i.e. $G \in sm\mathfrak{U}$. This contradiction completes the proof of 3). Statement 4) follows from 3). Prove Statement 5). Let $G$ be a group of the smallest order such that $G/\Phi(G) \in sm\mathfrak{U}$, but $G \not \in sm\mathfrak{U}$. Then $G$ is soluble in view of Corollary 2.1.1 and the solubility of $\Phi(G)$. Let $N$ be a minimal normal subgroups of $G$. From $\Phi(G)N/N \subseteq \Phi(G/N)$ and by Statement 2) of Theorem, we conclude that $G/N/\Phi(G/N) \in sm\mathfrak{U}$. Since $\|G/N\| < \|G\|$, $G/N \in sm\mathfrak{U}$. From Statement 3) it follows that $N$ is the only minimal subgroup of $G$, $\|N\|=p^n$ for some prime $p$ and $O_{p'}(G)=1$. Hence, $N \subseteq \Phi(G)$. Let $Q \in \text{Syl}_q(G)$. From $QN/N \in \text{Syl}_q(G/N)$ it follows that $QN/N$ is submodular in $G/N$. By Statement 2) of Lemma 1.1, $QN$ is submodular in $G$. If $p=q$, then $QN=Q$ is submodular in $G$. Let $p \neq q$. Let’s consider 2 cases: \(1) $\|\pi(G)\|=2$. Then $G/N = QN/N \cdot P/N$, where $P \in \text{Syl}_p(G)$. By Theorem 2.7, $G/N$ is strongly supersoluble. Since the class $s\mathfrak{U}$ of all strongly supersoluble groups is a saturated formation by Theorem 2.5, then from $G/\Phi(G) \simeq G/N/\Phi(G)/N \in s\mathfrak{U}$ it follows that $G \in s\mathfrak{U} \subseteq sm\mathfrak{U}$. This contradicts the choice of $G$. \(2) $\|\pi(G)\|>2$. Then $G/N \neq H/N = QN/N \cdot R/N$, where $R/N \in \text{Syl}_p(G/N)$. By Statement 1) of Theorem $H/N \in sm\mathfrak{U}$. So Sylow $q$-subgroup $QN/N$ and Sylow $p$-subgroup $R/N$ of $H/N$ are submodular in $H/N$. By Theorem 2.7, $H/N$ is strongly supersoluble. Since by Theorem 2.6 $s\mathfrak{U}$ is a local formation and $H=QR$ is a Hall $\{p,q\}$-subgroup of $G$, then apply Corollary 16.2.3 from [@shem]. From $H\Phi(G)/\Phi(G) \simeq H/H \cap \Phi(G) \simeq H/N/H \cap \Phi(G)/N \in s\mathfrak{U}$ it follows $H \in s\mathfrak{U}$. By Statement 1) of Theorem 2.5, $QN \in s\mathfrak{U}$. Then $Q$ is submodular in $QN$, and so, $Q$ is submodular in $G$. The arbitrariness of the choice of $Q$ implies $G \in sm\mathfrak{U}$. This contradiction completes the proof of Statement 5). Statement 6) follows from Statements 1)-3) and 5). Theorem is proved. [Lemma 3.2.]{} The following statements are held. $1)$ The class of groups $(G \in \mathfrak{S}\ \|\ \text{\rm Syl}(G) \subseteq \mathfrak{B})$ is a hereditary formation. $2)$ For any prime $p$ the class of groups $(G \in \mathfrak{S}\ \|\ \text{\rm Syl}(G) \subseteq \mathfrak{A}(p-1)\cap\mathfrak{B})$ is a hereditary formation. [Proof.]{} Prove Statement 1). Denote $\mathfrak{H}=(G \in \mathfrak{S}\ \|\ \text{\rm Syl}(G) \subseteq \mathfrak{B})$. Clearly, if $H \leq G \in \mathfrak{H}$ and $N \unlhd G$, then $H \in \mathfrak{H}$ and $G/N \in \mathfrak{H}$. Let’s show by induction on $\|G\|$, if $N_i \unlhd G$ and $G/N_i \in \mathfrak{H},\ i=1,2$, then $G/N_1 \cap N_2 \in \mathfrak{H}$. If $K = N_1 \cap N_2 \neq 1$, then from $\|G/K\| < G$ and $G/K/N_i/K \simeq G/N_i \in \mathfrak{H}$ it follows $G/K/N_1/K \cap N_2/K \simeq G/N_1 \cap N_2 \in \mathfrak{H}$. Let $N_1 \cap N_2 = 1$. Let $P \in \text{Syl}_p(G)$. From $G/N_i \in \mathfrak{H}$ it follows that $PN_i/N_i \simeq P/P \cap N_i$ is an elementary abelian group. Since the class of all abelian groups $\mathfrak{A}$ is a formation, then $P/(P \cap N_1) \cap (P \cap N_2) \simeq P \in \mathfrak{A}$. We will show that $P$ is an elementary abelian $p$-group. Let $z \in P$, $\|z\|=p^n$ and $Z = \langle z \rangle$. From $ZN_i/N_i \leq PN_i/N_i$ it follows that $\|ZN_i/N_i\|=\|Z/Z \cap N_i\| \leqslant p$. Since $N_1 \cap N_2 = 1$, then there exists $i \in \{1,2\}$ such that $Z \cap N_i = 1$. Then $\|Z\| = p^n,\ n=1$. Since $P$ is a direct product of cyclic subgroups, we get $P\in \mathfrak{B}$. So $G \in \mathfrak{H}$. Statement 2) is being proved similarly taking into account $\mathfrak{A}(p-1)\cap\mathfrak{B}$ is a hereditary formation. Lemma is proved. [Lemma 3.3.]{} [A local formation $LF(f)$ with a local screen $f$ such that $f(p)=(H \in \mathfrak{S}\ \|\ \text{\rm Syl}(H) \subseteq \mathfrak{A}(p-1)\cap\mathfrak{B})$ for any prime $p$, is a hereditary saturated formation.]{} [Proof.]{} By Theorem 1.6, $LF(f)$ is a saturated formation. Let’s prove the heredity of $LF(f)$. Let $G \in LF(f)$ и $R \leq G$. Then $G$ has a chief series $1=G_0 < G_1 < \ldots < G_{n-1} < G_n = G$ such that $G/C_G(G_i/G_{i-1}) \in f(p)$ for every $p \in \pi(G/G_{i-1})$ and $i=1,\ldots, n$. Let $R_{i-1}=R \cap G_{i-1},\ i=1,\ldots,n+1.$ Let $C_i=C_G(G_i/G_{i-1})$ and $C_i^=C_R(R_i/R_{i-1}),\ i=1,\ldots,n$. It is easy to see that $R \cap C_i \leq C_i^$. From $RC_i/C_i \leq G/C_i \in f(p)$ and the heredity of $f(p)$ it follows that $RC_i/C_i \simeq R/R \cap C_i \in f(p)$. Then $R/C_i^* \simeq R/R \cap C_i / C_i^/R \cap C_i \in f(p)$. Hence $R/C_R(H/K) \in f(p)$ for every chief factor $H/K$ of $R$ and $p \in \pi(H/K)$. So $R\in LF(f)$. Lemma is proved. [Theorem 3.4.]{} [Every minimal non $sm\mathfrak{U}$-group is biprimary minimal non $s\mathfrak{U}$-group.]{} [Proof.]{} Let $G \in {\cal{M}}(sm\mathfrak{U})$ and $q$ be the smallest prime divisor of $\|G\|$. Every subgroup $H$ of a group $G$ belongs $sm\mathfrak{U}$. By Corollary 2.1.1, $H$ is Ore disperive. So $H$ is $q$-nilpotent. Let’s consider two cases. 1\) $G$ is $q$-nilpotent. Then $G=QP$, where $Q \in \text{Syl}_q(G),\ P \unlhd G$ and $P$ is a Hall $q'$-subgroup of $G$. From $P \in sm\mathfrak{U}$ it follows the solubility of $P$. Then from $G/P \simeq Q$ we get the solubility of $G$. Suppose that $\Phi(G)=1$. Let $N$ be a minimal normal subgroup of $G$. Then $\|N\|=p^n$ for some $p \in \pi(G)$. Let $G=NM$, where $M$ is a maximal subgroup of $G$. In view of $G \in {\cal{M}}(sm\mathfrak{U})$ and $G/N \simeq M / M \cap N \in sm\mathfrak{U}$, $N$ is the only minimal normal subgroup of $G$. Let $R$ is an arbitrary Sylow $r$-subgroup of $G$. From $G/N \in sm\mathfrak{U}$ we conclude that $RN/N$ is submodular in $G/N$. By 3) of Lemma 1.1, $RN$ is submodular in $G$. If $RN \neq G$, then from $RN \in sm\mathfrak{U}$ we get $R$ is submodular in $G$. This contradicts with $G \not \in sm\mathfrak{U}$. Hence $RN = G$ is a biprimary group. Since every subgroup $T$ of $G$ belongs $sm\mathfrak{U}$, then, by Theorem 2.7, $T \in s\mathfrak{U}$. From $s\mathfrak{U} \subseteq sm\mathfrak{U}$ it follows that $G \not \in s\mathfrak{U}$, i.e. $G \in {\cal{M}}(s\mathfrak{U})$. Let $\Phi(G) \neq 1$. Then $\Phi(G/\Phi(G)) = 1$. Since $sm\mathfrak{U}$ is saturated, it follows $G/\Phi(G) \not \in sm\mathfrak{U}$. Then $G/\Phi(G) \in {\cal{M}}(sm\mathfrak{U})$. As proved above, $G/\Phi(G)$ is a biprimary group and $G/\Phi(G) \not \in s\mathfrak{U}$. Hence $G \in {\cal{M}}(s\mathfrak{U})$. 2\) $G$ is not $q$-nilpotent. By Theorem 5.4 of [@Hup гл. IV], $G$ is a Schmidt group. Since every subgroup $T$ of $G$ is nilpotent, then $T \in s\mathfrak{U}$. Then $G \in {\cal{M}}(s\mathfrak{U})$. Theorem is proved. Recall that a subgroup $H$ of a group $G$ is called [${\rm K}$-$\Bbb{P}$-subnormal [@vasTyt] in $G$]{} , if there exists a chain of subgroups $$H = H_0\leq H_1\leq\cdots\leq H_{n-1}\leq H_n=G \label{1}$$ such that either $H_{i-1}$ is normal in $H_{i}$, or $\|H_{i} : H_{i-1}\|$ is a prime for every $i = 1,\ldots, n$. If $H=G$ or in a chain the index $\|H_{i} : H_{i-1}\|$ is a prime for every $i = 1,\ldots, n$, then $H$ is called [$\Bbb{P}$-subnormal in $G$]{} [@vas]. [Lemma 3.5.]{} Let $H$ be a submodular Sylow subgroup of a group $G$. Then the following conditions are held:* $1)$ $H$ is ${\rm K}$-$\Bbb{P}$-subnormal in $G$; $2)$ if $G$ is soluble, then $H$ is $\mathbb{P}$-subnormal in $G$. [Proof.]{} Prove 1) by induction on $\|G\|$. We can suppose that $H\not=G$. Then $H$ is contained in a maximal modular subgroup $M$ of $G$. By 1) of Lemma 1.1 and $\|M\| < \|G\|$, it follows that $H$ is ${\rm K}$-$\Bbb{P}$-subnormal in $M$. By Lemma 1.2 either $M$ is normal in $G$, or $G/M_G$ is non-abelian of order $pq$, where $p$ and $q$ are primes. Hence, if $M_G\not= M$ we have $\|G : M\|= \|G/M_G: M/M_G\|$ is a prime. This means that $M$ is ${\rm K}$-$\Bbb{P}$-subnormal in $G$. So $H$ is ${\rm K}$-$\Bbb{P}$-subnormal in $G$. Statement 2) follows from 1), since in a soluble group ${\rm K}$-$\Bbb{P}$-subnormal subgroup is $\Bbb{P}$-subnormal. Lemma is proved. By Lemma 3.5, it follows that $sm\mathfrak{U}\subseteq \text{w}\mathfrak{U}$, where $\text{w}\mathfrak{U}$ is the class of all groups with $\mathbb{P}$-subnormal Sylow subgroups. Example 2.8 shows that $sm\mathfrak{U}\not= \text{w}\mathfrak{U}$. [Theorem 3.6.]{} [The class of all groups with submodular Sylow subgroups is a local formation and has a local screen $f$ such that $f(p)=(G \in \mathfrak{S}\ \|\ \text{\rm Syl}(G) \subseteq \mathfrak{A}(p-1)\cap\mathfrak{B})$ for every prime $p$.]{} [Proof.]{} Since $f(p)=(G \in \mathfrak{S}\ \|\ \text{\rm Syl}(G) \subseteq \mathfrak{A}(p-1)\cap\mathfrak{B})$ is a formation, $f$ is a local screen. Let a local formation $LF(f)$ be defined by a screen $f$. Denote $\mathfrak{F}=LF(f)$. By Theorem 2.10 [@vas], the class of groups $\text{w}\mathfrak{U}$ is a local formation and has a local screen $h$ such that $h(p) = (G \in \mathfrak{S}\ \|\ \text{\rm Syl}(G) \subseteq \mathfrak{A}(p-1))$ for every prime $p$. Hence $\mathfrak{F}\subseteq\text{w}\mathfrak{U}$. In view of Proposition 2.8 [@vas], $\mathfrak{F}$ consists of Ore dispersive groups. Show that $\mathfrak{F} \subseteq sm\mathfrak{U}$. Let $G$ be a group of the smallest order from $\mathfrak{F} \backslash sm\mathfrak{U}$. By Lemma 3.3, $\mathfrak{F}$ is a hereditary formation. Hence $G$ is a minimal non $sm\mathfrak{U}$-group. Since $\mathfrak{F} \subseteq \mathfrak{S}$ and $sm\mathfrak{U}$ is a hereditary formation by 6) of Theorem 3.1, then $G$ has the unique minimal normal subgroup $N$, $N=C_G(N)$ is an elementary abelian $p$-subgroup for some prime $p$, $\Phi(G)=1$. Then $G = NM$, where $M$ is a maximal subgroup of $G$. By Lemma 1.4, $O_p(M) = 1$. Since $G$ is Ore dispersive, it follows that $N \in \text{Syl}_p(G)$ and $p$ is the largest prime divisor of $\|G\|$. By Theorem 3.4, $G$ is a biprimary minimal non $s\mathfrak{U}$-group. Hence and from $G/C_G(N) \simeq M \in f(p)$ we get that $M$ is an elementary abelian $q$-group and $q$ divides $(p-1)$. Since $M$ is $\mathbb{P}$-subnormal in $G$ it follows that $\|G:M\|=p$ and $\|N\|=p$. Then $M \simeq G/N$ is isomorphically embedded in a cyclic group of order $p-1$. So $\|M\|=q$. Hence $G \in s\mathfrak{U} \subseteq sm\mathfrak{U}$. This contradicts the choice of $G$. So $\mathfrak{F} \subseteq sm\mathfrak{U}$. Prove that $sm\mathfrak{U} \subseteq \mathfrak{F}$. Let $G$ be a group of the smallest order from $sm\mathfrak{U} \backslash \mathfrak{F}$. Since $G \in sm\mathfrak{U}$, then $G$ is soluble by Corollary 2.1.1. Since $sm\mathfrak{U}$ and $\mathfrak{F}$ are saturated formations, then $\Phi(G)=1$. In $G$ there exists the unique minimal normal subgroup $N = C_G(N) = F(G)$, $\|N\|=p^n$ for some $p \in \pi(G)$. Then $G=NM$, where $M$ is a maximal subgroup in $G$, $N \cap M = 1$. By Corollary 2.1.1, $G$ is Ore dispersive. Then $p$ is the largest prime divisor of $\|G\|$. In view of Lemma 1.4, $N \in \text{Syl}_p(G)$ and $M$ is a $p'$-groups. Let $Q \in \text{Syl}_q(M)$. Then $Q \in \text{Syl}_q(G)$. Suppose that $QN \neq G$. In view of $QN \in sm\mathfrak{U}$ and by Theorem 2.7, $QN$ is strongly supersoluble. From $N \leq F(QN)$, $N=C_G(N)$ and $O_{p'}(QN)=1$, it follows that $N = F_p(QN)$. By Lemma 1.7, $Q \simeq QN/F_p(QN) \in \mathfrak{A}(p-1)\cap\mathfrak{B}$. Then $M \in f(p)$ in view of the arbitrariness of the choice of $Q$. By Lemma 1.7 $G \in \mathfrak{F}$. This contradicts the choice of $G$. Hence $QN=G$. By Theorem 2.7, $G$ is strongly supersoluble. By Lemma 2.6, $G/C_G(N)=G/N\in \mathfrak{A}(p-1)\cap\mathfrak{B}$. So $G \in \mathfrak{F}$. This contradicts the choice of $G$. Theorem is proved. [Theorem 3.7.]{} Let $G$ be a group in which every Sylow subgroup is submodular in $G$. Then the following conditions are held: $1)$ every metanilpotent subgroup of $G$ is strongly supersoluble; $2)$ every biprimary subgroup of $G$ is strongly supersoluble; $3)$ if $N$ is the smallest normal subgroups of $G$ such that $G/N$ is a group with elementary abelian Sylow subgroups, then $N$ is nilpotent. [Proof.]{} Statement 1) follows from the heredity of the class of groups $sm\mathfrak{U}$ and Theorem 3.6. Statement 2) follows from the heredity of the class of groups $sm\mathfrak{U}$ and Theorem 2.9. Prove Statement 3). By Lemma 3.2, the class of groups $\mathfrak{H}=(G \ \|\ \text{\rm Syl}(G) \subseteq \mathfrak{B})$ is a hereditary formation. By Statement 10 of [@shem с. 36], $\mathfrak{N}\mathfrak{H}$ has a local screen $f$ such that $f(p)=\mathfrak{H}$ for every prime $p$. In view of Theorem 3.6, $sm\mathfrak{U} \subseteq \mathfrak{N}\mathfrak{H}$. Then $N= G^{\mathfrak{H}}\in \mathfrak{N}$. Theorem is proved. [Theorem 3.8.]{} [Every Sylow subgroup of the group $G$ is submodular in $G$ if and only if the group $G$ is Ore dispersive and every its biprimary subgroup is strongly supersoluble.]{} [Proof.]{} Let the class of groups $\mathfrak{F}=(G\ \|$ the group $G$ is Ore dispersive and every biprimary subgroup of $G$ is strongly supersoluble). If $G \in sm\mathfrak{U}$, then $G \in \mathfrak{F}$ in view of Corollary 2.1.1 and by 2) of Theorem 3.7. Let $G$ be a group of the smallest order belonging $\mathfrak{F} \backslash sm\mathfrak{U}$. Since $s\mathfrak{U} \subseteq sm\mathfrak{U}$, then $\|\pi(G)\|>2$. Since $G$ is Ore dispersive, then for the largest prime $p \in \pi(G)$ and $P \in \text{Syl}_p(G)$, $P$ is normal in $G$. Since $G$ is soluble, there exists a Hall $p'$-subgroup $H$ of $G$. From $H \neq G$ and $H \in \mathfrak{F}$ if follows that $H \in sm\mathfrak{U}$. Note that $\|\pi(H)\| \geqslant 2$. Let $Q \in \text{Syl}_q(G),\ q \neq p$. Since $QP/P \in \text{Syl}_q(G/P)$ and $G/P \simeq H \in sm\mathfrak{U}$, then $QP/P$ is submodular in $G/P$. By 3) of Lemma 1.1, $QP$ is submodular in $G$. From $QP \in s\mathfrak{U}$, it follows that $Q$ is submodular in $QP$. So $Q$ is submodular in $G$. This means that $G \in sm\mathfrak{U}$. We got a contradiction with the choice of $G$. Theorem is proved. [16]{} Subgroup Lattices of Groups. — Berlin etc: Walter de Gruyter, 1994. Modulare Untergruppen endlicher Gruppen. // J. Ill. Math. — 1969. — Vol. 13. — P. 358–377. Submodular Subgroups in Finite Groups. // Math. Z. — 1989. — Vol. 202. — P. 545–557. Formations of finite groups. — Moscow: Nauka, 1987. (In Russian) Finite Soluble Groups. — Berlin etc: Walter de Gruyter, 1992. Products of finite nilpotent groups // Math. Ann. — 1990. — Vol. 287.—P. 643–652. ed. Between Nilpotent and Solvable / H.G. Bray \[and others\] — Passaic: Polygonal Publishing House, 1982. Über Produkte von nilpotenten Gruppen. III // J. Math. — 1958. — Bd. 2, No. 4B. — P. 611–618. Produkte nilpotenter Gruppen. // Arch. Math. — 1961. — Bd. 12, No. 2. — P. 90–93. Endliche Gruppen. I. — Berlin etc: Springer, 1967. On ${\rm K}$-$\Bbb{P}$-Subnormal Subgroups of Finite Groups // Math. Not. — 2014. — Vol. 95, No. 3-4. — P. 471–480. On the finite groups of supersoluble type // Siberian Math. J. — 2010. — Vol. 51, No. 6. — P. 1004–1012. V.A.Vasilyev Francisk Skorina Gomel State University, Sovetskaya str., 104, Gomel 246019, Belarus. E-mail address: [email protected]	{ "pile_set_name": "ArXiv" }
--- abstract: \| Ghys and Sergiescu proved in the $80$s that Thompson’s group $T$, and hence $F$, admits actions by $C^{\infty}$ diffeomorphisms of the circle . They proved that the standard actions of these groups are topologically conjugate to a group of $C^\infty$ diffeomorphisms. Monod defined a family of groups of piecewise projective homeomorphisms, and Lodha-Moore defined finitely presentable groups of piecewise projective homeomorphisms. These groups are of particular interest because they are nonamenable and contain no free subgroup. In contrast to the result of Ghys-Sergiescu, we prove that the groups of Monod and Lodha-Moore are not topologically conjugate to a group of $C^1$ diffeomorphisms. Furthermore, we show that the group of Lodha-Moore has no nonabelian $C^1$ action on the interval. We also show that many Monod’s groups $H(A)$, for instance when $A$ is such that ${\mathsf{PSL}}(2,A)$ contains a rational homothety $x\mapsto \tfrac{p}{q}x$, do not admit a $C^1$ action on the interval. The obstruction comes from the existence of hyperbolic fixed points for $C^1$ actions. With slightly different techniques, we also show that some groups of piecewise affine homeomorphisms of the interval or the circle are not smoothable. [^1] author: - 'Christian Bonatti[^2] [^3]' - 'Yash Lodha[^4]' - Michele Triestino title: 'Hyperbolicity as an obstruction to smoothability for one-dimensional actions' --- Introduction ============ A few examples are known of groups that admit no sufficiently smooth action on a one-dimensional manifold. Following the direction of Zimmer program, typical examples come from lattices in higher rank Lie groups [@burger-monod; @witte; @ghys], or more generally from groups with Kazhdan’s property $(T)$ [@navas(T); @navas-rel]. Other interesting examples appear in [@forcing; @calegari; @parwani; @navas-locind; @baik-kim-koberda]. In this work we address the problem of the existence of smooth actions of groups of piecewise projective homeomorphisms of the real line. Our principal interest comes from the existence of groups of this kind which are negative solutions to the so-called Day-von Neumann problem, as shown by Monod and Lodha-Moore [@monod; @LodhaMoore]. On the other hand, partially motivated by his work on Kazhdan groups acting on the circle, Navas raised the problem to find obstructions for a group of piecewise linear homeomorphisms of the interval to admit smooth actions (cf. [@navas-locind; @hyperbolic]). With this work, we illustrate relatively elementary tools which apply to a large variety of examples of such groups. Our techniques rely on some classical facts on one-dimensional dynamics and the recent work by Bonatti, Navas, Rivas and Monteverde on actions of abelian-by-cyclic groups [@hyperbolic]. A classical obstruction to have $C^1$ actions on the interval is Thurston’s Stability Theorem [@Th]: a group of $C^1$ diffeomorphisms of the interval is locally indicable, namely every finitely generated subgroup has a nontrivial morphism to ${\mathbb{Z}}$. This obstruction does not apply in our setting: the group of piecewise projective homeomorphisms of the real line is locally indicable. Therefore our results exhibit new examples of locally indicable groups that have no $C^1$ action on the interval. As an appetizer, even before introducing the notions and definitions which are necessary for presenting our main results, we start with two results whose statements are very easy to understand, and which illustrate the spirit of the paper. Fix $\lambda>1$ and consider: - the linear map $f_\lambda\colon {\mathbb{R}}\to{\mathbb{R}}$ defined as $x\mapsto \lambda x$, - the map $h_\lambda\colon {\mathbb{R}}\to {\mathbb{R}}$ defined as $h_\lambda(x)=\begin{cases} x&\text{if }x\le 0,\\ \lambda x&\text{if }x>0, \end{cases}$ - the translation $g\colon x\mapsto x+1$. Let $G_\lambda$ be the subgroup $\langle f_\lambda,g,h_\lambda\rangle\subset {\mathsf{Homeo}}_+({\mathbb{R}})$. \[t:main1\] For any $\lambda>1$ which is rational (in formula: $\lambda\in {\mathbb{Q}}\cap (1,+\infty)$) and any morphism $\rho\colon G_\lambda \to {\mathsf{Diff}}^1_+([0,1])$ one has : the commutator $[g,h_\lambda gh_\lambda^{-1}]$ belongs to the kernel of $\rho$. In particular $\rho$ cannot be injective. The same holds for any morphism $\varphi\colon G_\lambda\to {\mathsf{Diff}}^1_+({\mathbb{S}^1})$, where ${\mathbb{S}^1}$ is the circle. In fact, we get the stronger conclusion that for any representation $\rho\colon G_\lambda \to {\mathsf{Diff}}^1_+([0,1])$, the image $\rho(G_\lambda)$ is a metabelian group (that is, a solvable group with abelian derived subgroup). [[The same occurs for a more general class of algebraic numbers, that we call Galois hyperbolic (see Definition \[HyperbolicAlgebraic\] and Theorem \[t:main1Hyp\]).]{}]{} We do not know if the same occurs for $\lambda>1$ [[not Galois hyperbolic (see Remark \[couterex\])]{}]{}. Nevertheless, consider the natural realization $\rho_0\colon G_\lambda\to {\mathsf{Homeo}}_+({\mathbb{S}^1})$ defined as follows: - one considers ${\mathbb{S}^1}$ as being the projective space ${\mathbb{R}}{\mathsf{P}^1}$, - $\rho_0(f_\lambda)$ acts on ${\mathbb{S}^1}$ as the projective action of the matrix $\begin{pmatrix} \lambda&0\\ 0&1 \end{pmatrix}$, - $\rho_0(g)$ acts on ${\mathbb{S}^1}$ as the projective action of the matrix $\begin{pmatrix} 1&1\\ 0&1 \end{pmatrix}$, - $\rho_0(h_\lambda)$ coincides with $\rho_0(f_\lambda)$ on the half circle $[0, +\infty]$ and with the identity map on the half circle $[-\infty, 0]$. \[t:main2\] Fix an arbitrary real number $\lambda>1$. With the notation as above, it does not exist any homeomorphism $\phi\colon {\mathbb{S}^1}\to {\mathbb{S}^1}$ so that $\phi\rho_0(f_\lambda)\phi^{-1}$, $\phi\rho_0(h_\lambda)\phi^{-1}$ and $\phi\rho_0(g)\phi^{-1}$ belong to ${\mathsf{Diff}}_+^1({\mathbb{S}^1})$. In other words, the natural action of $G_\lambda$ on ${\mathbb{S}^1}$ is not smoothable, and furthermore, [[if $\lambda>1$ is Galois hyperbolic]{}]{}, then every $C^1$ action of $G_\lambda$ on the circle or the interval are (non-faithful) metabelian actions. For more precise statements, see Theorems \[t:affine\] and \[t:key\]. The paper is organized as follows. In Section \[s:Defs\] we introduce the basic objects and fix some notation. In Section \[s:Mech\] we roughly explain the different strategies that we develop in this work, showing which are the main applications. In Section \[s:Motiv\] we illustrate the main motivation of our work, which is the recent construction by Monod of nonamenable groups without free subgroups. In Section \[s:Monod\] we study the $C^1$ actions of Monod’s groups and the finitely presentable group defined by Lodha-Moore. Section \[s:Aff\] contains the main part of this work, namely the study of $C^1$ actions of the groups $G_\lambda$ introduced above. Finally, in Section \[s:C2\] we use different techniques that work in $C^2$ regularity. Some definitions and notation {#s:Defs} ============================= Let $M$ be a manifold and ${\mathsf{Homeo}}(M)$ the group of homeomorphisms of $M$. A subgroup $G\subset {\mathsf{Homeo}}(M)$ is $C^r$-smoothable ($r\ge 1$) if it is conjugate in ${\mathsf{Homeo}}(M)$ to a subgroup in ${\mathsf{Diff}}^r(M)$, the group of $C^r$ diffeomorphisms of $M$. \[r:abstract\] Even if a certain subgroup $G\subset {\mathsf{Homeo}}(M)$ is not $C^r$-smoothable, it is still possible that the group $G$, as abstract group, admits $C^r$ actions on the manifold $M$. Throughout this work we shall only be concerned by one-dimensional manifolds. We restrict our discussion to orientation-preserving homeomorphisms, which form a subgroup ${\mathsf{Homeo}}_+(M)$ of index two in ${\mathsf{Homeo}}(M)$. We will not make much distinction between the groups ${\mathsf{Homeo}}_+({\mathbb{R}})$ and ${\mathsf{Homeo}}_+([0,1])$. Notice however that the groups ${\mathsf{Diff}}_+^r({\mathbb{R}})$ and ${\mathsf{Diff}}^r_+([0,1])$ are different and for this reason we sometimes identify the interval $[0,1]$ to the compactified real line $[-\infty,+\infty]$. Choosing the affine chart $t\mapsto [t:1]$, we consider ${\mathbb{R}}$ as the affine line in the projective space ${\mathbb{R}}{\mathsf{P}^1}\cong{\mathbb{R}}\cup \{\infty\}$, which is topologically the circle ${\mathbb{S}^1}$. The group ${\mathsf{Homeo}}_+({\mathbb{R}})$ can be identified to a subgroup of ${\mathsf{Homeo}}_+({\mathbb{S}^1})$, for instance as the stabiliser of the point $\infty$ of ${\mathbb{S}^1}\cong{\mathbb{R}}{\mathsf{P}^1}$. The projective special linear group ${\mathsf{PSL}}(2,{\mathbb{R}})= {\mathsf{SL}}(2,{\mathbb{R}})/\{\pm id\}$ naturally acts on the projective real line ${\mathbb{R}}{\mathsf{P}^1}$ by Möbius transformations: from now on, we shall always suppose that ${\mathsf{PSL}}(2,{\mathbb{R}})$ acts on the circle in this way. A circle homeomorphism $h\in{\mathsf{Homeo}}_+({\mathbb{R}}{\mathsf{P}^1})$ is piecewise projective if there exists a finite partition ${\mathbb{R}}{\mathsf{P}^1}=I_1\cup\ldots \cup I_\ell$ of the circle into intervals, such that every restriction $h\vert_{I_k}$, $k=1,\ldots,\ell$, coincides with the restriction of a Möbius transformation. A breakpoint of $h$ is a point $b\in{\mathbb{R}}{\mathsf{P}^1}$ such that the restriction of $h$ to any neighbourhood of $b$ does not coincide with the restriction of a Möbius transformation. The group of all orientation-preserving piecewise projective homeomorphisms of the circle is denoted by ${\mathsf{P}}{\mathsf{P}}_+({\mathbb{R}}{\mathsf{P}^1})$. Similarly, we define the group of piecewise projective homeomorphisms of the real line ${\mathsf{P}}{\mathsf{P}}_+({\mathbb{R}})$, identifying it to the stabiliser of $\infty$ inside ${\mathsf{P}}{\mathsf{P}}_+({\mathbb{R}}{\mathsf{P}^1})$. We recall that a fixed point $p\in{\mathbb{R}}$ for a diffeomorphism $f\in{\mathsf{Diff}}^1_+({\mathbb{R}})$ is a hyperbolic fixed point if $f$ has derivative at $p$ which is not $1$. We shall say that a subgroup $G\subset {\mathsf{Diff}}^1_+({\mathbb{R}})$ has hyperbolic fixed points if there exists an element $f\in G$ with hyperbolic fixed points. This notion is related to the notion of hyperbolic elements in ${\mathsf{PSL}}(2,{\mathbb{R}})$. A nontrivial projective transformation in ${\mathsf{PSL}}(2,{\mathbb{R}})$ has at most two fixed points. If it has exactly two fixed points, it is called hyperbolic, and if it has only one fixed point it is called parabolic. A matrix $M$ in $\mathsf{SL}(2,{\mathbb{R}})$ is hyperbolic if $\|\operatorname{\mathsf{Tr}}(M)\|>2$, parabolic if $\|\operatorname{\mathsf{Tr}}(M)\|=2$ and elliptic if $\|\operatorname{\mathsf{Tr}}(M)\|<2$. Then the corresponding projective transformation is respectively hyperbolic, parabolic and elliptic. Given a subgroup $\Gamma\subset {\mathsf{PSL}}(2,{\mathbb{R}})$, we say that a real $r\in \mathbb{R}$ is a hyperbolic fixed point for $\Gamma$ if there is a $\gamma\in \Gamma$ such that $\gamma$ is hyperbolic and $\gamma(r)=r$. Similarly, we define the notion of a parabolic fixed point for $\Gamma$. We consider the sets ${\mathcal{H}}_\Gamma$ and ${\mathcal{P}}_\Gamma$ of hyperbolic fixed points and parabolic fixed points of elements of $\Gamma$, respectively. When $\Gamma={\mathsf{PSL}}(2,A)= {\mathsf{SL}}(2,A)/\{\pm Id\}$, for some subring $A\subset {\mathbb{R}}$, we simply write ${\mathcal{H}}_A$ and ${\mathcal{P}}_A$. Here ${\mathsf{SL}}(2,A)$ is the group of invertible $(2\times 2)$-matrices with determinant $1$ and coefficients in $A$. Let $\lambda\in {\mathbb{R}}$ be an algebraic real number of degree $d$ over ${\mathbb{Q}}$, and let $p_\lambda(t)=\frac{{\alpha}_0}{{\alpha}_d}+\frac{{\alpha}_1}{{\alpha}_d}t+\ldots +\frac{{\alpha}_{d-1}}{{\alpha}_d}t^{d-1}+t^d$, ${\alpha}_j\in {\mathbb{Z}}$, denote the associated minimal polynomial. The field ${\mathbb{Q}}(\lambda)$ is a ${\mathbb{Q}}$-vector space of dimension $d$, for which we fix $\{1,\lambda,\ldots,\lambda^{d-1}\}$ as preferred basis. With respect to this basis, multiplication by $\lambda$ on ${\mathbb{Q}}(d)$ is represented by the matrix $$\label{Companion} C_\lambda=\left (\begin{array}{ccc\|c} 0&\cdots&0 & -{\alpha}_0/{\alpha}_d \\ \hline &&& -{\alpha}_1/{\alpha}_d \\ &I_{d-1}& & \vdots \\ &&& -{\alpha}_{d-1}/{\alpha}_d \end{array}\right )$$ which is commonly named the Frobenius companion matrix of $\lambda$. When $\lambda\neq 0$ then ${\alpha}_0\neq 0$, so that $C_\lambda$ is an invertible $d\times d$ matrix with rational coefficients. The minimal polynomial of $C_\lambda$ is exactly $p_\lambda$, so the eigenvalues of $C_\lambda$ are exactly the Galois conjugates of $\lambda$, that is, all (complex) roots of $p_\lambda$. \[HyperbolicAlgebraic\] A nonzero real number $\lambda\in {\mathbb{R}}$ is Galois hyperbolic if it is algebraic and the companion matrix $C_\lambda$ has no eigenvalue of absolute value $1$. Equivalently, this means that all the Galois conjugates of $\lambda$ do not have absolute value $1$. For instance, any rational $\lambda\neq 0,\pm 1$ is Galois hyperbolic, as well as any quadratic integer $\sqrt{m}\neq 0,1$, $m\in {\mathbb{N}}$. However not every real number is Galois hyperbolic. As an explicit nontrivial example [@hyperbolic § 5], the polynomial $p(t)=1+4t+4t^2+4t^3+t^4$ is irreducible over ${\mathbb{Q}}$, has two positive real roots $\lambda$, $1/\lambda$ and two roots of absolute value $1$. Hence $\lambda,1/\lambda$ are not Galois hyperbolic. [[Theorem \[t:main1\] holds for this more general class of numbers.]{}]{} \[t:main1Hyp\] For any Galois hyperbolic $\lambda>1$ and any morphism $\rho\colon G_\lambda \to {\mathsf{Diff}}^1_+([0,1])$ one has : the commutator $[g,h_\lambda gh_\lambda^{-1}]$ belongs to the kernel of $\rho$. In particular $\rho$ cannot be injective. The same holds for any morphism $\varphi\colon G_\lambda\to {\mathsf{Diff}}^1_+({\mathbb{S}^1})$, where ${\mathbb{S}^1}$ is the circle. The mechanisms {#s:Mech} ============== The aim of this work is to present three different techniques which provide a variety of examples of non-smoothable groups in ${\mathsf{P}}{\mathsf{P}}_+({\mathbb{R}})$. The three techniques rely on the rigid hyperbolicity of the actions: there are subgroups $G\subset {\mathsf{Diff}}_+^1({\mathbb{R}})$ such that, no matter how one (topologically) conjugates them inside ${\mathsf{Diff}}^1_+({\mathbb{R}})$, will always have hyperbolic fixed points. More precisely, suppose that in $G\subset {\mathsf{Diff}}_+^1({\mathbb{R}})$ there is an element $f$ having a hyperbolic fixed point $p\in{\mathbb{R}}$. Consider another subgroup $\widetilde G\subset {\mathsf{Diff}}_+^1({\mathbb{R}})$ to which $G$ is topologically conjugate by some homeomorphism $\phi$: $\phi G\phi^{-1}=\widetilde G$. The point $\phi(p)$ is a fixed point for $\phi f \phi^{-1}$, but since $\phi$ is just a homeomorphism, we cannot ensure that it is a hyperbolic fixed point. However, there are some topological mechanisms that guarantee hyperbolicity. The first one is when there are linked pairs of fixed points in $G$. We now define this notion. Denote by ${\mathsf{Fix}}(g)$ the set of fixed points of a homeomorphism $g$. A pair of successive fixed points of $G$ is a pair $a,b\in {\mathbb{R}}$, $a<b$ such that there is an element $g\in G$ and $(a,b)$ is a connected component of ${\mathbb{R}}\setminus {\mathsf{Fix}}(g)$. A linked pair of fixed points consists of pairs $a,b$ and $c,d$ such that: 1) there are elements $f,g\in G$ such that $a,b$ is a pair of successive fixed points of $f$ and $c,d$ is a pair of successive fixed points of $g$; 2) either $\{a,b\}\cap (c,d)$ or $(a,b)\cap \{c,d\}$ is a point. In this case hyperbolicity is obtained by a probabilistic argument. Some element $h$ in the semigroup generated by $f$ and $g$ will have a hyperbolic fixed point somewhere. This is the so-called Sacksteder’s Theorem, in its version for $C^1$-pseudogroups [@DKNacta; @navas-book]. This method applies to large groups of piecewise projective homeomorphisms, as the Monod’s groups (see Definition \[d:Monod\]): \[t:H(A)\] The following holds for Monod’s groups $H(A)$ and $G(A)$. (1) \[i:Monod1\] For any subring $A\subset {\mathbb{R}}$, Monod’s groups $H(A)$ and $G(A)$ are not $C^1$-smoothable. (2) \[i:Monod2\] [[If $A$ contains $\sqrt{\lambda}^{\pm 1}$ for some Galois hyperbolic $\lambda>1$,]{}]{} then there exists no injective morphism $\rho:H(A)\to {\mathsf{Diff}}_+^1([0,1])$. [[ Condition on $A$ is equivalent to the fact that the group ${\mathsf{PSL}}(2,A)$ contains the homothety $$f_\lambda\colon x\mapsto \lambda x,$$ representing the matrix $\begin{pmatrix} \sqrt{\lambda} & 0 \\ 0 & \sqrt{\lambda}^{-1} \end{pmatrix}$.]{}]{} The second one is when there is an exponential growth of orbits. In this case we can ensure that a specific point is always a hyperbolic fixed point. This applies for example to the dyadic affine group $\langle t\mapsto t+1, t\mapsto 2t\rangle$, which is isomorphic to the solvable Baumslag-Solitar group $\mathsf{BS}(1,2)$, as described in [@hyperbolic]. From this, it is easy to build examples of finitely generated groups in ${\mathsf{P}}{\mathsf{P}}_+({\mathbb{R}})$ which are not $C^1$-smoothable. This method applies to the finitely presentable Lodha-Moore group (see § \[sc:LM\]), for which we do not only prove that its action is not $C^1$-smoothable, but also that it has no nontrivial $C^1$ action on the interval: \[t:LM\] Every morphism from the Lodha-Moore group $G_0$ to ${\mathsf{Diff}}_+^1([0,1])$ has abelian image. The third one relies on the nature of stabilisers, and here we require that the regularity of the group $G$ is $C^2$. If there exists a point $x\in{\mathbb{R}}$ such that the (right, for instance) germs of elements $g\in G$ fixing $x$ define a group which is dense in ${\mathbb{R}}$, then we can use the Szekeres vector field to have a well-defined local differentiable structure, by means of which we ensure that the hyperbolic nature of a fixed point cannot change after topological conjugacy to another $C^2$ action. This method applies to examples of groups in ${\mathsf{P}}{\mathsf{P}}_+({\mathbb{R}})$ that are naturally in ${\mathsf{Diff}}^1_+({\mathbb{R}})$, e.g. the group generated by Thompson’s group $F$ (which is $C^1$ in ${\mathsf{P}}{\mathsf{P}}_+({\mathbb{R}})$) together with $t\mapsto t+\frac12$, for which we establish that their actions are not $C^2$-smoothable. It also applies to the Thompson-Stein groups $F(n_1,\ldots, n_k)$ and $T(n_1,\ldots, n_k)$ (see Definition \[d:TS\]), extending a previous work by Liousse [@Liousse]: \[t:F23\] The Thompson-Stein groups $F(n_1,\ldots, n_k)$, $k\ge 2$, are not $C^2$-smoothable. \[t:T23\] (1) Assume $n_1=2$. The Thompson-Stein groups $T(2,n_2,\ldots,n_k)$, $k\ge 2$, have no faithful $C^2$ action on ${\mathbb{S}^1}$. (2) Assume $n_1=2,n_2=3$. Every $C^2$ action of $T(2,3,n_3,\ldots,n_k)$ on ${\mathbb{S}^1}$ is trivial. This holds in particular for $T(2,3)$. Historical motivations {#s:Motiv} ====================== Thompson’s groups $F$ and $T$ ----------------------------- In the 50s, Richard J. Thompson introduced three groups $F$, $T$ and $V$, which have many nice properties (cf. [@CFP]). These groups are finitely presented and $[F,F]$, $T$, $V$ are simple. They have been among the first known examples sharing these properties. Since only $F$ and $T$ act by homeomorphisms on the circle, we restrict our attention to them. Thompson’s group $T$ is the group of all piecewise linear homeomorphisms of the circle ${\mathbb{S}^1}\cong {\mathbb{R}}/{\mathbb{Z}}$ such that all derivatives are powers of $2$ and the breakpoints are dyadic rationals, i.e. points of the form $p/2^q$, $p,q\in{\mathbb{N}}$. Thompson’s group $F$ is the stabiliser of the point $0$ in $T$. It has been proved by Ghys and Sergiescu in [@GS] that the piecewise linear action of $T$ (and hence of $F$) on ${\mathbb{S}^1}$ is $C^\infty$-smoothable. On the other side, it is “not difficult” to find $C^\infty$ faithful actions (a priori not topologically conjugate to the standard one) of Thompson’s group. We recall Thurston’s interpretation of $T$ as a group of piecewise projective homeomorphisms of ${\mathbb{R}}{\mathsf{P}^1}$ (cf. [@CFP]). \[ThurstonT\] $T$ is the group of piecewise ${\mathsf{PSL}}(2,{\mathbb{Z}})$ homeomorphisms of ${\mathbb{R}}{\mathsf{P}^1}$ with breakpoints in ${\mathcal{P}}_{\mathbb{Z}}$ (which is the set of rational numbers together with the point at infinity). $T$ is generated by ${\mathsf{PSL}}(2,{\mathbb{Z}})$ and an additional element $c$ defined as $$c(t)=\begin{cases} t&\text{if }t\in [\infty,0],\\ \frac{t}{1-t}&\text{if }0\leq t\leq\frac{1}{2},\\ 3-\frac{1}{t}&\text{if }\frac{1}{2}\leq t\leq 1,\\ t+1&\text{if }t\in [1,\infty]. \end{cases}$$ It is particularly striking that the element $c$ has continuous first derivative. As the action of ${\mathsf{PSL}}(2,{\mathbb{Z}})$ is even real-analytic, Thurston’s interpretation gives a natural $C^1$-smoothing of $T$.[^5] In this model, $F$ is the group of piecewise ${\mathsf{PSL}}(2,{\mathbb{Z}})$ homeomorphisms of ${\mathbb{R}}{\mathsf{P}^1}$ with breakpoints in ${\mathcal{P}}_{\mathbb{Z}}$, that also fix infinity. So $F$ is the stabiliser of $\infty$ in $T$. $F$ is generated by $t\mapsto t+1$ together with $c$ from above. Recall that the group ${\mathsf{PSL}}(2,{\mathbb{Z}})$ is isomorphic to the free product ${\mathbb{Z}}_2{\mathbb{Z}}_3$, freely generated by the order two element $a: t\mapsto -\frac{1}{t}$ and the order three element $b: t\mapsto \frac{1}{1-t}$. Now we sketch a proof that $F$ admits a $C^{\infty}$ action inspired by [@chains] (see also [@fast]). Note that this is weaker than proving it is $C^{\infty}$-smoothable, which is a consequence of the theorem of Ghys-Sergiescu. Given any homeomorphism $h:[0,1]\to [0,2]$, if we define the element $$\widetilde c(t)=\begin{cases} t&\text{if }t\in [\infty,0],\\ h(t)&\text{if }t\in [0,1],\\ t+1&\text{if }t\in [1,\infty] \end{cases}$$ then the group generated by $t\to t+1$ and $\widetilde c$ is isomorphic to $F$. If we choose $h$ to be $C^\infty$, infinitely tangent to the identity at $0$ and to $t\mapsto t+1$ at $1$, then the modified element $\widetilde c$ is $C^\infty$. The algebraic properties of $F$ guarantee that the group generated by $t\to t+1$ and $\widetilde c$ is isomorphic to $F$.[^6] However, it is not guaranteed that one can choose $h$ and hence $\widetilde c$ such that the action of the group $\langle t\to t+1,\widetilde c\rangle$ is actually conjugate* to the standard action of $F$. A very important remark is that this strategy is morally possible because $0$ and $1$ are not hyperbolic fixed points (they are parabolic). This allows to slow-down the dynamics near these points and make $c$ infinitely tangent to the identity. This feature already appeared in the work of Ghys and Sergiescu. Hyperbolicity is a typical obstruction for such modifications in differentiable dynamics. One open problem: The Day-von Neumann problem for ${\mathsf{Diff}}^{2}_+({\mathbb{R}})$ --------------------------------------------------------------------------------------- One of the main motivations for our work is understanding amenable groups of diffeomorphisms of the circle. There are several equivalent definitions of amenability and an extensive literature on the topic (see [@tullio] for an elementary introduction). We provide one definition: A discrete group $G$ is amenable if it admits a finitely additive, left translation invariant probability measure. Here is an equivalent definition, à la Krylov-Bogolyubov, which is more natural from the viewpoint of dynamical systems: A discrete group $G$ is amenable if every continuous action on a compact space has an invariant probability measure. The class of amenable groups includes finite, abelian and solvable groups. Amenability is closed under extensions, products, direct unions and quotients. Subgroups of amenable groups are amenable. On the other hand, groups containing nonabelian free subgroups are nonamenable. The so-called Day-von Neumann problem (popularized by Day in the 50s) is about the converse statement: does every nonamenable group contain a nonabelian free subgroup? If one restricts the question to linear groups, then the well-known Tits alternative gives a positive answer: any non virtually solvable linear group contains nonabelian free subgroups. The problem has been solved with negative answers and currently various negative solutions are known. These include Tarski monsters, Burnside groups, and Golod-Shafarevich groups. In this article we are interested in a particular class of such groups, discovered by Monod [@monod] and Lodha-Moore [@LodhaMoore], which are subgroups of ${\mathsf{P}}{\mathsf{P}}_+({\mathbb{R}})$. Among them, there are examples that are in ${\mathsf{Diff}}_+^1({\mathbb{R}})$. For instance, the group generated by $t\to t+\frac{1}{2}$ together with the element $c$ from Definition \[ThurstonT\] above provides such an example. Interestingly, no negative solution to the Day-von Neumann problem is known among subgroups of ${\mathsf{Diff}}^2_+({\mathbb{R}})$. Motivated by this question, in this work we prove (Theorems \[t:Firr\] and \[t:Frat\]) that the natural actions of these groups are not $C^2$-smoothable. However, we have to stress that a priori there could be smooth actions of such nonamenable groups that are not topologically conjugate to the standard actions (cf. Remark \[r:abstract\]). The moral consequence of our results is that the Day-von Neumann problem in ${\mathsf{Diff}}_+^2({\mathbb{R}})$ is strictly harder than in ${\mathsf{Diff}}^1_+({\mathbb{R}})$. This is not so surprising, since there are important differences between $C^2$ and $C^1$ diffeomorphisms in one-dimensional dynamics. We end this section by recalling a couple of tantalising longstanding open questions in this direction. Is $F$ amenable? Does the Tits alternative hold for the group of real-analytic diffeomorphisms of the real line? A second open problem: Higher rank behaviour -------------------------------------------- \[d:TS\] Let $1<n_1<\cdots<n_k$ be natural numbers such that the group $\Lambda=\langle n_i\rangle\subset {\mathbb{R}}_+^$ is an abelian group of rank $k$. Denote by $A$ the ring ${\mathbb{Z}}[\frac1m]$, where $m$ is the least common multiple of the $n_i$’s. Thompson-Stein’s group* $T(n_1,\ldots,n_k)$ is the group of all piecewise linear homeomoprhisms of the circle ${\mathbb{S}^1}\cong{\mathbb{R}}/{\mathbb{Z}}$ such that all derivatives are in $\Lambda$ and the breakpoints are in $A$. Thompson-Stein’s group $F(n_1,\ldots,n_k)$ is the stabiliser of the point $0$ in $T(n_1,\ldots,n_k)$. With the above definition, the group $T(2)$ is the classical Thompson’s group $T$. It has been proved by Stein [@stein] that these groups share many group-theoretical properties with the classical Thompson’s groups, such as being finitely presentable (cf. [@bieri-strebel]). However there are important differences from the dynamical viewpoint. In [@minakawa1; @minakawa2], Minakawa discovers that $\mathsf{PL}_+({\mathbb{S}^1})$ contains “exotic circles”, namely topological conjugates of $\mathsf{SO}(2)$ that are not one-parameter groups inside $\mathsf{PL}_+({\mathbb{S}^1})$, in the sense that they are not $\mathsf{PL}$ conjugates of $\mathsf{SO}(2)$. In particular, Liousse shows in [@Liousse] that $T(n_1,\ldots,n_{k})$ contains an abelian group of rank $k-1$ that is contained in a topological conjugate of $\mathsf{SO}(2)$, but not in a $\mathsf{PL}$ conjugate of $\mathsf{SO}(2)$. Whence Navas suggested the following: Does $T(n_1,\ldots,n_k)$, $k\ge 2$, satisfy Kazhdan’s property $(T)$?[^7] On the other hand Navas proved in [@navas(T)] that the only groups of $C^{r}$ diffeomorphisms, $r> 3/2$, that have property $(T)$ are finite. Focusing our attention on one particular example, in [@Liousse], Liousse proves, among other things, that every action of $T(2,3)$ on ${\mathbb{S}^1}$ by $C^9$ diffeomorphisms is trivial and with Corollary \[t:T23\] we improve this result to $C^2$ regularity. It would be very interesting to prove that $T(2,3)$ has no $C^1$ action on the circle, as this would confirm that this group is a good candidate for finding an infinite Kazhdan group of circle homeomorphisms. Naturally, there could be also good candidates among groups of piecewise projective homeomorphisms. Nonamenable groups of piecewise projective homeomorphisms {#s:Monod} ========================================================= Monod’s groups -------------- Generalizing a well-known result by Brin and Squier [@brin-squier], Monod showed in [@monod] that ${\mathsf{P}}{\mathsf{P}}_+({\mathbb{R}})$ does not contain nonabelian free subgroups. One key feature is that given any $r\in \mathbb{R}$, the group of germs of elements in ${\mathsf{P}}{\mathsf{P}}_+({\mathbb{R}})$ fixing the point $r$ is isomorphic to the affine group. \[d:Monod\] Let $A$ be a subring of ${\mathbb{R}}$. $G(A)$ is defined as the group of all piecewise ${\mathsf{PSL}}(2,A)$ homeomorphisms of the circle with breakpoints in ${\mathcal{H}}_A$. The group $H(A)$ is the stabiliser of $\infty$ inside $G(A)$. Observe that the groups $G({\mathbb{R}})$ and $H({\mathbb{R}})$ coincide with ${\mathsf{P}}{\mathsf{P}}_+({\mathbb{R}}{\mathsf{P}^1})$ and ${\mathsf{P}}{\mathsf{P}}_+({\mathbb{R}})$ respectively. Relying on the fact that for any $A\neq {\mathbb{Z}}$, the group ${\mathsf{PSL}}(2,A)$ contains dense free subgroups, Monod proved in [@monod] that for any $A\neq {\mathbb{Z}}$, the group $H(A)$ is nonamenable. Therefore these groups give negative answer to the Day-von Neumann problem. The previous definition can be generalized, considering any subgroup $\Gamma\subset {\mathsf{PSL}}(2,{\mathbb{R}})$. Elements in $G(\Gamma)$ are piecewise $\Gamma$ and the breakpoints are in ${\mathcal{H}}_\Gamma$. For any non-discrete $\Gamma\subset{\mathsf{PSL}}(2,{\mathbb{R}})$, the group $H(\Gamma)$ does not contain free subgroups and is nonamenable. Theorem \[t:H(A)\] can be extended to these groups as well. We shall now demonstrate Theorem \[t:H(A)\], namely that Monod’s examples are not $C^1$-smoothable. For part , it is enough to prove the following: \[t:H(Z)\] Monod’s group $H({\mathbb{Z}})$ is not $C^1$-smoothable. On the other hand, part relies on Theorem \[t:main1Hyp\]. Let us first prove . Any subring $A\subset {\mathbb{R}}$ contains ${\mathbb{Z}}$, therefore ${\mathsf{PSL}}(2,{\mathbb{Z}})$ is a subgroup of any ${\mathsf{PSL}}(2,A)$. Therefore we have inclusions $H({\mathbb{Z}})\subset H(A)\subset G(A)$. As $H({\mathbb{Z}})$ is not $C^1$-smoothable (Theorem \[t:H(Z)\]), neither are $H(A)$ and $G(A)$. Next, we demonstrate part . Let $\lambda>1$ be a [[Galois hyperbolic]{}]{} number such that $$f_\lambda:x\mapsto \lambda x$$ belongs to ${\mathsf{PSL}}(2,A)$. As ${\mathbb{Z}}\subset A$, the translation $g:x\mapsto x+1$ belongs to ${\mathsf{PSL}}(2,A)$ as well. Moreover, $f_\lambda$ being a hyperbolic element in ${\mathsf{PSL}}(2,A)$, we have that ${\mathcal{H}}_A$ contains its fixed point $0$. Therefore Monod’s group $H(A)$ contains the piecewise defined element $$h_\lambda:x\mapsto\begin{cases} x&\text{if }x\le 0, \\ \lambda x&\text{if }x>0. \end{cases}$$ We have just shown that $G_\lambda=\langle f_\lambda,g,h_\lambda\rangle$ is a subgroup of $H(A)$. Let $\rho:H(A)\to {\mathsf{Diff}}_+^1([0,1])$ be a representation. Theorem \[t:main1Hyp\] implies that $[g,h_\lambda gh_\lambda^{-1}]$ is in the kernel of $\rho$. Therefore $\rho$ cannot be injective, as desired. The dynamical ingredient we need for Theorem \[t:H(Z)\] is the following Sacksteder-like result, originally due to Deroin, Kleptsyn, Navas [@DKNacta] (see [@navas-book Prop. 3.2.10] and also [@centralizers § 4.5] for a simplified proof). \[p:linked\_hyper\] Let $G=\langle f,g \rangle$ be a group acting by orientation-preserving $C^1$ diffeomorphisms on a compact one-dimensional manifold. If $\{a,b\}$ and $\{c,d\}$ are linked pairs of successive fixed points for $f,g$, then $G$ contains an element with a hyperbolic fixed point in $(a,b)\cap(c,d)$. Let us assume by way of contradiction that there is a homeomorphism $\phi:{\mathbb{R}}\to {\mathbb{R}}$ such that $G:=\phi H({\mathbb{Z}}) \phi^{-1}$ is a group of $C^{1}$ diffeomorphisms of $\mathbb{R}$. First we observe that there are elements $f,g\in H({\mathbb{Z}})$ that have linked pairs of fixed points. For example, consider the hyperbolic element $\gamma$ defined as the projective transformation $$\gamma=\begin{bmatrix} 2 & -1 \\ -1 & 1 \end{bmatrix},$$ whose fixed points $a,b$ satisfy that $a<-\frac{3}{2}<\frac{1}{2}<b$. Now define $$f(t)=\begin{cases} t&\text{if }t\notin [a,b],\\ \gamma(t)&\text{if }t\in [a,b], \end{cases} \qquad g(t)=f(t-1)+1.$$ Note that the pairs $a,b$ and $a+1,b+1$ are linked.[^8] Now the elements $$f_1=\phi f\phi^{-1},\qquad g_1=\phi g\phi^{-1}$$ in $G$ have fixed points $\phi(a),\phi(b)$ and $\phi(c),\phi(d)$ respectively. This forms a linked pair. By Proposition \[p:linked\_hyper\], there is an element $g\in G$ with a fixed point such that the derivative of $g$ at $x$ is not equal to $1$. Now let $g_1=\phi^{-1} g \phi$ be the corresponding element in $H({\mathbb{Z}})$. Note that $g_1$ fixes $y=\phi^{-1}(x)$. We claim that $y$ is a fixed point of a hyperbolic matrix in ${\mathsf{PSL}}(2,{\mathbb{Z}})$. If $y$ is a breakpoint of $g_1$, then this is true because the set of breakpoints of elements in $H({\mathbb{Z}})$ is exatly ${\mathcal{H}}_{{\mathbb{Z}}}$. We consider the case when $y$ is not a breakpoint of $g_1$, so there exists an element $\gamma_1\in{\mathsf{PSL}}(2,{\mathbb{Z}})$ whose restriction to a neighbourhood $U$ of $y$ coincides with the restriction $g_1\vert_U$. Observe that since $x$ is a hyperbolic fixed point for $g$, the corresponding point $y$ must be a topological attractor or repellor for $\gamma_1\in{\mathsf{PSL}}(2,{\mathbb{Z}})$ that acts locally like $g_1$ around $y$, and hence $g_1$ must be hyperbolic and $y$ is hence a hyperbolic fixed point for ${\mathsf{PSL}}(2,{\mathbb{Z}})$. Now consider an element $g_2\in H({\mathbb{Z}})$ which is the identity on $(-\infty, y)$ and agrees with $g_1$ on $[y,\infty)$. Then $g_3=\phi g_2\phi^{-1}\in G$ has right derivative $\lambda\neq 1$ at $x$ and a left derivative that equals $1$ at $x$. This contradicts the assumption that $g_3$ is $C^1$. Hence our original assumption that $H({\mathbb{Z}})$ is $C^1$-smoothable must be false. The Lodha-Moore example {#sc:LM} ----------------------- Lodha and Moore constructed a finitely presented subgroup $G_0$ of Monod’s group. This example provides the first torsion free finitely presentable example solving the Day-von Neumann problem. The group $G_0$ is generated by $t\mapsto t+1$ together with the following two homeomorphisms of ${\mathbb{R}}$: $$c(t)= \begin{cases} t&\text{ if }t\leq 0,\\ \frac{t}{1-t}&\text{ if }0\leq t\leq \frac{1}{2},\\ 3-\frac{1}{t}&\text{ if }\frac{1}{2}\leq t\leq 1,\\ t+1&\text{ if }1\leq t,\\ \end{cases} \qquad d(t)= \begin{cases} \frac{2t}{1+t}&\text{ if }0\leq t\leq 1,\\ t&\text{ if } t\notin [0,1].\\ \end{cases}$$ The following was proved in [@LodhaMoore]: The group $G_0$ is nonamenable and does not contain nonabelian free subgroups. Moreover, it is finitely presentable with $3$ generators and $9$ relations. In [@LodhaMoore] a combinatorial model for $G_0$ is constructed by means of a faithful action of $G_0$ by homeomorphisms of the Cantor set $\{{\mathtt{0}},{\mathtt{1}}\}^{\mathbb{N}}$. This model was used to prove that $G_0$ is finitely presentable. Here $\{{\mathtt{0}},{\mathtt{1}}\}^{{\mathbb{N}}}$ is the Cantor set of infinite binary sequences, viewed as the boundary of the infinite rooted binary tree. We denote by $\{{\mathtt{0}},{\mathtt{1}}\}^{<{\mathbb{N}}}$ as the set of all finite binary sequences, which are addresses of nodes in the infinite rooted binary tree. Consider the map $\Phi: \{{\mathtt{0}},{\mathtt{1}}\}^{\mathbb{N}}\to {\mathbb{R}}\cup \{\infty\}$ given by: $${\mathtt{1}}{\mathtt{1}}^{a_0}{\mathtt{0}}^{a_1}1^{a_2}...\mapsto a_0+\frac{1}{a_1+\frac{1}{a_2+\ldots}},\qquad {\mathtt{0}}{\mathtt{0}}^{a_0}{\mathtt{1}}^{a_1}{\mathtt{0}}^{a_2}...\mapsto -\left (a_0+\frac{1}{a_1+\frac{1}{a_2+\ldots}}\right ).$$ This function is one-to-one except on sequences $\xi$ which are eventually constant. On sequences which are eventually constant, the map is two-to-one: $\Phi(s{\mathtt{0}}{\mathtt{1}}^{\infty}) = \Phi(s{\mathtt{1}}{\mathtt{0}}^{\infty})$ and $\Phi({\mathtt{0}}^{\infty}) = \Phi({\mathtt{1}}^{\infty}) = \infty$. It was shown in [@LodhaMoore] that upon conjugating $a,b,c$ by $\Phi$ one obtains the following combinatorial model. We start with the following map $x:\{{\mathtt{0}},{\mathtt{1}}\}^{{\mathbb{N}}}\to \{{\mathtt{0}},{\mathtt{1}}\}^{{\mathbb{N}}}$ as: $$\begin{aligned} x({\mathtt{0}}{\mathtt{0}}\xi)&={\mathtt{0}}\xi,\\ x({\mathtt{0}}{\mathtt{1}}\xi)&={\mathtt{1}}{\mathtt{0}}\xi,\\ x({\mathtt{1}}\xi)&={\mathtt{1}}{\mathtt{1}}\xi\end{aligned}$$ and also, recursively, the pair of mutually inverse maps $y,y^{-1}:\{{\mathtt{0}},{\mathtt{1}}\}^{{\mathbb{N}}}\to \{{\mathtt{0}},{\mathtt{1}}\}^{{\mathbb{N}}}$ as: $$\begin{aligned} y({\mathtt{0}}{\mathtt{0}}\xi)&={\mathtt{0}}y(\xi),&y^{-1}({\mathtt{0}}\xi)&={\mathtt{0}}{\mathtt{0}}y^{-1}(\xi),\\ y({\mathtt{0}}{\mathtt{1}}\xi)&={\mathtt{1}}{\mathtt{0}}y^{-1}(\xi),&y^{-1}({\mathtt{1}}{\mathtt{0}}\xi)&={\mathtt{0}}{\mathtt{1}}y(\xi),\\ y({\mathtt{1}}\xi)&={\mathtt{1}}{\mathtt{1}}y(\xi),&y^{-1}({\mathtt{1}}{\mathtt{1}}\xi)&={\mathtt{1}}y^{-1}(\xi).\end{aligned}$$ From these functions, we define the functions $x_{s},y_s:\{{\mathtt{0}},{\mathtt{1}}\}^{{\mathbb{N}}}\to \{{\mathtt{0}},{\mathtt{1}}\}^{{\mathbb{N}}}$ for $s \in \{{\mathtt{0}},{\mathtt{1}}\}^{<{\mathbb{N}}}$ which act as $x$ and $y$ localised to binary sequences which extend $s$: $$x_s(\xi) = \begin{cases} s x(\eta) & \textrm{ if } \xi = s \eta, \\ \xi & \textrm{otherwise}, \end{cases} \qquad y_s(\xi) = \begin{cases} s y(\eta) & \textrm{ if } \xi = s \eta, \\ \xi & \textrm{otherwise}. \end{cases}$$ If $s$ is the empty-string, it will be omitted as a subscript. The group $G_0$ is generated by functions in the set $$S=\left \{x_t,y_s\mid s,t\in \{{\mathtt{0}},{\mathtt{1}}\}^{<{\mathbb{N}}},s\neq {\mathtt{0}}^k,s\neq {\mathtt{1}}^k, s\neq \emptyset\right \}$$ In fact, $G_0$ is generated by $x,x_{{\mathtt{1}}},y_{{\mathtt{1}}{\mathtt{0}}}$ which correspond respectively to conjugates of the functions $a,b,c$ defined above by $\Phi$. (See [@LodhaMoore] for details.) It is important to note that $G_0$ acts on the boundary of the infinite rooted binary tree, but not on the tree itself. Recall from the introduction that we are denoting by $G_2$ the group generated by $f_2,g,h_2$, where $f_2$ is the scalar multiplication by $2$, $g$ is the translation by $1$, and $h_2$ is the element which agrees with $f_2$ to the right of zero and is the identity elsewhere. We obtain the following obstruction to smoothability of $G_0$. \[l:bbs\_in\_G0\] The three elements $y_{{\mathtt{1}}{\mathtt{0}}{\mathtt{0}}}^{-1}y_{{\mathtt{1}}{\mathtt{0}}{\mathtt{1}}},y_{{\mathtt{1}}{\mathtt{0}}{\mathtt{1}}}$ and $x_{{\mathtt{1}}{\mathtt{0}}}$ generate an isomorphic copy of $G_2$ in the Lodha-Moore group $G_0$. It was demonstrated in [@LodhaMoore] that the elements $x$ and $y_{{\mathtt{0}}}^{-1}y_{\mathtt{1}}$ are conjugate respectively to $t\mapsto t+1$ and $t\mapsto 2t$ by $\Phi$. Hence they generate an isomorphic copy of ${\mathsf{BS}}(1,2)$. In particular, $y_{{\mathtt{0}}}^{-1}y_{{\mathtt{1}}},y_{{\mathtt{1}}},x_{{\mathtt{1}}{\mathtt{0}}}$ generate an isomorphic copy of $G_2$. It is easy to see that the groups $\left \langle y_{{\mathtt{1}}{\mathtt{0}}{\mathtt{0}}}^{-1}y_{{\mathtt{1}}{\mathtt{0}}{\mathtt{1}}},y_{{\mathtt{1}}{\mathtt{0}}{\mathtt{1}}},x_{{\mathtt{1}}{\mathtt{0}}}\right \rangle$ and $\left \langle y_{{\mathtt{0}}}^{-1}y_{{\mathtt{1}}},y_{\mathtt{1}},x\right \rangle$ are isomorphic, since their respective actions on boundaries of the binary trees, $T_1$ rooted at the empty sequence and $T_2$ rooted at the sequence ${\mathtt{1}}{\mathtt{0}}$, are the same. More explicitly, one can verify that the elements $y_{{\mathtt{1}}{\mathtt{0}}{\mathtt{0}}}^{-1}y_{{\mathtt{1}}{\mathtt{0}}{\mathtt{1}}},y_{{\mathtt{1}}{\mathtt{0}}{\mathtt{1}}},x_{{\mathtt{1}}{\mathtt{0}}}$ correspond via $\Phi$ to the following piecewise projective transformations: $$x_{{\mathtt{1}}{\mathtt{0}}}\sim\begin{cases} \begin{bmatrix} 1 & 0\\ -1 & 1 \end{bmatrix}&\text{on }\left[0,\frac{1}{3}\right ],\\ \\ \begin{bmatrix} 4 & -1\\ 5 & -1 \end{bmatrix}&\text{on }\left[\frac{1}{3},\frac12\right ],\\ \\ \begin{bmatrix} 0 & 1\\ -1 & 2 \end{bmatrix}&\text{on }\left[\frac{1}{2},1\right ],\\ \\ id&\text{on }{\mathbb{R}}\setminus [0,1], \end{cases}$$ $$\begin{aligned} y_{{\mathtt{1}}{\mathtt{0}}{\mathtt{1}}}\sim& \begin{cases} \begin{bmatrix} 3 & -1\\ 2 & 0 \end{bmatrix}&\text{on }\left [\frac{1}{2},1\right ],\\ \\ id&\text{on }{\mathbb{R}}\setminus \left [\frac12,1\right ], \end{cases} \\ \\ y_{{\mathtt{1}}{\mathtt{0}}{\mathtt{0}}}^{-1}y_{{\mathtt{1}}{\mathtt{0}}{\mathtt{1}}}\sim & \begin{cases} \begin{bmatrix} 1 & 0\\ -2 & 2 \end{bmatrix}&\text{on }\left [0,\frac{1}{2}\right ],\\ \\ \begin{bmatrix} 3 & -1\\ 2 & 0 \end{bmatrix}&\text{on }\left [\frac12,1\right ], \\ \\ id&\text{on }{\mathbb{R}}\setminus [0,1]. \end{cases}\end{aligned}$$ $$\includegraphics[scale=0.6]{BSinG0.png}$$ As a consequence of Lemma \[l:bbs\_in\_G0\], the group $G_0$ contains a subgroup $H$ isomorphic to $G_2$. Let $\rho:G_0\to{\mathsf{Diff}}_+^1([0,1])$ be a morphism. By a direct application of Theorem \[t:main1\], we obtain that the kernel of $\rho$ contains some nontrivial element of $H$. Thus $\rho$ is not injective. Now, it has been proven in [@commutators] that every proper quotient of $G_0$ is abelian, whence we get our result: as we have just shown that the kernel is not trivial, then the image must be abelian, as we wanted to prove. Further examples ---------------- An interesting family of nonamenable groups is obtained adding translations in top of $F$ (defined as in Definition \[ThurstonT\]). Mimicking Monod’s argument, it is not difficult to prove the following: For any $\alpha\in(0,1)$, the group of piecewise projective homeomorphisms generated by $F$ and the translation $t\mapsto t+\alpha$ is nonamenable. Observe that the groups $\langle F,t\mapsto t+\alpha\rangle$ appearing in the above statement are naturally of $C^1$ diffeomorphisms. \[t:Firr\] For any irrational $\alpha\in (0,1)$, the action of the group of piecewise projective homeomorphisms $\langle F,t\mapsto t+\alpha\rangle$ on the compactified real line $[-\infty,+\infty]$ is not $C^2$-smoothable. We denote by $T_\alpha$ the translation by $\alpha$. If $\alpha$ is irrational, then $T_1$ and $T_\alpha$ generate an abelian free group of rank $2$ of $C^2$ (even real analytic) diffeomorphisms of ${\mathbb{R}}$. The maps $f=T_{-1}$ and $g=T_{-\|\alpha\|}$ are contractions on ${\mathbb{R}}$. Consider any element $h\in \langle F,T_\alpha\rangle$ with a $C^2$ discontinuity point on ${\mathbb{R}}$. Then Theorem \[t:nonC2\] implies directly that the action of $\langle F,T_\alpha\rangle$ on $[-\infty,+\infty]$ is not $C^2$-smoothable. For rational translations $T_\alpha$, we can extend the previous argument and prove that even the action on the non-compactified real line $(-\infty,+\infty)$ is not $C^2$-smoothable. \[t:Frat\] For any rational $\alpha\in (0,1)$, the action of the group of piecewise projective homeomorphisms $\langle F,t\mapsto t+\alpha\rangle$ on ${\mathbb{R}}$ is not $C^2$-smoothable. We consider the conjugate of $c$ by $T_\alpha$: $$T_\alpha c T_{\alpha}^{-1}(t)=\begin{cases} t&\text{ if }t\leq \alpha,\\ \frac{t-\alpha}{1-(t-\alpha)}+\alpha&\text{ if }\alpha\leq t\leq \frac{1}{2}+\alpha,\\ 3-\frac{1}{t-\alpha}+\alpha&\text{ if }\frac{1}{2}+\alpha\leq t\leq 1+\alpha,\\ t+1&\text{ if }1+\alpha\leq t.\\ \end{cases}$$ In restriction to the interval $\left[\alpha,\frac{1}{2}+\alpha\right]$, the element $T_\alpha c T_{\alpha}^{-1}$ coincides with the projective transformation $$\begin{bmatrix} 1-\alpha & \alpha^2 \\ -1 & 1+\alpha \end{bmatrix},$$ which is a parabolic element in ${\mathsf{PSL}}(2,{\mathbb{Z}}[\alpha])$ fixing $\alpha$. It is not in ${\mathsf{PSL}}(2,{\mathbb{Z}})$. Inside Thompson’s $F$ we can find an element $f$ such that: - $f$ fixes $\alpha$, - the restriction of $f$ to the interval $\left[ \alpha,\tfrac12+\alpha\right ]$ is $C^2$, - $f$ is a contraction of the interval $\left[ \alpha,\tfrac12+\alpha\right ]$, namely $f(t)<t$ for any $t$ in the right neighbourhood $\left( \alpha,\tfrac12+\alpha\right )$ of $\alpha$. Indeed, since $\alpha$ is rational, there exists a parabolic element in ${\mathsf{PSL}}(2,{\mathbb{Z}})$ with $\alpha$ as fixed point, and we can take for $f$ any element of $F$ which coincides with this element (or its inverse) in restriction to $\left[ \alpha,\tfrac12+\alpha\right ]$. Finally, consider an element $h\in F$ which has a $C^2$ discontinuity point $p$ on $\left[ \alpha,\tfrac12+\alpha\right ]$, with $h(p)\in \left[ \alpha,\tfrac12+\alpha\right ]$. It is straightforward to verify that $f$, $g=\left (T_\alpha c T_{\alpha}^{-1}\right )^{-1}$ and $h$ satisfy the requirements of Theorem \[t:nonC2\] (when considering the interval $\left[ \alpha,\tfrac12+\alpha\right ]$ as the interval $[0,a]$ in the statement). Thus the theorem is proved. $C^1$ actions of affine and piecewise affine groups {#s:Aff} =================================================== [[Baumslag-Solitar groups and affine groups]{}]{} ------------------------------------------------- Let $n>1$ be an integer. The classical Baumslag-Solitar groups ${\mathsf{BS}}(1,n)$ are defined by the presentations $${\mathsf{BS}}(1,n)=\left \langle a,b\mid aba^{-1}=b^n\right \rangle.$$ They are naturally realized as subgroups of the affine group $\mathsf{Aff}_+({\mathbb{R}})\subset {\mathsf{PSL}}(2,{\mathbb{R}})$, generated by the homothety $a(x)=n x$, and the translation $b(x)=x+1$. Similarly, for any rational $\lambda=p/q>1$ there is a morphism from the Bausmlag-Solitar group $${\mathsf{BS}}(q,p)=\left \langle a,b\mid ab^qa^{-1}=b^p\right \rangle$$ to the subgroup $A_\lambda$ of $\mathsf{Aff}_+({\mathbb{R}})$ generated by $a(x)=\lambda x$ and $b(x)=x+1$. However, when $p/q>1$ is not an integer, this morphism is not an isomorphism. For general $\lambda>1$, we define $A_\lambda$ to be the subgroup of $\mathsf{Aff}_+({\mathbb{R}})$ generated by $a(x)=\lambda x$ and $b(x)=x+1$. Observe that the conjugate $aba^{-1}$ equals the translation $x\mapsto x+\lambda$, and hence $b$ and $ab a^{-1}$ commute in $A_\lambda$ (this is not true for nonsolvable Baumslag-Solitar groups ${\mathsf{BS}}(q,p)$, $p/q>1$ not integer). The group $A_\lambda$ is abelian by cyclic, abstractly isomorphic to the semi-direct product ${\mathbb{Z}}[\lambda,\lambda^{- 1}]\rtimes {\mathbb{Z}}$ (where ${\mathbb{Z}}$ acts on ${\mathbb{Z}}[\lambda,\lambda^{- 1}]$ by multiplication by $\lambda$). More precisely, for $\lambda$ transcendental, $A_\lambda$ is isomorphic to the wreath product ${\mathbb{Z}}\wr {\mathbb{Z}}\cong {\mathbb{Z}}[t,t^{-1}]\rtimes {\mathbb{Z}}$, whereas if algebraic the following properties hold. \[lem:abel\_aff\] Let $\lambda>1$ be an algebraic real number of degree $d$, and let $p_\lambda(t)=\frac{{\alpha}_0}{{\alpha}_d}+\frac{{\alpha}_1}{{\alpha}_d}t+\ldots +\frac{{\alpha}_{d-1}}{{\alpha}_d}t^{d-1}+t^d$, ${\alpha}_j\in {\mathbb{Z}}$, denote the associated minimal polynomial. (1) \[item:Qrank\] The group $H={\mathbb{Z}}[\lambda,\lambda^{-1}]$ has ${\mathbb{Q}}$-rank equal to $d$: it is an additive subgroup of ${\mathbb{Q}}(\lambda)\cong {\mathbb{Q}}^d$, and does not embed in a ${\mathbb{Q}}$-vector space of lower dimension. (2) \[item:matrix\] With respect to the basis $\{1,\lambda,\ldots,\lambda^{d-1}\}$, the homothety $a$ acts on $H$ as multiplication by the companion matrix $C_\lambda$, so that one has $A_\lambda\cong H\rtimes_{C_\lambda} {\mathbb{Z}}$. (3) The group $A_\lambda$ is a quotient of the finitely presented group $$\label{PresentationAff} \widehat A_\lambda=\left\langle \widehat a, b_0,\ldots,b_{d-1} \left\vert \begin{array}{lr} b_ib_j=b_jb_i& \text{for }i,j=0,\ldots,d-1,\\ \widehat ab_j\widehat a^{-1}=b_{j+1} & \text{for }j=0,\ldots, d-2,\\ \widehat ab_{d-1}^{{\alpha}_d}\widehat a^{-1}=b_0^{-{\alpha}_0}\cdots b_{d-1}^{-{\alpha}_{d-1}} & \\ \end{array} \right. \right\rangle,$$ where generators $b_j$ are mapped to the translations $x\mapsto x+\lambda^j$ in the standard affine action, and $\widehat a$ to the homothety of factor $\lambda$. (4) \[item:abelian\] The abelianization of $\widehat A_\lambda$ is the abelian group ${\mathbb{Z}}\times {\mathbb{Z}}/({\alpha}_d\, p_\lambda(1){\mathbb{Z}})$, where the factor ${\mathbb{Z}}$ is generated by the image of $\widehat a$, and the finite factor ${\mathbb{Z}}/({\alpha}_d\, p_\lambda(1){\mathbb{Z}})$ is generated by the image of any $b_j$. In particular, any torsion-free abelian quotient of $A_\lambda$ is either trivial or infinite cyclic. The proof being elementary, we rather omit it. For the last statement, observe that $p_\lambda$ is the minimal polynomial of $\lambda$ (which is a real algebraic number $\neq 1$), hence $1$ cannot be a root and therefore ${\alpha}_d\, p_\lambda(1)$ is always a nonzero integer. $C^1$ actions of [[abelian-by-cyclic groups]{}]{} ------------------------------------------------- [[In [@FarbFranksI] Farb and Franks, relying on Kopell’s lemma, show that every $C^2$ action of ${\mathsf{BS}}(q,p)$ on one-dimensional manifolds quotients through an action of its image ${\mathbb{Z}}[p/q,q/p]\rtimes {\mathbb{Z}}$ in $\mathsf{Aff}_+({\mathbb{R}})$. To the best of our knowledge, nothing appears in the literature about actions in lower regularity.]{}]{} [[The reason why actions of (solvable) Baumslag-Solitar groups are widely studied is because of the simple presentation, given by just one relation, $ab^ma^{-1}=b^n$, which has a dynamical meaning: $a$ conjugates a power of $b$ to another power. One of the first relevant works in this subject is the aforementioned [@FarbFranksI], where the authors study general actions of ${\mathsf{BS}}(q,p)$ on one-manifolds. This was pursued by Burslem-Wilkinson [@BW], where they study sufficiently regular actions of ${\mathsf{BS}}(1,n)$ on the circle. Later improvements are due to Guelman-Liousse [@GL], and finally to Bonatti-Monteverde-Navas-Rivas [@hyperbolic]. For actions on higher-dimensional manifolds, McCarthy [@McCarthy] proved that $C^1$ perturbations of the trivial action of torsion-free, finitely presented, abelian-by-cyclic groups are not faithful. Another example of rigidity result was obtained by Asaoka [@A1; @A2] for standard actions of the same class of groups on spheres and tori, and also by Wilkinson-Xue [@WX] for actions on tori. Finally, planar actions of ${\mathsf{BS}}(1,n)$ have been investigated by several authors [@GL2; @ARX; @AGRX].]{}]{} In relation with our work, Bonatti-Monteverde-Navas-Rivas study the $C^1$ actions on the interval of abelian-by-cyclic groups like $A_\lambda$. The following result appears in [@hyperbolic § 4.3] (even if not explicitly stated for general $\lambda>1$, the arguments in [@hyperbolic § 4.3] only use the condition $\lambda\ge 2$, which is always guaranteed, up to taking an integer power of $a$): \[p:affine\] Fix an arbitrary $\lambda>1$ and let $\phi:{\mathbb{R}}\to{\mathbb{R}}$ be a homeomorphism such that $\phi A_\lambda\phi^{-1}$ is in ${\mathsf{Diff}}_+^1([0,1])$. Then $\phi a\phi^{-1}$ has derivative equal to $\lambda$ at its interior fixed point $\phi(0)$. For [[Galois hyperbolic]{}]{} $\lambda>1$, for instance as $\lambda>1$ is rational, we obtain from [@hyperbolic] a much stronger statement: \[t:BMNR\] Let $\lambda>1$ be a Galois hyperbolic number and consider a $C^1$ action of the abelian-by-cyclic group $A_\lambda$ on the closed interval, without global fixed points in its interior. Then every nonabelian action of $A_\lambda$ is topologically conjugate to the standard affine action. Properties and in Lemma \[lem:abel\_aff\] guarantee that the hypotheses of [@hyperbolic Thm. 1.3] are satisfied for the group $A_\lambda$, provided $\lambda>1$ is Galois hyperbolic. This gives that any $C^1$ action of $A_\lambda$ on the closed interval, without global fixed points in its interior, is topological conjugate to a representation of $A_\lambda$ into the affine group $\mathsf{Aff}_+({\mathbb{R}})$. Representations $\psi:A_\lambda \to \mathsf{Aff}_+({\mathbb{R}})$ are classified by [@hyperbolic Prop. 2.1]: when the image of $\psi$ is nonabelian, (1) the generator $a$ of $A_\lambda$ (the homothety of factor $\lambda$) is mapped to itself and (2) the generator $b$ of $A_\lambda$ (the translation) is mapped to some translation. Therefore $\psi$ is conjugate to the standard affine action. \[couterex\] Observe that the statement above cannot be true as $\lambda>1$ is transcendental, because $A_\lambda\cong {\mathbb{Z}}\wr{\mathbb{Z}}$ has many distinct actions on the interval. Moreover, in [@hyperbolic § 5], even as $\lambda>1$ is algebraic but not Galois hyperbolic, there are further $C^1$ actions of $A_\lambda$. The groups $G_\lambda$ ---------------------- Inspired by the definition of Monod’s groups, we consider an analogous construction starting from these affine groups. Here we repeat the definition already given in the introduction: \[BrokeBS\] For any $\lambda>1$, we define $G_\lambda$ to be the subgroup of ${\mathsf{P}}{\mathsf{P}}_+({\mathbb{R}})$ generated by the elements $$a(x)=\lambda x,\quad a_+(x)=\begin{cases} x & \text{if }x\le 0\\ \lambda x & \text{if }x>0\end{cases},\quad b(x)=x+1.$$ We also set $a_-=aa_+^{-1}$, which agrees with $a$ to the left of $0$ and is the identity elsewhere. In the introduction, we were denoting $a,b,a_+$ by $f_\lambda,g,h_\lambda$ respectively. \[lem:abel\_G\] Let $\lambda>1$ be an algebraic number. The image of the generator $b\in G_\lambda$ is trivial in any torsion-free abelian quotient of $G_\lambda$. Indeed, any such quotient is either trivial, or infinite cyclic, or isomorphic to ${\mathbb{Z}}^2$, generated by the images of $a_\pm$. By in Lemma \[lem:abel\_aff\], every image of the generator $b$ in an abelian group must be of finite order. \[r:F\] The algebraic structure of $G_\lambda$ is highly complicated. For instance, in the case $\lambda=2$, inside the group $G_2$, the elements $b$ and $[a_+,b]$ are the generators of Thompson’s $F$, in its natural piecewise linear action on ${\mathbb{R}}$. [[In fact, every group $G_{\lambda}$ contains a copy of $F$. To see this, let $$f_1=b^{-1} a_+ b,\qquad f_2=b a_- b^{-1}.$$ The (open) support of $f_1$ is the half-line $J_1=(-1,+\infty)$ whereas the support of $f_2$ is the half-line $J_2=(-\infty,1)$. These supports form a chain $(J_1,J_2)$ in the sense of [@chains]. Then by [@chains Thm. 3.1], there exists $n\in \mathbf{N}$ such that $\langle f_1^n,f_2^n\rangle$ is isomorphic to Thompson’s group $F$.]{}]{} There are two canonical standard affine actions of the group $G_\lambda$ on the real line that factor through the affine group $A_\lambda$. First, as every element in $G_\lambda$ fixes $\pm\infty$, we can consider the germs of elements of $G_\lambda$ at these two points. This gives us two surjective homomorphisms $$\rho_\pm:G_\lambda\to A_\lambda.$$ It is clear from the definition of $G_\lambda$ that we have $$\rho_\pm(a_\mp)=id,\quad \rho_\pm(a_\pm)=\rho(a)$$ for these two morphisms. More generally, every element of $G_\lambda$ that is the identity outside a compact interval belongs to the kernels of both morphisms $\rho_\pm$. This is the case for the commutator $[b,a_+ba_+^{-1}]$ that appears in the statement of Theorem \[t:main1Hyp\]. [[On the other hand, as the abelianization of $G_\lambda$ is not trivial, there are plenty of abelian actions of $G_\lambda$ on the real line. Recall thats any group of orientation preserving homeomorphisms of the real line is torsion free, therefore as $\lambda>1$ is algebraic, Lemma \[lem:abel\_G\] implies that the generator $b$ must be in the kernel of any abelian action. In particular, as $\lambda>1$ is algebraic, the commutator $[b,a_+ba_+^{-1}]$ always acts trivially.]{}]{} $C^1$ actions of $G_\lambda$ ---------------------------- The following result is essentially the one contained in Theorem \[t:main2\]: \[t:affine\] For any $\lambda>1$, the natural action of $G_\lambda$ on the compactified real line $[-\infty,+\infty]$ is not $C^1$-smoothable. We argue by contradiction. After Proposition \[p:affine\], if there existed a homeomorphism $\phi:{\mathbb{R}}\to{\mathbb{R}}$ such that $\phi A_\lambda\phi^{-1}$ was in ${\mathsf{Diff}}_+^1([-\infty,+\infty])$, then $\phi a\phi^{-1}$ would have derivative equal to $\lambda$ at $p=\phi(0)$ and $\phi a_+ \phi^{-1}$ would not be $C^1$ at $p$. Hence the group is not $C^1$-smoothable. In the previous statement, it is fundamental to consider the action of $G_\lambda$ on the compactified line. Indeed, the statement is no longer true if one simply considers the action on ${\mathbb{R}}$ (see [@hyperbolic Remark 4.14]). Our second result, more precise than the statement in Theorem \[t:main1Hyp\], says that every $C^1$ action of $G_\lambda$ on the interval, for [[Galois hyperbolic]{}]{} $\lambda>1$, is always described by combining the examples above. \[t:key\] [[Let $\lambda>1$ be Galois hyperbolic and]{}]{} let $\rho:G_\lambda\to {\mathsf{Diff}}^1_+([0,1])$ be a nontrivial homomorphism. Then there exists finitely many pairwise disjoint subinterval $I_1,\ldots,I_n\subset [0,1]$ such that (1) for any $i=1,\ldots,n$, the image $\rho(G_\lambda)$ preserves the interval $I_i$, (2) for any $i=1,\ldots,n$, the restriction of $\rho(G_\lambda)$ to $I_i$ is topologically conjugate to one of the two canonical actions on ${\mathbb{R}}$. (3) the restriction of $\rho(G_\lambda)$ to the complement $[0,1]\setminus \bigcup_{i=1}^n I_i$ is abelian. In particular the group $G_\lambda$ admits no faithful $C^1$ action on the closed interval. Relying on [@hyperbolic Thm. 1.10] (cf. also [@GL]), we could provide a similar statement for $C^1$ actions of $G_\lambda$ on the circle ${\mathbb{S}^1}$. Indeed, every nonabelian action of $A_\lambda$ has a global [[finite orbit]{}]{} so, [[up to passing to a finite-index subgroup]{}]{}, every nonabelian action of $G_\lambda$ reduces to an action on the interval. \[r:C1+\] The proof of Theorem \[t:key\] would be much simpler for representations $\rho:G_\lambda\to {\mathsf{Diff}}^{1+\alpha}_+([0,1])$ of the group $G_\lambda$ into the group of $C^1$ diffeomorphisms with $\alpha$-Hölder continuous derivative. Indeed, it is classical that any $C^1$ element commuting with a $C^{1+\alpha}$ hyperbolic contraction of an interval lies in a one parameter flow (cf. Theorem \[t:Szekeres\]: when the fixed point of the contraction is hyperbolic, Szekeres theorem requires only $C^{1+\alpha}$ regularity). Let us sketch the proof under the assumption of $C^{1+\alpha}$ regularity. Assume that the image $\rho(G_\lambda)$ is nonabelian. Then the image $\rho(A_\lambda)$ is also nonabelian (cf. Lemma \[l:equiv1\]). From Theorem \[t:BMNR\] and Proposition \[p:affine\] we deduce that the element $\rho(a)$ behaves as the corresponding scalar multiplication in restriction to some interval $I\subset [0,1]$ and has a hyperbolic fixed point $s\in I$. As the elements $\rho(a_\pm)$ commute with $\rho(a)$, we deduce from Szekeres theorem that in restriction to the interval $I$, also these elements behave like scalar multiplications (as the one parameter flow containing a scalar multiplication is exactly the one parameter flow of all scalar multiplications). This implies that the group $\rho(G_\lambda)$ acts like an affine group in restriction to the interval $I$. The proof of Theorem \[t:key\] will occupy the rest of the section. Elementary ingredients ---------------------- When working with $C^1$ actions on the interval, hyperbolic fixed points do not often give rigidity (one usually needs $C^{1+\alpha}$ regularity, cf. Remark \[r:C1+\]). Indeed there are only a few dynamical tools that work in $C^1$ regularity. For this reason our proof relies mainly on very elementary arguments. A first tool is the following: \[FinitelyMany\] Let $\alpha\in{\mathsf{Diff}}_+^1([0,1])$ be a diffeomorphism. For any $\delta>1$, there are only finitely many points $s\in [0,1]$ that are fixed by $\alpha$ and such that $\alpha'(s)>\delta$. Suppose $\alpha$ has infinitely many fixed points $\{s_n\mid n\in {\mathbb{N}}\}$ in $[0,1]$, such that $\alpha'(s_n)>\delta$ for any $n\in{\mathbb{N}}$. Let $s_\in [0,1]$ be an accumulation point of the sequence $\{s_n\}$. By continuity of the derivative, we must have $\alpha'(s_)\ge\delta$. On the other hand, let $\{s_{n_k}\}$ be a subsequence converging to $s_$; by the very definition of the derivative we must have $\alpha'(s_)=1$. This is a contradiction. Then we state and prove a second crucial elementary fact. \[CommutingHyp\] Let $\alpha,\beta \in {\mathsf{Diff}}_+^1([0,1])$ be two commuting $C^1$ diffeomorphisms. Let $s\in [0,1]$ be a hyperbolic fixed point of $\alpha$. Then $\beta$ fixes $s$. Let us assume by way of contradiction that $\beta$ does not fix $s$. For each $n\in {\mathbb{Z}}$ we have $$\alpha(\beta^n(s))=\beta^n(\alpha(s))=\beta^n(s)$$ This means that $\alpha$ fixes each point in the set $S=\{\beta^n(s)\mid n\in {\mathbb{Z}}\}$. Each $t\in S$ is a hyperbolic fixed point of $\alpha$ and $\alpha'(t)=\alpha'(s)$ for all $t\in S$. Let $\lambda_n$ be the formal word $\beta^{-n} \alpha \beta^{n}$. Using the chain rule, we find $$\lambda_n'(s)=\alpha'(\beta^n(s)).$$ However, since $\alpha$ and $\beta$ commute, indeed $\lambda_n=\alpha$ and hence $\lambda_n'(s)=\alpha'(s)$. It follows that $\alpha'(s)=\alpha'(\beta^n(s))$ for each $n\in \mathbb{N}$. Since the set $S$ is infinite, the claim is in contradiction with Lemma \[FinitelyMany\]. A particular case: no global fixed points for $A_\lambda$ --------------------------------------------------------- Before dealing with a general statement as in Theorem \[t:key\], we study actions on the interval without global fixed points. For the statement, recall that we denote by $A_\lambda\subset G_\lambda$ the subgroup generated by $a$ and $b$. \[p:noglobal\] Let $\lambda>1$ be [[Galois hyperbolic]{}]{} and $\rho:G_\lambda\to {\mathsf{Diff}}_+^1([0,1])$ be a morphism satisfying the following: (1) the image $\rho(A_\lambda)$ is nonabelian, (2) the action of $\rho(A_\lambda)$ has no global fixed point in $(0,1)$. Then $\rho(G_\lambda)$ is topological conjugate to one of the two canonical representations $\rho_\pm:G_\lambda\to A_\lambda$. In the following, we let $s_0$ denote the hyperbolic fixed point of $\rho(a)$ in $(0,1)$, ensured by Theorem \[t:BMNR\] and Proposition \[p:affine\] above. For simplicity of notation, we also write $$\rho(a)=f,\quad \rho(b)=g, \quad \rho(a_-)=h,\quad \rho(a_+)=k.$$ \[l:bbs-derivative\] With the notation as above, the elements $h$ and $k$ fix the point $s_0$ and we have $$h'(s_0)\cdot k'(s_0)=\lambda.$$ By Lemma \[CommutingHyp\], the two elements $h,k$ fix the point $s_0$, as they commute with $f$. Remarking that $hk=f$, applying the chain rule we deduce that $$\begin{aligned} \lambda&=f'(s_0)\\ &=h'(k(s_0))\cdot k'(s_0)\\ &=h'(s_0)\cdot k'(s_0),\end{aligned}$$ as wanted. The previous lemma implies that $s_0$ is a hyperbolic fixed point for at least one among $h$ and $k$. Without loss of generality, we assume $h'(s_0)>1$. The following lemma says that $h$ behaves like a hyperbolic element on the whole interval $[0,1]$: \[l:bbs2\] With the notation as above, suppose $h'(s_0)>1$. Then $s_0$ is the only point of $(0,1)$ which is fixed by $h$. If $h$ had a fixed point $s$ different from $s_0$, since $h$ and $f$ commute, the images $f^{-n}(s)$ would form a sequence of fixed points for $h$ that converge to $s_0$. This would imply that the derivative of $h$ at $s_0$ should be equal to $1$. Contradiction. Given [[$r\in {\mathbb{Z}}[\lambda,\lambda^{-1}]$]{}]{}, we denote by $g_r$ the image by $\rho$ of the translation by $r$ when thinking of $A_\lambda$ as an affine group. We have $g_{-r}=g_r^{-1}$. By Theorem \[t:BMNR\] these elements are actually topologically conjugate to the corresponding translations. \[l:bbs3\] Take [[a positive $r\in {\mathbb{Z}}[\lambda,\lambda^{-1}]$]{}]{}. The conjugate $g_{r}kg_r^{-1}$ commutes with $h$. This is actually a statement about relations of the group $G_\lambda$: we prove the relation looking at the standard action of $G_\lambda$ on the real line. The support of $a_+$ is $[0,+\infty)$, therefore the support of the conjugate of $a_+$ by the translation by $r$ is $[r,+\infty)$, which is disjoint from $(-\infty,0]$, which is the support of $a_-$. \[l:bbs4\] With the notation as above, the restriction of $k$ to $[0,s_0]$ is the identity. The element $k_r=g_{r}kg_r^{-1}$ commutes with $h$ by Lemma \[l:bbs3\]. [[As $s_0$ is a hyperbolic fixed point for $h$, Lemma \[CommutingHyp\] implies that $k_r$ fixes $s_0$ for each $r>0$.]{}]{} In particular, it follows that $k$ fixes $g_{-r}(s_0)$ for every positive $r\in {\mathbb{Z}}[\lambda,\lambda^{-1}]$. By density of [[${\mathbb{Z}}[\lambda,\lambda^{-1}]$]{}]{} in $(0,+\infty)$, we obtain the statement. The end of the proof is inspired by [@centralizers]: in a centraliser of a hyperbolic element, like $h$, there cannot be elements with hyperbolic fixed points (different from the fixed points of the hyperbolic element), and therefore by Proposition \[p:linked\_hyper\], there cannot be linked pairs of successive fixed points. \[l:bbs5\] Suppose that $k$ is not the identity and let $s, t\in [s_0,1]$ be a pair of successive fixed points of $k$. There exists a [[positive $r\in {\mathbb{Z}}[\lambda,\lambda^{-1}]$]{}]{} such that the pair $g_r(s),g_r(t)$ together with $s,t$ defines a linked pair of fixed points for $g_rkg_r^{-1}$ and $k$. For any [[positive $r\in {\mathbb{Z}}[\lambda,\lambda^{-1}]$]{}]{}, the points $g_r(s),g_r(t)$ define a pair of successive fixed points for the conjugate $g_rkg_r^{-1}$. An element $g_r$, $r>0$, moves every point in $(s_0,1)$ to the right, and using the fact that $g_r$ is topologically conjugate to the translation by $r$, we can choose $r>0$ sufficiently small such that $$s<g_r(s)<t\le g_r(t)$$ (with equality $t=g_r(t)$ if and only if $t=1$). Suppose that $k$ is not the identity. Then, from Lemma \[l:bbs5\] and Proposition \[p:linked\_hyper\], we realize that the subgroup $\langle g_rkg_r^{-1},k\rangle$ contains an element $\gamma$ with a hyperbolic fixed point $p$ in $(s_0,1)$. Since $g_rkg_r^{-1},k$ commute with $h$, it follows that $\gamma$ commutes with $h$. So by Lemma \[CommutingHyp\] $h$ must fix the point $p$. This contradicts the conclusion of Lemma \[l:bbs2\]. Therefore we must have that $k=\rho(a_+)$ is the identity, and so $\rho(a_-)=\rho(a)$. Thus the representation $\rho:G_\lambda\to {\mathsf{Diff}}_+^1([0,1])$ is topologically conjugate to the canonical representation $\rho_-:G_\lambda\to A_\lambda$. This finishes the proof of Proposition \[p:noglobal\]. Equivalent properties --------------------- Now we consider almost the same statement as in Proposition \[p:noglobal\], but we only make assumptions on the global dynamics of $G_\lambda$, rather than on the one of $A_\lambda$. \[p:noglobal2\] Let $\lambda>1$ be [[Galois hyperbolic]{}]{} and $\rho:G_\lambda\to {\mathsf{Diff}}_+^1([0,1])$ be a morphism satisfying the following: (1) the image $\rho(G_\lambda)$ is nonabelian, (2) the action of $\rho(G_\lambda)$ has no global fixed point in $(0,1)$. Then $\rho(G_\lambda)$ is topological conjugate to one of the two canonical representations $\rho_\pm:G_\lambda\to A_\lambda$. The proof follows directly from the following two lemmas [[and from Proposition \[p:noglobal\]]{}]{}. \[l:equiv1\] [[Let $\lambda>1$ be algebraic]{}]{} and $\rho:G_\lambda\to {\mathsf{Diff}}_+^1([0,1])$ be a morphism. Then the following properties are equivalent: (1) the image $\rho(G_\lambda)$ is nonabelian, (2) the image $\rho(A_\lambda)$ is nonabelian. Clearly (2) implies (1). On the other hand, if the image $\rho(A_\lambda)$ is abelian, [[ in Lemma \[lem:abel\_aff\] implies that the translation $b$ is in the kernel of $\rho$, and as in Lemma \[lem:abel\_G\]]{}]{}, $\rho(G_\lambda)$ itself is abelian. \[l:equiv2\] Let $\lambda>1$ be [[Galois hyperbolic]{}]{} and $\rho:G_\lambda\to {\mathsf{Diff}}_+^1([0,1])$ be a morphism with nonabelian image. Then the following properties are equivalent: (1) the action of $\rho(G_\lambda)$ has no global fixed point in $(0,1)$, (2) the action of $\rho(A_\lambda)$ has no global fixed point in $(0,1)$. Again, (2) easily implies (1). Assume (1). Since the image $\rho(G_\lambda)$ is nonabelian, by Lemma \[l:equiv1\] also the image $\rho(A_\lambda)$ is nonabelian. Using Theorem \[t:BMNR\], we find at least one interval $I=[x,y]\subset [0,1]$ such that (1) $I$ is preserved by $\rho(A_\lambda)$, (2) $\rho(A_\lambda)$ has no global fixed point in the interior of $I$, (3) the restriction $\rho(A_\lambda)\vert_I$ is nonabelian and therefore it is topologically conjugate to the standard affine action on ${\mathbb{R}}$. Moreover, by Proposition \[p:affine\], there exists a unique point $s_0\in (x,y)$ in the interior of $I$ which is a hyperbolic fixed point for $\rho(a)$, with derivative $\rho(a)'(s_0)=\lambda$. Proceeding as in Lemma \[l:bbs-derivative\], the elements $\rho(a_\pm)$ must fix the point $s_0$ and we can suppose that $s_0$ is a hyperbolic fixed point for $\rho(a_-)$, with derivative $\rho(a_-)'(s_0)>1$. Let $s_-$ be the first fixed point of $\rho(a_-)$ which lies to the left of $s_0$. If $s_-\in (x,s_0)$, then $\{\rho(a)^{-n}(s_-)\mid n\in {\mathbb{N}}\}$ is a sequence of fixed points for $\rho(a_-)$ that converges to $s_0$ as $n\to\infty$. But this is not possible because the derivative of $\rho(a_-)$ at $s_0$ is not $1$ (cf. Lemma \[CommutingHyp\]). Similarly, if $s_-\in [0,x)$, then $\{\rho(a_-)^{-n}(x)\mid n\in {\mathbb{N}}\}$ is a sequence of fixed points for $\rho(a)$ that converges to $s_0$ as $n\to\infty$. Again, this is not possible. Thus $s_-=x$ and so $x$ is a global fixed point for $\rho(G_\lambda)$. As we are assuming (1), this implies $x=0$. Similarly, denoting by $s_+$ the first fixed point of $\rho(a_-)$ which lies to the right of $s_0$, we obtain that $s_+=y$ and so $y=1$. This is what we wanted to prove. The statement follows directly from Lemmas \[l:equiv1\] and \[l:equiv2\], [[and from Proposition \[p:noglobal\]]{}]{}. General case ------------ We proceed now to the proof of Theorem \[t:key\]. Let $\rho:G_\lambda\to {\mathsf{Diff}}^1_+([0,1])$ be a homomorphism. If the image $\rho(G_\lambda)$ is abelian, there is nothing to prove. Hence we can assume that $\rho(G_\lambda)$ is nonabelian and by Lemma \[l:equiv1\] this is equivalent to saying that $\rho(A_\lambda)$ is nonabelian, where $A_\lambda$ is denoting the subgroup generated by $a$ and $b$. If $\rho(A_\lambda)$ is nonabelian, then after Theorem \[t:BMNR\], there exists at least an interval $I\subset [0,1]$ which is preserved by $\rho(A_\lambda)$ and such that $\rho(A_\lambda)$ is topologically conjugate to the standard affine action. There are only finitely many pairwise disjoint intervals $I_1,\ldots, I_n$ that are preserved by $\rho(A_\lambda)$ and such that for any $i=1,\ldots,n$ the restrictions $\rho(A_\lambda)\vert_{I_i}$ is nonabelian. Let $I$ be an interval preserved by $\rho(A_\lambda)$ and such that $\rho(A_\lambda)\vert_I$ is nonabelian. By Theorem \[t:BMNR\], the action is topologically conjugate to the standard action of $A_\lambda$ and by Proposition \[p:affine\] there exists a point $s\in I$ which is fixed by $\rho(a)$ and such that $\rho(a)'(s_n)=\lambda>1$. Then Lemma \[FinitelyMany\] implies that there can only be finitely many such intervals, whence the first statement. Let $I_1,\ldots,I_n$ be the intervals provided by the previous claim. Then $\rho(G_\lambda)$ preserves $I_i$ for any $i=1,\ldots,n$. Let $I$ be an interval as above. Let $J\subset I$ be a interval which is preserved by $\rho(G_\lambda)$ and such that $\rho(G_\lambda)$ has no global fixed point in its interior. By Proposition \[l:equiv2\], we must have the equality $I=J$. After Proposition \[p:noglobal2\], we deduce that the restriction of the action of $G_\lambda$ to any of the intervals $I_1,\ldots,I_n$ is topologically conjugate to one of the two canonical affine actions that filters through a quotient $\rho_\pm:G_\lambda\to A_\lambda$. This is what we wanted to prove. $C^2$ actions with locally non-discrete stabilisers {#s:C2} =================================================== Szekeres vector field --------------------- The method that we present in this section is inspired by [@hyperbolic Prop. 4.17] and relies on the following important result in one-dimensional dynamics, due to Szekeres [@szekeres]. Here we state it as in [@navas-book § 4.1.3]: \[t:Szekeres\] Let $f$ be a $C^2$ diffeomorphism of the half-open interval $[0,1)$ with no fixed point in $(0,1)$. Then there exists a unique $C^1$ vector field $\mathcal X$ on $[0,1)$ with no singularities on $(0,1)$ such that 1) $f$ is the time-one map of the flow $\{\phi_{\mathcal X}^s\}$ generated by $\mathcal X$, 2) the flow $\{\phi_\mathcal X^s\}$ coincides with the $C^1$ centraliser of $f$ in ${\mathsf{Diff}}_+^1([0,1))$. An obstruction to $C^2$ smoothability ------------------------------------- The criterion we provide holds in a framework that is far more general that the one of piecewise-projective dynamics. First we need a statement of differentiable rigidity for the conjugacy of some particular actions. \[p:nonC2\] Take $a\in (0,1)$ and assume that two homeomorphisms $f,g\in {\mathsf{Homeo}}_+([0,1])$ satisfy the following properties: 1) the restrictions of $f$ and $g$ to $[0,a]$ are $C^2$ contractions, namely the restrictions are $C^2$ diffeomorphisms such that $$f(x)<x,\quad g(x)<x\quad\text{for every }x\in (0,a],$$ 2) $f$ and $g$ commute in restriction to $[0,a]$, that is $$fg(x)=gf(x)\quad\text{for every }x\in [0,a],$$ 3) the $C^2$ germs of $f$ and $g$ at $0$ generate an abelian free group of rank $2$. Then for every homeomorphism $\varphi\in {\mathsf{Homeo}}_+([0,1])$ such that $\varphi f\varphi^{-1}$, $\varphi g\varphi^{-1}$ are $C^2$ in restriction to $[0,\varphi(a)]$, one has that the restriction of $\varphi$ to $(0,a]$ is $C^2$. Before giving the proof of the proposition, which encloses the main arguments, we present the main result of the section: \[t:nonC2\] Assume $a\in (0,1)$, $f,g\in {\mathsf{Homeo}}_+([0,1])$ satisfy the properties of Proposition \[p:nonC2\] above. Moreover, assume that there exists $h\in {\mathsf{Homeo}}_+([0,1])$ such that there exists $t\in (0,a)$ which is a $C^2$ discontinuity point of $h$ and $h(t)\in (0,a)$. Then the natural action of $\langle f,g,h\rangle \subset {\mathsf{Homeo}}_+([0,1])$ on $[0,1]$ is not $C^2$-smoothable. We argue by contradiction. If there was a homeomorphism $\varphi\in {\mathsf{Homeo}}_+([0,1])$ such that $\varphi\langle f,g,h\rangle\varphi^{-1}$ is in ${\mathsf{Diff}}^2_+([0,1])$, then $\varphi$ would satisfy the requirements of Proposition \[p:nonC2\], whence $\varphi$ would be $C^2$ in restriction to $(0,a]$. However $h$ has a $C^2$ discontinuity point $t\in (0,a)$ and hence the conjugate $\varphi h\varphi^{-1}$ would have $\varphi(t)$ as $C^2$ discontinuity point, against the assumption that $\varphi h\varphi^{-1}$ is $C^2$. As $f$ is a $C^2$ contraction when restricted to $[0,a]$, Szekeres’ Theorem \[t:Szekeres\] applies: we denote by $\mathcal X$ the Szekeres vector field of $f$, which is $C^1$, defined on $[0,a)$ and with no singularities on $(0,a)$. We have the assumption 2) that $f$ and $g$ commute in restriction to $[0,a]$, so by Szekeres’ theorem $g$ belongs to the Szekeres flow $\{\phi^s_{\mathcal X}\}$. Let $\lambda>0$ be such that $g=\phi^\lambda_{\mathcal X}$. Then by assumption 3), the subgroup $A:=\langle 1,\lambda\rangle\subset {\mathbb{R}}$ is a dense abelian group of rank $2$. As $f$ and $g$ are contractions, we have that for any positive power $n\in {\mathbb{N}}$, the restrictions of the iterates $f^n$ and $g^n$ to the interval $[0,a]$ coincide with the times $\phi^n_{\mathcal X}$ and $\phi^{n\lambda}_{\mathcal{X}}$ respectively (however such a statement is in general not true for negative powers of $f$ and $g$). More generally, we have the following: Denote by $A$ the rank $2$ abelian subgroup of ${\mathbb{R}}$ generated by $1$ and $\lambda$. For every $\alpha\in A, \alpha>0$ there exists an element $h_\alpha\in \langle f,g\rangle$ such that $$h_\alpha\vert_{[0,a]}(x)=\phi^\alpha_{\mathcal X}(x)\quad \text{for any }x\in [0,a].$$ Moreover, $f$ and $h_\alpha$ commute on $[0,a]$: $[f,h_\alpha]\vert_{[0,a]}=[h_\alpha,f]\vert_{[0,a]}=id\vert_{[0,a]}$. Let $\ell,m\in{\mathbb{Z}}$ be such that $\alpha= \ell+m\lambda$. There exists $y>0$ such that the element $f^\ell g^m$ is equal to $\phi^\alpha_{\mathcal X}$ on the right neighbourhood $[0,y]$. If $y\ge a$, then we set $h_\alpha=f^\ell g^m$ and we are done. Otherwise, we have $y<a$. As $f$ is a contraction on $[0,a]$, there exists a positive integer $N\in {\mathbb{N}}$ such that $f^{N}([0,a])=[0,\phi^N_{\mathcal X}(a)]\subset [0,y]$. Define $h_\alpha=f^{-N}f^\ell g^m f^N$. Then for any $x\in [0,a]$ we have $$\begin{aligned} h_\alpha\vert_{[0,a]}(x)=\,&f^{-N}f^\ell g^m f^N\vert_{[0,a]}(x)\\ =\,&f^{-N}f^\ell g^m \vert_{[0,\phi^N_{\mathcal X}(a)]} (\phi^N_{\mathcal X}(x))\\ =\,&f^{-N}\vert_{[0,\phi^{\alpha+N}_{\mathcal X}(a)]}(\phi^{\alpha+N}_{\mathcal X}(x)).\end{aligned}$$ The element $f^{-N}$ equals $\phi^{-N}_{\mathcal X}$ on the interval $[0,\phi^N_{\mathcal X}(a)]$. Here $\alpha>0$ hence $[0,\phi^{\alpha+N}_{\mathcal X}(a)]\subset [0,\phi^N_{\mathcal X}(a)]$. We conclude that for any $x\in [0,a]$ we have $$h_\alpha\vert_{[0,a]}(x)=\phi^{-N}_{\mathcal X}\phi^{\alpha}_{\mathcal X}\phi^{N}_{\mathcal X}(x)=\phi^\alpha_{\mathcal X}(x),$$ as desired. The claim implies that the group generated by $f$ and $g$ contains a one-parameter flow in its local $C^0$-closure. Suppose that $\varphi$ is a homeomorphism such that $\varphi \langle f,g,h\rangle\varphi^{-1}$ is in ${\mathsf{Diff}}_+^2([0,1])$. The element $\varphi f\varphi^{-1}$ is a $C^2$ contraction on a right neighbourhood $[0,\varphi(a)]$ of $0$, thus Szekeres’ theorem applies again. Let $\mathcal{Y}$ denote the Szekeres vector field of $\varphi f\varphi^{-1}$ and let $\{\phi_{\mathcal Y}^s\}$ be the associated one-parameter flow defined on $[0,\varphi(a)]$. The elements $\varphi h_\alpha\varphi^{-1}$’s commute with $\varphi f\varphi^{-1}$ on $[0,\varphi(x)]$, hence by Szekeres’ theorem we must have that their restrictions to $[0,\varphi(a)]$ are contained in the flow $\{\phi_{\mathcal Y}^s\}_{s\ge 0}$. Moreover they are densely contained because $A$ is dense in ${\mathbb{R}}$. The restriction of $\varphi$ to $(0,a]$ is $C^1$ and takes the Szekeres vector field $\mathcal X$ of $f$, defined on $[0,a]$, to $\mathcal{Y}$: $$\varphi_\mathcal X=\mathcal Y.$$ For any $x\in [0,a]$ and $\alpha\in A$, $\alpha>0$, we have $$\phi_{\mathcal Y}^\alpha\left (\varphi(x)\right )=\varphi\left (\phi_{\mathcal X}^\alpha(x)\right ).$$ Now, $\phi_{\mathcal X}^\alpha(x)\neq x$ for any $\alpha>0$, $x\in (0,a]$. Thus for any $x\in (0,a]$ and $\alpha\in A$, $\alpha>0$, we have $$\begin{aligned} \frac{\phi_{\mathcal Y}^\alpha\left (\varphi (x)\right )-\varphi(x)}{\alpha}&= \frac{\varphi\left (\phi_{\mathcal X}^\alpha(x)\right )-\varphi(x)}{\alpha} \\ &= \frac{\varphi\left (\phi_{\mathcal X}^\alpha(x)\right )-\varphi(x)}{\phi^\alpha_{\mathcal X}(x)-x}\cdot \frac{\phi^\alpha_{\mathcal{X}}(x)-x}{\alpha}.\end{aligned}$$ Hence $$\frac{\varphi\left (\phi_{\mathcal X}^\alpha(x)\right )-\varphi(x)}{\phi^\alpha_{\mathcal X}(x)-x} = \frac{\phi_{\mathcal Y}^\alpha\left (\varphi (x)\right )-\varphi(x)}{\alpha} \cdot \left(\frac{\phi^\alpha_{\mathcal{X}}(x)-x}{\alpha}\right )^{-1}.$$ Taking the limit on both sides as $\alpha\in A$, $\alpha>0$, goes to $0$ (recall that $A$ is dense in ${\mathbb{R}}$), we get on the left hand side the derivative $\varphi'(x)$ and on the right hand side the ratio $\mathcal Y\left (\varphi(x)\right )/\mathcal{X}(x)$ (here we identify $\mathcal X,\mathcal Y$ to $C^1$ functions). Observe that the ratio $\mathcal Y\left (\varphi(x)\right )/\mathcal{X}(x)$ is well-defined because $\mathcal{X}$ has no singularities on $(0,a)$. This gives the two desired statements. Indeed, as $\mathcal Y\left ( \varphi(x)\right )/\mathcal{X}(x)$ is $C^0$ on $(0,a]$, so is $\varphi'$. Hence $\varphi$ is $C^1$ on $(0,a]$. Moreover, taking $\mathcal X(x)$ to the left hand side, we get $$\varphi'(x)\cdot \mathcal{X}(x)=\mathcal Y\left (\varphi(x)\right )\quad\text{for any }x\in [0,a],$$ that is $\varphi_\mathcal X=\mathcal Y$, as wanted. Now we can conclude the proof. From the previous Claim, we write $$\varphi'(x)=\frac{\mathcal Y\left (\varphi(x)\right )}{\mathcal X(x)}\quad\text{for every }x\in (0,a].$$ Moreover the previous Claim gives that the right hand side in the last expression is at least $C^1$ on $(0,a]$ and therefore the same holds for $\varphi'$. This implies that $\varphi$ is $C^2$ on $(0,a]$, as desired. Thompson-Stein groups --------------------- We finally apply the previous result to prove that the Thompson-Stein groups are not $C^2$-smoothable. In the group $F(n_1,\ldots,n_k)$, $k\ge 2$, it is possible to find two elements $f,g$ fixing $0$ such that $f'(0)=1/n_1$, $g'(0)=1/n_2$. Let $a\in (0,1)$ be such that $f,g$ are linear contractions in restriction to $[0,a]$. Consider any element $h$ which is the identity in restriction to $[0,a/2]$ but not in restriction to $[0,a]$. Then there exists $t\in [a/2,a)$ which is a $C^1$ discontinuity point of $h$ with $h(t)\in [a/2,a)$ (actually we may take for $t$ the leftmost point in the support of $h$). Thus we apply Theorem \[t:nonC2\] and conclude that the natural action on $[0,1]$ of the group generated by $f,g$ and $h$ is not $C^2$-smoothable. In particular the natural action of $F(n_1,\ldots,n_k)$ on $[0,1]$ is not $C^2$-smoothable. From [@Liousse Theorem 3.A], every faithful $C^2$ action of $T(2,n_2,\ldots,n_k)$ on ${\mathbb{S}^1}$ is topologically conjugate to its standard piecewise linear action. However $T(2,n_2,\ldots,n_k)$ contains $F(2,n_2,\ldots,n_k)$ as a subgroup, whose standard action on ${\mathbb{S}^1}$ cannot be conjugate to a $C^2$ action, after our Theorem \[t:F23\]. Therefore a $C^2$ action of $T(2,n_2,\ldots,n_k)$ cannot be faithful. As in [@Liousse Theorem 3.B’], if we assume furthermore $n_2=3$, the simplicity of $T(2,3,n_3,\ldots,n_k)$ allows to conclude that every $C^2$ action of such a group is trivial. Acknowledgements {#acknowledgements .unnumbered} ================ The authors thank Isabelle Liousse for her valuable comments on the first version of the paper (Corollary \[t:T23\] is due to her). [[Many thanks also go to the referee for his/her suggestions and for pointing out the many imprecise earlier algebraic statements about $A_\lambda$.]{}]{} This work has been done during visits of the authors to the Institut de Mathématiques de Bourgogne and the École Polytechnique Fédérale de Lausanne. The authors thank these institutions for the welcoming atmosphere. We are also grateful to the members of the IMB participating to the working seminar “Dynamique des actions de groupes” during Autumn 2017, as well as to the organizers of the conference Dynamics Beyond Uniform Hyperbolicity held in BYU, Provo, June 2017. M.T. was partially supported by PEPS – Jeunes Chercheur-e-s – 2017 (CNRS). [^1]: MSC: Primary 37C85, 57M60. Secondary 43A07, 37D40, 37E05. [^2]: CNRS. [^3]: Institut de Mathématiques de Bourgogne, Université Bourgogne Franche-Comté, CNRS UMR 5584, 9 av. Alain Savary, 21000 Dijon, France. [^4]: EPFL SB MATH EGG MA C3 584 (Batiment MA) Station 8 CH-1015 Lausanne, CH-1015, Switzerland. [^5]: Another way of seeing this is that $C^1$ continuity follows from the choice of $\mathcal P_{{\mathbb{Z}}}$ for the set of breakpoints. [^6]: To see this, first check that the relations of $F$ are satisfied and conclude using the property $F$ satisfies that every proper quotient is abelian. [^7]: We do not define property $(T)$ here, we refer the reader to [@kazhdan]. [^8]: We remark that in general a linked pair may not look like it does in this situation, for instance such maps may have components of support lying outside $(a,b)$ and $(a+1,b+1)$ respectively.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Let $G=(V,E)$ be a finite graph and $M_G$ be the centered Hardy-Littlewood maximal operator defined there. We find the optimal value $\bf{C}_{G,p}$ such that the inequality $${{\rm Var\,}}_{p}(M_{G}f)\le {\bf C}_{G,p}{{\rm Var\,}}_{p}(f)$$ holds for every $f:V\to \mathbb{R},$ where ${{\rm Var\,}}_p$ stands for the $p$-variation, when: (i) $G=K_n$ (complete graph) and $p\in [\frac{\log(4)}{\log(6)},\infty)$ or $G=K_4$ and $p\in (0,\infty)$; (ii) $G=S_n$(star graph) and $1\ge p\ge \frac{1}{2}$; $p\in (0,\frac{1}{2})$ and $n\ge C(p)$ or $G=S_3$ and $p\in (1,\infty).$ We also find value of the norm $\\|M_{G}\\|_{2}$ when: (i) $G=K_n$ and $n\ge 3$; (ii) $G=S_n$ and $n\ge 3.$' address: - \| IMPA - Instituto de Matemática Pura e Aplicada\ Rio de Janeiro - RJ, Brazil, 22460-320. - 'Department of Mathematics, University of California, Los Angeles (UCLA), Portola Plaza 520, Los Angeles, California, 90095, USA' author: - 'Cristian González-Riquelme and José Madrid' title: 'Sharp p-bounds for maximal operators on finite graphs' --- Introduction ============ A brief historical overview and background ------------------------------------------ The study of maximal operators is a central theme in analysis. Since the beginning of the past century many properties of these operators have been useful in several areas of mathematics. In general, properties related with the behavior of the norm of these operators have been the main interest of study, until the work of Kinnunen [@Ki] where he observed that it was possible to prove the boundedness of the map $$f\to Mf$$ from $W^{1,p}(\mathbb{R}^d)\to W^{1,p}(\mathbb{R}^d),$ when $p>1$, (where $M$ stands for the centered Hardy-Littlewood maximal function). He also showed meaningful applications of this property. This work was the first to study maximal operators at a derivative level. Since then many authors followed this path and proved several results concerning these [[derivative level questions]{}]{} in a broad class of contexts and for several kinds of maximal operators, see for instance [@AlPe; @BM; @BM2; @CGR; @CaMa; @CMP; @CaSv; @GR; @KiSa; @Ku; @L2; @LM; @R]. An interesting framework of study is the following. Let $G=(V,E)$ be a graph with $f:V\to{\mathbb{R}}$ a real valued function. We define the Hardy-Littlewood maximal function of $f$ along $G$ at the point $e\in V$ by $$\label{classical} M_{G}f(e):=\max_{r\geq0}\frac{1}{\|B(e,r)\|}\sum_{m\in B(e,r)}\|f(m)\|,$$ where $B(e,r)=\{m\in V;d_{G}(e,m)\leq r\}$, where $d_{G}$ is the metric induced by the edges of $G$ (that is, the distance between two vertices is the number of edges in a shortest path connecting them). A more general version of this, is the so called fractional maximal function defined by $$M_{\alpha,G}f(e):=\max_{r\geq0}\frac{1}{\|B(e,r)\|^{1-\alpha}}\sum_{m\in B(e,r)}\|f(m)\|$$ for all $\alpha\in(0,1]$. Both operators have uncentered versions defined by $$\widetilde{M}_{\alpha,G}f(e)=\max_{B(v,r)\ni e}\frac{1}{\|B(v,r)\|^{1-\alpha}}\sum_{m\in B(v,r)}\|f(m)\|$$ for the fractional one, and $\widetilde{M}_{G}=\widetilde{M}_{0,G}$ for the classical one. In this paper we study the regularity properties of these objects acting on $l^p-$spaces and bounded $p-$variation spaces. We focus on the classical maximal function defined in . Discrete properties for maximal operators have caught significant attention along the last years, whether in a discrete context (see, for instance [@CaHu; @Ma; @Ma2; @Te]) or as an intermediary step towards the solution of a continuous problem (see, for instance [@Me]). The most natural context for the discrete version of the [derivative level]{} questions mentioned above is the following, given $p\in (0,\infty)$ we define the $p$-variation of a function $f:V\to \mathbb{R}$ as follows $${{\rm Var\,}}_{p}f:=\left(\frac{1}{2}\sum_{n}\sum_{\substack{m \\ d_{G}(n,m)=1}}\|f(n)-f(m)\|^p\right)^{1/p}.$$ The first work to address a result concerning the derivative level (in this case, the variation) of a maximal operator in a discrete setting was [@BCHP]; where they, among other things, found sharp constant for the $1$-variation of the uncentered Hardy-Littlewood maximal operator $\widetilde{M}_{\mathbb{Z}}$, where in $\mathbb{Z}$ we take the usual distance. That is, they proved that for every $f:\mathbb{Z}\to \mathbb{R},$ we have $${{\rm Var\,}}_{1}(\widetilde{M}_{\mathbb{Z}}f)\le {{\rm Var\,}}_{1}(f),$$ and that the constant in front of ${{\rm Var\,}}_{1}f$ ($1$ in this case) is sharp. In the following we use the notation ${{\rm Var\,}}_{1}=:{{\rm Var\,}}.$ Other kind of graphs were studied by Soria and Tradacete in [@ST], where sharp $l^p-$bounds for maximal operators on finite graphs were first obtained. Later, some other regularity properties of maximal functions on graphs were studied by those authors in [@ST2]. More recently, bounds for the $p-$variation of the maximal functions on finite graphs were established by Liu and Xue in [@LX]. Finding optimal bounds for both the $l^p-$norm of the maximal functions and the $p$-variation of the maximal functions acting on finite graphs is a very interesting and challenging problem. In this paper we make progress on this kind of problem. Conjectures and results for the $p$-variation in finite graphs. --------------------------------------------------------------- For a given graph $G=(V,E)$ and $0<p<\infty$, we define $${\bf C}_{G,p}:=\sup_{f:V\to \mathbb{R};{{\rm Var\,}}_{p}f>0}\frac{{{\rm Var\,}}M_{G}f}{{{\rm Var\,}}_{p}f}.$$ Liu and Xue ([@LX]) obtained optimal results for $n=3$ and for the general case $n>3$ they found some bounds and posed some interesting conjectures. More precisely, they proved that if $G$ is the complete graph with $n$ vertices $K_n$ or the star graph with $n$ vertices $S_n$, then $$1-\frac{1}{n}\leq {\bf C}_{G,p}\leq 1$$ for $0<p<\infty$, and for $n=3$ the lower bound becomes an equality. Moreover, Liu and Xue posed the following conjectures \[See [@LX], Conjecture 1\]. [[Conjecture A (for the complete graph $K_n$): For every $n\ge 2$ and $p\in (0,\infty)$ we have $${\bf C}_{K_n,p}=1-\frac{1}{n}.$$ ]{}]{} In this paper we give a positive answer to this conjecture for all $p\ge \frac{\log 4}{\log 6}\approx 0.77$. This range is certainly not optimal and is an interesting problem to try to extend it. Also, we prove the conjecture for every $0<p<1$ when $n=4.$ That is the content of our Theorem \[theo 1\]. \[theo 1\] Let $0<p\leq\infty$ and $K_{n}=(V,E)$ a complete graph with $n$ vertices $(a_{1},a_{2},\dots,a_{n})$. Then - If $p\ge 1$, then $${\bf C}_{K_n,p}=1-\frac{1}{n}$$ - If $0<p<1$ and $n=4$, or $n\ge 3$ and $1>p\ge \frac{\log(4)}{\log(6)}\approx 0.77$, then $${\bf C}_{K_n,p}=1-\frac{1}{n}.$$ Moreover, in both cases the function $\delta_{a_2}$ is an extremizer. We notice that given the different behavior of the function $x\to x^p$ when $p\ge 1$ and $p<1$ very contrasting techniques are needed in each case. Also, we observe that proving in a larger range implies a proof of the second assertion of Theorem \[theo 1\] (ii) in the same range. This is the case because the remaining of the proof is independent of the condition $p\ge \frac{\log(4)}{\log(6)}.$\ The second conjecture that they posed is the following. [[Conjecture B (for the star graph $S_n$): For any $n\ge 2$ and $p\in (0,1]$ we have $${\bf C}_{S_n,p}=1-\frac{1}{n}.$$ ]{}]{} In this case we prove that, in fact, this equality is not true for $p>1$. In fact, for $n=3$, we find some bounds different to the ones conjectured in that case. However, we give a positive answer to this conjecture when $1/2\leq p\leq1$ for all $n\geq 2$. Moreover, we give a positive answer to the conjecture when $0<p<1/2$ if $n$ is sufficiently large, this is the content of our Theorem \[theo 2\]. \[theo 2\] Let $S_{n}=(V,E)$ be a start graph with $n$ vertices $(a_{1},a _{2},\dots,a_n)$, with center at $a_1$. Then, the following hold. - For all $1<p\leq\infty$ we have that $$\begin{aligned} \label{star graph n=3} {\bf C}_{S_3,p}=\frac{(1+2^{p/(p-1)})^{(p-1)/p}}{3}.\end{aligned}$$ - If $p=1$, then $$\label{star graph p=1} {\bf C}_{S_n,p}=1-\frac{1}{n}.$$ - If $n=4$ and $0<p<1,$ or $n\ge 5$ and $\frac{1}{2}\le p\le 1,$ then $$\label{star graph p<1} {\bf C}_{S_n,p}=1-\frac{1}{n}.$$ Moreover, holds for every $\frac{1}{2}>p>0$ when $n\ge C(p),$ for some finite constant $C(p)$ depending only on $p.$ The range $(\frac{1}{2},1)$ in (iii) is certainly not optimal, to find improvements on this range is an interesting problem.\ [[Conjecture C (boundedness and continuity): Let $0<p,q\le \infty $ and $0\le \alpha<1.$ The operator $M_{\alpha,G}$ is bounded and continuous from $BV_{p}(V)$ to $BV_{q}(V)$, where $BV_{p}(V):=\{f:V\to \mathbb{R}\}$ is endowed with $\\|f\\|_{BV_{p}(V)}:={{\rm Var\,}}_{p}f$ (note that $\\|\cdot\\|_{BV_{p}(V)}$ depends on $G).$]{}]{} We prove that with a slight modification this affirmation is true. We also prove that a modification is strictly required. That is the content of our next theorem. \[theo 3\] Let $G_{n}=(V,E)$ be a graph with $n$ vertices $(a_{1},a _{2},\dots,a_n)$. The following statements hold. - [[\[Boundedness\]]{}]{} Let $\alpha\in[0,1)$. For all $0<p,q\leq\infty$ there exists a constant $C(n,p,q)>0$ such that $$\label{boundedness} {{\rm Var\,}}_q M_{\alpha,G_{n}}f\leq C(n,p,q){{\rm Var\,}}_p f.$$ for all functions $f:V\to{\mathbb{R}}$. - [[\[Continuity\]]{}]{} Let $0<p,q\leq\infty$. Consider a sequence of functions $f_j:V\to{\mathbb{R}}$ such that $\\|f_j-f_0\\|_{BV_p(V)}\to0$ as $j\to\infty$. 1. Assuming that $\lim_{j\to\infty}\min_{x\in V}\|f(x)-f_j(x)\|=0$. Then $$\label{continuity} {{\rm Var\,}}_q (M_{\alpha,G_{n}}f-M_{\alpha,G_n}f_j)\to0 \ \text{as}\ j\to\infty.$$ 2. could fail to be true without the extra assumption that $\lim_{j\to \infty} \min_{x\in V}\|f(x)-f_j(x)\|=0$. Optimal $l^p$ bounds for maximal operators on finite graphs. ------------------------------------------------------------ We are also interested in the $l^p$ norm of $M_G$ when acting on finite graphs. That is, to find the exact value of the expression $$\sup_{f:V\to \mathbb{R}, f\neq 0}\frac{\\|M_{G}f\\|_{p}}{\\|f\\|_{p}}=:\\|M_{G}\\|_{p},$$ where $\\|g\\|_{p}:=(\sum_{e\in V}\|g(e)\|^p)^{\frac{1}{p}},$ for $g:V\to \mathbb{R}.$ These norms were first treated by Soria and Tradacete, who found $\\|M_{G}\\|_p$ when $G=S_n$ and $G=K_n$, where $p\in (0,1)$ (see [@ST Proposition 2.7] and [@ST Theorem 3.1]). Their results rely strongly in Jensen’s inequality for the function $x\to x^p$ where $p\le 1$, so those methods are not available when $p>1$. In fact, they claimed that this problem was difficult when $p>1$ (see [@ST Remark 2.8]). The following inequality was proved by Soria and Tradacete \[See [@ST], Proposition 2.7\] $$\left(1+\frac{n-1}{n^2}\right)^{1/2}\leq \\|M_{K_n}\\|_2\leq \left(1+\frac{n-1}{n}\right)^{1/2}.$$ Our next result is a formula for the precise value of $\\|M_{K_n}\\|_{2}$ for $n\ge 2$. We also find extremizers for all $n\geq2$. Moreover, we prove that $\\|M_{K_{3n}}\\|_{2}=\\|M_{K_{3}}\\|_{2},$ for all $n\geq 2$. We list these results as follows. \[p=2, complete graph\] Let $K_n=(V,E)$ be the complete graph with $n$ vertices $V=\{a_1,a_2,\dots,a_n\}$. Then we have $$\\|M_{K_n}\\|_{2}= \max_{k\in\{\lfloor\frac{n}{3}\rfloor,\lceil\frac{n}{3}\rceil\}}\left(1-\frac{k}{2n}+\frac{(4kn-3k^2)^{1/2}}{2n}\right)^{1/2}.$$ In particular, we have. \[p=2, n=3m\] If $n=3m$ for some $m\in{\mathbb{N}}$, then $\\|M_{K_{3m}}\\|_2=\left(\frac{4}{3}\right)^{1/2}$. For $n=2$ we have $\\|M_{K_2}\\|_2=\frac{(3+5^{1/2})^{1/2}}{2}$. Similarly, the following inequality was also proved by Soria and Tradacete \[See [@ST], Proposition 3.4\] $$\left(1+\frac{n-1}{4}\right)^{1/2}\leq \\|M_{S_n}\\|_2\leq \left(\frac{n+5}{2}\right)^{1/2}.$$ Our next result is a formula for the precise value of $\\|M_{S_n}\\|_2$. Moreover, we find some extremizers. \[p=2, star graph\] Let $n\geq4$ and $S_n=(V,E)$ be the star graph with $n$ vertices $V=\{a_1,a_2,a_3,\dots,a_{n}\}$ and center at $a_1$. Then, the following holds. $$\label{p=2,star} \\|M_{S_n}\\|_2= \left(1+\frac{n-4}{8}+\frac{(n^2+8n)^{1/2}}{8}\right)^{1/2}.$$ It was observed by Soria and Tradecete that in the case $n=2$ the optimal constant is $\frac{[3+5^{1/2}]^{1/2}}{2}$ \[See remark 2.8 in [@ST]\], this coincides with our formula .\ Proof of optimal bounds for the $p$-variation of maximal functions. =================================================================== We start by proving our results on $K_n.$\ Optimal bounds for the $p$-variation on $K_n$: Proof of Theorem \[theo 1\] -------------------------------------------------------------------------- For every result listed in Theorem \[theo 1\] we can see that, taking $f=\delta_{a_1}$ in the definition of ${\bf C}_{K_n,p},$ we have the following. $${\bf C}_{K_n,p}\ge 1-\frac{1}{n}.$$ In the following we prove, in each case, that $$\begin{aligned} \label{grupperbound} {\bf C}_{K_n,p}\le 1-\frac{1}{n}.\end{aligned}$$ We assume without loss of generality that $f$ is non-negative. Let $$m:=m_n:=\frac{\sum_{i=1}^{n}f(a_{i})}{n},$$ and for all $k\in\{1,2,\dots,n-1\}$ we define $$m_k=\frac{\sum_{i=1}^{k}f(a_i)}{k}.$$ Reordering if necessary, we can assume without loss of generality that $$f(a_{n})\geq f(a_{n-1})\geq\dots\geq f(a_{r})\geq m> f(a_{r-1})\geq\dots\geq f(a_{1}),$$ thus we have that $$M_{K_n}f(a_{i})=f(a_{i}) \ \forall\ i\geq r\ \ \ \text{and}\ \ \ M_{K_n}f(a_{i})=m \ \forall\ i<r.$$ Observe that $m_1\leq m_2\leq m_3\leq\dots\leq m_{n-1}\leq m$. Therefore $$\begin{aligned} ({{\rm Var\,}}_p M_{K_n}f)^p&\leq &(n-1)(f(a_{n})-m)^p +(n-2)(f(a_{n-1})-m)^p\nonumber\\ &&\dots +(r-1)(f(a_{r})-m)^{p}\nonumber\\ &\leq &(n-1)(f(a_{n})-m)^p +(n-2)(f(a_{n-1})-m_{n-1})^p\nonumber\\ &&\dots +(r-1)(f(a_{r})-m_{r})^{p}\label{key ineq 1}.\end{aligned}$$ Then, we note that by Hölder inequality $$\begin{aligned} f(a_i)-m_i\leq\frac{\sum_{t=1}^{i-1}{\|f(a_i)-f(a_t)\|}}{i}\leq \frac{\left(\sum_{t=1}^{i}{\|f(a_i)-f(a_t)\|^p}\right)^{1/p}(i-1)^{1/{p'}}}{i}.\end{aligned}$$ Combining the two previous estimatives we obtain $$\begin{aligned} ({{\rm Var\,}}_p M_{K_n}f)^p &\leq &(n-1)(f(a_{n})-m)^p +(n-2)(f(a_{n-1})-m_{n-1})^p\\ &&\dots +(r-1)(f(a_{r})-m_{r})^{p}\\ &\leq& (n-1)\frac{\left(\sum_{t=1}^{n-1}{\|f(a_n)-f(a_t)\|^p}\right)^{p/p}(n-1)^{p/{p'}}}{n^p}\\ &&+ (n-2)\frac{\left(\sum_{t=1}^{n-2}{\|f(a_{n-1})-f(a_t)\|^p}\right)^{p/p}(n-2)^{p/{p'}}}{(n-1)^p}\\ &&\dots+(r-1)\frac{\left(\sum_{t=1}^{r-1}{\|f(a_r)-f(a_t)\|^p}\right)^{p/p}(r-1)^{p/{p'}}}{r^p}\\ &\leq& \left(\frac{n-1}{n}\right)^{p}{\sum_{t=1}^{n-1}{\|f(a_{n})-f(a_t)\|^p}}\\ &&+ \left(\frac{n-2}{n-1}\right)^{p}{\sum_{t=1}^{n-2}{\|f(a_{n-1})-f(a_t)\|^p}}\\ &&+\dots+ \left(\frac{r-1}{r}\right)^{p}{\sum_{t=1}^{r-1}{\|f(a_{r})-f(a_t)\|^p}}\\ &\leq& \left(\frac{n-1}{n}\right)^{p}({{\rm Var\,}}_pf)^p. \end{aligned}$$ From where we conclude in this case. Concluding the proof of this assertion of Theorem \[theo 1\]. We keep the notation of the previous proof and the assumption that $$f(a_n)\ge \dots f(a_r)\ge m> \dots f(a_1).$$ We assume for the remaining of the proof that $0<p<1.$ The simplest case of the theorem (that holds in full generality for $0<p<1$), that is when $r=n,$ can be proved directly by $$\begin{aligned} (n-1)\|f(a_n)-m\|^p\le &(n-1)\left\|\frac{n-1}{n}(f(a_n)-f(a_1))\right\|^p\\ &\le \left(\frac{n-1}{n}\right)^p(\|f(a_n)-f(a_1)\|^p+\sum_{i=2}^{n-1} (\|f(a_n)-f(a_i)\|^p+\|f(a_i)-f(a_1)\|^p)\\ &\le \left(\frac{n-1}{n}\right)^p ({{\rm Var\,}}_p f)^p,\end{aligned}$$ where, in the second inequality, we used that if $a,b\ge 0$ then $(a+b)^p\le a^p+b^p.$ So, in the following we assume that $r<n.$ Now we prove the assertion for $n=4.$ [Case $n=4.$]{} Since the case $r=4$ was already solved, we have two cases left. First we treat the case $r=3.$ [Case $r=3$. ]{}We have the following inequality. $$\begin{aligned} \label{grline} \left(\frac{3}{4}\right)^p(\|f(a_4)-f(a_3)\|^p+\|f(a_3)-f(a_2)\|^p+\|f(a_2)-f(a_1)\|^p)\ge \|f(a_4)-f(a_3)\|^p+\|f(a_3)-m\|^p.\end{aligned}$$\ [Step 1: Proving .]{} In order to prove this, we write $f(a_3)-f(a_2)=x$ and $f(a_4)-f(a_3)=y,$ then $m=\frac{f(a_1)+3f(a_2)+2x+y}{4}$ and $$\frac{f(a_1)+3f(a_2)+2x+y}{4}\le f(a_2)+x \implies f(a_1)+y\le f(a_2)+2x,$$ also $$m\ge f(a_2)\implies f(a_2)\le f(a_1)+2x+y.$$ Then $$\begin{aligned} \left(\frac{3}{4}\right)^p(\|f(a_4)-f(a_3)\|^p+\|f(a_3)-f(a_2)\|^p+\|f(a_2)-f(a_1)\|^p)= \left(\frac{3}{4}\right)^p(y^p+(f(a_2)-f(a_1))^p+x^p),\end{aligned}$$ Consider first the case where $f(a_2)-f(a_1)+2x\le 4y.$ Here, by Karamata’s inequality we have $$(3y)^p+(f(a_2)-f(a_1)+2x)^p\ge (4y)^p+(f(a_2)-f(a_1)+2x-y)^p,$$ then since $(3x)^p+(3(f(a_2)-f(a_1)))^p\ge (f(a_2)-f(a_1)+2x)^p$ we have $$\begin{aligned} \label{grline2} (3y)^p+(3x)^p+(3(f(a_2)-f(a_1)))^p\ge (4y)^p+(f(a_2)-f(a_1)+2x-y)^p,\end{aligned}$$ from where follows. In the other case, where $f(a_2)-f(a_1)+2x\ge 4y,$ considering that $$(4^p-3^p)\left(\frac{f(a_2)-f(a_1)}{4}+\frac{x}{2}\right)^p+(f(a_2)-f(a_1)+2x)^p\ge (4^p-3^p)(y)^p+(f(a_2)-f(a_1)+2x-y)^p,$$ we have that follows by $$\begin{aligned} (3x)^p+(3(f(a_2)-f(a_1)))^p&\ge \left(\frac{4^p-3^p}{4^p}+1\right)(f(a_2)-f(a_1))^p+\left(\frac{4^p-3^p}{2^p}+2^p\right)x^p\\ &\ge(4^p-3^p)\left(\frac{f(a_2)-f(a_1)}{4}+\frac{x}{2}\right)^p+(f(a_2)-f(a_1)+2x)^p,\end{aligned}$$ where in the first inequality we use that $3^p\ge \frac{4^p-3^p}{4^p}+1$ and $3^p\ge \frac{4^p-3^p}{2^p}+2^p,$ both consequences of the inequality $$6^p+3^p=e^{\log(6)p}+e^{\log(3)p}\ge 2e^{\frac{\log(18)}{2}p}\ge 2(4)^p.$$ In the second inequality we use $$(4^p-3^p)\left(\frac{f(a_2)-f(a_1)}{4}+\frac{x}{2}\right)^p\le \frac{4^p-3^p}{4^p}(f(a_2)-f(a_1))^p+\frac{4^p-3^p}{2^p}x^p$$ and $$(f(a_2)-f(a_1)+2x)^p\le (f(a_2)-f(a_1))^p+2^px^p.$$ In the following, we also need the inequality $$\begin{aligned} \label{gr2} \left(\frac{3}{4}\right)^p(\|f(a_4)-f(a_2)\|^p+\|f(a_3)-f(a_1)\|^p)\ge \|f(a_4)-m\|^p+\|f(a_3)-m\|^p.\end{aligned}$$\ [Step 2: Proving .]{} We have that is equivalent to $$\begin{aligned} (3x+3y)^p+(3x+3(f(a_2)-f(a_1)))^p\ge (f(a_2)-f(a_1)+2x+3y)^p+(f(a_2)-f(a_1)+2x-y)^p. \end{aligned}$$ Here we distinguish among two cases, the first when $x+4y\ge f(a_2)-f(a_1).$ Here, by the concavity of the function $x\to x^p$, since $$4x+2(f(a_2)-f(a_1))+4y\ge 3x+3y\ge 2x+f(a_2)-f(a_1)-y,$$ we have $$\begin{aligned} (3x+3y)^p+(3x+3(f(a_2)-f(a_1)))^p&\ge (4x+2(f(a_2)-f(a_1))+4y)^p+(2x+(f(a_2)-f(a_1))-y)^p\\ &\ge ((f(a_2)-f(a_1))+2x+3y)^p+(f(a_2)-f(a_1)+2x-y)^p, \end{aligned}$$ from where follows. In the other case, where $x+4y\le f(a_2)-f(a_1),$ we can prove that $$\begin{aligned} \label{gr3} (f(a_2)-f(a_1)+2x+2y)^p+(f(a_2)-f(a_1)+2x)^p\ge (f(a_2)-f(a_1)+2x+3y)^p+(f(a_2)-f(a_1)+2x-y)^p, \end{aligned}$$ by Karamata’s inequality. Also, since $y\le \frac{f(a_2)-f(a_1)}{4}$ and $2x+y\ge f(a_2)-f(a_1),$ we have $x\ge \frac{3(f(a_2)-f(a_1))}{8},$ therefore we obtain $$\begin{aligned} \label{grlnineq} \log(3x)+\log(3x+3(f(a_2)-f(a_1)))\ge \log\left(\frac{3}{2}(f(a_2)-f(a_1))+2x\right)+\log(f(a_2)-f(a_1)+2x). \end{aligned}$$ Let us observe that $$\log(3x)\le \log(f(a_2)-f(a_1)+2x)\le \log\left(\frac{3}{2}(f(a_2)-f(a_1)+2x\right),$$ let us take then $v:=\log(f(a_2)-f(a_1)+2x)+\log(\frac{3}{2}(f(a_2)-f(a_1)+2x)-\log(3x),$ by Karamata’s inequality (now applied to the convex function $x\to e^{px}$) and we have $$\begin{aligned} e^{p\log (\frac{3}{2}(f(a_2)-f(a_1))+2x)}+e^{p\log(f(a_2)-f(a_1)+2x)}\le e^{p\log(3x)}+e^{pv}\le e^{p\log(3x)}+e^{p\log(3x+3(f(a_2)-f(a_1)))} \end{aligned}$$ and therefore: $$\begin{aligned} (3x+3y)^p+(3x+3(f(a_2)-f(a_1)))^p &\ge (3x)^p+(3x+3(f(a_2)-f(a_1)))^p=e^{p\log(3x)}+e^{p\log(3x+3(f(a_2)-f(a_1)))}\\ &\ge e^{p\log(\frac{3}{2}(f(a_2)-f(a_1))+2x)}+e^{p\log(f(a_2)-f(a_1)+2x)}\\ &=(\frac{3}{2}(f(a_2)-f(a_1))+2x)^p+(f(a_2)-f(a_1)+2x)^p\\ &\ge (f(a_2)-f(a_1)+2x+2y)^p+(f(a_2)-f(a_1)+2x)^p, \end{aligned}$$ from where we obtain, by combining this with , the desired inequality .\ [Step 3: Conclusion case $r=3$.]{} The case $r=3$ then follows by combining , and the inequality $\frac{3}{4}(f(a_4)-f(a_1))\ge f(a_4)-m.$ [Case $r=2$.]{} Here, we have that $m\le f(a_2)\rightarrow 2x+y\le f(a_2)-f(a_1).$ We prove first the inequality $$\begin{aligned} \label{grliner=2} \|f(a_4)-f(a_3)\|^p+\|f(a_3)-f(a_2)\|^p+\|f(a_2)-m\|^p&\le \left(\frac{3}{4}\right)^p(\|f(a_4)-f(a_3)\|^p\\ &+\|f(a_3)-f(a_2)\|^p+\|f(a_2)-f(a_1)\|^p)\nonumber,\end{aligned}$$\ [Step 1: Proving .]{} Our desired inequality is equivalent to $$(4y)^p+(4x)^p+(f(a_2)-f(a_1)-2x-y)^p\le (3y)^p+(3x)^p+(3(f(a_2)-f(a_1)))^p.$$ We observe that $$(4y)^p+(4x)^p+(f(a_2)-f(a_1)-2x-y)^p\le (4y)^p+(4x)^p+(f(a_2)-f(a_1))^p,$$ also, since $x\le \frac{f(a_2)-f(a_1)}{2}$ and $y\le f(a_2)-f(a_1)$, we have $$(4^p-3^p)(x^p+y^p)\le (4^p-3^p)\left[\left(\frac{1}{2}\right)^p+1\right](f(a_2)-f(a_1))^p\le (3^p-1)(f(a_2)-f(a_1))^p,$$ because $4^p+8^p+2^p\le 2(6)^p+3^p$ by Jensen inequality, from where it follows . Also, we have that $$\begin{aligned} \label{grlast} (\|f(a_4)-f(a_2)\|^p+\|f(a_3)-f(a_1)\|^p)\left(\frac{3}{4}\right)^p\ge (\|f(a_4)-f(a_2)\|^p+\|f(a_3)-m\|^p),\end{aligned}$$ [Step 2: Proving .]{} We have that is equivalent to $$(4x+4y)^p+(f(a_2)-f(a_1)+2x-y)^p\le (3x+3y)^p+(3(x+f(a_2)-f(a_1)))^p,$$ but, since by Jensen $4^p+2^p\le 2(3)^p$ and $x+y\le f(a_2)-f(a_1)$, we have $$\begin{aligned} (4^p-3^p)(x+y)^p+(2(f(a_2)-f(a_1)))^p\le (3(f(a_2)-f(a_1)))^p\le (3(f(a_3)-f(a_1)))^p, \end{aligned}$$ from where follows by observing that $2x\le f(a_2)-f(a_1).$\ [Step 3: Conclusion of $r=2$ and $n=4.$]{} The case $r=2,$ and thus our result in $n=4$ follows by combining , and the inequality $f(a_4)-m\le (\frac{3}{4})(f(a_4)-f(a_1)).$ We conclude this part.\ Now we prove our assertion for general $n$ and $p\in [\frac{\log(4)}{\log(6)},1).$\ [[Case $n\ge 5$ and $1>p\ge \frac{\log(4)}{\log(6)}$.]{}]{} We write $x_i:=f(a_i)-f(a_r)$ for $i=n,..,r+1$, $u=f(a_r)-m$, $y_i=f(a_r)-f(a_i)$ for $i=r-1,..,1$. We have then that $\sum x_i+nu=\sum y_i.$ One key step is to prove the following. $$\begin{aligned} \label{grmaintool} \sum_{i=r+1}^{n}x_i^p+(r-1)u^p\le \left(1-\frac{1}{n}\right)^p\left(\sum_{i=r+1}^{n}x_i^p+\sum_{i=1}^{r-1}y_i^p\right).\end{aligned}$$ In order to prove that let us see that, by Karamata’s inequality (since $y_i\ge u$) $$\sum_{i=1}^{r-1}y_i^p\ge (r-2)u^p+\left(\sum_{i=1}^{r-1}y_i-(r-2)u\right)^p=(r-2)(u)^p+\left((n-r+2)u+\sum_{i=r+1}^{n}x_i\right)^p,$$ also, by Jensen’s inequality we have $((n-r+2)u+\sum_{i=r+1}^{n}x_i)^p\ge 2^{p-1}(((n-r+2)u)^p+(\sum_{i=r+1}^nx_i)^p.$ Therefore, $$\begin{aligned} \left(1-\frac{1}{n}\right)^p\left(\sum_{i=r+1}^{n}x_i^p+\sum_{i=1}^{r-1}y_i^p\right)\ge \left(1-\frac{1}{n}\right)^p\left(\sum_{i=r+1}^nx_i^p+(r-2)u^p+2^{p-1}((n-r+2)u)^p+2^{p-1}(\sum_{i=r+1}^nx_i)^p\right),\end{aligned}$$ so, in order to get is enough (since $(\sum x_i)^p\ge (n-r)^{p-1}(\sum x_i^p)$ by Jensen) $$\begin{aligned} \label{grmaintoolpart1} 1\le \left(1-\frac{1}{n}\right)^p(1+2^{p-1}(n-r)^{p-1})\end{aligned}$$ and $$\begin{aligned} \label{grmaintoolpart2} r-1\le \left(1-\frac{1}{n}\right)^p(r-2+2^{p-1}(n-r+2)^p).\end{aligned}$$ In order to prove is enough (since $r\le n-1$ and $n\ge 3$) $$\begin{aligned} 1\le \left(\frac{2}{3}\right)^p(1+2^{p-1}) \end{aligned}$$ and that is equivalent to $2(3)^p\le 2^p3,$ an elementary fact. Now, to prove , we observe that (since $p\ge \frac{\log(4)}{\log(6)}$), $$2^{p-1}(n-r+2)^p\ge 2^{p-1}(3)^p\ge 2,$$ and therefore $$\left(1-\frac{1}{n}\right)^p(r-2+2^{p-1}(n-r+2)^p)\ge \left(1-\frac{1}{n}\right)^p(r)\ge \left(1-\frac{1}{n}\right)r\ge (r-1),$$ from where follows. Now we do an inductive argument. Assume that inequality holds for every $f$ in $K_{n-1},$ whenever $p\ge \frac{\log(4)}{\log(6)}$ (it holds for $n=3,4$). Then, if $b_1,\dots,b_{n-1}$ are the vertex of the $K_{n-1}$ graph, defining $\widetilde{f}$ as $\widetilde{f}(b_i)=f(a_{i+1})$ for $i=r,\dots,n-1$ and $\widetilde{f}(b_i)=f(a_{i})$ for $i=1,\dots,r-1.$ We write $\widetilde{m}=\frac{\sum_{i=1}^{n-1}\widetilde{f}}{n-1},$ is clear that $\widetilde{m}\le m.$ Then, by the inductive hypothesis, we have $$\sum_{i,j\in \{r+1,..,n\}}\|f(a_i)-f(a_j)\|^p+(r-1)\sum_{i=r+1}^n \|f(a_i)-\widetilde{m}\|^p\le \left(1-\frac{1}{n-1}\right)^p\left(\sum_{i,j\in \{1,..r-1,r+1,..,n\}}\|f(a_i)-f(a_j)\|^p\right),$$ therefore (since $f(a_i)-m\le f(a_i)-\widetilde{m}$ for $i=r+1,..n.$) we have $$\sum_{i,j\in \{r+1,..,n\}}\|f(a_i)-f(a_j)\|^p+(r-1)\sum_{i=r+1}^n \|f(a_i)-m\|^p\le \left(1-\frac{1}{n}\right)^p\left(\sum_{i,j\in \{1,..r-1,r+1,..,n\}}\|f(a_i)-f(a_j)\|^p\right),$$ Combining this with we conclude $$\sum_{i,j\in \{r,..,n\}}\|f(a_i)-f(a_j)\|^p+(r-1)\sum_{i=r}^n \|f(a_i)-m\|^p\le \left(1-\frac{1}{n}\right)^p\left(\sum_{i,j\in \{1..,n\}}\|f(a_i)-f(a_j)\|^p\right),$$ that is equivalent to in this case. This concludes the proof of our theorem. 0ptimal bounds for the $p$-variation on $S_n$: Proof of Theorem \[theo 2\] -------------------------------------------------------------------------- Now we deal with the problems related to the $p$-variation of the maximal operator in $S_n.$ We assume without loss of generality that $f$ is non-negative. We analyse three different cases. [[Case 1: $f(a_1)\geq\max\{f(a_2),f(a_3)\}$ ]{}]{}.\ In this case we have that $M_{S_3}f(a_1)=f(a_1)$, then $$\begin{aligned} ({{\rm Var\,}}_p M_{S_{3}}f)^p&\leq \left(f(a_1)-\frac{f(a_1)+f(a_2)}{2}\right)^p+\left(f(a_1)-\frac{f(a_1)+f(a_3)}{2}\right)^p\\ &\leq \frac{1}{2^{p}}({{\rm Var\,}}_p f)^p.\end{aligned}$$ [[Case 2: $f(a_1)\leq\min\{f(a_2),f(a_3)\}$ ]{}]{}. We assume without loss of generality that $f(a_1)\leq f(a_3)\leq f(a_2)$. Then $$\begin{aligned} ({{\rm Var\,}}_p M_{S_{3}}f)^p&= \left(f(a_2)-\frac{f(a_1)+f(a_2)+f(a_3)}{3}\right)^p+\left(\left[f(a_3)-\frac{f(a_1)+f(a_2)+f(a_3)}{3}\right]_{+}\right)^p\\ &=\left(\frac{f(a_2)-f(a_1)+f(a_2)-f(a_3)}{3}\right)^p+\left(\left[\frac{f(a_3)-f(a_1)-(f(a_2)-f(a_3))}{3}\right]_{+}\right)^p\\ &=\left(\frac{2(f(a_2)-f(a_1))-(f(a_3)-f(a_1))}{3}\right)^p+\left(\left[\frac{2(f(a_3)-f(a_1))-(f(a_2)-f(a_1))}{3}\right]_{+}\right)^p\\ &\leq \left(\frac{2(f(a_2)-f(a_1))-(f(a_3)-f(a_1))}{3}+\left[\frac{2(f(a_3)-f(a_1))-(f(a_2)-f(a_1))}{3}\right]_{+}\right)^p \\ &\leq \frac{(1+2^{p'})^{p/p'}}{3^p}({{\rm Var\,}}_p f)^p.\end{aligned}$$ Where we have used the fact that $p>1$ in the fourth line and the final step follows by Hölder’s inequality.\ [[Case 3: $\min\{f(a_2),f(a_3)\}<f(a_1)<\max\{f(a_2),f(a_3)\}$]{}]{}. We assume without loss of generality that $f(a_3)< f(a_1)< f(a_2)$. Then, since $p>1$, by Holder inequality we have $$\begin{aligned} ({{\rm Var\,}}_p M_{S_{3}}f)^p&= \left(f(a_2)-M_{S_3}f(a_1)\right)^p+\left(M_{S_3}f(a_1)-\frac{f(a_1)+f(a_2)+f(a_3)}{3}\right)^p\\ &\leq \left(f(a_2)-\frac{f(a_1)+f(a_2)+f(a_3)}{3}\right)^p\\ &=\left(\frac{2(f(a_2)-f(a_1))+(f(a_1)-f(a_3))}{3}\right)^p\\ &\leq \frac{(1+2^{p'})^{p/p'}}{3^p}({{\rm Var\,}}_p f)^p.\end{aligned}$$ This conclude the proof of $$\begin{aligned} {\bf C}_{S_3,p}\le \frac{(1+2^{p/(p-1)})^{(p-1)/p}}{3}. \end{aligned}$$ in . Finally, we observe that $$\begin{aligned} \label{grupperboundn=3} {\bf C}_{S_3,p}\ge \frac{(1+2^{p/(p-1)})^{(p-1)/p}}{3}. \end{aligned}$$ For that we consider the function $f:V\to{\mathbb{R}}$ defined by $$f(a_3)=2, f(a_1)=3 \ \text{and}\ f(a_2)=3+2^{\frac{1}{p-1}}.$$ Then, ${{\rm Var\,}}_pf=(1+2^{\frac{p}{p-1}})^{\frac{1}{p}}$. Moreover, $$M_{S_3}f(a_2)=f(a_2)=3+2^{\frac{1}{p-1}}\ \text{and}\ \ M_{S_3}(a_3)=M_{S_3}f(a_1)=\frac{2+3+3+2^{\frac{1}{p-1}}}{3}.$$ Thus $${{\rm Var\,}}_{p} M_{S_3}f=M_{S_3}f(a_2)-M_{S_3}f(a_1)=\frac{1+2^{\frac{p}{p-1}}}{3}.$$ Therefore $$\frac{{{\rm Var\,}}_{p}M_{S_3}f}{{{\rm Var\,}}_{p}f}=\frac{(1+2^{p'})^{\frac{1}{p'}}}{3}.$$ So, we obtain and thus . The proof of the previous result provides an example where the value $$\sup_{f:V\to \mathbb{R}; {{\rm Var\,}}_{p}f>0}\frac{{{\rm Var\,}}_{p}M_{G}f}{{{\rm Var\,}}_{p}f}$$ is not attained by any [Dirac delta.]{} This is a sign of the complexity of this problem when $p>1,$ since is not clear how the extremizers should behave for $n>3.$ In the case $p=2$, an interesting example is the following: let $S_n=(V,E)$ as in the Theorem \[theo 2\], consider the function $f:V\to{\mathbb{R}}$ defined by $$f(a_1)=n, \ f(a_2)=n+(n-1),\ \ \text{and}\ \ f(a_i)=n-1 \ \ \text{for all}\ \ 3\leq i\leq n.$$ In this case $$M_{S_n}f(a_2)=n+(n-1)\ \ \text{and}\ \ M_{S_n}f(a_i)=n+\frac{1}{n} \ \text{for all}\ \ i\neq 2.$$ Then $$\frac{{{\rm Var\,}}_{2}M_{S_n}f}{{{\rm Var\,}}_2f}=\frac{n-1-\frac{1}{n}}{[(n-1)^2+(n-2)]^{1/2}}=\frac{[(n-1)^2+(n-2)]^{1/2}}{n}>\frac{n-1}{n}.$$ This provides further evidence to the fact that in general the extremizers on $S_n$ are different when $p>1$ than when $p\le 1.$\ Now we deal with the next assertion of our theorem. Taking $f=\delta_{a_2}$ on the definition of ${\bf C}_{S_n,p}$ we have that $${\bf C}_{S_n,p}\ge 1-\frac{1}{n}.$$ In the following we prove the inequality $$\begin{aligned} \label{grupperboundestrella} {\bf C}_{S_n,p}\le 1-\frac{1}{n},\end{aligned}$$ from where both assertion follows. This inequality is equivalent to $$\begin{aligned} \label{grinequalitystar} {{\rm Var\,}}_{p} M_{S_{n}}f\le (1-\frac{1}{n}) {{\rm Var\,}}_{p}f, \end{aligned}$$ for all functions $f:V\to \mathbb{R}.$ We assume without loss of generality that $f$ is non-negative. Let $$m=\frac{1}{n}\sum_{i=1}^{n}f(a_i).$$ Then $$\begin{aligned} {{\rm Var\,}}M_{S_n}f&=\sum_{i=2}^{n}\|M_{S_n}f(a_i)-M_{S_n}f(a_1)\|\\ &=\sum_{M_{S_n}f(a_i)>M_{S_n}f(a_1)}M_{S_n}f(a_i)-M_{S_n}f(a_1)+\sum_{M_{S_n}f(a_1)>M_{S_n}f(a_i)}M_{S_n}f(a_1)-M_{S_n}f(a_i)\\ &= \sum_{M_{S_n}f(a_i)>M_{S_n}f(a_1)}f(a_i)-M_{S_n}f(a_1)+\sum_{M_{S_n}f(a_1)>M_{S_n}f(a_i)}f(a_1)-M_{S_n}f(a_i)\\ &\leq \sum_{M_{S_n}f(a_i)>M_{S_n}f(a_1)}f(a_i)-m+\sum_{M_{S_n}f(a_1)>M_{S_n}f(a_i)}f(a_1)-m\\ &=\sum_{M_{S_n}f(a_i)>M_{S_n}f(a_1)}\left[\frac{n-1}{n}(f(a_i)-f(a_1))+\sum_{j\neq i}\frac{f(a_1)-f(a_j)}{n}\right]\\ &\ \ \ \ \ \ \ \ \ \ +\sum_{M_{S_n}f(a_1)>M_{S_n}f(a_i)}\sum_{k=2}^{n}\frac{f(a_1)-f(a_k)}{n}\\ &\leq \frac{n-1}{n}{{\rm Var\,}}f, \end{aligned}$$ from where follows and therefore our result. We write $f(a_2)\ge \dots \ge f(a_r)\ge m>f(a_{r+1})\ge \dots\ge f(a_{n}).$ We distinguish among two cases, the first being $f(a_1)\le m.$\ [Case 1: $f(a_1)\le m.$]{} In this case is enough to prove inequality when $f(a_i)<f(a_1)$ for $i>r$. In fact, if fails for some $f$ with $f(a_i)>f(a_1)$ and $i>r$, it also fails for the function $\widetilde{f}$ defined by $\widetilde{f}(e)=f(e)$ for every $e\notin \{a_2,a_i\}$, $\widetilde{f}(a_i)=2f(a_1)-f(a_i)$ and $\widetilde{f}(a_2)=f(a_2)+f(a_i)-\widetilde{f}(a_i).$ This holds because $$(1-\frac{1}{n})^p(f(a_2)-f(a_1)^p)-(f(a_2)-m)^p\ge (1-\frac{1}{n})^p(\widetilde{f}(a_2)-\widetilde{f}(a_1))^p-(\widetilde{f}(a_2)-m)^p,$$ and this is the case because $$\|\widetilde{f}(a_2)-\widetilde{f}(a_1)\|^p-\|f(a_2)-f(a_1)\|^p\le \|\widetilde{f}(a_2)-m\|^p-\|f(a_2)-m\|^p,$$ inequality that follows because $\widetilde{f}(a_1)=f(a_1)$, the concavity of the function $x\to x^p$ (and thus the function $x\to (x+c)^p-x^p$ is decreasing for $x,c>0$) and the fact that $f(a_1)\le m.$ Iterating the previous argument we get the desired reduction. We write $f(a_i)-m=x_i$ for $i=2,...r;$ $m-f(a_1)=u$ and $y_i=f(a_1)-f(a_i)$ for $i=r+1,...,n.$ Observe that given our reduction we have $y_i\ge 0.$ We observe that $$\begin{aligned} \sum_{i=2}^r x_i=u+\sum_{i=r+1}^{n}(u+y_i), \end{aligned}$$ from where we obtain $u\le \frac{\sum_{i=2}^r x_i}{n-r+1}.$ Also, let us observe that is equivalent in this case to $$\sum_{i=2}^{r}\|x_i\|^p\le \left(1-\frac{1}{n}\right)^p\left(\sum_{i=2}^r\|x_i+u\|^p+\sum_{i=r+1}^{n}\|y_i\|^p\right),$$ but since $\sum_{i=r+1}^{n}\|y_i\|^p\ge \|\sum_{i=r+1}^{n}y_i\|^p=\|\sum_{i=2}^{r} x_i-(n-r+1)u\|^p.$ Then, we observe that the function $g(u):=\sum_{i=2}^r\|x_i+u\|^p+\|\sum_{i=2}^{r} x_i-(n-r+1)u\|^p,$ for $u\in [0, \frac{\sum_{i=2}^r x_i}{n-r+1}]$ is concave, therefore $g\ge \min \{g(\frac{\sum_{i=2}^r x_i}{n-r+1}),g(0)\},$ so, it is enough to prove that $$\begin{aligned} \label{grborder1} \sum_{i=2}^{r}\|x_i\|^p\le \left(1-\frac{1}{n}\right)^p\left(\sum_{i=2}^{r}\|x_i\|^p+\|\sum_{i=2}^{r}x_i\|^p\right),\end{aligned}$$ and $$\begin{aligned} \label{grborder2} \sum_{i=2}^{r}\|x_i\|^p\le \left(1-\frac{1}{n}\right)^p\left(\sum_{i=2}^{r}\|x_i+\frac{\sum_{i=2}^r x_i}{n-r+1}\|^p\right),\end{aligned}$$ for we observe that $\|\sum_{i=2}^r x_i\|^p\ge \max_{i=2,..r}\|x_i\|^p\ge \frac{\sum_{i=2}^r\|x_i\|^p}{r-1}$, so $$\left(1-\frac{1}{n}\right)^p\left(\sum_{i=2}^{r}\|x_i\|^p+\|\sum_{i=2}^{r}x_i\|^p\right)\ge \left(\sum_{i=2}^r\|x_i\|^p\right)\left(1-\frac{1}{n}\right)^p\left(1+\frac{1}{r-1}\right)\ge \left(\sum_{i=2}^r\|x_i\|^p\right),$$ concluding this inequality. For , we notice that $x_i+\frac{\sum_{i=2}^{r}x_i}{n-r+1}\ge x_i(1+\frac{1}{n-r+1}),$ so, since $(1-\frac{1}{n})^p(1+\frac{1}{n-r+1})\ge 1$ for $r\ge 2,$ we conclude this inequality, and therefore this case. Notice that this argument holds for every $p\in (0,1).$\ [Case 2: $f(a_1)>m.$]{} Here, we observe that if $f(a_2)\leq f(a_1)$ then $\|M_{S_n}f(a_1)-M_{S_n}f(a_i)\|\le \frac{\|f(a_1)-f(a_i)\|}{2}$ for all $i$ and thus follows in this case. So we can assume that $f(a_2)>f(a_1).$ Let us take $k$ such that $f(a_2)\ge f(a_3)\ge \dots f(a_k)\ge f(a_1)\ge f(a_{k+1}),$ and $s$ is the minimum such that $f(a_1)+f(a_{s+1})\ge 2m.$ Let us write $u=f(a_1)-m,$ $f(a_i)-f(a_1)=x_i$ for $i=2,..k$ and $y_i=f(a_1)-f(a_i)$ for $i=k+1,\dots n.$ We observe that $\sum_{i=2}^k x_i+nu=\sum_{k+1}^{n}y_i.$ Then is equivalent to $$\begin{aligned} \label{grequivalentstar} \sum_{i=2}^k(x_i)^p+\sum_{i=k+1}^{s}\left(\frac{y_i}{2}\right)^p+\sum_{s+1}^{n}u^p\le \left(1-\frac{1}{n}\right)^p\left(\sum_{i=2}^kx_i^p+\sum_{i=k+1}^{n}y_i^p\right).\end{aligned}$$ It is useful to solve first the case $k=n-1$. In this case we observe that $y_{n}=\sum_{i=2}^{k}x_i+nu,$ here $\sum_{i=2}^kx_i^p+u^p\le (1-1/n)^p(\sum_{i=2}^k x_i^p+(\sum_{i=2}^k x_i+nu)^p),$ from where we conclude this case. This claim follows by $$\begin{aligned} \sum_{i=2}^k(nx_i)^p+(nu)^p&\le \sum_{i=2}^k(nx_i)^p+(\sum_{i=2}^k x_i+nu)^p\\ &\le \sum_{i=2}^k((n-1)x_i)^p+\left((\sum_{i=2}^k x_i+nu)(n-1)\right)^p,\end{aligned}$$ since $(n-1)^p-1\ge (n^p-(n-1)^p),$ by Jensen’s inequality. So, we assume in the following that $k\le n-2.$ We observe that $u\le \frac{y_i}{2}$ for $i=s+1,..n,$ and thus $$\sum_{i=2}^k(x_i)^p+\sum_{i=k+1}^{s}\left(\frac{y_i}{2}\right)^p+\sum_{s+1}^{n}u^p\le \sum_{i=2}^k(x_i)^p+\sum_{i=k+1}^{n}\left(\frac{y_i}{2}\right)^p,$$ therefore would follow if $$\sum_{i=2}^k (x_i)^p\left(1-\left(1-\frac{1}{n}\right)^p\right)\le \left(\left(1-\frac{1}{n}\right)^p-\frac{1}{2^p}\right)\left(\sum_{i=k+1}^{n} y_i^p\right),$$ but by Jensen’s inequality $\sum_{i=k+1}^{n} y_i^p\ge (\sum_{i=k+1}^{n}y_i)^p\ge (\sum_{i=2}^k x_i)^p\ge (k-1)^{p-1}(\sum_{i=2}^kx_i^p)$, so, we need $(k-1)^{1-p}(1-(1-\frac{1}{n})^p)\le (1-\frac{1}{n})^p-\frac{1}{2^p}.$ Since $k-1\le n-3$ is enough $$\begin{aligned} \label{grgoal} (n-3)^{1-p}\left(1-\left(1-\frac{1}{n}\right)^p\right)\le \left(1-\frac{1}{n}\right)^p-\frac{1}{2^p},\end{aligned}$$ but that is equivalent to $$(n-3)^{1-p}(n^p-(n-1)^p)\le (n-1)^p-\left(\frac{n}{2}\right)^p,$$ the, is enough to prove $$\begin{aligned} \label{gr00}(n-3)^{1-p}p(n-1)^{p-1}\le (n-1)^p-\left(\frac{n}{2}\right)^p,\end{aligned}$$ or, the stronger bound, $p\le (n-1)^p-(\frac{n}{2})^p.$ Fixed $p$, is possible to observe that this last inequality holds for $n$ big enough. Therefore, we conclude the last statement of Theorem \[theo 2\] (iii). Now we assume that $1>p\ge \frac{1}{2}.$ First observe that for $n\ge 6$ we have that $p\le (n-1)^p-(\frac{n}{2})^p,$ in fact $g(n)=(n-1)^p-(\frac{n}{2})^p$ is increasing for $n\ge 6$ and $p>0.$ So, we need to prove $p\le 5^p-3^p$, but $g(p)=5^p-3^p-p$ is increasing for $p\ge \frac{1}{2}$ and $\sqrt{5}-\sqrt{3}-\frac{1}{2}\ge 0.$ Then, considering [@LX Theorem 1.4], the only cases left are $n=4$ and $n=5.$ For $n=4$, considering , we just need $$\begin{aligned} \left(\frac{1}{3}\right)^{1-p}p\le 3^p-2^p, \end{aligned}$$ or, equivalently, $p\le 3-3(\frac{2}{3})^p$, but $g(p)=3-3(\frac{2}{3})^p-p$ is concave in $(0,1),$ so, since $g(0)=0=g(1),$ we conclude in this case. Notice that this argument holds for every $1>p>0,$ and therefore the case $n=4$ is completed. Finally, for $n=5,$ we just need $$\left(\frac{1}{2}\right)^{1-p}p\le 4^p-\left(\frac{5}{2}\right)^p,$$ or equivalently $$\frac{p}{2}\le 2^p-\left(\frac{5}{4}\right)^p,$$ but $g(p)=2^p-(\frac{5}{4})^p-\frac{p}{2}$ is increasing for $p\ge \frac{1}{2},$ then since $\sqrt{2}-\sqrt{\frac{5}{4}}-\frac{1}{4}\ge 0$ we conclude this case. Since we finished the analysis of cases, we conclude the proof of the theorem. It is possible, in fact, to prove for every $0<p<1$ when $n=5,$ thus proving Theorem \[theo 2\](iii) for every $0<p<1$ in this case. We omit the details for simplicity. Qualitative results: Proof of Theorem \[theo 3\] ------------------------------------------------ In the last part of this section we prove our versions of the qualitative results conjectured in Conjecture C. We assume without loss of generality that $f$ is non-negative. Also, in the following we assume that $G$ is connected, since the general case follows from there. Given $u,v\in G_n:=\{a_1,a_2,\dots,a_n\}$, such that $M_{\alpha,G_n}f(u)>M_{\alpha,G_n}f(v)$, we observe that there exists $k\leq n-1$ such that $$M_{\alpha,G_n}f(u)=\frac{k^{\alpha}}{\|B(u,k)\|}\sum_{a_i\in B(u,k)}f(a_i),$$ then $$\begin{aligned} M_{\alpha,G_n}f(u)-M_{\alpha,G_n}f(v)&\leq\frac{k^{\alpha}}{\|B(u,k)\|}\sum_{a_i\in B(u,k)}f(a_i)-\frac{n^\alpha}{n}\sum_{i=1}^{n}f(a_i)\\ &\leq n^{\alpha}\left[\frac{1}{\|B(u,k)\|}\sum_{a_i\in B(u,k)}f(a_i)-\frac{1}{n}\sum_{i=1}^{n}f(a_i)\right]\\ &\leq n^{\alpha}(f(x)-f(y))\\ &\leq n^{\alpha}(n-1)^{\max\{1-\frac{1}{p},0\}}{{\rm Var\,}}_{p}f \end{aligned}$$ Where, in the third line $x\in G_n$ is choose such that $f(x):=\max\{f(a_i);a_i\in B(u,k)\}$ and $y\in G_n$ is choose such that $f(y):=\min\{f(a_i);a_i\in G_n\}$. In the fourth line we used Hölder inequality. Therefore $$\begin{aligned} {{\rm Var\,}}_{q}M_{\alpha,G_n}&=\left(\frac{1}{2}\sum_{u\in G_n}\sum_{v\in N_{G_n}(u)}\|M_{\alpha,G_n}f(u)-M_{\alpha,G_n}f(v)\|^q\right)^{1/q}\\ &\leq \left(\frac{n(n-1)}{2}\right)^{1/q}n^{\alpha}(n-1)^{\max\{\frac{p-1}{p},0\}}{{\rm Var\,}}_pf\\ &=C(n,p,q){{\rm Var\,}}_pf.\end{aligned}$$ We start observing that for all $j\geq 1$ $$\begin{aligned} \\|f-f_j\\|_{l^\infty(G_n)}&=\max_{y\in V}\|f(y)-f_j(y)\|-\min_{x\in V}\|f(x)-f_j(x)\| +\min_{x\in V}\|f(x)-f_j(x)\|\\ &\leq {{\rm Var\,}}(f-f_j)+\min_{x\in V}\|f(x)-f_j(x)\|\\ &\leq n^{\max\{1-1/p,0\}}{{\rm Var\,}}_p(f-f_j)+\min_{x\in V}\|f(x)-f_j(x)\|.\end{aligned}$$ Then, assuming that $\lim_{j\to\infty}\min_{x\in V}\|f(x)-f_j(x)\|=0$, we have that $$\\|f-f_j\\|_{l^\infty(G_n)}\to 0 \ \ \text{as} \ \ j\to\infty.$$ Moreover, for any $u,v\in G_n$ we have that $$\begin{aligned} M_{\alpha,G_n}f(u)-M_{\alpha,G_n}f_j(u)-[M_{\alpha,G_n}f(v)-M_{\alpha,G_n}f_j(v)]&\leq M_{\alpha,G_n}(f-f_j)(u)+M_{\alpha,G_n}(f-f_j)(v)\\ &\leq 2\\|f-f_j\\|_{l^1(G_n)}\\ &\leq 2n\\|f-f_j\\|_{l^{\infty}(G_n)}\to 0 \ \ \text{as}\ \ j\to\infty.\end{aligned}$$ Therefore $$\begin{aligned} {{\rm Var\,}}_q(M_{\alpha,G_n}f-M_{\alpha,G_n}f_j)\leq \left(\frac{n(n-1)}{2}\right)^{1/q}2n\\|f-f_j\\|_{l^{\infty}(G_n)}\to 0 \ \ \text{as}\ \ j\to\infty.\end{aligned}$$ Finally, we observe that without the assumption that $\lim_{j\to\infty}\min_{x\in V}\|f(x)-f_j(x)\|=0$ the continuity property could fail, with this purpose in mind consider the following situation: Let $G_n=S_n$ the star graph with $n$ vertices $V=\{a_1,a_2,\dots,a_n\}$ and center at $a_1$, for simplicity we take $\alpha=0$ and $p=q=1$. We define the function $f$ by $f(a_1)=2$ and $f(a_i)=1$ for all $i\neq 1$ thus $M_{S_n}f(a_1)=2$ and $M_{S_n}f(a_i)=3/2$ for all $i\neq1$. Then, we consider the sequence of functions $(f_{j})_{j\in{\mathbb{N}}}$ defined by $f_j(a_i)=f(a_i)-3$ for all $a_i\in V$ and for all $j\in {\mathbb{N}}$. Then ${{\rm Var\,}}(f-f_j)=0$ for all $j\in{\mathbb{N}}$, moreover $M_{S_n}f_j(a_1)=\frac{1+2(n-1)}{n}$ and $M_{S_n}f_j(a_i)=2$ for all $i\neq 1$. Therefore $$\begin{aligned} {{\rm Var\,}}(M_{S_n}f-M_{S_n}f_j)&\geq M_{S_n}f(a_1)-M_{S_n}f_j(a_1)-[M_{S_n}f(a_2)-M_{S_n}f_j(a_2)]\\ &=2-\frac{1+2(n-1)}{n}-[3/2-2]\\ &=\frac{1}{n}+\frac{1}{2}\ \ \text{for all}\ \ j\in{\mathbb{N}}.\end{aligned}$$ Then ${{\rm Var\,}}(M_{S_n}f-M_{S_n}f_j)\nrightarrow 0$ as $j\to \infty$. Proof of optimal bounds for the $2$-norm of maximal functions ============================================================= In this subsection we prove our results concerning the values $\\|M_{G}\\|_2$ for our graphs of interest. 2-norm of the maximal operator in $K_n$: Proof of Theorem \[p=2, complete graph\] and Corollary \[p=2, n=3m\] ------------------------------------------------------------------------------------------------------------- We start by proving that Corollary \[p=2, n=3m\] follows by Theorem \[p=2, complete graph\]. The inequality $$\\|M_{K_n}f\\|_2\leq \left(\frac{4}{3}\right)^{1/2}\\|f\\|_2$$ follows from the Theorem \[p=2, complete graph\], since $k=n/3$ in the right hand side. On the other hand, we consider the following example: we define $f:V\to{\mathbb{R}}$ by $$f(a_i)=4\ \text{for all}\ 1\leq i\leq \frac{n}{3} \ \text{and} \ f(a_i)=1\ \text{for all}\ \frac{n}{3}+1\leq i\leq n.$$ Then, in this case we have $$M_{K_n}f(a_i)=4\ \text{for all} \ 1\leq i\leq \frac{n}{3} \ \text{and}\ M_{K_n}f(a_i)=2 \ \text{for all}\ \frac{n}{3}+1\leq i\leq n.$$ Therefore $$\\|M_{K_n}f\\|_2=\left(\frac{\frac{16n}{3}+\frac{4(2n)}{3}}{\frac{16n}{3}+\frac{2n}{3}}\right)^{1/2}\\|f\\|_2=\left(\frac{4}{3}\right)^{1/2}\\|f\\|_2.$$ Now we prove our bound that holds for $K_n$ for every $n\ge 2.$ We assume without loss of generality that $f$ is no-negative. Consider the case $$f(a_1)\geq f(a_2)\geq\dots\geq f(a_k)\geq m\geq f(a_{k+1})\geq\dots\geq f(a_n).$$ Then, in this case $$M_{K_n}f(a_i)=f(a_i) \ \text{for all}\ 1\leq i\leq k,\ \text{and}\ M_{K_n}f(a_i)=m \ \text{for all} \ k+1\leq i\leq n.$$ Therefore $$\begin{aligned} \label{ineq:1 p=2 complete graph} \\|M_{K_n}f\\|^{2}_2&=\sum_{i=1}^{k}f(a_i)^{2}+(n-k)m^2\nonumber\\&=\left(1+\frac{n-k}{n^2}\right)\sum_{i=1}^{k}f(a_i)^{2}+\frac{n-k}{n^2}\sum_{i=k+1}^{n}f(a_i)^2\nonumber\\ &\ \ \ +\frac{2(n-k)}{n^2}\sum_{\substack{1\leq i<j\leq k}}f(a_i)f(a_j) +\frac{2(n-k)}{n^2}\sum_{\substack{k+1\leq i<j\leq n}}f(a_i)f(a_j)\nonumber\\ &\ \ \ +\frac{2(n-k)}{n^2}\sum_{\substack{1\leq i\leq k\\ k+1\leq j\leq n}}f(a_i)f(a_j)\nonumber\\ &\leq \left(1+\frac{n-k}{n^2}\right)\sum_{i=1}^{k}f(a_i)^{2}+\frac{n-k}{n^2}\sum_{i=k+1}^{n}f(a_i)^2\\ &\ \ \ +\frac{(n-k)(k-1)}{n^2}\sum_{i=1}^{k}f(a_i)^2+\frac{(n-k)(n-k-1)}{n^2}\sum_{i=k+1}^{n}f(a_i)^2\nonumber\\ &\ \ \ +\frac{2(n-k)}{n^2}\sum_{\substack{1\leq i\leq k\\ k+1\leq j\leq n}}f(a_i)f(a_j)\nonumber\\ &=A_k\sum_{i=1}^{k}f(a_i)^{2}+B_K\sum_{i=k+1}^{n}f(a_i)^2+\frac{2(n-k)}{n^2}\sum_{\substack{1\leq i\leq k\nonumber\\ k+1\leq j\leq n}}f(a_i)f(a_j),\end{aligned}$$ where $A_k:=1+\frac{(n-k)k}{n^2}$ and $B_k:=\frac{(n-k)^2}{n^2}$. Observe that $A_k-B_k=\frac{3nk-2k^2}{n^2}$ and by the AM-GM inequality $$\begin{aligned} \label{ineq:2 p=2 complete graph} \\|M_{K_n}f\\|^{2}_2&\leq A_k\sum_{i=1}^{k}f(a_i)^{2}+B_k\sum_{i=k+1}^{n}f(a_i)^2+\frac{2(n-k)}{n^2}\sum_{\substack{1\leq i\leq k\\ k+1\leq j\leq n}}f(a_i)f(a_j)\nonumber\\ &\leq A_k\sum_{i=1}^{k}f(a_i)^{2}+B_k\sum_{i=k+1}^{n}f(a_i)^2+\frac{1}{n^2}\sum_{\substack{1\leq i\leq k\\ k+1\leq j\leq n}}(xf(a_i)^2+yf(a_j)^2)\\ &=\left(A_k+\frac{(n-k)x}{n^2}\right)\sum_{i=1}^{k}f(a_i)^{2}+\left(B_k+\frac{ky}{n^2}\right)\sum_{i=k+1}^{n}f(a_i)^2\nonumber\end{aligned}$$ for all $0<x,y$ such that $xy=(n-k)^2$. Then, we choose $x,y$ such that $$A_k+\frac{(n-k)x}{n^2}=B_k+\frac{ky}{n^2}.$$ So, $x$ is the positive solution for the equation $$(3nk-2k^2)x+(n-k)x^2=k(n-k)^2.$$ More precisely $$x:=\frac{-(3nk-2k^2)+(4kn^3-3n^2k^2)^{1/2}}{2(n-k)}.$$ Therefore, combining and we obtain $$\begin{aligned} \\|M_{K_n}f\\|^{2}_2&\leq\max_{k\in[1,n-1]}\left(A_k+\frac{(n-k)x}{n^2}\right)\sum_{i=1}^{n}f(a_i)^2\\ &=\max_{k\in[1,n-1]}\left(1+\frac{(n-k)k}{n^2}+\frac{(4kn^3-3n^2k^2)^{1/2}-(3nk-2k^2)}{2n^2}\right)\sum_{i=1}^{n}f(a_i)^2\\ &=\max_{k\in[1,n-1]}\left(1-\frac{k}{2n}+\frac{(4kn-3k^2)^{1/2}}{2n}\right)\sum_{i=1}^{n}f(a_i)^2.\end{aligned}$$ Then, we consider the function $g:[1,n-1]\to{\mathbb{R}}$ defined by $g(t):=-t+(4tn-3t^2)^{1/2}$. Observe that $$\max_{t\in[1,n-1]}g(t)=g\left(\frac{n}{3}\right).$$ Moreover, $g$ is increasing in $[1,n/3]$ and decreasing in $[n/3,n-1]$. Therefore $$\label{ineq:3 p=2 complete graph} \\|M_{K_n}f\\|^{2}_2\leq \max_{k\in\{\lfloor\frac{n}{3}\rfloor,\lceil\frac{n}{3}\rceil\}}\left(1-\frac{k}{2n}+\frac{(4kn-3k^2)^{1/2}}{2n}\right)\\|f\\|^2_2.$$ Finally, observe that in order to have an equality in it is enough to have equality in and . Moreover, the equality in is attained if and only if $f(a_i)=f(a_1)=\gamma$ for all $1\leq i\leq k$, and $f(a_j)=f(a_{k+1})=\eta$ for all $k+1\leq j\leq n$, for some $0<\eta<\gamma$. We can assume without loss of generality that $\eta=1$. On the other hand, the equality in is attained if and only if $y^{1/2}=x^{1/2}\gamma=(n-k)^{1/2}\gamma^{1/2}$, or equivalently $\gamma=\frac{n-k}{x}$. Therefore, in order to obtain an equality in for $k \in\{\lfloor\frac{n}{3}\rfloor,\lceil\frac{n}{3}\rceil\}$ we consider the function $g_{k}:V\to{\mathbb{R}}$ defined by $$g_{k}(a_i)=\gamma:=\frac{2(n-k)^2}{(4kn^3-3n^2k^2)^{1/2}-(3nk-2k^2)} \ \ \ \text{for all} \ \ 1\leq i \leq k,$$ and $g_k(a_j)=1$ for all $k+1\leq j\leq n$. Then, by construction $$\\|M_{K_n}\\|_2=\max_{k\in\{\lfloor \frac{n}{3}\rfloor,\lceil \frac{n}{3}\rceil \}}\frac{\\|M_{K_n}g_{k}\\|_2}{\\|g_k\\|_2}.$$ this shows that our bound is optimal, moreover we have found extremizers. Observe that, in the particular case when $n=3k$, we obtain $\gamma=4$ as in the Corollary \[p=2, n=3m\]. 2-norm of the maximal operator in $S_n$: Proof of Theorem \[p=2, star graph\]. ------------------------------------------------------------------------------ Now we prove our result concerning the $2$-norm of our maximal operator on $S_n.$ As usual we assume without loss of generality that $f$ is no negative and we denote by $m$ the average of $f$ along $V$ [[i.e.]{}]{} $m=\frac{\sum_{i=1}^{n}f(a_i)}{n}$. We observe that $M_{S_n}f(a_1)=f(a_1)$ or $M_{S_n}f(a_1)=m.$ We study this two cases separately.\ [[Case 1: $M_{S_n}f(a_1)=f(a_1).$]{}]{} Assume without loss of generality that $M_{S_n}f(a_i)=f(a_i)$ for all $1\leq i\leq k$, $M_{S_n}f(a_i)=\frac{f(a_i)+f(a_1)}{2}$ for all $k+1\leq i\leq k+r$, and $M_{S_n}f(a_i)=m$ for all $k+r+1\leq i\leq n$. By Cauchy-Schwarz inequality we have $$m^2\leq \frac{\sum_{i=1}^{n}f(a_i)^2}{n}.$$ Then $$\begin{aligned} \\|M_{S_n}f\\|^2_2&\le \left(1+\frac{r}{4}\right)f(a_1)^2+\sum_{i=2}^{k}f(a_i)^2+\frac{1}{4}\sum_{i=k+1}^{k+r}f(a_i)^2+\frac{2}{4}\sum_{i=k+1}^{k+r}f(a_i)f(a_1) +\frac{s}{n}\sum_{i=1}^{n}f(a_i)^2\\ &=\left(1+\frac{r}{4}+\frac{s}{n}\right)f(a_1)^2+\left(1+\frac{s}{n}\right)\sum_{i=2}^{k}f(a_i)^2+\left(\frac{1}{4}+\frac{s}{n}\right)\sum_{i=k+1}^{k+r}f(a_i)^2\\ &\ \ \ +\frac{2}{4}\sum_{i=k+1}^{k+r}f(a_i)f(a_1) +\frac{s}{n}\sum_{i=k+r+1}^{n}f(a_i)^2.\end{aligned}$$ where $s:=n-k-r$. Moreover, for all $k+1\leq i\leq k+r$, we have that $$\frac{2}{4}f(a_i)f(a_1)\leq xf(a_1)^2+yf(a_i)^2$$ for all $x,y>0$ such that $xy\geq\frac{1}{16}$. We can choose $x$ and $y$ such that $$y-rx=1+\frac{r-1}{4} \ \text{and} \ \ xy=\frac{1}{16}.$$ or equivalently $$x:=\frac{[(r+9)(r+1)]^{1/2}-(r+3)}{8r}.$$ Therefore, for all $n\geq 4$ we have $$\begin{aligned} \\|M_{S_n}f\\|^2_2&\le \max_{\{k,r\in{\mathbb{N}};1\leq k+r\leq n\}}\left(1+\frac{n-k-r}{n}+\frac{r}{4}+\frac{[(r+9)(r+1)]^{1/2}-(r+3)}{8}\right)\\|f\\|^2_2\\ &\le \left(1+\frac{n-1}{4}+\frac{(n^2+8n)^{1/2}-(n+2)}{8}\right)\\|f\\|^2_2.\end{aligned}$$ [[Case 2: $M_{S_n}f(a_1)=m.$]{}]{} In this case $k\geq 2$. Following the same strategy (and notation), for all $n\geq 4$ we obtain that $$\begin{aligned} \\|M_{S_n}f\\|^2_2 &\le \left(\frac{r}{4}+\frac{s+1}{n}\right)f(a_1)^2+\left(1+\frac{s+1}{n}\right)\sum_{i=2}^{k}f(a_i)^2+\left(\frac{1}{4}+\frac{s+1}{n}\right)\sum_{i=k+1}^{k+r}f(a_i)^2\\ &\ \ \ +\frac{2}{4}\sum_{i=k+1}^{k+r}f(a_i)f(a_1) +\frac{s+1}{n}\sum_{i=k+r+1}^{n}f(a_i)^2.\\ &\leq \max_{\{k,r\in{\mathbb{N}};1\leq k+r\leq n\}}\left\{\frac{n-k-r+1}{n}+\frac{r+1}{4},\frac{n-k-r+1}{n}+1\right\}\\|f\\|^2_2\\ &=\max_{\{k,r\in{\mathbb{N}};1\leq k+r\leq n\}}\left\{\frac{n-k-r+1}{n}+\frac{r+1}{4},\frac{n-1}{n}+1\right\}\\|f\\|^2_2.\end{aligned}$$ The inequality $$\\|M_{S_n}\\|_2\le \left(1+\frac{n-1}{4}+\frac{(n^2+8n)^{1/2}-(n+2)}{8}\right)^{1/2}:=C_n$$ follows from these two estimates.\ Finally, we observe that $\\|M_{S_n}\\|_2=C_n$. Consider the function $g:V\to{\mathbb{R}}$ defined by $g(a_i)=1$ for all $1\leq i\leq n-1$ and $g(a_0)=\gamma$, where we choose $ \gamma$ to be a positive real number larger than 1, such that $\gamma$ is a solution for the quadratic equation $$aX^2+bX+c:=\left( C^2_{n}-1-\frac{(n-1)}{4}\right)x^2-\frac{n-1}{2}x+C^2_{n}(n-1)-\frac{n-1}{4}=0.$$ The existence of $\gamma$ follows from the definition of $C_n$, since we can see that $b^2-4ac=0$ and $\frac{-b}{2a}>1$. More precisely $$\gamma=-\frac{b}{2a}=\frac{2(n-1)}{(n^2+8n)^{1/2}-(n+2)}.$$ For this particular function we have $$\frac{\\|M_{S_n}g\\|_2}{\\|g\\|_2}=\left(\frac{\gamma^2+(n-1)\left(\frac{\gamma+1}{2}\right)^2}{\gamma^2+(n-1)}\right)^{1/2}= C_n.$$ This concludes the proof of our theorem. Acknowledgements. ================= The authors are thankful to Emanuel Carneiro and Terence Tao for very helpful comments. C.G.R was supported by CAPES-Brazil. [99]{} J.M. Aldaz and J. Pérez Lázaro, Functions of bounded variation, the derivative of the one-dimensional maximal function, and applications to inequalities, 2443–2461. J. Bober, E. Carneiro, K. Hughes and L. B. Pierce, On a discrete version of Tanaka’s theorem for maximal functions, 1669–1680. D. Beltran and J. Madrid, Endpoint Sobolev continuity of the fractional maximal function in higher dimensions, . D. Beltran and J. Madrid, Regularity of the centered fractional maximal function on radial functions, . E. Carneiro and C. González-Riquelme, Gradient bounds for radial maximal functions, E. Carneiro and K. Hughes, On the endpoint regularity of discrete maximal operators, 1245–1262. E. Carneiro and J. Madrid, Derivative bounds for fractional maximal operators, 4063–4092. E. Carneiro, J. Madrid and L. B. Pierce, Endpoint Sobolev and BV Continuity for Maximal Operators, E. Carneiro and B. F. Svaiter, On the variation of maximal operators of convolution type, 837–865. C. González-Riquelme, Sobolev regularity of polar fractional maximal functions, J. Kinnunen, The Hardy-Littlewood maximal function of a Sobolev-function, 117–124. J. Kinnunen and E. Saksman, Regularity of the fractional maximal function, 529–535. O. Kurka, On the variation of the Hardy-Littlewood maximal function, 109–133. H. Luiro, The variation of the maximal function of a radial function, H. Luiro and J. Madrid, The variation of the fractional maximal function of a radial function, Feng Liu and Qingying Xue, , J. Madrid, Sharp inequalities for the variation of the discrete maximal function, 94–107. J. Madrid, Endpoint Sobolev and BV Continuity for Maximal Operators, II, , To appear in Rev. Mat. Iberoam. A.D. Melas, The best constant for the centered [H]{}ardy-[L]{}ittlewood maximal inequality, 647–688. J. P. G. Ramos, Sharp total variation results for maximal functions, 41–64. Javier Soria and Pedro Tradacete, , Javier Soria and Pedro Tradacete, , F. Temur, On regularity of the discrete Hardy-Littlewood maximal function, preprint at https://arxiv.org/pdf/1303.3993.pdf.	{ "pile_set_name": "ArXiv" }
=6.5in \#1[Eq. (\[\#1\])]{} \#1[section \[\#1\]]{} \#1[Fig. (\[\#1\])]{} PUPT-1422\ hep-th/9310098 [Scaling Functions for Baby Universes]{} [in Two-Dimensional Quantum Gravity]{} [Steven S. Gubser and Igor R. Klebanov]{} Joseph Henry Laboratories\ Princeton University\ Princeton, New Jersey 08544 [Abstract]{} We apply the recently proposed transfer matrix formalism to 2-dimensional coupled to $(2, 2k-1)$ minimal models. We find that the propagation of a parent universe in geodesic (Euclidean) time is accompanied by continual emission of s and derive a distribution function describing their sizes. The $k\to \infty~ (c\to -\infty)$ limit is generally thought to correspond to classical geometry, and we indeed find a classical peak in the universe distribution function. However, we also observe dramatic quantum effects associated with s at finite length scales. October 1993 Introduction {#Intro} ============ One fascinating aspect of is the possibility of topology changing processes where a compact connected 3-geometry (a “baby universe”) splits off or joins a large “parent universe.” Such processes, peculiar as they may seem, are quite natural at the Planckian distance scales, where geometry undergoes large quantum fluctuations. Even if the topology changing processes are confined to the smallest distance scales, they may drastically affect the observed physics. As shown in ref. [@CGS], summation over all possible emissions and absorptions of the Planck-size baby universes effectively gives rise to averaging over the fundamental constants of nature. If the weight in the average is sharply peaked, then the fundamental constants are determined [@SC]. The wormhole theory of fundamental constants shows that, via some unexpected connections, the short-distance properties of can affect the universe at large in crucial ways. The status of this theory when applied to 4-dimensional is somewhat uncertain due to the instability of the Euclidean path integral. It may be, for instance, that the weight in the average over the coupling constants is not sharply peaked [@US]. If, on the other hand, the weight is sharply peaked, then it is not quite clear what prevents the emission and absorption of macroscopic baby universes [@VK]. While these puzzles await their resolution through a better understanding of 4-dimensional , we may try to gain some intuition about the topology changing processes from low-dimensional models. The 2-dimensional Euclidean coupled to $c\leq 1$ conformal matter appears to provide some promising toy models which are not plagued by any perturbative instabilities. Our understanding of these theories has dramatically improved recently, in part due to the success of the matrix model techniques [@GM]. With their help we can perform summations over discretized random surfaces and extract the exact results in the universal continuum limit. Until recently, however, the available exact solutions have not shed any light on the physics of topology changing processes. The problem is that the matrix models do not give any direct information on the internal geometry of 2-dimensional space-times. Some first pieces of information were extracted via direct Monte Carlo studies of triangulated random surfaces [@AM]. It was noted that the set of points at a given geodesic distance $D$ from some point $P$ typically consists of many disconnected loops, whose number grows rapidly with $D$. The study in ref. [@AM] gave a first indication that, as a 1-dimensional universe (a string) propagates forward in geodesic (Euclidean) time, it is very likely to continually emit baby universes. This phenomenon received a detailed quantitative confirmation in a recent very interesting paper by Kawai, Kawamoto, Mogami and Watabiki [@Kawai1]. Relying on new combinatorial techniques that they invented, these authors derived a formula for the average number of loops with lengths between $L$ and $L+dL$ located at a geodesic distance $D$ from some point on the surface, $$\rho(L, D) dL = {3\over 7 \sqrt{\pi} D^2} \left(x^{-5/2}+{1\over 2} x^{-3/2}+{14\over 3} x^{1/2} \right) e^{-x} dL \label{Scaling}$$ where $x=L/D^2$. A number of beautiful conclusions follow from this formula. First of all, $\rho(L, D)$ is not normalizable at small $L$, which means that the microscopic s are overwhelmingly likely to be emitted. Secondly, apart from an overall dimensionful factor, $\rho$ is a function of the dimensionless scaling variable $x$. Introducing the moments, $$\langle L^n \rangle_D=\int_0^\infty \rho(L, D) L^n dL \ ,$$ one finds that for $n<2$ they are dominated by the non-universal short-distance cut-off. But for $n\geq 2$ there are simple scaling relations $\langle L^n \rangle_D \sim D^{2n}$, whose form follows essentially from $D$ having the dimension of (Length)$^{1/2}$. The distribution function $\rho(L, D)$ is very important because it quantifies the effects of s in pure Euclidean 2-dimensional quantum gravity. The next question is how the coupling to matter affects the topology changing processes. In this paper we begin to address this question by considering the emission of baby universes in 2-dimensional coupled to the $(2, 2k-1)$ minimal models, whose central charges are $c=1-3(2k-3)^2/(2k-1)$. These theories have been identified [@Staud] with the multicritical points of the one-matrix model [@Kaz], [@GM]. The $k=2$ theory corresponds to pure gravity, where it suffices to consider the discretizations of random surfaces with squares only. For $k=3~(c=-22/5)$ one needs squares and hexagons; for $k=4$ one needs squares, hexagons and octagons; etc. For $k\geq 3$ some polygons enter with negative weights, which is not surprising given that these models are not unitary. We find that, for odd $k$, the negative weights have so much effect that there is no sensible positive $\rho(L, D)$. For even $k$, however, there does exist a positive $\rho(L, D)$, whose calculation constitutes the main result of this paper. We find that it depends on the scaling variable $x=L/D^{1/(k-3/2)}$ and for small $x$ diverges as $x^{-k-1/2}$. Thus, the emission of microscopic s becomes more enhanced with increasing $k$. For large $k$, $c\to -\infty$ and the sum over surfaces is expected to be dominated by classical geometry. In this limit we indeed find that $\rho(L, D)$ exhibits a sharp peak that corresponds to the macroscopic classical geometry. Surprisingly, we also find that baby universes remain prevalent for all length scales less than a fixed constant, $s_c$, times the macroscopic (classical) length scale, and calculate the critical value $s_c$. As a further application of these ideas, we discuss the production of s by a very large parent universe of length $L_0$. We find that their distribution is governed by a very simple scaling law, $$\rho (L_0\to \infty, L, D)\sim L_0 D L^{-k-1/2} \ ,$$ which applies to both even and odd $k$. The structure of our paper is as follows. In we review and rederive the necessary matrix model results. In we extend the transfer matrix formalism of ref. [@Kawai1] to general discretizations, which are necessary to describe the theories with matter. In we calculate the transfer matrix in some limiting cases, and in we study the branching structure of space-time by deriving the distribution functions for s. In we conclude with a few remarks. The Disk Amplitudes {#MMloops} =================== For the calculation of the transfer matrix, the only result needed from the matrix models is the disk amplitude—that is, the partition function for surfaces with one boundary loop. It was shown in [@MSS] that in the $k^{\rm th}$ multicritical theory, the universal part of the disk amplitude is $$F_k(L, \tau) = {1 \over L} (\sqrt{\tau})^{k-1/2} K_{k-1/2} (\sqrt{\tau} L) \label{ExactFL}$$ where $\tau$ is the cosmological constant and $K_{k-1/2}$ is a modified Bessel function. We will need to keep track of certain non-singular parts of the disk amplitudes because they will be important in the calculation of the transfer matrix. Thus, we present our own calculation of the disk amplitudes. Throughout this paper, we will be working with surfaces without handles, so it is sufficient to use saddle point techniques in the matrix models. The partition function of the multicritical model (that is, the sum of the weights of all vacuum diagrams, including the disconnected ones) is $${\cal Z} = \int d\Phi \exp \left[ -\beta \, \Tr V(\Phi ) \right]$$ where $\Phi$ is a $N \times N$ matrix with $N \sim \beta \to \infty$, and $$V(\Phi) = \sum_{m \geq 1} {c_m \over 2 m} \Phi^{2 m}$$ is a potential energy function. (We restrict ourselves to even $V(\Phi)$ because they are more easily analyzed, but we believe that more general $V(\Phi)$ do not exhibit any more general behavior.) Let $u(\lambda)$ be the density of eigenvalues of $\Phi$, normalized so that $\int d\lambda \, u(\lambda) = 1$. As shown in [@IZ78], $u(\lambda)$ has support $[-\sqrt{z},\sqrt{z}\,]$ for some $z$, and $$V'(\lambda ) = 2 {N \over \beta} \pvint_{-\sqrt{z}}^{\sqrt{z}} d \mu {u(\mu) \over \lambda - \mu}\ .$$ For $\lambda \in {\bf C} - [-\sqrt{z},\sqrt{z}\,]$, define $$F(\lambda ) = 2 {N \over \beta} \int_{-\sqrt{z}}^{\sqrt{z}} d \mu {u(\mu) \over \lambda - \mu} \ .$$ $F(\lambda)$ is analytic except for a branch cut along $[-\sqrt{z},\sqrt{z}\,]$: for $\lambda \in [-\sqrt{z},\sqrt{z}\,]$ and $\epsilon > 0$ infinitesimally small, $F(\lambda \pm i \epsilon) = V'(\lambda) \mp 2 \pi i (N/\beta) u(\lambda)$. Also, $F(\lambda) = 2 (N/\beta) / \lambda + O(\lambda^{-2})$ for large $\lambda$. It is not hard to prove that these properties uniquely determine $F(\lambda)$ and hence $u(\lambda)$. It turns out that $F(\lambda)$ has the form $$F(\lambda) = V'(\lambda) - f(\lambda) \lambda \sqrt{1 - {z \over \lambda^2}}$$ where $f(\lambda)$ is a polynomial. Both $z$ and $f(\lambda)$ are determined by the large $\lambda$ expansion of $F(\lambda)$. Defining for convenience the functions $g(\lambda) = \lambda V'(\lambda)$ and $h(\lambda) = \lambda^2 f(\lambda)$, we find that $$g(\lambda) \left(1 - {z \over \lambda^2} \right)^{-1/2} = \left( \sum_{m \geq 1} c_m \lambda^{2 m} \right) \left( \sum_{m \geq 0} {(2 m)! \over m!^2 4^m} {z^m \over \lambda^{2 m} } \right) = h(\lambda) + 2 {N \over \beta} + O(\lambda^{-1}).$$ The $O(\lambda^0)$ terms determine $z$: $$\sum_{m \geq 1} c_m {(2 m)! \over m!^2 4^m} z^m = 2 {N \over \beta}. \label{zd1}$$ We can express the partition function ${\cal Z}$ as a power series in the couplings $c_m$ and the parameter $N/\beta$. If we consider the potential $V$ and hence the $c_m$ to be fixed, then ${\cal Z}$ has a radius of convergence in $N/\beta$; we construct the potential so that this radius of convergence is $1$. For $N/\beta > 1$—that is, for (renormalized) temperatures higher than a critical temperature—the series expansion of ${\cal Z}$ diverges. The behavior of ${\cal Z}$ near the critical point is what describes the quantum geometry of random surfaces, since at the critical point the behavior of ${\cal Z}$ is dominated by Feynman graphs of large order. It turns out that the singular behavior of ${\cal Z}$ at the critical point is determined by the dependence of $z$ on $N/\beta$ in . The multicritical model is constructed by finding $c_m$ such that takes the form $$1 - \left( 1 - {z \over z_c} \right)^k = {N \over \beta} \equiv 1 - \mu_0$$ for some $z_c$. We shall use $$c_m = \left\{ \begin{array}{ll} \displaystyle{(-1)^{m+1} {2 k!\, m! \over (2m)!\, (k-m)!\, k^m}} \qquad & {\rm for} \quad m \leq k \\ \displaystyle{0} \qquad & {\rm for} \quad m > k \end{array} \right.$$ which gives $z_c = 4 k$. Let $\epsilon$ be the lattice spacing of our random lattice. Then area is measured in units of $\epsilon^2$, which will be taken to zero as the average number of plaquettes in the discretizations of the random surface diverges, so as to yield finite area. The parameter $\mu_0$ corresponds to the lowest dimension operator in the multicritical theory, and it was at first believed that if one wrote $\mu_0 = (2 \epsilon)^k t_0$, then $t_0$ would be the cosmological constant. This is false for the $k>2$ models, which are non-unitary: $t_0$ has dimensions of $({\rm Length})^{-k}$, and is the coupling constant for the gravitationally dressed conformal field of the lowest dimension. The true cosmological constant is by definition conjugate to area, so it has dimensions of $({\rm Length})^{-2}$. We must consider a more general perturbation around criticality. Specifically, we perturb the potential by replacing $c_m \to c_m + {2 \, m!^2 \over (2 m)! k^m} \mu_m$ for all $m>0$. then reduces to $$\left( 1 - {z \over z_c} \right)^k = \sum_{m \geq 0} \mu_m \left( {z \over z_c} \right)^m.$$ Higher dimension scaling operators are introduced by choosing the $\mu_m$ so the that right hand side becomes $t_l \epsilon^{k-l} (1 - z/z_c)^l$ for some $l < k$, where $t_l$ is the coupling to the $l^{\rm th}$ scaling operator. As was shown in [@MSS], the couplings to the operators corresponding to conformal fields in the Liouville theory are analytical combinations of $t_l$ with a definite overall dimension. The cosmological constant $\tau$ corresponds to a perturbation with $t_l = \alpha_l 2^{k-l} \tau^{(k-l)/2}$ for $l-k$ even and $0 \leq l \leq k-2$. The $\alpha_l$ are real numbers chosen so that the singular part of the one-loop amplitude obeys a Bessel equation, which emerges from the Wheeler de Witt equation. $$\left( 1 - {z \over z_c} \right)^k = \sum_{l} \alpha_l (2 \epsilon \sqrt{\tau})^{k-l} \left( 1 - {z \over z_c} \right)^l\ , \label{CosmCnst}$$ and by requiring $\sum_l \alpha_l = 1$ (which is equivalent to fixing a normalization for $\tau$) we find $z = z_c (1 - 2 \epsilon \sqrt{\tau})$. For $k=2$ and $k=3$ there is only one $\alpha_l$, so $\tau \propto t_{k-2}$. For $k=4$, it is found that $\alpha_0 = -1/5$ and $\alpha_2 = 6/5$. We do not know a general way to determine the $\alpha_l$ except by straight calculation. The partition function for surfaces with one boundary is the continuum limit of the Green’s function of the field theory above. The $l$-point Green’s function is $$G_l = \langle \Tr \Phi^l \rangle = \int_{-\sqrt{z}}^{\sqrt{z}} d\lambda \, u(\lambda) \lambda^l.$$ It is convenient to introduce $G(y) = \sum_{l \geq 0} G_l\, y^l$, since it is easy to see that $$G(y) = {\beta \over N} {1 \over 2 y} F\left( {1/ y} \right) = {\beta \over N} {1 \over 2} \left( g\left( {1/ y} \right) - h\left( {1/ y} \right) \sqrt{1 - z y^2} \right).$$ Since at $z = z_c$ the radius of convergence of $G(y)$ is $y_c = 1/\sqrt{z_c}$, the continuum limit is taken by setting $y = y_c \exp (-\epsilon \zeta)$, where now $\zeta$ is conjugate to the length of the boundary. The continuum limit ($\epsilon \to 0$) of $G(y)$ was calculated using Mathematica, with the following results: $$\begin{array}{ll} k=2: \ & G(\zeta) = {4 \over 3} - {8 \over 3} \zeta \epsilon + {16 \sqrt{2} \over 3} \epsilon^{3/2} f_2(\zeta,\tau) + O(\epsilon^2) \\ & \qquad\quad {\rm where} \quad f_2(\zeta,\tau) = 2 \left( \zeta - {1 \over 2} \sqrt{\tau} \right) \sqrt{\zeta + \sqrt{\tau}} \\ k=3: \ & G(\zeta) = {6 \over 5} - {4 \over 5} \zeta \epsilon + {4 (-4 \tau + 7 \zeta^2) \over 5} \epsilon^2 + {96 \sqrt{2} \over 5} \epsilon^{5/2} f_3(\zeta,\tau) + O(\epsilon^3) \\ & \qquad\quad {\rm where} \quad f_3(\zeta,\tau) = {2\over 3}\left( -\zeta^2 + {1 \over 2} \zeta \sqrt{\tau} + {1 \over 4} \tau \right) \sqrt{\zeta + \sqrt{\tau}} \\ k=4: \ & G(\zeta) = {8 \over 7} - {16 \over 35} \zeta \epsilon + {16 (-2 \tau + 3 \zeta^2) \over 35} \epsilon^2 + {32 \zeta (174 \tau - 185 \zeta^2) \over 525} \epsilon^3 + {512 \sqrt{2} \over 7} \epsilon^{7/2} f_4(\zeta,\tau) + O(\epsilon^4) \\ & \qquad\quad {\rm where} \quad f_4(\zeta,\tau) = {2\over 5}\left( \zeta^3 - {1 \over 2} \zeta^2 \sqrt{\tau} - {1 \over 2} \zeta \tau + {1 \over 8} \tau^{3/2} \right) \sqrt{\zeta + \sqrt{\tau}} \end{array} \label{ContG1}$$ The coefficient $f_k(\zeta,\tau)$ of the leading nonanalytic term in $\epsilon$ is the universal part of the disk amplitude. $f_k(\zeta,\tau)$ is, up to a numerical factor, the Laplace transform of the disk amplitude $F_k(L,\tau)$ in . The lower order analytic terms in $\zeta$ and $\tau$ correspond to zero length and zero area terms, and most of them can be dropped, for the following reason. We are free to adjust $\zeta$ by an analytic function of $\zeta$ and $\tau$, which, in order to avoid trivial additive and multiplicative rescalings of $\zeta$, and to preserve the dimension of $\zeta$, should have the form $$\zeta \to \zeta + r_1 \epsilon \tau + r_2 \epsilon \zeta^2 + r_3 \epsilon^2 \tau \zeta + r_4 \epsilon^2 \zeta^3 + \ldots \label{Redef}$$ where the $r_i$ are $c$-numbers. Such a redefinition preserves the universal term $f_k(\zeta,\tau)$ and corresponds simply to a different way of treating zero area and zero length terms, as pointed out in [@MSS]. By an appropriate choice of the $r_i$ in , we can absorb all but the first two analytic terms of the expressions in into $\zeta$. We then find the simple form $$G(\zeta) = {2k \over 2k-1} \left( 1 - {\zeta \epsilon\over \sigma} \right) + \alpha \epsilon^{k-1/2} f_k(\zeta,\tau) + O(\epsilon^k) \label{ContG2}$$ where $$\sigma = k - 3/2$$ and $\alpha$ is a numerical factor chosen for each $k$ so that the leading term of a small $\tau$ expansion of $f_k(\zeta,\tau)$ is $(-1)^k \zeta^{\sigma +1}/\sigma$. For later convenience, we mention one more mathematical point: since $z/z_c = 1 + O(\epsilon)$, we can alter the perturbation to read $$\left( 1 - {z \over z_c} \right)^k = \sum_{l} \alpha_l (2 \epsilon \sqrt{\tau})^{k-l} \left( 1 - {z \over z_c} \right)^l \left( z \over z_c \right)^n$$ where $n$ is a fixed integer, and the equation $z = z_c (1 - 2 \epsilon \sqrt{\tau})$ will receive corrections that are analytic in $\epsilon$ and of order $\epsilon^2$ and higher. The non-universal terms in will change, but the leading analytic and leading nonanalytic terms will be unaffected, so by making an appropriate redefinition of the form , we still arrive at . If we take $n=2$, then $\mu_0 = 0$ and $\mu_1 = 0$. This is desirable because it means that the weights of planar Feynman diagrams do not depend on the number of edges. The transfer matrix {#TransferM} =================== The focus of [@Kawai1] is the analysis of the evolution of a loop through some fixed geodesic distance on the surface, where on a discretized surface geodesic distance is defined as the minimal number of plaquettes one must traverse to get from one point to another. The goal is to calculate the partition function of a tube with one entrance loop and one exit loop, such that each point on the exit loop is a fixed geodesic distance $D$ from the entrance loop. On a discretized surface, one starts with an entrance loop $\gamma$ (see Fig. 1), defines a “forward” direction for geodesic distance (inward in Fig. 1), and thinks of advancing the loop along the lattice one step at a time. To accomplish this “one-step deformation,” as it was called in [@Kawai1], one first removes any double links that may exist on $\gamma$ and then moves each remaining link across the plaquette it borders in the forward direction. Clearly, this process will sometimes split $\gamma$ into several loops. To phrase it another way, once we have removed all double links from $\gamma$, we color all the plaquettes that the remaining links border in the forward direction. Each side of a colored plaquette that is not part of the entrance loop $\gamma$ becomes a link in one of the exit loops. Since we are interested in having just one exit loop, we designate one of the loops evolved from $\gamma$ as [the]{} exit loop, and close off the other loops with disks generated by the disk amplitude of the matrix model. A series of one-step deformations gives an exit loop which is a fixed geodesic distance from the entrance loop. We now want to ask the following combinatorical question: given that the entrance loop $\gamma$ has $l$ links and the exit loop $\gamma'$ has $l'$ links, how many ways are there to construct a discretization between $\gamma$ and $\gamma'$, using a given number of squares, hexagons, octagons, etc., such that $\gamma'$ is the exit loop evolved from $\gamma$ in the course of a one-step deformation, as defined above? Or, in terms of the Feynman diagram which is the dual graph of the discretization, how many planar diagrams can be drawn on the surface of a cylinder with $l$ external legs pointing “down” and $l'$ external legs pointing “up”, with specified numbers of vertices of order four, six, eight, etc., and such that every “up” leg is connected to the same vertex as some “down” leg (but not necessarily vice versa)? To answer this question, we assign a multiplicative weight $g_m$ to each $2m$-gon (equivalently, to each vertex of order $2m$ on the dual lattice), assign as an overall weight to each of the discretizations between $\gamma$ and $\gamma'$ the product of the weights of the plaquettes used to build it, and let $N_{l,l'}$ be the sum of the weights of all such discretizations. $N_{l,l'}$ is the generating function that answers the above combinatorical question, but of course its more interesting property is that it is the partition function of tubes, or rather ribbons, of geodesic width one (in lattice units) and with entrance and exit loops of $l$ and $l'$ links, respectively. We propose to evaluate $$N(y,y') = \sum_{l,l' \geq 1} N_{l,l'} \, y^l (y')^{l'}$$ using combinatorics and the disk amplitude from the matrix models. For calculational convenience we will designate the entrance loop as unmarked and the exit loop as marked. On the dual lattice, this corresponds to considering the $l$ external legs of the planar diagram that point “down” to be distinguishable only up to cyclic permutations and the $l'$ external legs that point “up” to be completely distinguishable. A different convention of marking would change $N_{l,l'}$ only by a factor of $l$ or $l'$. The Green’s function $G(y)$ of the matrix field theory refers to diagrams in which the external legs are distinguishable—that is, $G(y)$ is the amplitude of a disk with a marked boundary. The weight $g_m$ assigned to a $2m$-gon by the Feynman rules for the matrix field theory is $$g_m = -c_m - {2 \, m!^2 \over (2 m)! k^m} \mu_m \ .$$ Let us first consider the case where only squares are used in the discretizations. All the discretizations contributing to $N(y,y')$ may be built up as follows (here we are paralleling closely the work of [@Kawai1], only using squares instead of triangles). The marked point on the exit loop must border one of the shapes in list a) of Fig. 2: the curved loops on the last three shapes denote the insertion of a disk. As we follow the exit loop around its length, we can encounter any of the shapes in list b) in any order: the entrance and exit loops are connected from the corner of one shape to the next, and in the end the last shape is connected back to the original member of list a), closing the loop. It is perhaps not transparent why the weighting of each shape is what it is, so let us work through the example shown in Fig. 3. The loop above the square indicates the insertion of a disk amplitude, which must be marked because something must pick out the point where the disk boundary meets the square. The disk discretizations with less than two sides on the boundary are omitted because there have to be two boundary sides that match with the two free sides of the square. Let us think now in terms of the dual lattice. Given any Feynman graph contributing to the disk amplitude, we form a graph of the desired type by attaching to an adjacent pair of its external legs another vertex. The weight of the original disk amplitude graph must be multiplied by $g_2$ to get the weight of the new graph, because we have added one vertex and tied up two external legs while adding two new external legs. Summing over all allowable disk amplitude graphs (those with at least two external legs) then gives a total weight of $g_2 (G(y) - G_0 - G_1 y)$, as claimed. Incidentally, the first, “undrawable” term $g_2 y^2$ in Fig. 3 corresponds to the two free sides of the square being identified. It might be helpful for the reader to identify which disk amplitude diagrams correspond to each term in Fig. 3. Now it is not hard to write down the weight that would be assigned to a $2m$-gon bordered on $n$ sides by the exit loop. Starting with a disk amplitude graph with at least $2m-n-2$ external legs, we add to it one $2m$-vertex, tying up $2m-n-2$ of its external legs and adding $n+2$ new external legs, $2$ of which are entrance legs and $n$ of which are exit legs. Thus the total weight is $$y'^n y^{n+4-2m} g_m \left( G(y) - \sum_{l=0}^{2m-n-3} G_l y^l \right).$$ Note that because of the planarity of the surface, the sides of the $2m$-gon bordered by the exit loop must be contiguous. When $2m$-gons are allowed in the discretization, we include all the $2m$-gon shapes in list b), and also in list a) with a multiplicity determined by the number of exit links on the shape. The combinatorical problem is now quite simple: we construct an arbitrary discretization from one member of list a) and some sequence of members of list b). The total amplitude when only squares are used is $$\begin{aligned} N(y,y') &=& \left( 3 y'^3 y g_2 + 2 y'^2 y^2 g_2 G(y) + y' y g_2 (G(y) - G_0) \right) \times \nonumber \\ & & \sum_{n=0}^{\infty} [y'^3 y g_2 + y'^2 y^2 g_2 G(y) + y' y g_2 (G(y) - G_0) + g_2 (G(y) - G_0 - G_1 y) + y^2 G(y) ]^n \nonumber\end{aligned}$$ $$= {3 y'^3 y g_2 + 2 y'^2 y^2 g_2 G(y) + y' y g_2 (G(y) - G_0) \over 1 - y'^3 y g_2 - y'^2 y^2 g_2 G(y) - y' y g_2 (G(y) - G_0) - g_2 (G(y) - G_0 - G_1 y) - y^2 G(y) }\ . \label{N1k4}$$ Expanding out the first expression gives terms corresponding to individual discretizations. Let us define $S$ to be $1$ minus the sum of weights in list b): $$S = 1 - y'^3 y g_2 - y'^2 y^2 g_2 G(y) - y' y g_2 (G(y) - G_0) - g_2 (G(y) - G_0 - G_1 y) - y^2 G(y).$$ Recalling that each shape from list b) enters into list a) with a multiplicity equal to the number of exit links on the shape, we find that $$N(y,y') = -{\partial \log S \over \partial \log y'}\ . \label{NS}$$ In the general case where arbitrary $2m$-gons are allowed, list a) and list b) are appropriately expanded, and $$\begin{aligned} S &=& 1 - y^2 G(y) - \sum_{m \geq 1} \left( y'^{2m-1} y g_m + \sum_{n=0}^{2m-1} y'^n y^{n+4-2m} g_m \left( G(y) - \sum_{l=0}^{2m-n-3} G_l y^l \right) \right) \nonumber \\ &=& 1 - y y' + {y \over y'} g(y') - 2 y^2 G(y) - y^4 \sum_{m \geq 1} {g_m \over y^{2 m}} \sum_{n=0}^{2m-2} (y' y)^n \left( G(y) - \sum_{l=0}^{2m-n-3} G_l y^l \right), \label{GenS}\end{aligned}$$ and still holds. The continuum limit of $N(y,y')$ is taken by expanding about the radius of convergence of each of the variables. The convergence of $S$ is determined by the convergence of $G(y)$, which we explained in Section \[MMloops\]: $y_c = 1/\sqrt{z_c}$, and we set $y = y_c \exp (-\epsilon \zeta)$. For fixed $y \leq y_c$, $S$ is entire in $y'$, but $N(y,y')$ is analytic in $y'$ only up to the magnitude of the zero of $S$ nearest $y' = 0$. It turns out that for $\zeta = 0$ and $\tau = 0$, this zero is at $y' = y_c^{-1}$. Hence we set $y' = y_c^{-1} \exp (-\epsilon \zeta')$. An analytic redefinition of $\zeta'$ of the form $$\zeta' \to \zeta' + r_1 \epsilon \tau + r_2 \epsilon \zeta^2 + r_3 \epsilon \zeta \zeta' + r_4 \epsilon \zeta'^2 + \ldots$$ is allowed, for the same reasons as for the redefinition . Thus when we expand $N(y,y')$ in $\epsilon$, we need only retain the leading analytic and leading nonanalytic terms in $\zeta'$. Writing $N(\zeta,\zeta')$ in place of $N\left( y_c \exp (-\epsilon \zeta), y_c^{-1}\exp (-\epsilon \zeta') \right)$, we find remarkably simple results: $$N(\zeta,\zeta') = {1 \over \epsilon} \left({1\over \zeta + \zeta' - \alpha' \epsilon^{\sigma} f(\zeta,\tau)} + O(\epsilon^{k-1}) \right) \label{ContN}$$ where $\alpha'$ is another numerical factor. Let $N_{l,l'}(d)$ be the lattice partition function of tubes of geodesic length $d$ with an unmarked entrance loop of $l$ links and a marked exit loop of $l'$ links. The great insight of [@Kawai1] is that this object has a simple composition law: $$N_{l,l'} (d_1 + d_2) = \sum_{l'' = 1}^{\infty} N_{l,l''} (d_1) N_{l'',l'} (d_2)\ . \label{Comp1}$$ In rough terms, we can cut a tube in two at some geodesic time and find its amplitude by summing the products of the amplitudes of the pieces over all possible lengths of the intermediate loop (see Fig. 4). The motivation for the convention of taking the entrance loop to be unmarked and the exit loop to be marked lies in the fact that there are $l''$ ways to glue the exit loop of one tube to the entrance loop of another—$l''$ being the number of links on each loop—but we want to avoid factors of $l''$ in the composition law. Suppose we mark the entrance loop of the rightmost tube in Fig. 4; such tubes would have amplitude $l'' N_{l'',l'} (d_2)$. We now can glue the two tubes together in such a way that the two marks are at any of $l''$ positions relative to each other, so the amplitude for the resulting surface would be $l'' \left( N_{l,l''} (d_1) l'' N_{l'',l} (d_2) \right)$. But that surface would still have the two marks on the intermediate loop, and we must delete them to get a surface of the type shown on the left side of Fig. 4. Thus we divide our last expression by $l''^2$ to get $N_{l,l''} (d_1) N_{l'',l'} (d_2)$ and sum over $l''$ to get . Since is just a matrix product, it is clear that $N_{l,l'} (d)$ is the time evolution kernel of some Hamiltonian. We will find the continuum limit of this Hamiltonian and then calculate $N_{l,l'} (d)$ by solving the corresponding Schrödinger equation. In terms of $N(y,y',d) = \sum_{l,l' \geq 1} N_{l,l'} (d) \, y^l (y')^{l'}$, takes the form $$N(y,y',d_1 + d_2) = {1 \over 2 \pi i } \oint {dx \over x} N(y,x,d_1) N(1/x,y',d_2)$$ where the integral is taken along a contour around the origin. Making the change of variables $x = y_c^{-1} \exp (-\epsilon \xi)$, we find $$N(\zeta,\zeta',d_1 + d_2) = {\epsilon \over 2 \pi i} \int_{-i \pi / \epsilon}^{i \pi / \epsilon} d\xi N(\zeta,\xi,d_1) N(-\xi,\zeta',d_2) \to {\epsilon \over 2 \pi i} \int_{-i \infty}^{i \infty} d\xi N(\zeta,\xi,d_1) N(-\xi,\zeta',d_2) \label{Comp2}$$ for small $\epsilon$. In particular, to the first nontrivial order in $\epsilon$, $$\begin{aligned} N(\zeta,\zeta',d+1) &=& {\epsilon \over 2 \pi i} \int_{-i \infty}^{i \infty} d\xi N(\zeta,\xi) N(-\xi,\zeta',d) \\ &=& {1 \over 2 \pi i} \int_{-i \infty}^{i \infty} d\xi {1 \over \zeta + \xi - \alpha' \epsilon^{\sigma} f(\zeta,\tau)} N(-\xi,\zeta,d) \\ &=& N\left( \zeta - \alpha' \epsilon^{\sigma} f(\zeta,\tau), \zeta',d \right) = N(\zeta,\zeta',d) - \alpha' \epsilon^{\sigma} f(\zeta,\tau) {\partial \over \partial \zeta} N(\zeta,\zeta',d)\ ,\end{aligned}$$ where we have closed the contour to the left because the function $N(-\xi,\zeta,d)$ has a branch cut for positive real $\xi > \sqrt{\tau}$. Setting $D = \alpha' \epsilon^{\sigma} d$ and taking $\epsilon \to 0$, we find that $N(\zeta,\zeta',D)$ is a solution to the equation $${\partial \over \partial D} \psi(\zeta,D) = -f(\zeta,\tau) {\partial \over \partial \zeta} \psi(\zeta,D) \label{DiffEQ}$$ with the initial condition read off from , $N(\zeta,\zeta',0) = 1/(\zeta + \zeta')$.[^1] In [@Kawai1], this equation was derived for the case $k=2$, but using only triangles in the discretizations, rather than squares; thus for this case we have a universality check. What is interesting is that the same equation holds for higher $k$, with the disk amplitude $f(\zeta,\tau)$ adjusted appropriately, and with a new scaling behavior for macroscopic geodesic distance $D$. The definition $D = \alpha' \epsilon^{\sigma} d$ is very interesting, and in a way unexpected, because it means that $D$ has the dimension of $L^{\sigma}$ instead of just $L$ (recall $\sigma = k - 3/2$). In what follows it is convenient to introduce $N(\zeta,L',D)$, the inverse Laplace transform of $N(\zeta,\zeta',D)$ with respect to $\zeta'$. Since involves only $\zeta$ and $D$, $N(\zeta,L',D)$ is also a solution of this differential equation, only with different initial conditions, namely $N(\zeta,L',0) = \exp(-\zeta L')$. Our next step will be to find an approximate method for solving the differential equation and to use the solutions, as was done in [@Kawai1], to determine universal functions describing the way random surfaces branch. These functions directly depend on $N(L,L',D)$, the inverse Laplace transform of $N(\zeta,L',D)$ with respect to $\zeta$. $N(L,L',D)$ is the continuum amplitude for tubes of length $D$ with boundaries of length $L$ and $L'$, and it satisfies the initial condition $N(L,L',0) = \delta(L-L')$. Calculating the tube amplitude {#TubeAmp} ============================== A solution to may be found by the method of characteristic curves: if the function $\zeta_0 (\zeta,D)$ solves the ODE $${d \over d D} \zeta_0(\zeta,D) = -f(\zeta_0(\zeta,D),\tau)\ , \qquad \zeta_0(\zeta,0) = \zeta \label{CCEQ1}$$ then $\psi(\zeta,D) = \psi(\zeta_0(\zeta,D),0)$ is a solution to . Geometrically speaking, $\psi(\zeta,D)$ is constant along a family of (non-intersecting) characteristic curves in $\zeta$-$D$ space; $\zeta_0(\zeta,D)$ is the $\zeta$-coordinate of the point where the characteristic curve passing through $(\zeta,D)$ meets the line $D = 0$. From it follows that $\zeta_0(\zeta,D)$ is implicitly determined by $$\int_{\zeta_0}^{\zeta} d\xi {1\over f(\xi,\tau)} = D\ . \label{IntEQ}$$ Now $N(\zeta, L',D)$, the solution of with initial conditions $N(\zeta,L',0) = \exp (-\zeta L')$, is simply given by $$N(\zeta,L',D) = \exp \left( -\zeta_0 (\zeta,D) L' \right)\ . \label{DSol}$$ The inverse Laplace transform of this function gives the desired tube amplitude $N(L, L', D)$. Thus, the problem is reduced to the explicit determination of $\zeta_0(\zeta,D)$. For $k=2$, $\zeta_0(\zeta,D)$ can be expressed in terms of elementary functions. Indeed, after evaluating the integral in with $f(\zeta,\tau) = \left(2 \zeta - \sqrt{\tau} \right) \sqrt{\zeta + \sqrt{\tau}} $, and some algebra, we arrive at $$\zeta_0 = -\sqrt\tau +{3\over 2}\sqrt\tau \coth^2 \left (\sqrt{3\over 2}\tau^{1/4} D-{1\over 2}\log {\sqrt{\zeta+\sqrt\tau}-\sqrt{3\over 2}\tau^{1/4}\over \sqrt{\zeta+\sqrt\tau}+\sqrt{3\over 2}\tau^{1/4} }\right ) \label{Exact}$$ Unfortunately, it seems impossible to find the exact inverse Laplace transform of . Instead, we will look for the behavior of $N(L,L',D)$ in the limit of a point-like entrance loop ($L\to 0$), as well as in the limit of a very large entrance loop ($L\to\infty $). In the first case, it is sufficient to study the dominant behavior of for large $\zeta$, and in the second case, for small $\zeta$. Let us first consider the limit of a point-like entrance loop. Substituting into and expanding for large $\zeta$, we find $$N(\zeta,L',D) = e^{L'\sqrt\tau} e^{-{3\over 2}L'\sqrt\tau \coth^2 \left (\sqrt{3\over 2}\tau^{1/4} D\right )}+ {1\over\sqrt \zeta} e^{L'\sqrt\tau} {\partial\over\partial D} e^{-{3\over 2}L'\sqrt\tau \coth^2\left (\sqrt{3\over 2}\tau^{1/4} D\right)}+ O(1/\zeta) \label{Largez}$$ Performing the inverse Laplace transform, we find $$N(L\to 0, L',D) = {1\over\sqrt {\pi L}} e^{L'\sqrt\tau} {\partial\over\partial D} e^{-{3\over 2}L'\sqrt\tau \coth^2\left (\sqrt{3\over 2}\tau^{1/4} D\right)}+ O(L^0) \label{SmallL}$$ This formula is valid for $L\ll D^2, ~\tau^{-1/2}$. Nothing is assumed, however, about the relative magnitudes of $D^2$ and $\tau^{-1/2}$. In [@Kawai1] the calculation was performed in the regime $L \ll D^2 \ll \tau^{-1/2}$ so that the total area of the surface could be sent to infinity while $D$ and $L'$ were kept finite. Our generalizes the result of [@Kawai1] and reduces to it when expanded for small $\tau$. For $k>3$, however, we cannot find an explicit form of $\zeta_0(\zeta,D)$ analogous to . For that reason we will work in the limit $1/\zeta \ll D^{1/\sigma} \ll \tau^{-1/2}$ (recall $\sigma = k - 3/2$), and expand $\zeta_0$ in powers of $1/\zeta$ and $\tau$. Since $f_k(\xi,\tau) = (-1)^k~ {\xi^{\sigma+1}\over \sigma} (1 + O(\tau / \xi^2))$, as our first approximation we have from , $$\zeta_0 (\zeta,D) = \left( (-1)^k D + \zeta^{-\sigma} \right)^{-1/\sigma} + O(\tau) \ . \label{SickApp}$$ From this formula we can see a sickness in the case of odd $k$: as $D \to \zeta^{-\sigma}$ from below, $\zeta_0 (\zeta,D) \to \infty$. What is happening is that the solution to with $\tau = 0$ is running off to $\infty$ in finite time. becomes complex for $D > \zeta^{-\sigma}$, but the real solution to just stays at $\infty$. Letting $\tau$ become finite does not help the situation except when the initial $\zeta$ is less than the rightmost zero of $f(\zeta,\tau)$, which is at $\zeta\sim\sqrt{\tau}$. Since $N(\zeta, L', D)$ vanishes for large enough real $\zeta$, the inverse Laplace transform $N(L,L',D)$ cannot be positive definite for small $L$. This seems to lead to meaningless results, and we will not consider the limit of a point-like entrance loop in the odd $k$ models. It may be that if we worked with fixed $L$ from the beginning, instead of with Laplace-tranformed equations like , we could avoid the sickness we have observed. The even $k$ models avoid this sickness, as is clear from Fig. 5 where we show the qualitative behavior of $\zeta_0 (\zeta,D)$ for $k=3$ and $k=4$ and finite $\tau$. To determine the baby universe distribution function in the case of even $k$, we need to continue the expansion of in powers of $\tau$ and find the coefficient of the leading fractional power. First we expand $${1\over f_k(\xi, \tau)}={k-3/2\over\xi^{k-1/2}} +\sum_{n=1}^{k-1} a_n\tau^n {k+2n-3/2\over\xi^{k+2n-1/2}} +b\tau^{k-1/2}{3k-5/2\over\xi^{3k-3/2}} +O(\tau^{k})\ , \label{App1}$$ where $a_n$ and $b$ are numerical coefficients which can be found from . Substituting into and integrating term-by-term, we obtain the following relation: $$\begin{aligned} \zeta_0 (\zeta,D)& = \left( D + \zeta^{-\sigma} \right)^{-1/\sigma} +\tau{a_1\over\sigma} \left( D +\zeta^{-\sigma} \right)^{-1-1/\sigma}\left (\zeta_0^{-\sigma-2} -\zeta^{-\sigma-2}\right )+\ldots+O(\tau^{k-1}) \cr & +\tau^{k-1/2}~ {b\over\sigma} \left( D +\zeta^{-\sigma} \right)^{-1-1/\sigma} \left (\zeta_0^{-3\sigma-2}-\zeta^{-3\sigma-2}\right )+O(\tau^k) \ . \label{ItApp}\end{aligned}$$ Now the expansion of $\zeta_0$ in powers of $\tau$ can be found iteratively. The coefficient of the leading fractional power, $\tau^{k-1/2}$, is actually obtained after one iteration. Thus, the desired expansion has the form $$\begin{aligned} \zeta_0 (\zeta,D) &= \left( D + \zeta^{-\sigma} \right)^{-1/\sigma} +\tau{a_1\over\sigma} \biggl [\left( D +\zeta^{-\sigma} \right)^{1/\sigma}-\zeta^{-\sigma-2} \left( D +\zeta^{-\sigma} \right)^{-1-1/\sigma}\biggr ] +\ldots+O(\tau^{k-1}) \cr & +\tau^{k-1/2}~ {b\over\sigma} \biggl [\left( D +\zeta^{-\sigma} \right)^{2+1/\sigma}-\zeta^{-3\sigma-2} \left( D +\zeta^{-\sigma} \right)^{-1-1/\sigma}\biggr ]+O(\tau^k) \ . \label{CompExp}\end{aligned}$$ Substituting this into , we have $$\begin{aligned} &N(\zeta,L',D) = e^{- L'\left( D + \zeta^{-\sigma} \right)^{-1/\sigma}}\biggl\{ 1 -\tau L'{a_1\over\sigma} \biggl [\left( D +\zeta^{-\sigma} \right)^{1/\sigma}-\zeta^{-\sigma-2} \left( D +\zeta^{-\sigma} \right)^{-1-1/\sigma}\biggr ] +\ldots\cr & +O(\tau^{k-1}) -\tau^{k-1/2} L'{b\over\sigma} \biggl [\left( D +\zeta^{-\sigma} \right)^{2+1/\sigma}-\zeta^{-3\sigma-2} \left( D +\zeta^{-\sigma} \right)^{-1-1/\sigma}\biggr ]+O(\tau^k) \biggr\} \ .\label{CompTrans}\end{aligned}$$ Now we expand this for large $\zeta$ and inverse Laplace transform term-by-term, discarding the $O(\zeta^0)$ pieces, which give zero length terms. We find $$\begin{aligned} N(L\to 0,L',D) &=& {L^{\sigma - 1} \over \Gamma (\sigma)} e^{-x} \biggl( {L' \over \sigma D^{1+1/\sigma}} + O(\tau) + \ldots+O(\tau^{k-1}) + \nonumber \\ & & \gamma_1 D L' ((2k-2) D^{1/\sigma} + L') \tau^{k-1/2} + O(\tau^k)\biggr) + O(L^{2 \sigma-1}) \label{NLL}\end{aligned}$$ where $x = L'/D^{1/\sigma}$ and $\gamma_1=-b/\sigma^2$. The branching structure of random surfaces {#BStructure} ========================================== We now wish to extract from $N(L,L',D)$ a universal function which describes the branching structure of a large surface of planar topology and one boundary loop of length $L_0$. We will find that such surfaces have lots of little protuberances—hair, if you will. Let $R(D)$ be the part of the surface which is at a geodesic distance $D$ or less from the boundary loop. Let $\rho(L_0, L,D)$ be the distribution of boundary loop lengths for $R(D)$: on average, $R(D)$ has $\rho(L_0, L,D) dL$ boundary loops with length in the interval $(L,L+dL)$. This quantity was introduced and calculated for $k=2$ and $L_0\to 0$ in [@Kawai1]. The strategy for calculating it is as follows. The partition function for all surfaces is the disk amplitude $F(L_0)/L_0$. Let us think for a moment of the discrete version $\rho(l_0, l,d)$ of $\rho(L_0, L,D)$. This function is the statistical average over all disks with boundary length $l_0$ of the number of loops with $l$ links at geodesic distance $d$ from the boundary. According to [@Kawai1], in the limit of large disk area $$\rho(l_0,l,d) = \lim_{\tau\to 0} {\displaystyle{\left({\partial \over \partial \tau} \right)^n \left[ N_{l_0,l}(d) G_l/l \right]} \over \displaystyle{\left({\partial \over \partial \tau} \right)^n \left[ G_{l_0}/l_0 \right]} }\ . \label{DiscRho}$$ $N_{l_0,l}(d) G_l/l$ generates only surfaces with at least one boundary loop of the desired sort. Moreover, if there are $p$ such boundary loops on a particular surface, $N_{l_0,l}(d) G_l/l$ will generate that surface $p$ times: each time, the tube generated by $N_{l_0,l}(d)$ will have a different one of its boundary loops left open for $G_l/l$ to plug. Hence $N_{l_0,l}(d) G_l/l$ is the sum over surfaces generated by $G_{l_0}/l_0$ of the weight of each surface times the number of boundary loops at distance $d$ and of length $l$ on that surface. To understand the presence of the puncture operators $\partial / \partial \tau$, think for a moment of surfaces with fixed area instead of fixed $\tau$. To get to the fixed area representation, we would separately carry out inverse Laplace transforms on the numerator and denominator. The large area behavior of the fixed area quantities would be controlled by the leading singular term in the small $\tau$ expansion of the fixed $\tau$ quantities. Inserting enough puncture operators $\partial / \partial \tau$ in the numerator and denominator of to eliminate the leading analytic terms in $\tau$ thus provides a convenient way to isolate the terms that survive in the large area limit. In order to obtain a generalization of to disks of finite area, we would have to replace the numerator and denominator by their inverse Laplace transforms. The continuum limit of is obviously $$\rho(L_0, L,D) = \lim_{\tau\to 0} {\displaystyle{\left({\partial \over \partial \tau} \right)^n \left[ N(L_0,L,D) F(L)/L \right]} \over \displaystyle{\left({\partial \over \partial \tau} \right)^n \left[ F(L_0)/L_0 \right]} }\ . \label{ContRho}$$ Let us first consider the limit $L_0\to 0$ where the entrance loop is shrunk to a point. We expand $$F_k(L, \tau)/L = {1 \over L^{k+3/2}} + O(\tau) + \ldots + O(\tau^{k-1}) + \gamma_2 L^{k - 5/2} \tau^{k-1/2} + O(\tau^k), \label{FL}$$ where $\gamma_2$ is another numerical factor and we have adjusted $F_k(L, \tau)/L$ by an overall multiplicative factor for convenience. In this expansion, as in , only the $O(\tau^0)$ term and the leading nonanalytic term are relevant to $\rho(L_0, L, D)$. This is because both the numerator $N(L_0,L,D) F(L)/L$ and the denominator $F(L_0)/L_0$ of have $k$ leading terms analytic in $\tau$, followed by an $O(\tau^{k-1/2})$ term. The leading terms are deleted by making $k$ punctures. If one made fewer punctures, a positive overall power of $L_0$ would make $\rho(L_0, L,D)$ identically $0$. If one made more punctures, the leading nonanalytic term would still be dominant as $\tau \to 0$, so the final result would be unchanged. In the $\tau \to 0$, $L_0\to 0$ limit, the only finite physical quantities in $\rho(L_0,L,D)$ are $L$ and $D$, so the only possible dimensionless scaling parameter is $x = L / D^{1/\sigma}$. Thus $\rho(L_0\to 0, L, D)$ is a function only of $x$, up to a dimensionful overall factor: $$\rho(L_0 \to 0, L, D) = {1 \over D^{1/\sigma}} \left[ {\gamma_1 \over\gamma_2 \Gamma(\sigma)} x^{-\sigma-2} (2\sigma+1+ x) + {x^\sigma \over \Gamma(\sigma + 1)} \right] e^{-x}\ , \label{RhoSol}$$ where $\sigma=k-{3\over 2}$, and the numbers $\gamma_1$ and $\gamma_2$ can be calculated explicitly: $$\gamma_1 = {2^{3 - 2k} \over (6k - 5) (2k - 3)}\ ,\qquad\qquad \quad \gamma_2 = {1 \over (2k-1)!! (2k-3)!!}\ .$$ The terms in square brackets in separate beautifully as $k \to \infty$: the first term gives a non-integrable divergence at $x = 0$, and the second term gives a Poisson distribution, normalized to one and peaked at $x = \sigma$. The first term shows that the random surface has huge numbers of protuberances whose circumferential length is small compared to their geodesic distance from a given point; this is what we mean by the surface being hairy. The profusion of “microscopic” boundary loops suggests that as a one-dimensional universe propagates through geodesic time, it emits a divergent number of baby universes. For large $x$, the second term in square brackets is dominant, and it shows that there is exactly one “macroscopic” boundary loop to our region, and for large $k$ its length is sharply peaked about $\sigma D^{1/\sigma}$. Thinking again of a one-dimensional universe propagating through geodesic time, we would interpret the one macroscopic boundary loop as the parent universe, which survives the emission of its numerous baby universes. To determine at what $x$ the first term becomes significant, we note that $${1\over \sigma} \log \left[ {\gamma_1\over \gamma_2\Gamma(\sigma)} x^{-\sigma-2} (2\sigma + 1 + x) e^{-x} \right] \to -(1+s+\log s)$$ as $\sigma = k-3/2 \to \infty$ with $s \equiv x/\sigma$. This is positive for $s<s_c$ and negative for $s>s_c$, where $s_c$ satisfies $$1+s_c+\log s_c=0\ ,$$ the solution to which is $s_c \approx 0.2784645428$. Let us define $\rho(s)$ by $$\rho(s) \, ds = \rho(L_0 \to 0,L,D) \, dL = \left[ {\gamma_1\over \gamma_2\Gamma(\sigma)} x^{-\sigma-2} (2\sigma+1 + x) + {x^\sigma \over \Gamma(\sigma + 1)} \right] e^{-x}\, dx\ {}.$$ It follows from the above discussion that, as $k \to \infty$, $\rho(s)$ converges in the weak sense to a distribution which is $+\infty$ for $s < s_c$ and $\delta (s - 1)$ for $s > s_c$. In Fig. 6 we show plots of $\rho(s)$ for $k = 2$, where there is no separation of the “microscopic” and “macroscopic” terms; for $k=6$, where the separation is significant; and for $k=100$, where the convergence to the limiting case is very clear. Large $k$ corresponds to the central charge going to $-\infty$, which is held to be the semiclassical limit where quantum fluctuations vanish and the surface is smooth, with constant scalar curvature. In the thermodynamic (infinite area) limit, the surface would be a plane, and we would have $\rho(L_0 \to 0,L,R) = \delta(L - 2\pi R)$, so $\rho(s) = \delta(s - 1)$ with $s = L/(2 \pi R)$. It is intriguing that this classical term is present in semiclassical limit we found for $\rho(s)$. The different scaling law, $$s = {L\over \sigma D^{1/\sigma}}\ , \label{Slaw}$$ and the profusion of microscopic boundary loops for $s < s_c$, are striking features of the quantum case which we cannot conceive of predicting by quasi-classical arguments. The scaling law might seem to be an artifact of the combinatorics: on the discrete surface, the matter fields are incorporated into the manifold by polygons with different numbers of sides, so the notion of defining geodesic distance as the minimal number of polygons one must traverse to get from point to point is suspect. So, couldn’t we just define $R = \sigma D^{1/\sigma} / 2\pi$ and claim that $R$ is the “real” geodesic distance? The problem with this approach is that $D$ enjoys a linearity property that seems essential to the notion of geodesic distance. Namely, if $\gamma$ is a loop each of whose points is a geodesic distance $D_1$ from a given point $P$ on a random surface, and if $\gamma'$ is a loop each of whose points is a geodesic distance $D_2$ from $\gamma$, then each point on $\gamma'$ will be a geodesic distance $D_1 + D_2$ from $P$. Any increasing function of $D$ with the same property would have to be linear. The authors feel that a resolution of this question will have to come from a continuum formalism where the matter fields are more easily distinguished from the metric on the manifold. Another interesting limit where exact calculations are possible is that of an extremely long entrance loop, $L_0\to \infty$. Here it is appropriate to calculate $N(\zeta,L',D)$ in the limit where $\zeta^{-1}$ and $\tau^{-1/2}$ are much larger than $D^{1/\sigma}$ and $L'$. Now the characteristic curve equation, , has a very simple approximate solution, $$\zeta_0 \approx \zeta- Df_k(\zeta, \tau)\ , \label{SimpSol}$$ so that, from , $$N(\zeta,L',D) =1-\zeta L' + L'Df_k(\zeta, \tau)+\ldots \label{Smallz}$$ Performing the inverse Laplace transform, we obtain $$N(L\to\infty, L', D)= L' D F_k(L, \tau)\ .$$ Using this and in , we arrive at a remarkably simple formula: $$\rho(L_0\to\infty, L, D)= L_0 D L^{-k-1/2}\ . \label{LargeL}$$ This distribution is valid whenever $\tau^{-1/2} \gg L_0$ and both are much larger than $L$ and $D^{1/\sigma}$. The fact that the number of emitted s scales as $L_0$ could have been anticipated: this is simply due to the fact that a can split off anywhere along the parent. It is also clear why the number grows linearly with the elapsed geodesic time $D$, since in the discrete case the number of s that split off at each step of evolution should be constant as long as the length of the parent universe does not change appreciably. Note that the distribution in $L$ is again non-integrable for small $L$: the emitted s are overwhelmingly likely to be microscopic. Another consequence of the preceding discussion is that $$N(L, L'\to 0, D\to 0)= L' D F_k(L, \tau)\ . \label{TubeD}$$ This formula is interesting because it establishes a connection between the tube amplitude and the disk amplitude. $N(L, L'\to 0, D\to 0)$ is the sum over disks of boundary length $L$ with a marked point located at a vanishing geodesic distance $D\to 0$ from the boundary. Thus, we expect that in the $L' \to 0$, $D \to 0$ limit the tube amplitude reduces to the disk amplitude with a marked boundary point. This is indeed what happens, according to . We speculate that or some more general form of it could be used as the basis of a continuum derivation of . Discussion {#Discuss} ========== In ref. [@Kawai2] a string field theory formalism for $c=0$ gravity was introduced. From this formalism was elegantly derived. In fact, was derived first in ref. [@Kawai1] via a careful combinatorial analysis of discretizations, and the string field theory was tailored to reproduce this result. In this paper we extended the combinatorial analysis to arbitrary discretizations with $2m$-gons and established the validity of for 2-dimensional gravity coupled to the $(2, 2k-1)$ minimal models. This strongly suggests that the string field theory formalism of ref. [@Kawai2] encompasses all these theories. We choose a particular theory only through its disk amplitude $f(\zeta, \tau)$, which is the background value of the string field. In solving we found a drastic difference between the even and odd $k$. In other calculations no such major differences were noted. For instance, the disk amplitudes of are positive for all $k$. Thus, we may have the first indication of a serious difference between even and odd $k$ occurring for spherical surfaces. It would be nice to understand a deeper reason behind this. The non-integrable divergence of $\rho(L,D)$ for small $L$ indicates that random surfaces are very hairy, even when restricted to spherical topology. In terms of the propagation through geodesic distance of a loop along its world-sheet, we take this to mean that tiny loops—the baby universes—are constantly splitting off the main loop. The conclusion that the microscopic s are overwhelmingly more likely to split off than the macroscopic ones is quite intriguing in light of the large wormhole problem [@VK]. One might think of the microscopic s as analogous to the soft photons which create the well-known infrared problem in the bremsstrahlung cross-section. Therefore, only the inclusive probabilities, where we sum over all possible splittings, are non-vanishing. We have verified that the probability for a loop to propagate any finite $D$ along its world sheet without splitting is zero. The limit $k \to \infty$ (corresponding to the central charge decreasing without bound) bears a subtle relationship to classical gravity which we do not fully understand. Suggestions of a manifold that is smooth at length scales large compared to $D^{1/\sigma}$ emerge from the $k \to \infty$ limit of $\rho(L_0 \to 0,L,D)$. But baby universes still play an important role in this limit, as evidenced by the divergence of $\rho(L_0 \to 0,L,D)$ as $k \to \infty$ for $L < s_c \sigma D^{1/\sigma}$. Perhaps the proliferation of s up to this critical scale is related to the presence of many operators of negative dimension. It would be very interesting to study models with unitary conformal fields coupled to gravity, to see whether the surfaces they produce are more regular or more wild. To calculate the transfer matrix with the same methods as described here, however, one would need the disk amplitudes of the model with arbitrary boundary conditions on the matter fields, and these are not available even for such simple models as the Ising model. Acknowledgements {#acknowledgements .unnumbered} ================ We thank A. Polyakov and N. Seiberg for useful discussions. This work was supported in part by DOE grant DE-AC02-76WRO3072, the NSF Presidential Young Investigator Award PHY-9157482, James S. McDonnell Foundation grant No. 91-48, and an A. P. Sloan Foundation Research Fellowship. [9]{} S. Coleman, Nucl. Phys. [B307]{}, 864 (1988); S. Giddings and A. Strominger, Nucl. Phys. [B307]{}, 854 (1988). S. Coleman, Nucl. Phys. [B310]{}, 643 (1988); W. Fischler, I. Klebanov, J. Polchinski and L. Susskind, Nucl. Phys. [B327]{}, 157 (1989). V. Kaplunovsky, unpublished; W. Fischler and L. Susskind, Phys. Lett. [B217]{}, 48 (1989). D. J. Gross and A. A. Migdal, Phys. Rev. Lett. [64]{}, 127 (1990); M. Douglas and S. Shenker, Nucl. Phys. [B335]{}, 635 (1990); E. Brezin and V. Kazakov, Phys. Lett. [B236]{}, 144 (1990). M. Agishtein and A. A. Migdal, Nucl. Phys. [B350]{}, 690 (1991). H. Kawai, N. Kawamoto, T. Mogami, and Y. Watabiki, Phys. Lett. [B306]{}, 19 (1993). M. Staudacher, Nucl. Phys. [B336]{}, 349 (1990). V. Kazakov, Mod. Phys. Lett. [A4]{}, 2125 (1989). G. Moore, N. Seiberg, and M. Staudacher, Nucl. Phys. [B362]{}, 665 (1991). N. Ishibashi and H. Kawai, “String Field Theory of Noncritical Strings”, KEK–TH–364, UT–649 (1993). E. Brézin, C. Itzykson, G. Parisi, and J. B. Zuber, Commun. Math. Phys. [59]{}, 35 (1978). D.J. Gross and A.A. Migdal, Nucl. Phys. [B340]{}, 333 (1990). Figures {#figures .unnumbered} ======= 1\. Small example of a one-step evolution of a loop on a random surface tiled with squares. 2\. Basic shapes for $k=2$. 3\. The first few terms contributing to one basic shape, drawn on both the polygonal and dual lattices. 4\. The composition law. 5\. Characteristic curves for $k=3$ (top left) and $k=4$ (top right) above graphs of $f_3(\zeta,\tau)$ (bottom left) and $f_4(\zeta,\tau)$ (bottom right). 6\. The scaling function $\rho(s)$ for $k=2$, $k=6$, and $k=100$. [^1]: In order to work with finite quantities in the continuum limit, we rescale $N(\zeta,\zeta',D)$ to absorb the factor $1/\epsilon$ present in the lattice definition.	{ "pile_set_name": "ArXiv" }
--- author: - 'S. Hekker' bibliography: - 'bibinvited.bib' title: 'Observations and interpretation of solar-like oscillations in red-giant stars' --- Introduction ============ All stars with masses between roughly 0.6 and 6 solar masses evolve through a red-giant phase. Stars in this evolutionary phase follow a relatively narrow track through the H-R diagram, i.e., during the ascent and descent of the giant branch and their time on the horizontal branch, while their internal structures are different as a function of e.g., mass, metallicity and primordial rotational velocity. The large number of observable red giants, and the drastic changes in their internal structures on relatively fast timescales make these stars ideal for studying stellar evolution. The only way to unravel their internal structures is through study of stellar oscillations. For years however, late-G and early-K red giants had been considered to be pulsationaly stable, see e.g. Eyer & Grenon (1997), Jorissen et al. (1997). Therefore these stars were selected as candidates to act as reference stars for space missions, such as SIM (Space Interferometry Mission (SIM PlanetQuest), @frink2001). Indeed stable stars with long term (months to years) radial velocity variations of the order of 20 ms$^{-1}$ or less were discovered during the SIM preparation survey [@hekker2006a]. On a much shorter time scale of the order of hours to days most of these stars do however show pulsational variability. This variability originates from stochastically excited p-mode oscillations in the turbulent outer layers of the stars. Through the study of these, so-called solar-like oscillations, the internal structure of these stars can be investigated. However, until about ten years ago, the quality of observational data was insufficient to perform such study on many giants. Only for significantly evolved giants, which have oscillations with large amplitudes and relatively long periods of the order of days, such as Arcturus, oscillations could be detected. See for a review of these results @merline1999. The oscillations in less evolved red giants became detectable with the development of high-precision instrumentation and oscillations in $\alpha$ Uma (K0III) from photometry with the WIRE satellite [@buzasi2000] and in $\xi$ Hydrae (G7III) from radial velocity measurements [@frandsen2002] are among the earlier results mentioned. These observations were followed by theoretical computations by Dziembowski et al. (2001) for $\alpha$ Uma and, by among others, Houdek & Gough (2002), @teixeira2003 for $\xi$ Hydrae to interpret these observations. These early observations of solar-like oscillations in red giants initiated radial velocity multi-site campaigns on similar targets [@barban2004; @deridder2006], observations with non-dedicated space instruments, such as the star tracker on the WIRE satellite [@buzasi2000] and the Solar Mass Ejection Imager (SMEI) on board the Coriolis spacecraft, and observations with dedicated satellites such as MOST [@matthews2000], CoRoT Baglin et al. (2006) and recently Kepler [@borucki2009]. Results of these observations are presented by e.g., @tarrant2007 using SMEI data, @barban2007 [@kallinger2008] using MOST observations, @deridder2009 [@hekker2009] using CoRoT data, and @bedding2010 [@hekker2010a] using Kepler data. The interpretation of the observations of solar-like oscillations in red-giant stars from early observations and theoretical computations appeared to be ambiguous. Different interpretations resulted for instance in differences in mode lifetimes of $\xi$ Hydrae of an order of magnitude. This was later followed by a related discussion on the presence and observability of non-radial modes. Solar-like oscillations ======================= The oscillations in red giants originate from the same mechanism as the oscillations observed for the Sun, i.e. they are stochastically excited in the turbulent outer layer of the star. In many stars the observed oscillations follow a regular pattern and it seems to be justified to use the asymptotic approximation developed by @tassoul1980 for high-order low-degree modes: $$\nu_{n,\ell} \approx \Delta \nu \left (n+\frac{1}{2}\ell+\epsilon \right )-\ell(\ell+1)D_0, \label{asymptot}$$ with $\nu$ the oscillation frequency, $n$ the radial order, $\ell$ the angular degree, $\Delta \nu$ the large separation between modes of same degree and consecutive orders, which is inversely proportional to the sound travel time through the star, i.e. a proxy for the average density of the star. $\epsilon$ is sensitive to the surface layers and $D_0$ to deeper layers in the star. $D_0$ can be measured from any of the three small separations which are defined as: $\delta \nu_{02}$ the spacing between adjacent modes with $\ell$ = 0 and $\ell$ = 2, $\delta \nu_{13}$ the spacing between adjacent modes with $\ell$ = 1 and $\ell$ = 3, and $\delta \nu_{01}$ the amount by which the $\ell$ = 1 modes are offset from the midpoints between the $\ell$ = 0 mode on either side. If Eq. \[asymptot\] holds than $\delta \nu_{02}$ = 6$D_0$, $\delta \nu_{13}$ = 10$D_0$ and $\delta \nu_{01}$ = 2$D_0$ [@tassoul1980], in which case the $\ell$ = 1 modes are located on the left side of the midpoint between adjacent $\ell$ = 0 modes. However there is evidence that for red giants the relations between the three small separations and $D_0$ do not hold, as negative values for $\delta \nu_{01}$ have been observed [@carrier2010; @bedding2010]. In these cases the $\ell$ = 1 modes are located on the right side of the midpoints between the adjacent $\ell$ = 0 modes. A negative $\delta \nu_{01}$ is most likely a trace of stellar evolution (see Montalban et al. 2010, these proceedings). When the oscillation frequencies follow a regular pattern as predicted by Eq. \[asymptot\], a so-called échelle diagram (Grec et al. 1983) can be used to investigate the degree of the mode. In an échelle diagram the frequencies are represented as a function of the frequency modulo $\Delta \nu$. Oscillation modes of the same degree are then visible in vertical ridges, with the modes with $\ell$ = 0 and $\ell$ = 1 separated by approximately 0.5$\cdot \Delta \nu$ and the $\ell$ = 2 modes close to the $\ell$ = 0 mode (see for an example the bottom panel of Fig. \[deridder\]). Mode lifetimes ============== Solar-like oscillations are stochastically excited and have finite mode lifetimes, i.e., the modes are damped, with a rate $e^{-\eta t}$, with $\eta$ the damping rate, and re-excited before the oscillation mode is damped out. Due to this effect both the amplitude and phase of the oscillation are time dependent and the frequency peaks in the power spectrum have Lorentzian shapes with a width inversely proportional to the mode lifetime ($\tau$ $\sim$ $\eta^{-1}$). Therefore, the mode lifetime provides information on the damping processes. ![Theoretical damping rate $\eta$ as a function of frequency for $\xi$ Hydrae. The solid line shows the raw damping rates while the other lines show smoothed rates over the number of radial orders $n$ as indicated in the legend (Figure taken from @houdek2002). The large dip in the raw damping rates is related to the properties of the superadiabatic boundary layer.[]{data-label="houdek"}](m0_eta_579.eps){width="\linewidth"} @houdek2002 modelled the excitation and damping of the stochastic modes observed in $\xi$ Hydrae (Frandsen et al. 2002). From the damping rates (see Fig. \[houdek\]) obtained from their models, @houdek2002 computed the mode lifetimes of the observed oscillations. They obtained a mode lifetime of the order of 20 days. For the same star, @stello2006 developed a new method to measure the mode lifetimes directly from the width of the oscillation features in the power spectrum. For stars with short mode lifetimes the peaks in the power spectrum are wider than for stars with long mode lifetimes. By measuring the scatter of the observed frequencies around the expected regular pattern (Eq. \[asymptot\] and Fig. \[stello\]) and comparing this scatter with extensive simulations they found that the mode lifetime for $\xi$ Hydrae had to be of the order of 2 days. These results were confirmed by @brewer2009 using Gaussian process modelling. Observations with the CoRoT satellite have revealed that some red giants have long mode lifetimes ($>$ 50 days) (Baudin et al. 2010, De Ridder et al. 2009), while other red giants have shorter mode lifetimes of the order 15 days [@carrier2010]. Although the mode lifetimes of $\xi$ Hydrae could not be verified, these space-based photometric data did show that red giants can have mode lifetimes from the order of 15 days up to a few tens of days or even longer, possibly depending on their evolutionary state. ![Schematic illustration of the frequency scatter due to a finite mode lifetime. Dashed lines are the predicted mode frequencies and the solid peaks are the measured frequencies (Figure taken from @stello2006).[]{data-label="stello"}](stello.eps){width="\linewidth"} Non-radial modes ================ The Brunt-Väisälä ($N$) and Lamp ($L_{\ell}$) frequencies are huge in the dense core of red giants. Because of these huge values, all non-radial modes with frequencies lower than the cut-off frequency and larger than the fundamental radial mode have mixed character. In the envelope the oscillations behave like acoustic modes, while in the core the same modes behave like gravity modes. The evanescent region between the p- and g-mode cavities is at the origin of mode trapping. Some modes have significant kinetic energy in the g-cavity and low ones in the p-cavity; they are trapped in the core. Others have low kinetic energy in the g-cavity, but high ones in the envelope; they are trapped in the envelope @dupret2009. The theoretical computations of oscillations of $\alpha$ Uma and other red giants by @dziembowski2001 have been the standard theory work for red-giant asteroseismology for nearly a decade. In this work they predict that only radial modes will be observable on the surface of red-giant stars due to the trapping of non-radial oscillation modes in the core of the star. Consequently, the observed oscillation modes of $\xi$ Hydrae [@frandsen2002] and $\epsilon$ Ophiuchi (De Ridder et al. 2006) were interpreted as radial oscillations. Hekker et al. (2006a) were the first to investigate the nature of the modes using line profile analysis of the spectra of $\xi$ Hydrae, $\epsilon$ Ophiuchi, $\eta$ Serpentis and $\delta$ Eridani. They concluded that non-radial oscillations were present, although a definite mode identification could not be made. An updated version of this analysis has recently been published [@hekker2010b], which supports the general conclusions of Hekker et al. (2006a). The analysis of the data taken during the spectroscopic multi-site campaign presented by @deridder2006 revealed two possible models for $\epsilon$ Ophiuchi. Therefore this star was also observed for three weeks with the MOST satellite. These results were interpreted as radial modes with mode lifetimes of the order of a few days [@barban2007] and could narrow the solution for $\epsilon$ Ophiuchi to one specific model. Subsequently, @kallinger2008 combined the frequencies obtained from spectroscopic and photometric observations and interpreted them as long lived (order tens of days) radial and non-radial modes. These results were in agreement with @hekker2006b, but not with @deridder2006 and @barban2007. From the latter example it is clear that the interpretation of the mode lifetime and the degree of the mode are not independent. On the one hand an interpretation of short mode lifetimes (wide frequency peaks in the power spectrum) might incorporate a possible $\ell$ = 2 mode as part of the $\ell$ = 0 mode. On the other hand an interpretation of long lifetimes (narrow peaks in the power spectrum) might lead to the interpretation of peaks as $\ell$ = 2 modes, which are in reality part of the neighbouring $\ell$ = 0 mode. With recent CoRoT and Kepler observations (Bedding et al. 2010, De Ridder et al. 2009) the degree of the oscillation modes of many red giants could be investigated through the asymptotic relation (Eq. \[asymptot\] and Fig. \[deridder\]). These results show unambiguous evidence for the presence of non-radial oscillations in red giants. Also more recent theoretical work by @dupret2009 predicts the presence of non-radial oscillations in red giants with observable heights in the power spectrum, i.e., these are non-radial modes trapped in the envelope. ![Power spectrum (top) and échelle diagram of a red giant observed with CoRoT, which shows evidence of non-radial modes. (Figure taken from @deridder2009).[]{data-label="deridder"}](powerspectrumnature.eps){width="\linewidth"} ![Power spectrum (top) and échelle diagram of a red giant observed with CoRoT, which shows evidence of non-radial modes. (Figure taken from @deridder2009).[]{data-label="deridder"}](echellenature.eps){width="\linewidth"} Current challenges ================== The field of red-giant asteroseismology has not yet matured, and is both observationally as well as theoretically a challenging field. To infer the internal structure of the stars a number of parameters need to be determined, such as the frequency of maximum oscillation power and large frequency separation ($\Delta \nu$ in Eq. \[asymptot\]), which provides information on the stellar radius [@kjeldsen1995], which increases with age, the oscillation frequencies, small frequency spacing and mode identification. When knowing the individual frequencies, their degrees and stellar parameters such as effective temperature, surface gravity and metallicity, an exact model for the star can in principle be computed. So far, both theoretical and observational problems have prevented the determination of the internal structure of red giants in detail. Theoretical challenges ---------------------- One of the main challenges from the theoretical side is the computation of convection. The often used classical mixing-length parametrisation [@bohmvitense1958] is computationally attractive, but does not provide a physical description of the convection processes within a star. A further theoretical dilemma is that the numerical stability of some evolution codes decreases when modelling a helium flash scenario at the tip of the giant branch. This can be solved by computationally expensive algorithms or by circumventing the numerical problems by restarting the computation at the horizontal branch (see e.g. @mazumdar2009). This has the disadvantage that the internal structure changes taking place during the He-flash cannot be followed. Another numerical difficulty originates from the very large number of nodes in the eigenfunctions in the g-mode cavity. The huge number of nodes originates from the high density contrast between the core and the envelope and leads to radiative damping. Due to the mixed nature of the non-radial modes, the amount and exact region in the star where this damping occurs may have influence on the observed oscillations. Therefore the inner region of the star needs to be probed with a dense numerical grid, which is computationally challenging. See @dupret2009 and Montalban et al. (2010) (these proceedings) for state-of-the-art models of oscillating red-giant stars. Observational challenges ------------------------ First of all, the finite mode lifetimes of stochastically driven and damped oscillations are predicted to be different for modes of different degrees, depending on their mode inertia. Higher mode inertia for non-radial modes lead to longer lifetimes, generally of the order of tens of days, but they can be of the order of hundreds of days or more (see e.g. @dupret2009). If the duration of the observations is shorter than these lifetimes, the modes are unresolved and have lower heights in the power spectrum than resolved modes. In order to resolve, and thus observe, non-radial modes, time series should be used with a length that exceeds the expected lifetime of the observed star. To be able to obtain a reliable measure of the lifetime, the timespan of the timeseries should be of the order of ten times as long as the lifetimes of the oscillation modes [@hekker2010c]. The relatively short timespan of about a month for the single-site observations of $\xi$ Hydrae are most likely the reason for the different results regarding the mode lifetimes (see Section 3). A second observational challenge relates to the trapping of the non-radial modes. The size of the evanescent region and thus the efficiency of trapping increases with the degree $\ell$ of the modes. When only modes trapped in the envelope are considered, the frequency spectrum shows a regular pattern similar to solar-type main-sequence stars which can be interpreted using Eq. \[asymptot\]. However, for stars at a certain evolutionary phase, such as the bottom of the giant branch, the trapping is not efficient and the interaction between the p- and g-mode cavities leads to many avoided crossings. Hence several modes with $\ell$ = 1 with nearly the same frequency but different $n$ can be observable [@dupret2009]. This leads to a complex frequency pattern in which there might exist ambiguity between the interpretation of an oscillation feature as a single wide mode with short mode lifetime or several mixed modes with the same degree and longer lifetimes. Thirdly, the large radii of red-giant stars adjust the pulsation periods from minutes, for the Sun, to hours. This complicates ground-based single-site efforts considerably, and calls for either multi-site campaigns or observations from space. Furthermore, like for the Sun, turbulent motions in the convective envelope cause granulation. The granulation time scales depend on the size of the granules, and their velocities, and are generally of the order of days for red giants. For the most evolved and luminous stars, which have the largest radii, the granulation time scales are of the same order as the oscillation timescales. This causes difficulties in disentangling granulation and oscillation features in the power spectrum. Increase in frequency resolution, i.e. timeseries with longer timespan, are needed to investigate the features at low frequencies in more detail. The realisation of the CoRoT satellite, and more recently the Kepler satellite, are very important for tackling the observational challenges of problems related to observing solar-like oscillations in red giants. These satellites observe the same field with high-precision photometry for 150 days and 3.5 years, respectively. This means that non-radial modes can be resolved in which case they have detectable heights in the power spectrum. The increased frequency resolution will also be important to attempt to disentangle granulation and oscillation features in the power spectrum. Future prospects ================ So far the CoRoT data have already greatly increased our knowledge of solar-like oscillations in red-giant stars, both from an observational as well as from a theoretical point of view. Apart from the ensemble studies [@hekker2009; @mosser2010] and the unambiguous determination of long lived non-radial oscillations [@deridder2009], detailed studies of individual stars, an investigation into mass and radius determination from oscillation properties (Kallin- ger et al. 2010) and a statistical investigation in mode lifetimes (Baudin et al. 2010) have been carried out. These observational efforts have been accompanied by theoretical investigations using adiabatic (e.g., @dupret2009) and non-adiabatic codes to simulate the stars, as well as stellar population studies [@miglio2009]. Also, first results from the Kepler satellite on low-mass low-luminosity red giants [@bedding2010] enabled for an extended statistical study in a frequency range where CoRoT is less sensitive. Notwithstanding these efforts, there still remain many open questions about the internal structures of red giants, and the different processes taking place in different layers and between layers in these stars. The excitation and damping of these oscillations, the time scales at which these process occur and their connection with convection and granulation are not well understood. Furthermore, it would be very interesting to be able to infer whether the star is in the hydrogen-shell burning or helium-burning phase. With a detailed picture of the internal structure and mass of the red giants, it might also be possible to trace back their evolution and determine their primordial rotation velocity or infer the difference in internal structure due to different chemical composition.The future looks very promising with respect to observations from dedicated instruments for asteroseismology. The CoRoT satellite is still taking data and will do so for the next couple of years and Kepler has now been taking data for nearly a year, with a 3.5 year nominal duration of the mission. Both satellites observe many relatively faint stars (roughly 9 - 16 mag in V), which has the advantage of providing a statistically significant sample, but the disadvantage that stellar parameters, such as effective temperature, surface gravity and metallicity can not be obtained with high accuracy. For bright targets, for which it is relatively easy to obtain stellar parameters, the MOST satellite (Micro variability and Oscillations of STars) has been taking valuable time series data of a couple of weeks for several red giants. Furthermore, dedicated instrumentation is currently under development. BRITE-Constellation consists of at least two small satellites with different colour filters, which aim to take simultaneous two-colour photometry of bright targets (2-6 mag in V). Depending on the exact launch time, which is scheduled for 2011, there are fields that can be observed year round, while other fields can be observed for at least half a year. Also PLATO (PLAnetary Transits and Oscillations of stars), a medium-sized mission, is in preparation and among three missions of which two will be launched in the ESA Cosmic Vision 2015 - 2025 programme. Another development is SONG (Stellar Observations Network Group), which aims for a dedicated network of 1-m class telescopes with high-resolution spectrographs for radial-velocity measurements. The prototype of these telescopes is currently being build at Tenerife, Spain. Progress is also being made on the theoretical interpretation of the solar-like oscillations in red giants. Stellar evolution codes are constantly improved and many more are now extended to also incorporate evolution of giant stars. Detailed modelling of red giants for which oscillations are now observed with high accuracy (for example HR 7349 @carrier2010) are now carried out. Also ensemble studies such at the one by @miglio2009 are developed further.It is clear that with the increased frequency resolution and accuracy of the data from recent and future instruments and the simultaneous developments in theory, significant progress in the field of red-giant asteroseismology has recently been made and can be expected in the (near) future. SH wants to thank Maarten Mooij and Bill Chaplin for useful discussions that improved the manuscript considerably. SH acknowledges financial support from the UK Science and Technology Facilities Council (STFC).	{ "pile_set_name": "ArXiv" }
--- abstract: 'Experimental observations of turbulent buoyant plumes, produced by a constant source of buoyancy, have been described with great success using a horizontally averaged model for the conservation of mass, momentum and buoyancy flux. However, experimental observations of plumes with time-dependent buoyancy fluxes has proved more challenging for quantitative models. At each level in the plume, the horizontal variation in velocity leads to an along-axis shear and hence dispersive transport relative to the mean. With a time-dependent source, axial dispersion of the dynamic properties of the plume has a key role in the evolution of the flow. Using ideas of mixing length theory, we introduce a model for this axial dispersion, and test the model by comparison with experimental observations of plumes in which the buoyancy flux is suddenly decreased or suddenly increased. In both cases, we show the transition from one buoyancy flux to another is self-similar, and using the data, we find the axial dispersion may be expressed as $\beta U b$ where $\beta$ lies in the range $0.70$–$0.88$ and $U$ and $b$ are respectively the horizontally averaged vertical velocity and radius of the plume at height $z$ and time $t$. Our model also reduces to that of a classical plume when the buoyancy flux is steady' author: - 'Matthew M. Scase$^{1, 2}$ & Andrew W. Woods$^{1, 3}$' title: Dynamical dispersion in turbulent plumes --- \[sec:int\]Introduction ======================= Turbulent buoyant plumes arise in many environmental and geophysical situations where there is a localised source of buoyancy. In the special case in which there is a steady source of buoyancy, Morton Taylor & Turner (1956) tested the predictions of an integral model for the conservation of mass, momentum and buoyancy in a plume, coupled with an entrainment hypothesis, by using the results of a series of laboratory experiments. Applications ranging from volcanic eruption columns [see e.g., @woods88; @scase09], ocean currents (see e.g., Holland 2011; Holland, Hewitt & Scase 2014), hydrothermal plumes, fire plumes through to convective plumes in buildings have been studied, advancing this original work [@woods10]. In some situations, the buoyancy flux may vary in time, and this leads to interesting questions about how the evolving buoyancy flux migrates through the plume and provides a new test for the model, and in particular the entrainment hypothesis. In this context, @scase06, Scase, Caulfield & Dalziel (2008), Scase, Aspden & Caulfield (2009) and @scaseHewitt12 have explored the dynamics of plumes in which the source buoyancy flux changes suddenly from one constant value to another. @scd08 present a series of laboratory experiments which identify that the change in source conditions leads to a region of adjustment which migrates through the plume, such that near the source a steady plume with the new buoyancy flux becomes established, while further from the source, the flow adjusts over some region to that of the original plume. However, the theoretical description of this transition region has to date posed some challenges (e.g., Scase & Hewitt 2012; Woodhouse, Phillips & Hogg 2016). Parallel to this work, there has been some investigation of the mixing of passive tracer within plumes and jets. Landel, Caulfield & Woods (2012) and @roccoWoods15 demonstrated that in a steady jet or plume, a pulse of tracer becomes axially dispersed owing to the variation in the velocity with horizontal position across the flow. Guided by the ideas of shear dispersion, they demonstrated that this axial dispersion could be described by using an axial dispersivity based on mixing length theory. In that work, the tracer was dynamically passive. In the present contribution, we explore the effect of such dispersion on the dynamically active properties of the flow, and test the model using the experimental data presented by @scd08 [@sac09]. In §\[sec:model\], we propose a model of a time dependent plume, including an idealised model of the axial dispersion. We demonstrate that the model is analogous to the classical model of turbulent buoyant plumes when there is a steady buoyancy flux, and we then illustrate that the model admits similarity solutions to describe the transition in the flow following a change in the source buoyancy flux from one value to a second. We re-analyse the experimental data originally presented by @scd08 [@sac09] and demonstrate that the transition region does in fact evolve in a self-similar manner. By comparison with our model, we determine the empirical constant $\beta$, which constrains the magnitude of the axial dispersivity, $\beta U b$, where $U$ and $b$ are the characteristic vertical plume velocity and plume radius. Finally, we discuss some possible future avenues for research. \[sec:model\]Model ================== ![\[fig:cartoon\]Image of a plume rising through a homogeneous ambient fluid. The colours represent buoyancy. Fluid at the centre of the plume rises more rapidly than fluid at the edges of the plume leading to a Taylor-like shear dispersion. The images shown are from Movie 2 and correspond to times $t=-8.417$s, $t = -8.083$s and $t=-7.750$s.](scaseWoods_fig1-eps-converted-to.pdf) For a Boussinesq flow, the governing equations for the rate of change in time of the horizontally integrated mass, momentum and buoyancy may be expressed in terms of the gradient of the horizontally integrated mass, momentum and buoyancy fluxes, combined with (i) a model for the entrainment of ambient fluid, (ii) the buoyancy force and (iii) a model to represent the axial dispersion associated with the turbulent mixing coupled with the shear [cf. @LandelEtAl]. In building the model, we work with ensemble averages, so that the properties are averaged over the turbulent fluctuations. For each of the properties $f$ of the flow, we write $f = \bar f + \hat f$, where $\bar f$ is the ensemble average, and $\hat f$ denotes the turbulent fluctuations. The horizontally integrated, ensemble averaged properties for the vertical velocity $u(r, \theta, z, t)$ and buoyancy $g'(r, \theta, z, t)$ are then given by $$\renewcommand{\theequation}{\arabic{section}.\arabic{equation}\,$a$,$b$} \pi U b^2 = Q = \int \bar{u} \, {\textrm{d}}A, \qquad \pi S_m U^2 b^2 = S_m M = \int \left(\bar{u}^2+\hat{u}^2\right) {\textrm{d}}A,$$ $$\renewcommand{\theequation}{\arabic{section}.\arabic{equation}\,$c$} \addtocounter{equation}{-1} \pi S_f U G' b^2 = S_f F = \int \left(\bar{u} \bar{g}' + \hat{u}\hat{g'}\right) {\textrm{d}}A,$$ where $U$, $b$ and $G'$ are the characteristic vertical velocity, radius and buoyancy of the plume, such that the ensemble-averaged volume, momentum and buoyancy, per unit height, are given by $\pi b^2$, $\pi b^2 U$ and $\pi b^2 G'$. The shape factors $S_f$ and $S_m$ are associated with the radial distribution of the velocity and buoyancy [cf. @turner79]. The dispersive flux arises from the turbulent fluctuations in the flow coupled with the axial shear, and, using mixing length arguments for such a highly turbulent flow, we expect the associated dispersivity to be proportional to the characteristic speed and radius of the flow, $\beta U b$ [cf. @scaseHewitt12; @roccoWoods15], where $\beta$ is a constant. This leads to fluxes of the horizontal integral of the dispersive buoyancy and momentum of the form $-\beta U b\ \partial (G' b^2)/\partial z$ and $-\beta U b\ \partial(U b^2)/\partial z$ respectively. As well as the dispersive transport of momentum and buoyancy, the eddies tend to transport material near the edge of the plume outwards and downwards relative to the mean flow, engulfing ambient fluid. The mixture is then swept back into the plume, driven by the shear and becomes incorporated into the plume leading to the turbulent entrainment of ambient fluid into the plume. Indeed, figure \[fig:cartoon\] shows three instantaneous images of the concentration field in a turbulent plume taken in a plane which cuts through the centre of the plume and which is perpendicular to the line of sight. The plume has a constant source of buoyancy. The frames are taken from the experiment shown in Movie 1 at times $t=-8.417$s, $t = -8.083$s and $t=-7.750$s. The plume’s concentration field has been passed through an edge-detection algorithm [@canny86] to highlight the coherent structures within the flow and this is presented in Movie 2. @mtt56 proposed that the volume flux of ambient fluid which is entrained into the plume is proportional to the vertical speed of the plume and the plume circumference, and this model is quantitatively consistent with experimental data. However, the turbulent eddies also tend to smooth out discontinuities in the width of the plume, since eddies which form in the wider parts of the plume partially mix back into the adjacent narrower parts of the plume, owing to the shear in the flow. This can lead to a net dispersive transport of volume in regions where there is a vertical gradient of radius, and, for consistency with the momentum and buoyancy transport, we hypothesize that this dispersive volume flux has the form $-\beta Ub\ \partial(b^2)/\partial z$. We now write down a model for the time-dependent plume, and then develop solutions of this model to test for consistency with (i) the classical steady plume solutions presented by @mtt56 and (ii) the experimental data monitoring the progression of the transition region following a discrete increase or decrease in the plume buoyancy flux [@scd08; @sac09]. Combining the processes above, we hypothesize that the mass, momentum and buoyancy fluxes are governed by the relations \[eq:sys\] $$\label{eq:massCons} {\frac{\partial}{\partialt}}\left(\frac{Q^2}{M}\right) + {\frac{\partialQ}{\partialz}} = 2\alpha M^{1/2} + \beta{\frac{\partial}{\partialz}}\left[M^{1/2}{\frac{\partial}{\partialz}}\left(\frac{Q^2}{M}\right)\right],$$ $$\label{eq:momCon} {\frac{\partialQ}{\partialt}} + S_m{\frac{\partialM}{\partialz}} = \frac{QF}{M} + \beta{\frac{\partial}{\partialz}}\left(M^{1/2}{\frac{\partialQ}{\partialz}}\right),$$ $$\label{eq:buoyCon} {\frac{\partial}{\partialt}}\left(\frac{QF}{M}\right) +S_f{\frac{\partialF}{\partialz}} = \beta{\frac{\partial}{\partialz}}\left[M^{1/2}{\frac{\partial}{\partialz}}\left(\frac{QF}{M}\right)\right].$$ Similarity solutions for a steady plume --------------------------------------- A steady plume, generated by a constant source of buoyancy issuing from a point source follows a self-similar structure as a function of height [@mtt56]. The time-dependent system of equations should admit steady solutions of the same form. We therefore seek solutions $$Q = (2\alpha)^{4/3}F_0^{1/3}z^{5/3} q_0, \quad M = (2\alpha)^{2/3}F_0^{2/3}z^{4/3} m_0, \quad F = F_0,$$ where $F_0$ is the source buoyancy flux and $\alpha$ is the entrainment coefficient. Consistency conditions following from then require that $$\label{eq:q0m0} q_0 = \frac{12S_m m_0^2}{20\beta^* m_0^{3/2} + 9} \quad\textnormal{and}\quad m_0 = \left\{ \frac{9}{40}\frac{S_m\left[1-\left(1-24\beta^/5\right)^{1/2}\right]-2\beta^}{\beta^\left[S_m(6S_m/5 - 1)+\beta^\right]}\right\}^{2/3},$$ where $\beta^* = 2\alpha\beta$. The plume radius is defined as $$\label{eq:plumeRadius} b = \frac{Q}{M^{1/2}} = 2\alpha z\frac{q_0}{m_0^{1/2}} = \frac{6\alpha z}{5}\frac{2}{1 + \left(1 - 24\beta^/5\right)^{1/2}} , $$ this solution requires that $\beta^<5/24$, and so we can write $b = 6\alpha z/5\left(1 + 5\beta^/5 + O(\beta^{2})\right)$. This is analogous to the classical result $$b = \frac{6\alpha_c z}{5}$$ provided the classical entrainment coefficient $\alpha_c$ is related to the present entrainment coefficient, $\alpha$, and dispersion coefficient, $\beta^$ according to $$\alpha_c = \alpha \left[1 + \frac{6\beta^}{5}+O\left(\beta^{2}\right)\right]$$ The equivalence of the two models implies that the present entrainment coefficient $\alpha$ is a fractionally smaller than the classical value [@mtt56] owing to the dispersion which in fact has the effect of carrying some of the entrained material to points in the plume ahead of the height at which it was entrained: the classical entrainment coefficient, which implicitly includes such dispersion, therefore appears a little larger. The magnitude of this difference can only be determined by measurement of the dispersion coefficient $\beta$ which forms the subject of the main part of this paper. Figure \[fig:abc\] shows contours of the equivalent classical entrainment coefficient $\alpha_c$ for varying entrainment coefficient $\alpha$ and scaled dispersion coefficient $\beta^ = 2\alpha\beta$. For a steady source of buoyancy, the dispersive flux of buoyancy exactly vanishes so that the buoyancy flux at each height is a constant. The dispersive transport of mass is a constant fraction of the entrained mass flux, while the dispersive transport of momentum scales with height exactly as the buoyancy force. Note also that the plume radius is independent of both shape factors $S_m$ and $S_f$. The effects of the axial dispersion can be much more important on the evolution of flows in which the source is time-dependent. As mentioned in the introduction, we test the new model by examining solutions for the case in which the steady flux supplying the plume changes from one value to another. There is already detailed experimental data available for the this problem [@scd08; @sac09], and we use the data to test and calibrate the present model. ![\[fig:abc\]Contours of the classical entrainment coefficient $\alpha_c$ for varying entrainment coefficent $\alpha$ and scaled dispersion coefficient $\beta^* = 2\alpha\beta$.](scaseWoods_fig2-eps-converted-to.pdf) Similarity solutions for plumes involving a sudden change in buoyancy flux -------------------------------------------------------------------------- Guided by the classical solutions for a steady state plume, we now seek solutions to describe the evolution of the transition region in which a plume adjusts from one steady buoyancy flux to another. Far from the region of transition, we expect the flow to asymptote to the solution for a steady plume with constant buoyancy flux. These limiting solutions have the same structure but a different value of the buoyancy flux and so given that there is no external length scale regulating the along-plume extent of the transition region, we expect that the transition region will be described by similarity solutions of the form $$\label{eq:simSols} Q = (2\alpha)^{4/3}F_0^{1/3}z^{5/3}q(\eta), \quad M = (2\alpha)^{2/3}F_0^{2/3}z^{4/3}m(\eta), \quad F = F_0f(\eta),$$ where the functions $q$, $m$ and $f$ vary across the transition region, and where $F_0$ is the initial source buoyancy flux, $F_1$ is the new source buoyancy flux, $\alpha$ is the entrainment coefficient and $\eta$ is the nondimensional similarity variable defined by $$\eta = \frac{(2\alpha)^{1/2}}{F_0^{1/4}}\frac{z}{t^{3/4}}.$$ This similarity variable, $\eta$, is analogous to that identified by @woodhouseEtAl but preserves a linear relationship between height and $\eta$. The boundary conditions require that $q(0)=q_0(F_1/F_0)^{1/3}$, $m(0)=m_0(F_1/F_0)^{2/3}$, $f(0)=F_1/F_0$ while in the far-field, as $\eta \rightarrow \infty $, $q \rightarrow q_0$, $m \rightarrow m_0$ and $f\rightarrow 1$, where $q_0$ and $m_0$ are the nondimensional factors defined in associated with steady plumes. The similarity solutions are therefore governed by the coupled, non-linear, dimensionless set of ordinary differential equations \[eq:simEqs\] $$\label{eq:massSim} -\frac{3\eta^{7/3}}{4}{\frac{{\textrm{d}}}{{\textrm{d}}\eta}}\left(\frac{q^2}{m}\right) + \frac{1}{\eta^{2/3}}{\frac{{\textrm{d}}}{{\textrm{d}}\eta}}\left(\eta^{5/3}q\right) = m^{1/2} + \frac{\beta^}{\eta^{2/3}} {\frac{{\textrm{d}}}{{\textrm{d}}\eta}}\left[\eta^{2/3}m^{1/2}{\frac{{\textrm{d}}}{{\textrm{d}}\eta}}\left(\frac{\eta^2q^2}{m}\right)\right],$$ $$\label{eq:momSim} -\frac{3\eta^{7/3}}{4}{\frac{{\textrm{d}}q}{{\textrm{d}}\eta}} + \frac{S_m}{\eta^{1/3}}{\frac{{\textrm{d}}}{{\textrm{d}}\eta}}\left(\eta^{4/3}m\right) = \frac{qf}{m} + \frac{\beta^}{\eta^{1/3}} {\frac{{\textrm{d}}}{{\textrm{d}}\eta}}\left[\eta^{2/3}m^{1/2}{\frac{{\textrm{d}}}{{\textrm{d}}\eta}}\left(\eta^{5/3}q\right)\right],$$ $$\label{eq:buoySim} -\frac{3\eta^{4/3}}{4}{\frac{{\textrm{d}}}{{\textrm{d}}\eta}}\left(\frac{qf}{m}\right) + S_f{\frac{{\textrm{d}}f}{{\textrm{d}}\eta}} = \beta^{\frac{{\textrm{d}}}{{\textrm{d}}\eta}}\left[\eta^{2/3}m^{1/2}{\frac{{\textrm{d}}}{{\textrm{d}}\eta}}\left(\eta^{1/3}\frac{qf}{m}\right)\right].$$ We can solve these equations numerically, and determine solutions for $q, m$ and $f$ as functions of $\eta$ for given values of $\alpha$, $\beta$ and $F_1/F_0$. The boundary value problem defined by and the stated boundary conditions on the domain $[0, \infty)$ was solved by formulating the problem in terms of Chebyshev polynomials [see @driscollEtAl for details]. By comparison of the model prediction of the transition region with the data presented by [@scd08; @sac09] for plumes in which there is (i) a decrease in the buoyancy flux and (ii) an increase in the buoyancy flux, we can then obtain an estimate for $\beta$. In comparing the model with the experimental observations we follow the evolution of the concentration of fluorscein dye integrated across a horizontal plane through the centre of the plume. The evolution of the concentration of the dye is analogous to the evolution of the buoyancy, as may be seen by noting that if a passive tracer has flux $C = C_0 c(\eta)$ (cf. \[eq:simSols\]$c$) then $c$ is governed by an equation identical to (\[eq:simEqs\]$c$) with $f$ replaced by $c$. Comparison of solutions with experimental data ============================================== A decrease in buoyancy flux --------------------------- In @scd08 a series of experiments describing the effect of a rapid reduction in the source buoyancy flux of a plume were presented. A strong initial turbulent plume was established via two supplies of buoyant fluid to a single source. At $t = 0$s one of the supplies of buoyant fluid was switched off, reducing the buoyancy flux at the source. This reduction in buoyancy flux was due to a reduction in the source velocity. The source diameter, density difference (reduced gravity) and tracer concentration (fluoroscein) were unchanged. Data was presented from 100 experiments in which observations of the horizontally averaged concentration of dye as a function of height were made. By averaging the data from the nominally identical experiments, an ensemble average of the concentration as a function of height was produced at a series of fixed times after the change in the buoyancy flux. The flux of tracer concentration at the source may be written as $C_0 = c_0 b_0^2 U_0$ and the source buoyancy flux is given by $G'_0 b_0^2 U_0$, where $c_0$ is the source concentration of tracer (not to be confused with the similarity solution for concentration [flux]{}, $c$), $G'_0$ is the reduced gravity at the source, $b_0$ is the source radius and $U_0$ is the source velocity. Hence, reducing the buoyancy flux at the source by reducing the source velocity, but maintaining both the reduced gravity (density difference or buoyancy equivalently) and the source diameter, leads to a reduction in the source concentration flux, $C_0$, of exactly the same proportion. For the experiments described in @scd08, the source buoyancy flux and hence concentration flux was reduced by a factor of 4.90. The concentration of tracer at a given height, $z$, in the plume, $c/q$ is proportional to $C_0F_0^{-1/3}z^{-5/3}$. Hence when $F_0$ is reduced by a factor of 4.90, the concentration is correspondingly reduced by a factor of $4.90^{2/3}\approx2.88$. By rescaling the vertical coordinate with $t^{3/4}$, the concentration $c/q$ can be presented as a function of $\eta$, and if the transition is self-similar, the profiles should collapse to a universal curve. In figure \[fig:data\]a, we present the experimental data, rescaled in this manner, for times $t= 0.83$s – $16.29$s after the change in the buoyancy flux. The grey shaded zone represents one standard deviation of the concentration relative to the mean as a function of $\eta$. It is seen that the concentration profile does collapse to a fixed profile, with the standard deviation being on average $1\%$, and no more that $4\%$, of the signal. We consider the experimental data in a fixed interrogation window where the response of the light sheet to tracer concentration is linear and we observe the anticipated reduction in tracer concentration after the transient has passed. The interrogation window had a vertical extent of 5cm and was centred on the axis of the plume at a height of 40cm above the plume source. The design of the experiments means that there is a necessary change in the non-zero virtual origin [see @huntKaye] between the initially forced jet-like and subsequently distributed, or ‘lazy’, plume source conditions. Here we assume that these effects are negligible within the interrogation window, far from the source, for comparison with the similarity solutions that satisfy . We have solved the new model equations as presented in and now compare these solutions with the experimental data. The nondimensional boundary condition for the source buoyancy flux that corresponds to the dimensional parameters in the experiments is $F_1/F_0 = 4.90^{-1} = 0.20$, and, following @woodhouseEtAl, values for the shape factor of both $S_m = 1.08$ and $S_m = 1.00$ were used together with $S_f = 1.00$ to assess the relative importance of the dispersion and the shear. The value of $\beta^$ was varied over the range 0 and $5/24$ and the root mean square (RMS) error between the model similarity solution and the experimental data calculated. In figure \[fig:data\]a, we illustrate the best fit model prediction for concentration, $c/q$, as a function of $\eta$, and this is given by $\beta^=0.14$. The model prediction for $S_m = 1.08$ is shown solid, and the model prediction for $S_m = 1.00$ is shown dashed, though it is difficult to distinguish the solutions at this scale. The RMS error between the model prediction and the experimental data as $\beta^$ is varied is shown in figure \[fig:data\]c for $S_m = 1.08$ (blue solid line) and $S_m = 1.00$ (blue dashed line). In both cases the error is minimized for $\beta^* \approx 0.14$. We observed that as $\beta$ decreased to very small values, the model predicts a sharp transition region as the model prediction attempts to adjust from one buoyancy flux to another, somewhat reminiscent of the shock-like solutions presented by @woodhouseEtAl. However, for larger values of $\beta$ this transition region is predicted to spread out vertically, and the model prediction provides a closer fit to the experimental data. It is of relevance to note that @woodhouseEtAl proposed that the transition region spreads in time in a self-similar fashion owing to the different advection speeds of the momentum and buoyancy fluxes. These different speeds are associated with the different shape factors for the velocity and buoyancy profile in the plume. The difference in the shape factor does indeed lead to a gradual spreading of the transition region, but, in the absence of the dispersion modelled herein, the model also leads to the prediction of discontinuities in the concentration profile that were not observed in the experiments. Owing to the nonlinearity of the governing equations, the transition region spans a finite range of values of $\eta$ and within this region the plume adjusts smoothly from the original to the new steady buoyancy flux. In figure \[fig:data\]d we illustrate the minimum and maximum values of $\eta$, referred to as $\eta_1$ and $\eta_2$ respectively, across the transition zone as a function of $\beta^$. We present solutions for the case in which the advection speed of momentum and buoyancy are the same (dashed), corresponding to the case $S_m= S_f =1$, and also for the case in which $S_m=1.08$, $S_f=1$ (solid), values consistent with experimental data for plumes with a constant buoyancy flux [cf. @woodhouseEtAl; @turner79]. It is seen that when $\beta^$, and hence $\beta$, is very small, the two models lead to somewhat different ranges for $\eta$ across the transition zone. However, as $\beta$ approaches the range of values which provide the best fit with the experimental data of @scd08 ($\beta^=0.14$), there is less than 1.2$\%$ difference in range $\eta_2 - \eta_1$ across the transition region between the model predictions using the different values for the shape factors. This suggests that the effect of the different advection speeds of momentum and buoyancy in causing the transition zone to spread out is somewhat secondary to the effect of the dispersion, as parameterised by $\beta$. An increase in buoyancy flux ---------------------------- In @sac09 the results of a series of numerical experiments in which the source buoyancy flux of an established plume was suddenly increased were reported. These data provide a second, and independent test of the similarity model . Again, the data describing the vertical variation of the horizontal average of the concentration of passive tracer collapses to a universal curve when we rescale height to the similarity variable $\eta = z/t^{3/4}$. In these numerical experiments the buoyancy flux was increased by simultaneously changing both the reduced gravity and the velocity of the source fluid. This was performed in such a way that the buoyancy flux at the source increased by a factor of 20, but the ratio of reduced gravity of the source fluid to the square of the velocity of the source fluid was constant. This implies that the source velocity was increased by a factor of $20^{1/3}$ and the source reduced gravity was increased by a factor of $20^{2/3}$. This choice leads to a constant value $\varGamma_0 = 0.26$ so that the location of the virtual origin of the plume is a constant [cf. @huntKaye]. A consequence of this choice of boundary conditions is that for a steady plume, the tracer concentration at a height $z$ scales as $U_0^{2/3}G'_0\,\!^{-1/3}z^{-5/3}$. As a result, the expected concentration at each height $z$ returns to its initial value after the transient has passed through. Figure \[fig:data\]b illustrates the collapse of the experimental data to a universal curve, again with the grey zone representing the standard deviation of the concentration relative to the ensemble average at each height, as derived from the numerical simulations. The red curve is the best-fit prediction of the model for the concentration as a function of $\eta$, optimized over $\beta^\in[0, 5/24]$. Again, with small values of $\beta^$ a shock like feature develops in the model predictions, but as $\beta^$ increases to values of order $0.14$, the comparison with the data improves, as shown in figure \[fig:data\]c (red) where we illustrate the RMS error between the numerical data and the solutions of . There is a significant overlap in the best fit range of values of $\beta^$ for the cases of increasing and decreasing buoyancy flux, with the optimal value for both sets of data being $0.14$. For $\alpha$ in the range $0.08$–$0.10$ this value of $\beta^$ corresponds to values of $\beta$ in the range 0.70–0.88. ![\[fig:data\](a) The experimental plume data from @scd08 vs. solution of . The individual rescaled experimental concentration contours from each time step in the ensemble are in grey. A standard deviation either side of the mean is shown (black dashed) and the mean is the solid black line. The best theoretical fit, corresponding to $\beta^* = 0.14$ is shown in solid blue for $S_m = 1.08$ and dashed blue for $S_m = 1.00$. (b) The numerical plume data from @sac09 vs. solution of . The individual rescaled numerical concentration contours are in grey. A standard deviation either side of the mean is shown (black dashed) and the mean is the solid black line. The best theoretical fit, corresponding to $\beta^* = 0.14$ is shown in solid red for $S_m = 1.08$. (c) The RMS Error for varying $\beta^$ between the theoretical model and the experimental data (buoyancy turn-down, left-hand vertical axis) for $S_m = 1.08$ (solid blue), $S_m = 1.00$ (dashed blue) and between the theoretical model and the numerical data (buoyancy turn-up, right-hand vertical axis) in red. (d) The range over which the plume radius varies by a factor of $1 \pm 10^{-5}$ of its steady value $S_m = 1.08$ (blue solid) and $S_m = 1$ (blue dashed). The lower two curves are $\eta_1$ and the upper two curves are $\eta_2$. The solid data points, corresponding to $\beta^=0$ are given by the similarity solutions of @woodhouseEtAl (blue solid) and the solutions of @scase06 (blue dashed). As $\beta^$ approaches the experimentally determined value of approximately $0.14$ the effect of $S_m$ being greater than 1 becomes negligible.](scaseWoods_fig3-eps-converted-to.pdf) \[sec:conc\]Discussion and Conclusions ====================================== The quantitative description of the evolution of a turbulent buoyant plume with a variable source flux has attracted a considerable amount of attention, motivated by the modelling and experimental data presented by @scase06 [@scld07; @scd08; @sac09] in which the transition in the structure of a plume as the buoyancy flux is either increased or decreased was described. New analysis of this data, as presented herein, suggests that the transition in the plume is relatively smooth, and involves a region whose vertical extent grows in time at a rate proportional to $F_0^{1/4} t^{3/4}{\cal F}(F_0/F_1)$, where ${\cal F}(F_0/F_1)$ is a function of the initial to final buoyancy fluxes. Guided by this data, we propose a new model for the time dependent evolution of a turbulent buoyant plume, in which the mass, momentum and buoyancy are assumed to disperse along axis, with a turbulent dispersion coefficient $\beta U b$. We show the new model is consistent with the classical solutions for a steady turbulent buoyant plume, and that it admits self-similar solutions for the evolution of the transition zone in a plume if the flux is adjusted from one value to another. By comparison with the experimental data of @scd08 [@sac09] we predict that $\beta$ has a value in the range 0.70–0.88. These solutions represent the long-term asymptotic flow in the case that the flux steadily adjusts from one value to another over a finite time. We plan to explore the case in which the flow continually increases or decreases with time, in which case new experiments are required to describe the dynamics of both the associated starting plume [cf. @turner62] and the time-evolving flow at each point below the starting plume [cf. @deli79]. [20]{} <span style="font-variant:small-caps;">Canny, J.</span> 1986 A computational approach to edge-detection. IEEE Trans. Patt. An. Mach. Int.* 8 (6), 679–698. doi: 10.1109/TPAMI.1986.4767851. <span style="font-variant:small-caps;">Delichatsios, M. A.</span> 1979 Time similarity analysis of unsteady buoyant plumes in neutral surroundings. J. Fluid Mech. 93 (2), 241–250. <span style="font-variant:small-caps;">Driscoll, T. A., Hale, N. & Trefethen, L. N.</span> (eds.) 2014 Chebfun Guide. Pafnuty Publications. <span style="font-variant:small-caps;">Holland, P. R.</span> 2011 Oscillating dense plumes. J. Phys. Oceanog. 41, 1465–1483. <span style="font-variant:small-caps;">Holland, P. R., Hewitt, R. E. & Scase, M. M.</span> 2014 Wave breaking in dense plumes. J. Phys. Oceanog., 44, 790–800. <span style="font-variant:small-caps;">Hunt, G. R. & Kaye, N. B.</span> 2001 Virtual origin correction for lazy plumes. J. Fluid Mech. 435, 377–396. <span style="font-variant:small-caps;">Landel, J. R., Caulfield, C. P. & Woods, A. W.</span> 2012 Streamwise dispersion and mixing in quasi-two-dimensional steady turbulent jets. J. Fluid Mech. 711, 212–258. <span style="font-variant:small-caps;">Morton, B. R., Taylor, G. I. & Turner, J. S.</span> 1956 Turbulent gravitational convection from maintained and instantaneous sources. Proc. Roy. Soc. Lond. A 234, 1–32. <span style="font-variant:small-caps;">Rocco, S. & Woods A. W.</span> 2015 Dispersion in two-dimensional turbulent buoyant plumes. J. Fluid Mech. 774, R1. <span style="font-variant:small-caps;">Scase, M. M.</span> 2009 Evolution of volcanic eruption columns. J. Geophys. Res. 114 F04003. <span style="font-variant:small-caps;">Scase, M. M. & Hewitt R. E.</span> 2012 Unsteady turbulent plume models. J. Fluid Mech. 697, 455–480. <span style="font-variant:small-caps;">Scase, M. M., Caulfield, C. P., Dalziel, S. B. & Hunt, J. C. R.</span> 2006 Time-dependent plumes and jets with decreasing source strengths. J. Fluid Mech. 563, 443–461. <span style="font-variant:small-caps;">Scase, M. M., Caulfield, C. P., Linden, P. F. & Dalziel, S. B.</span> 2007 Local implications for self-similar turbulent plume models. J. Fluid Mech. 575, 257–265. <span style="font-variant:small-caps;">Scase, M. M., Caulfield, C. P. & Dalziel, S. B.</span> 2008 Temporal variation of non-ideal plumes with sudden reductions in buoyancy flux. J. Fluid Mech. 600, 181–199. <span style="font-variant:small-caps;">Scase, M. M., Aspden, A. J. & Caulfield, C. P.</span> 2009 The effect of sudden source buoyancy flux increases on turbulent plumes. J. Fluid Mech. 635, 137–169. <span style="font-variant:small-caps;">Turner, J. S.</span> 1962 The ‘starting plume’ in neutral surroundings. J. Fluid Mech. 13, 356–368. <span style="font-variant:small-caps;">Turner, J. S.</span> 1979 Buoyancy Effects in Fluids. Cambridge University Press. <span style="font-variant:small-caps;">Woodhouse, M. J, Phillips, .J. C. & Hogg, A. J.</span> 2016 Unsteady turbulent buoyant plumes. J. Fluid Mech. 794, 595–638. <span style="font-variant:small-caps;">Woods, A. W.</span> 1988 The fluid dynamics and thermodynamics of eruption columns. Bull. Volcanol. 50, 169–193. <span style="font-variant:small-caps;">Woods, A. W.</span> 2010 Turbulent plumes in nature. Ann. Rev. Fluid Mech. 42 (1), 391–412.	{ "pile_set_name": "ArXiv" }
--- author: - \| \ CERN, PH Department, 1211 Geneva, Switzerland\ E-mail: title: Perturbative probes of QCD matter at the LHC --- Introduction ============ Nucleus-nucleus collisions (A-A) at ultrarelativistic energies provide the experimental means to study the (thermo)dynamics of quarks and gluons under extreme conditions of temperature and density. Head-on collisions of Pb ions at LHC energies produce very hot and dense matter by concentrating a substantial amount of energy $\mathscr{O}$(2 TeV) in an extended volume $\mathscr{O}$(150 fm$^3$) at thermalisation times of $\tau_0$ = 1 fm/c [@Chatrchyan:2012mb]. Such energy densities are more than one order of magnitude above the critical value, ${\varepsilon_{\mbox{\tiny{\rm crit}}}}\approx$ 1 GeV/fm$^3$, predicted by lattice quantum chromodynamics (QCD) calculations [@Karsch:2001cy] for the formation of a deconfined system of bare-mass quarks and gluons (Quark Gluon Plasma, QGP) [@shuryak77]. ![Left: Examples of the sensitivity of various perturbative probes to quark-gluon matter properties in A-A collisions [@d'Enterria:2006su]. Right: “Jet quenching” event displays in central Pb-Pb collisions at the LHC: monojet-like event [@Aad:2010bu] (top) and $\gamma$-jet event with a recoiling jet with reduced energy [@Chatrchyan:2012gt] (bottom).[]{data-label="fig:qgp_probes"}](atlas_pbpb_dijet_evtdisplay.eps "fig:"){width="0.90\columnwidth" height="4.cm"} ![Left: Examples of the sensitivity of various perturbative probes to quark-gluon matter properties in A-A collisions [@d'Enterria:2006su]. Right: “Jet quenching” event displays in central Pb-Pb collisions at the LHC: monojet-like event [@Aad:2010bu] (top) and $\gamma$-jet event with a recoiling jet with reduced energy [@Chatrchyan:2012gt] (bottom).[]{data-label="fig:qgp_probes"}](cms_pbpb_gammajet_evtdisplay.eps "fig:"){width="0.88\columnwidth" height="4.2cm"} Among all experimental observables, particles with large transverse momentum $p_T$ and/or high mass $m$ (“hard probes”) are useful tomographic tools of the produced medium (Fig. \[fig:qgp\_probes\] left) as: (i) they originate from partonic scatterings with large momentum transfer $Q^2$ and couple directly to the fundamental quark and gluon degrees of freedom; (ii) their production time-scale is very short, $\tau\approx 1/(p_T,m)\lesssim$ 0.1 fm/c, allowing them to propagate through (and be potentially affected by) the medium, and (iii) their cross sections can be theoretically computed in perturbative QCD calculations as a convolution of parton distribution (PDFs, $f_{a/A}$) and fragmentation (FFs, $D_{c\rightarrow h}$) functions times the subprocess parton-parton scattering cross section: $d\sigma^{hard}_{AB\rightarrow h}= f_{a/A}(x,Q^2)\otimes f_{b/B}(x,Q^2)\otimes d\sigma_{ab\rightarrow c}^{hard} \otimes D_{c\rightarrow h}(z,Q^2)$. In A-A collisions in the absence of final-state effects, the parton flux in a nucleus $A$ is the same as that of a superposition of $A$ independent nucleons, $f_{a/A}\approx A\cdot f_{a/N}$, and thus $d\sigma^{hard}_{AA\rightarrow h} \approx A^2 \cdot f_{a/N}(x,Q^2)\otimes \;f_{a/N}(x,Q^2)\otimes d\sigma_{ab\rightarrow c}^{hard}\otimes D_{c\rightarrow h}(z,Q^2)$. The standard method to quantify the effects of the medium on a given hard probe is via the ratio of A-A yields over p-p cross sections (scaled by the nuclear overlap function $T_{AA}(b)$ at impact parameter $b$): $R_{AA}(p_{T},y;b)\,=\,\frac{d^2N_{AA}/dy dp_{T}}{\langle T_{AA}(b)\rangle\,\times\, d^2 \sigma_{pp}/dy dp_{T}},$ which measures the deviation of A-A at $b$ from an incoherent superposition of nucleon-nucleon collisions. The study of the suppression (or enhancement) factors as a function of various variables provides information on the medium transport coefficients, its temperature, energy density, etc. (Fig. \[fig:qgp\_probes\], left). Unsuppressed electro-weak particle production ============================================= Weakly-interacting particles such as isolated photons, W, Z, and Drell-Yan pairs – whose rates are unaffected by final-state interactions in the produced medium – are valuable benchmark processes in A-A collisions at collider energies. On the one hand, they allow one to experimentally confirm the validity of the perturbative ($A^2$) scaling of the p-p cross sections and, on the other, they provide constraints on the nuclear PDFs which, in particular for the Pb case, are barely known from deep-inelastic e-A data [@Eskola:2009uj]. At the LHC, prompt-$\gamma$ have been measured above 20 GeV/c applying isolation cuts on background-subtracted Pb-Pb collisions [@Chatrchyan:2012vq], Z bosons have been measured in the dimuon [@Aad:2010aa; @Chatrchyan:2011ua] and dielectron [@atlas:Z12] decay modes, and W bosons in events with a high-$p_T$ muon and large missing transverse momentum from the undetected $\nu$ [@Chatrchyan:2012nt] (Fig. \[fig:WZ\]). ![Measured mass distributions for Z bosons ($\mu^+\mu^-$ and $e^+e^-$ invariant masses, left) [@atlas:Z12] and W (transverse mass, right) [@Chatrchyan:2012nt] in Pb-Pb (and p-p) collisions at 2.76 TeV.[]{data-label="fig:WZ"}](cms_pbpb_W_minv.eps){width="0.48\columnwidth" height="5.cm"} In all three cases, the measured yields for all Pb-Pb centralities are consistent with the corresponding p-p cross sections scaled by the nuclear overlap function ($R_{AA}\approx$ 1, Fig. \[fig:worldRAA\] left). Comparisons to NLO pQCD calculations indicate small nuclear PDF modifications as expected in the ranges of parton fractional momentum $x$ and energy scale $Q^2$ probed [@Eskola:2009uj]. Interestingly, inclusive W production shows $R_{AA}$(W) = 1.04 $\pm$ 0.07 $\pm$ 0.12 while individual W$^\pm$ yields show differences due to isospin effects (unequal $u$ and $d$-quark content in Pb and p): $R_{AA}(\bar u d \to $W$^-)$ = 1.46 and $R_{AA}(u\bar d \to $W$^+)$ = 0.8 [@Chatrchyan:2012nt]. Future analyses of larger $\gamma,\gamma^\star$, W, Z data samples with reduced uncertainties, will provide enhanced constraints on the parton densities in nuclei. Suppression of high-$p_T$ hadrons ================================= One of the first proposed smoking guns of QGP formation was “jet quenching” [@Bjorken:1982tu] i.e. the attenuation or disappearance of the spray of hadrons resulting from the fragmentation of a parton having lost energy in the dense plasma produced in the reaction. Among the most exciting results from RHIC is the observation of large high-$p_T$ hadron suppression ($R_{AA}\approx$ 0.2) in central Au-Au compared to p-p or d-Au collisions, in agreement with jet quenching expectations [@Adcox:2004mh; @Adams:2005dq]. At the LHC, charged hadrons appear to be further suppressed, by up to factor of 6 at $p_T$ = 7 GeV/c [@Aamodt:2010jd; @Appelshauser:2011ds] but the amount of suppression is slowly reduced with increasing $p_T$ and plateaus at $R_{AA}\approx$ 0.5 at $p_T\approx$ 40$-$100 GeV/c [@CMS:2012aa] (Fig. \[fig:worldRAA\], left). ![Left: Compilation of measured $R_{AA}(p_{T})$ in central Pb-Pb collisions at the LHC for inclusive charged hadrons [@Aamodt:2010jd; @CMS:2012aa], $\pi^\pm$ [@Appelshauser:2011ds], D [@ALICE:2012ab] and B [@Chatrchyan:2012np] mesons, isolated-$\gamma$ [@Chatrchyan:2012vq], and W [@Chatrchyan:2012nt] and Z [@Chatrchyan:2011ua] bosons. Right: Jet spectra (anti-$k_T$ algorithm, $R$ = 0.2) measured in various Pb-Pb centralities scaled by their corresponding $T_{AA}$ [@atlas:jets12]. The bands around all data points indicate the associated systematic uncertainties.[]{data-label="fig:worldRAA"}](RAA_systematics_LHC.eps "fig:"){width="0.60\columnwidth" height="6.2cm"} ![Left: Compilation of measured $R_{AA}(p_{T})$ in central Pb-Pb collisions at the LHC for inclusive charged hadrons [@Aamodt:2010jd; @CMS:2012aa], $\pi^\pm$ [@Appelshauser:2011ds], D [@ALICE:2012ab] and B [@Chatrchyan:2012np] mesons, isolated-$\gamma$ [@Chatrchyan:2012vq], and W [@Chatrchyan:2012nt] and Z [@Chatrchyan:2011ua] bosons. Right: Jet spectra (anti-$k_T$ algorithm, $R$ = 0.2) measured in various Pb-Pb centralities scaled by their corresponding $T_{AA}$ [@atlas:jets12]. The bands around all data points indicate the associated systematic uncertainties.[]{data-label="fig:worldRAA"}](atlas_pbpb_jet_spectra.eps "fig:"){width="0.40\columnwidth" height="6.2cm"} The observed $p_T$ dependence provides strong discrimination power for parton radiative energy loss models. Once the underlying energy loss mechanism responsible for the observed high-$p_T$ hadron deficit is confirmed [@Milhano:2012gf] the properties of the medium (e.g. its ${\hat{q}}$ transport coefficient) can be properly derived taking into account the expansion of the produced matter in 3D-hydrodynamics calculations [@Armesto:2009zi]. Jet quenching in dijet and $\gamma$-jet events ============================================== High-$p_T$ jets have been fully reconstructed in heavy-ion collisions for the first time at the LHC using modern algorithms for jet finding and background subtraction [@Cacciari:2010te]. The strongest evidence so far for jet quenching is the observation of a large dijet momentum imbalance for increasingly central Pb-Pb collisions [@Aad:2010bu; @Chatrchyan:2011sx] up to very high momenta ($p_T\approx$ 350 GeV/c) [@Chatrchyan:2012nia]. In some cases, the quenched jet is not even visible above the underlying event background (Fig. \[fig:qgp\_probes\], right). As observed for the inclusive (leading) hadrons above 40 GeV/c, the jet spectra in central Pb-Pb are suppressed by a factor of two compared to peripheral Pb-Pb collisions (Fig. \[fig:worldRAA\], right) [@atlas:jets12]. The lost energy of the jet appears to be emitted in the form of soft particles ($p_T~<$ 4 GeV/c) outside of jet cone ($R >$ 0.8) [@Chatrchyan:2011sx]. Interestingly, even if the quenched jet loses a large fraction of its energy, the dijet event keeps its back-to-back topology without significant azimuthal decorrelation [@Aad:2010bu; @Chatrchyan:2011sx]. Fragmentation functions have been constructed using tracks with $p_T~>$ 2(4) GeV/c within $R <$ 0.2(0.3) of the jet-axis combined with the reconstructed (but quenched) jet energy [@atlas:jets12; @Chatrchyan:2012gw]. The resulting FFs, shown in Fig. \[fig:jets\] as a function of the longitudinal fraction of the jet momentum carried by the charged particles $z = p_{\|\|}^{\rm track}/p^{\rm jet}$ (left) and of $\xi = \ln(1/z)$ (center), are similar in Pb-Pb (central and peripheral) and p-p collisions indicating that after energy loss the quenched jet fragments in the vacuum with the same pattern as “normal” jets. ![Jet fragmentation functions reconstructed in Pb-Pb at 2.76 TeV as a function of the scaling variables $z$ (left panel, for central and peripheral collisions) [@atlas:jets12] and $\xi$ (mid panel, for central collisions) [@Chatrchyan:2012gw]. Right: Ratio of jet to photon momenta in central Pb-Pb collisions compared to [pythia]{} p-p predictions [@Chatrchyan:2012gt].[]{data-label="fig:jets"}](cms_pbpb_FF.eps "fig:"){width="5.4cm" height="4.9cm"}![Jet fragmentation functions reconstructed in Pb-Pb at 2.76 TeV as a function of the scaling variables $z$ (left panel, for central and peripheral collisions) [@atlas:jets12] and $\xi$ (mid panel, for central collisions) [@Chatrchyan:2012gw]. Right: Ratio of jet to photon momenta in central Pb-Pb collisions compared to [pythia]{} p-p predictions [@Chatrchyan:2012gt].[]{data-label="fig:jets"}](cms_pbpb_xJgamma.eps "fig:"){width="5.cm" height="4.5cm"} Jet quenching has also been studied in “cleaner” photon-jet events with $p_T^\gamma >$ 60 GeV/c and $p_T^{\rm jet} >$ 30 GeV/c [@Chatrchyan:2012gt]. The balance ratio $x_{\rm J\gamma} = p_T^{\rm jet}/p_T^\gamma$ is seen to decrease significantly with increasing centrality (see Fig. \[fig:jets\] right, for central Pb-Pb) as does the fraction of $\gamma$-jet associations found in the data, compared to p-p and [pythia]{} [@Sjostrand:2006za] results. Yet, the $\gamma$-jet azimuthal distribution measured in Pb-Pb events with large $p_T$ imbalance is consistent with that measured in p-p collisions and simulated with [pythia]{}. The fact that the quenched jet is back-to-back to the photon confirms that the energy is not lost in single hard gluon radiation. Heavy-flavour suppression ========================= In QCD, gluon-strahlung from heavy quarks is expected to be less important than for light quarks due to the “dead-cone” effect [@Dokshitzer:2001zm]. Thus, a robust prediction of jet quenching models is the hierarchy $\Delta E_{Q} < \Delta E_{q} < \Delta E_{g}$ accounting for mass and colour-charge dependencies of radiative energy loss. Surprisingly, RHIC measurements of high-$p_T$ electrons from semileptonic $D$ and $B$ decays seem to indicate a charm+bottom suppression comparable to that of light quarks: $R_{AA}(c,b)\approx R_{AA}(q,g)\approx$ 0.2 [@Adare:2010de; @Abelev:2006db]. At the LHC, D-mesons have been for the first time directly measured in Pb-Pb events with displaced-vertices in three hadronic decay channels ($D^0\to K\pi$, $D^\pm \to K\pi\pi$, $D^\star \to D^0\pi$ for $p_T >$ 2, 4 and 6 GeV/c respectively) [@ALICE:2012ab], and B-meson production has been observed via secondary ${J/\psi}$ for which a clean separation of the displaced vertex is also possible for $p_T >$ 6.5 GeV/c [@Chatrchyan:2012np]. Within uncertainties, the measured suppression factors, $R_{AA}$(D) $\approx$ 0.3 and $R_{AA}$(B) $\approx$ 0.4 are larger than $R_{AA}$(h$^\pm$,$\pi^\pm$) $\approx$ 0.15–0.2 (Fig. \[fig:worldRAA\], left) and support the expected radiative energy loss hierarchy. Coming measurements with larger data samples and reduced uncertainties will allow one to better understand the behaviour of heavy-quarks in the QGP, including the relative contribution of radiative/elastic losses, and determine e.g. their drag coefficients $\eta_{D}$ in the medium [@Gossiaux:2012th]. Quarkonia dissociation ====================== The study of heavy-quark bound states in high-energy A-A collisions has a long story as a sensitive probe of the thermodynamical properties of deconfined QCD matter [@Matsui:1986dk]. Analysis of quarkonia correlators and potentials in finite-$T$ lattice QCD [@Datta:2003ww] indicate that the different $c$-$\bar c$ and $b$-$\bar b$ states dissociate at temperatures for which the colour (Debye) screening radius of the medium falls below their corresponding $Q\bar{Q}$ binding radius. The surprisingly similar amount of ${J/\psi}$ suppression observed at SPS [@Alessandro:2004ap] and RHIC [@Adare:2006ns; @Adare:2011yf] energies (with expected larger QGP temperature in the latter) has been interpreted as due to a partial compensation at RHIC of the reduced ${J/\psi}$ yields by $c\bar{c}$ recombination in the medium [@BraunMunzinger:2000px; @Thews:2000rj]. In this scenario, the large charm pair production expected at the LHC would further enhance charmonium regeneration [@Andronic:2007bi]. LHC data on ${J/\psi}$ production [@Aad:2010aa; @Chatrchyan:2012np; @Abelev:2012rv] show a different suppression pattern than observed at RHIC (Fig. \[fig:jpsi\_ups\], left). At low $p_T$, ${J/\psi}$ are less suppressed ($R_{AA}\approx$ 0.5) than measured at lower c.m. energies ($R_{AA}\approx$ 0.2–0.3), whereas at high-$p_T$ they are more suppressed ($R_{AA}\approx$ 0.2) than at RHIC ($R_{AA}\approx$ 0.6). Approaches based on deconfinement followed by charm recombination can reproduce the observed trends in the data although the model parameters ($\sigma_{c\bar c}$, density, ...) need to be validated with other LHC observations. ![Left: Inclusive ${J/\psi}$ $R_{AA}$ as a function of the central charged-particle density in Pb-Pb at 2.76 TeV compared to PHENIX results in Au-Au at 200 GeV at central and forward rapidities [@Abelev:2012rv]. Right: $\Upsilon$(nS) states measured in Pb-Pb (data points) compared to the fitted p-p distribution (dashed red line) at 2.76 TeV [@cms:Ups12].[]{data-label="fig:jpsi_ups"}](alice_pbpb_jpsi_vs_dNchdeta.eps "fig:"){width="0.55\columnwidth" height="6.2cm"} ![Left: Inclusive ${J/\psi}$ $R_{AA}$ as a function of the central charged-particle density in Pb-Pb at 2.76 TeV compared to PHENIX results in Au-Au at 200 GeV at central and forward rapidities [@Abelev:2012rv]. Right: $\Upsilon$(nS) states measured in Pb-Pb (data points) compared to the fitted p-p distribution (dashed red line) at 2.76 TeV [@cms:Ups12].[]{data-label="fig:jpsi_ups"}](cms_upsilon_pbpb_pp_2012.eps "fig:"){width="0.45\columnwidth" height="6.1cm"} The abundant production of the ${\Upsilon}(1S,2S,3S)$ states at the LHC opens up a unique opportunity to study also the dissociation of the $b$-quark bound states. The ${\Upsilon}$ is expected to survive up to $4\,{T_{\mbox{\tiny{\rm crit}}}}$ whereas the less tightly-bound ${\Upsilon^{'}}$ and ${\Upsilon^{''}}$ resonances should “melt” at lower temperatures. The data confirms that the ${\Upsilon}$ states are suppressed in central Pb-Pb relative to p-p collisions: $R_{AA}({\Upsilon}) \approx$ 0.56, $R_{AA}({\Upsilon^{'}}) \approx$ 0.12 and $R_{AA}({\Upsilon^{''}}) <$ 0.1 (Fig. \[fig:jpsi\_ups\], right) [@Chatrchyan:2011pe; @cms:Ups12]. The deficit in the ${\Upsilon}$ production yields coincides with the expected fraction from the feed-down contributions of the heavily-suppressed ${\Upsilon}(2S,3S)$ excited states. The overall scenario is consistent with sequential suppression of the bottomonium family. Summary ======= The main findings in the perturbative sector of Pb-Pb collisions at ${\sqrt{s_{_{\ensuremath{\it{NN}}}}}}$ = 2.76 TeV after 2 years of operation at the LHC can be summarised as follows: - Electroweak probes ($\gamma$, W, Z) unaffected by final-state interactions can provide, with larger datasets and reduced uncertainties, valuable constraints of the nuclear PDFs [@Chatrchyan:2012vq; @Aad:2010aa; @Chatrchyan:2011ua; @atlas:Z12; @Chatrchyan:2012nt]. - High-$p_T$ (leading) hadron production is suppressed compared to p-p collisions by factors ranging from 6 (at 7 GeV/c) to 2 (above 40 GeV/c), indicating the formation of a dense opaque medium that suppresses the energy of the parent fragmenting partons [@Aamodt:2010jd; @Appelshauser:2011ds; @CMS:2012aa]. - Fully reconstructed jets are quenched in dijet and $\gamma$-jet events and show (i) a large $p_T$ imbalance, (ii) enhanced soft off-cone radiation, (iii) preserved back-to-back azimuthal correlations, and (iv) vacuum-like fragmentation functions (reconstructed using the quenched jet energy) [@Aad:2010bu; @Chatrchyan:2012gt; @atlas:jets12; @Chatrchyan:2011sx; @Chatrchyan:2012nia; @Chatrchyan:2012gw], which suggest that parton energy loss occurs via soft multi-gluon emission followed by vacuum fragmentation. - D and B mesons are suppressed by factors of $\sim$4 and $\sim$3 respectively relative to p-p collisions [@ALICE:2012ab; @Chatrchyan:2012np], in agreement with the expected mass and colour-charge dependencies of radiative energy loss. - The deficit of ${J/\psi}$ yields at the LHC is smaller than at lower energies. The suppression is weaker (stronger) at low (high) $p_T$ [@Aad:2010aa; @Chatrchyan:2012np; @Abelev:2012rv] and is suggestive of model predictions based on colour screening followed by $c\bar c$ recombination. - The ${\Upsilon^{'}}$ and ${\Upsilon^{''}}$ states are strongly suppressed, whereas the ${\Upsilon}$ ground-state yield is depleted by about 40% (the same fraction expected from feed-down contribution of the two excited states) relative to p-p collisions [@Chatrchyan:2011pe; @cms:Ups12], in agreement with sequential dissociation scenarios. Figure \[fig:worldRAA\] (left) compiles the $R_{AA}(p_T)$ factors for most of the high-$p_T$ and large-mass particles measured in central Pb-Pb collisions at the LHC. The mechanisms of initial production of all particles are well understood within pQCD. The final reduced yields observed for all particles containing colour degrees of freedom clearly point to strong final-state effects (radiative energy loss, colour screening, ...). Final extraction of the thermodynamical and transport properties of the hot and dense matter produced at the LHC will require detailed calculations including mechanisms of particle production+“destruction” coupled with a (3D viscous) hydrodynamics description of the expansion of the system.\ I am grateful to Bernard Pire and Michel Garçon for their invitation to this well-organised and interesting multidisciplinary conference. [99]{} S. Chatrchyan [et al.]{} \[CMS Collaboration\], arXiv:1205.2488 \[nucl-ex\] F. Karsch, Lect. Notes Phys. [583]{} (2002) 209 E. V. Shuryak, Sov. Phys. JETP [47]{} (1978) 212 \[Zh. Eksp. Teor. Fiz. [74]{} (1978) 408\] D. d’Enterria, J. Phys. G [34]{} (2007) S53 G. Aad [et al.]{} \[Atlas Collaboration\], Phys. Rev. Lett. [105]{} (2010) 252303 S. Chatrchyan [et al.]{} \[CMS Collaboration\], arXiv:1205.0206 \[nucl-ex\] K. J. Eskola, H. Paukkunen and C. A. Salgado, JHEP [0904]{} (2009) 065 S. Chatrchyan [et al.]{} \[CMS Collaboration\], Phys. Lett. B [710]{} (2012) 256 G. Aad [et al.]{} \[Atlas Collaboration\], Phys. Lett. B [697]{} (2011) 294 S. Chatrchyan [et al.]{} \[CMS Collaboration\], Phys. Rev. Lett. [106]{} (2011) 212301 Atlas Collaboration, ATLAS-CONF-2011-052. S. Chatrchyan [et al.]{} \[CMS Collaboration\], arXiv:1205.6334 \[nucl-ex\] J. D. Bjorken, FERMILAB-PUB-82-059-THY (1982) K. Adcox [et al.]{} \[PHENIX Collaboration\], Nucl. Phys. A [757]{} (2005) 184 J. Adams [et al.]{} \[STAR Collaboration\], Nucl. Phys. A [757]{} (2005) 102 K. Aamodt [et al.]{} \[ALICE Collaboration\], Phys. Lett. B [696]{} (2011) 30 H. Appelshauser \[ALICE Collaboration\], J. Phys. G [38]{} (2011) 124014 S. Chatrchyan [et al.]{} \[CMS Collaboration\], Eur. Phys. J. C [72]{} (2012) 1945 B. Abelev \[ALICE Collaboration\], arXiv:1203.2160 \[nucl-ex\] S. Chatrchyan [et al.]{} \[CMS Collaboration\], JHEP [1205]{} (2012) 063 Atlas Collaboration, ATLAS-CONF-2011-075 J. G. Milhano, EPJ Web Conf. [28]{} (2012) 03001 N. Armesto, M. Cacciari, T. Hirano, J. L. Nagle and C. A. Salgado, J. Phys. G [37]{} (2010) 025104 M. Cacciari, J. Rojo, G. P. Salam and G. Soyez, Eur. Phys. J. C [71]{} (2011) 1539 S. Chatrchyan [et al.]{} \[CMS Collaboration\], Phys. Rev. C [84]{} (2011) 024906 S. Chatrchyan [et al.]{} \[CMS Collaboration\], Phys. Lett. B [712]{} (2012) 176 S. Chatrchyan [et al.]{} \[CMS Collaboration\], arXiv:1205.5872 \[nucl-ex\] T. Sjostrand, S. Mrenna and P. Z. Skands, JHEP [0605]{} (2006) 026 Y. L. Dokshitzer and D. E. Kharzeev, Phys. Lett. B [519]{} (2001) 199 A. Adare [et al.]{} \[PHENIX Collaboration\], Phys. Rev. C [84]{} (2011) 044905 B. I. Abelev [et al.]{} \[STAR Collaboration\], Phys. Rev. Lett. [106]{} (2011) 159902 P. Gossiaux, J. Aichelin and T. Gousset, Prog. Theor. Phys. Suppl. [193]{} (2012) 110 T. Matsui and H. Satz, Phys. Lett. B [178]{} (1986) 416 S. Datta, F. Karsch, P. Petreczky and I. Wetzorke, Phys. Rev. D [69]{} (2004) 094507 B. Alessandro [et al.]{} \[NA50 Collaboration\], Eur. Phys. J. C [39]{} (2005) 335 A. Adare [et al.]{} \[PHENIX Collaboration\], Phys. Rev. Lett. [98]{} (2007) 232301 A. Adare [et al.]{} \[PHENIX Collaboration\], Phys. Rev. C [84]{} (2011) 054912 P. Braun-Munzinger and J. Stachel, Phys. Lett. B [490]{} (2000) 196 R. L. Thews, M. Schroedter and J. Rafelski, Phys. Rev. C [63]{} (2001) 054905 A. Andronic, P. Braun-Munzinger, K. Redlich and J. Stachel, Phys. Lett. B [652]{} (2007) 259 B. Abelev [et al.]{} \[ALICE Collaboration\], arXiv:1202.1383 \[hep-ex\] T. Dahms \[CMS Collaboration\], Proceeds. Hard Probes’12, CMS-PAS-HIN-11-011 S. Chatrchyan [et al.]{} \[CMS Collaboration\], Phys. Rev. Lett. [107]{} (2011) 052302	{ "pile_set_name": "ArXiv" }
--- abstract: 'Initial-boundary value problems in a bounded rectangle with different types of boundary conditions for two-dimensional Zakharov–Kuznetsov equation are considered. Results on global well-posedness in the classes of weak and regular solution are established. As applications of the developed technique results on boundary controllability and long-time decay of weak solutions are also obtained.' address: 'RUDN University, 6 Miklukho–Maklaya Street, Moscow, 117198, Russia' author: - 'Andrei V. Faminskii' title: 'Initial-Boundary Value Problems in a Rectangle for Two-Dimensional Zakharov–Kuznetsov Equation' --- [^1] Introduction. Description of main results {#S1} ========================================= The two dimensional Zakharov–Kuznetsov equation (ZK) $$\label{1.1} u_t+bu_x+u_{xxx}+u_{xyy}+uu_x=f(t,x,y)$$ ($b$ is a real constant) is one of the variants of multi-dimensional generalizations of Korteweg–de Vries equation (KdV) $u_t+bu_x+u_{xxx}+uu_x=f(t,x)$. For the first time it was derived in the three-dimensional case in [@ZK] for description of ion-acoustic waves in magnetized plasma. The equation, considered is the present paper, is known as a model of two-dimensional nonlinear waves in dispersive media propagating in one preassigned ($x$) direction with deformations in the transverse ($y$) direction. A rigorous derivation of the ZK model can be found, for example, in [@H-K; @LLS]. From the point of view of solubility and well-posedness the most significant results for ZK equation and its generalizations were obtained for the initial value problem. In the two-dimensional case the corresponding results in different functional spaces can be found in [@S; @F89; @F95; @BL; @LP09; @LP11; @RV; @FLP; @BJM; @GH; @MP; @FP; @G]. For initial-boundary value problems such a theory is most developed for domains, where the variable $y$ is considered in the whole line, ([@F02; @F07-1; @FB08; @F08; @ST; @F12; @DL]). Initial-boundary value problems posed on domains, where the variable $y$ is considered on a bounded interval, are studied less, although from the physical point of view they seem at least the same important. Certain technique developed for the case $y\in\mathbb R$ (especially related to the investigation of the corresponding linear equation) up to this moment is extended to the case of bounded $y$ only partially. An initial-boundary value problem in a strip $\mathbb R\times (0,L)$ with periodic boundary conditions was considered in [@LPS] for ZK equation and local well-posedness result was established in the spaces $H^s$ for $s>3/2$. This result was improved in [@MP] where $s\geq 1$, in addition, in the space $H^1$ appropriate conservation laws provided global well-posedness. Initial-boundary value problems in such a strip with homogeneous boundary conditions of different types – Dirichlet, Neumann or periodic – were considered in [@BF13; @F15-2] and results on global well-posedness in classes of weak solutions with power and exponential weights at $+\infty$ were established. Global well-posedness results for ZK equation with certain parabolic regularization also for the initial-boundary value problem in a strip $\mathbb R\times (0,L)$ with homogeneous Dirichlet boundary conditions can be found in [@F15-1; @F15-2; @L15; @L16]. Similar results on global well-posedness in weighted spaces for initial-boundary value problems in a half-strip $\mathbb R_+\times (0,L)$ were obtained in [@LT; @L13; @F17]. Initial-boundary value problems in a bounded rectangle were studied in [@STW; @DL]. In [@STW] either homogeneous Dirichlet or periodic boundary conditions with respect to $y$ were considered and results on global existence and uniqueness of weak solutions were established. In [@DL] similar results in more regular classes for homogeneous Dirichlet boundary conditions were obtained. In both papers boundary conditions with respect to $x$ were homogeneous. In the present paper we consider initial-boundary value problems in a domain $Q_T=(0,T)\times\Omega$, where $\Omega=(0,R)\times (0,L)=\{(x,y): 0<x<R, 0<y<L\}$ is a bounded rectangle of given length $R$ and width $L$, $T>0$ is arbitrary, for equation with an initial condition $$\label{1.2} u(0,x,y)=u_0(x,y),\qquad (x,y)\in\Omega,$$ boundary conditions for $(t,y)\in B_{T}=(0,T)\times (0,L)$ $$\label{1.3} u(t,0,y)=\mu_0(t,y), \quad u(t,R,y)=\nu_0(t,y),\quad u_x(t,R,y)=\nu_1(t,y)$$ and boundary conditions for $(t,x)\in (0,T)\times (0,R)$ of one of the following four types: $$\label{1.4} \begin{split} \mbox{whether}\qquad &a)\mbox{ } u(t,x,0)=u(t,x,L)=0,\\ \mbox{or}\qquad &b)\mbox{ } u_y(t,x,0)=u_y(t,x,L)=0,\\ \mbox{or}\qquad &c)\mbox{ } u(t,x,0)=u_y(t,x,L)=0,\\ \mbox{or}\qquad &d)\mbox{ } u \mbox{ is an $L$-periodic function with respect to $y$.} \end{split}$$ We use the notation “problem –” for each of these four cases. The main results consist of theorems on global well-posedness in classes of weak and regular solutions. Besides that, certain results on large-time decay of small solutions and boundary controllability, when $\mu_0=\nu_0\equiv 0$, $f\equiv 0$, are established. In what follows (unless stated otherwise) $j$, $k$, $l$, $m$, $n$ mean non-negative integers, $p\in [1,+\infty]$, $s\in\mathbb R$. Let $[s]$ be the integer part of $s$ ($s-[s]\in [0,1)$). For any multi-index $\alpha=(\alpha_1,\alpha_2)$ let $\partial^\alpha =\partial^{\alpha_1}_{x}\partial^{\alpha_2}_{y}$, let $$\|D^k\varphi\|=\Bigl(\sum_{\|\alpha\|\leq k}(\partial^\alpha \varphi)^2\Bigr)^{1/2}, \qquad \|D\varphi\|=\|D^1\varphi\|.$$ Let $L_p=L_p(\Omega)$, $W_p^k=W_p^k(\Omega)$, $H^s=H^s(\Omega)$. Introduce special function spaces taking into account boundary conditions . Let $\Sigma= \mathbb R\times (0,L)$, $\widetilde{\EuScript S}(\overline{\Sigma})$ be a space of infinitely smooth on $\overline{\Sigma}$ functions $\varphi(x,y)$ such that $\displaystyle{(1+\|x\|)^n\|\partial^\alpha\varphi(x,y)\|\leq c(n,\alpha)}$ for any $n$, multi-index $\alpha$, $(x,y)\in \overline{\Sigma}$ and $\partial_y^{2m}\varphi\big\|_{y=0} =\partial_y^{2m}\varphi\big\|_{y=L}=0$ in the case a), $\partial_y^{2m+1}\varphi\big\|_{y=0} =\partial_y^{2m+1}\varphi\big\|_{y=L}=0$ in the case b), $\partial_y^{2m}\varphi\big\|_{y=0} =\partial_y^{2m+1}\varphi\big\|_{y=L}=0$ in the case c), $\partial_y^{m}\varphi\big\|_{y=0} =\partial_y^{m}\varphi\big\|_{y=L}$ in the case d) for any $m$. Let $\widetilde H^s(\Sigma)$ be the closure of $\widetilde{\EuScript S}(\overline{\Sigma})$ in the norm $H^s(\Sigma)$ and $\widetilde H^s(I\times (0,L))$ be the restriction of $\widetilde H^s(\Sigma)$ on $I\times (0,L)$ for any interval $I\subset \mathbb R$, $\widetilde H^s =\widetilde H^s(\Omega)$. It is easy to see, that $\widetilde H^0=L_2$; $\widetilde H^s=H^s$ if $s<0$; for $j \geq 1$ in the case a) $\widetilde H^j=\{\varphi\in H^j: \partial_y^{2m}\varphi\|_{y=0}=\partial_y^{2m}\varphi\|_{y=L}=0, \ 2m<j\}$, in the case b) $\widetilde H^j=\{\varphi\in H^j: \partial_y^{2m+1}\varphi\|_{y=0}=\partial_y^{2m+1}\varphi\|_{y=L}=0,\ 2m+1<j\}$, in the case d) $\widetilde H^j=\{\varphi\in H^j: \partial_y^{m}\varphi\|_{y=0}=\partial_y^{m}\varphi\|_{y=L}, \ m<j\}$. We also use an anisotropic Sobolev space $\widetilde H^{(0,k)}$ which is defined as the restriction on $\Omega$ of a space $\widetilde H^{(0,k)}(\Sigma)$, where the last space is the closure of $\widetilde{\EuScript S}(\overline{\Sigma})$ in the norm $\sum\limits_{m=0}^k \\|\partial_y^m\varphi\\|_{L_2(\Sigma)}$. We construct solutions to the considered problems in spaces $X^k(Q_T)$ for $k=0$ and $k=3$, consisting of functions $u(t,x,y)$, such that $$\label{1.5} \partial_t^j u\in C([0,T]; \widetilde H^{k-3j})\cap L_2(0,T;\widetilde H^{k-3j+1})$$ if $k-3j\geq 0$, let $X(Q_T)=X^{0}(Q_T)$. For description of properties of the boundary data introduce anisotropic functional spaces. Let $B=\mathbb R^t \times (0,L)$. Define the functional space $\widetilde{\EuScript S}(\overline{B})$ similarly to $\widetilde{\EuScript S}(\overline{\Sigma})$, where the variable $x$ is substituted by $t$. Let $\widetilde H^{s/3,s}(B)$ be the closure of $\widetilde{\EuScript S}(\overline{B})$ in the norm $H^{s/3,s}(B)$. More exactly, let $\psi_l(y)$, $l=1,2\dots$, be the orthonormal in $L_2(0,L)$ system of the eigenfunctions for the operator $(-\psi'')$ on the segment $[0,L]$ with corresponding boundary conditions $\psi(0)=\psi(L)=0$ in the case a), $\psi'(0)=\psi'(L)=0$ in the case b), $\psi(0)=\psi'(L)=0$ in the case c), $\psi(0)=\psi(L),\psi'(0)=\psi'(L)$ in the case d), $\lambda_l$ be the corresponding eigenvalues. Such systems are well-known and are written in trigonometric functions. For any $\mu\in \widetilde{\EuScript S}(\overline{B})$, $\theta\in\mathbb R$ and $l$ let $$\label{1.6} \widehat\mu(\theta,l) \equiv \iint_B e^{-i\theta t}\psi_l(y)\mu(t,y)\,dtdy.$$ Then the norm in $H^{s/3,s}(B)$ is defined as $\displaystyle\Bigl(\sum\limits_{l=1}^{+\infty} \bigl\\| (\|\theta\|^{2/3}+l^2)^{s/2}\widehat\mu(\theta,l)\bigr\\|_{L_2(\mathbb R^\theta)}^2\Bigr)^{1/2}$ and the norm in $H^{s/3,s}(I\times (0,L))$ for any interval $I\subset \mathbb R$ as the restriction norm. The use of these norm is justified by the following fact. Let $v(t,x,y)$ be the appropriate solution to the initial value problem $$v_t+v_{xxx}+v_{xyy}=0,\qquad v\big\|_{t=0}=v_0.$$ Then according to [@F08] uniformly with respect to $x\in \mathbb R$ $$\label{1.7} \bigl\\|D_t^{1/3}v\bigr\\|_{H_{t,y}^{s/3,s}(\mathbb R^2)}^2+ \bigl\\|\partial_x v\bigr\\|_{H_{t,y}^{s/3,s}(\mathbb R^2)}^2+ \bigl\\|\partial_y v\bigr\\|_{H_{t,y}^{s/3,s}(\mathbb R^2)}^2 \sim \\|v_0\\|_{H^s(\mathbb R^2)}^2.$$ Introduce the notion of weak solutions to the considered problems. \[D1.1\] Let $u_0\in L_2$, $\mu_0,\nu_0,\nu_1\in L_2(B_T)$, $f\in L_1(0,T;L_2)$. A function $u\in L_\infty(0,T;L_2)$ is called a generalized solution to problem – if for any function $\phi\in L_2(0,T;\widetilde H^2)$, such that $\phi_t, \phi_{xxx}, \phi_{xyy}\in L_2(Q_T)$, $\phi\big\|_{t=T}\equiv 0$, $\phi\big\|_{x=0} =\phi_x\big\|_{x=0} =\phi\big\|_{x=R}\equiv 0$, the following equality holds: $$\begin{gathered} \label{1.8} \iiint_{Q_T}\Bigl[u(\phi_t+b\phi_x+\phi_{xxx}+\phi_{xyy}) +\frac 12 u^2 \phi_x +f\phi\Bigr]\,dxdydt +\iint_{\Omega} u_0\phi\big\|_{t=0}\,dxdy \\+ \iint_{B_T} \Bigl[\mu_0\phi_{xx}\big\|_{x=0}-\nu_0\phi_{xx}\big\|_{x=R}+ \nu_1\phi_x\big\|_{x=R}\Bigr]\,dydt =0.\end{gathered}$$ \[R1.1\] Note that the integrals in are well defined (in particular, since $\phi_x\in L_2(0,T;H^2)\subset L_2(0,T;L_\infty)$). Now we can formulate the main results of the paper concerning well-posedness, which means existence, uniqueness of solutions and Lipschitz continuity of the map $(u_0,\mu_0,\nu_0,\nu_1,f)\mapsto u$ in the corresponding norms on any ball in the space of the input data. \[T1.1\] Let $u_0\in L_2$, $f\in L_1(0,T; L_2)$ for certain $T>0$, $\mu_0,\nu_0\in \widetilde H^{s/3,s}(B_T)$ for certain $s>3/2$, $\nu_1\in L_2(B_T)$. Then problem – is well-posed in the space $X(Q_T)$. \[R1.2\] In the cases a) and d) for $\mu_0=\nu_0=\nu_1\equiv 0$ similar result was established in [@STW]. In the last paper certain properties of traces of $u_x$ with respect to $x$ were also obtained. \[T1.2\] Let $u_0\in \widetilde H^3$, $f\in C([0,T];L_2)\cap L_2(0,T; \widetilde H^{(0,2)})$, $f_t\in L_1(0,T;H^{-1})$ for certain $T>0$, $\mu_0,\nu_0\in \widetilde H^{4/3,4}(B_T)$, $\nu_1\in \widetilde H^{1,3}(B_T)$, $\mu_0(0,y)\equiv u_0(0,y)$, $\nu_0(0,y)\equiv u_0(R,y)$, $\nu_1(0,y)\equiv u_{0x}(R,y)$. Then problem – is well-posed in the space $X^3(Q_T)$. \[R1.3\] According to the assumptions on the boundary data $\mu$ are natural. In [@DL] for construction of regular solutions only homogeneous Dirichlet boundary conditions were considered. Moreover, in that paper for $u_{yyy}$ was established only that $u_{yyy}\in L_2(Q_T)$. Estimates on solutions, established in the proof of Theorem \[T1.1\], provide the following result on the large-time decay of small solutions. Let $B_+ = \mathbb R_+^t\times (0,L)$. \[T1.3\] Let there exists $\delta\in (0,1)$ such that $\varkappa>0$, where $$\label{1.9} \varkappa = -b+\left\{ \begin{aligned} \pi^2(1-\delta)\bigl(\frac 3{R^2} + \frac 1{L^2}\bigr) \qquad&\mbox{in the case a)},\\ \pi^2(1-\delta)\bigl(\frac 3{R^2} + \frac 1{4L^2}\bigr) \qquad&\mbox{in the case c)},\\ \pi^2(1-\delta)\frac 3{R^2} \qquad&\mbox{in the cases b) and d)}. \end{aligned} \right.$$ Let $$\label{1.10} \epsilon_0 = \frac{3^{5/4}\pi\delta}{4} \times\left\{ \begin{aligned} \max\Bigl(\frac{\sqrt{3}}{R},\frac 1{L}\Bigr) \qquad&\mbox{in the case a)},\\ \max\Bigl(\frac{\sqrt{3}}{R},\frac 1{2L}\Bigr)\qquad&\mbox{in the case c)},\\ \frac{\sqrt{3}}{R}\times\frac {3^{1/4}(\pi L)^{1/2}}{R^{1/2}+3^{1/4}(\pi L)^{1/2}} \qquad&\mbox{in the cases b) and d)}. \end{aligned} \right.$$ Let $u_0\in L_2$, $\nu_1\in L_2(B_+)$, $$\\|u_0\\|^2_{L_2}+\\|\nu_1\\|^2_{L_2(B_+)}\leq \epsilon_0^2,$$ $f\equiv 0$, $\mu_0=\nu_0\equiv 0$. Then the corresponding unique weak solution $u(t,x,y$) to problem – from the space $X(Q_T)$ $\forall T>0$ satisfies an inequality $$\label{1.11} \\|u(t,\cdot,\cdot)\\|^2_{L_2}\leq(1+R) e^{-\varkappa t/(1+R)}\Bigl[\\|u_0\\|_{L_2}^2 + \bigl\\|e^{\varkappa\tau/(2(1+R))}\nu_1\bigr\\|^2_{L_2(B_t)}\Bigr] \quad \forall t\geq 0.$$ \[R1.4\] In the case a) if $b=1$, $\nu_1\equiv 0$ a similar result for regular solutions in a slightly different form was previously established in [@DL]. On the basis of ideas and results from [@R] as an application of the developed technique we obtain the following result on the controllability problem for system – with the unknown boundary control $\nu_1$ and with the condition of final overdetermination $$\label{1.12} u(T,x,y)=u_T(x,y),\qquad (x,y)\in\Omega.$$ \[T1.4\] Let for any natural $l$, such that $\lambda_l <b$ (where $\lambda_l$ are the aforementioned eigenvalues of the operator $(-\psi'')$ on $(0,L)$ with corresponding boundary conditions), $$\label{1.13} R \ne 2\pi \Bigl(\frac{k^2+km+m^2}{3(b-\lambda_l)}\Bigr)^{1/2}\qquad \forall k,m\in\mathbb N.$$ Let $T>0$, $f\equiv 0$, $\mu_0=\nu_0\equiv 0$, $u_0, u_T\in L_2$. Then there exists $\varepsilon>0$, such that if $\\|u_0\\|_{L_2}, \\|u_T\\|_{L_2} <\varepsilon$ there exists a function $\nu_1\in L_2(B_T)$, such that there exists a unique solution $u\in X(Q_T)$ to problem –, satisfying . \[R1.5\] In comparison with Theorem \[T1.3\] the constant $\varepsilon$ is not evaluated explicitly. Further, let $\eta(x)$ denotes a cut-off function, namely, $\eta$ is an infinitely smooth non-decreasing function on $\mathbb R$ such that $\eta(x)=0$ when $x\leq 0$, $\eta(x)=1$ when $x\geq 1$, $\eta(x)+\eta(1-x)\equiv 1$. We drop limits of integration in integrals over the rectangle $\Omega$. The following interpolating inequality specifying the one from [@LSU] is crucial for the study. \[L1.1\] Let $\varphi(x,y)\in H^1$ satisfy $\varphi\big\|_{x=0}=0$ or $\varphi\|_{x=R}=0$, then the following inequalities hold: $$\begin{gathered} \label{1.14} \iint \varphi^4 dxdy \leq 4\Bigl(\iint \varphi_x^2\,dxdy \iint \varphi_y^2\,dxdy\Bigr)^{1/2} \iint \varphi^2\,dxdy \\ +\frac{4\sigma}{L} \Bigl(\iint \varphi_x^2\,dxdy\Bigr)^{1/2} \Bigl(\iint \varphi^2\,dxdy\Bigr)^{3/2},\end{gathered}$$ $$\begin{gathered} \label{1.15} \iint \|\varphi\|^3 dxdy \leq 2\Bigl(\iint \varphi_x^2\,dxdy \iint \varphi_y^2\,dxdy\Bigr)^{1/4} \iint \varphi^2\,dxdy \\ +\frac{2\sigma}{L^{1/2}} \Bigl(\iint \varphi_x^2\,dxdy\Bigr)^{1/4} \Bigl(\iint \varphi^2\,dxdy\Bigr)^{5/4},\end{gathered}$$ where $\sigma=0$ if $\varphi\big\|_{y=0}=0$ or $\varphi\big\|_{y=L}=0$ and $\sigma=1$ in the general case. We follow the argument from [@LSU] and start with the following inequality: $$\label{1.16} \iint \varphi^2\,dxdy \leq \iint \|\varphi_x\|\,dxdy \Bigl( \iint \|\varphi_y\|\,dxdy + \frac{2\sigma}{L} \iint \|\varphi\|\,dxdy\Bigr).$$ In fact, $$\sup\limits_{x\in (0,R)} \|\varphi(x,y)\| \leq \int_0^R \|\varphi_x(x,y)\|\,dx;$$ in the general case $\varphi(x,y) = \displaystyle \varphi(x,y)\frac{y}{L} + \varphi(x,y)\frac{L-y}{L} \equiv \varphi_1(x,y)+\varphi_2(x,y)$, where $$\sup\limits_{y\in (0,L)} \|\varphi_j(x,y)\| \leq \int_0^L \|\varphi_y(x,y)\|\alpha_j(y)\,dy +\frac{1}{L}\int_0^L \|\varphi(x,y)\|\,dy,$$ where either $\alpha_j(y)\equiv y/L$, or $\alpha_j(y)\equiv (L-y)/L$, therefore, $$\sup\limits_{y\in (0,L)} \|\varphi(x,y)\| \leq \int_0^L \|\varphi_y(x,y)\|\,dy + \frac{2\sigma}{L} \int_0^L \|\varphi(x,y)\|\,dy.$$ Since $$\iint \varphi^2(x,y)\,dxdy \leq \int_0^L \sup_{x\in (0,R)} \|\varphi(x,y)\|\,dy \int_0^R \sup_{y\in (0,L)} \|\varphi(x,y)\|\,dx,$$ we obtain . Therefore, $$\iint \varphi^4\,dxdy \leq \iint \bigl\|(\varphi^2)_x\bigr\|\,dxdy \Bigl( \iint \bigl\|(\varphi^2)_y\bigr\|\,dxdy + \frac{2\sigma}{L}\iint \varphi^2\,dxdy\Bigr),$$ whence succeeds. Inequality obviously follows from and Hölder’s inequality. For the decay results, we need Steklov’s inequalities in the following form: for $\psi\in H_0^1(0,L)$, $$\label{1.17} \int_0^L \psi^2(y)\,dy \leq \frac{L^2}{\pi^2} \int_0^L \bigl(\psi'(y)\bigr)^2\,dy,$$ for $\psi\in H^1(0,L)$, $\psi\big\|_{y=0}=0$, $$\label{1.18} \int_0^L \psi^2(y)\,dy \leq \frac{4L^2}{\pi^2} \int_0^L \bigl(\psi'(y)\bigr)^2\,dy.$$ In the following obvious interpolating results values of constants are indifferent for our purposes: for $\varphi\in H^1$ $$\label{1.19} \sup\limits_{x\in [0,R]}\int_0^L \varphi^2(x,y)\,dy \leq c \Bigl(\iint \varphi_x^2\,dxdy \iint \varphi^2\,dxdy\Bigr)^{1/2}+ c\iint \varphi^2\,dxdy,$$ $$\label{1.20} \\|\varphi\\|_{L_4} \leq c\\|\varphi\\|^{1/2}_{H^1}\\|\varphi\\|^{1/2}_{L_2}$$ and for $\varphi\in H^2$ $$\label{1.21} \\|\varphi\\|_{L_{\infty}} \leq c \\|\varphi\\|_{H^2}.$$ \[L1.2\] For $k=1$ and $k=2$ introduce functional spaces $$H^{(-k,0)}= \{\varphi = \sum\limits_{m=0}^k \partial_x^m \varphi_m: \varphi_m \in L_2\}$$ endowed with the natural norms. Then for $j=1$ and $j=2$ $$\label{1.22} \\|\partial_x^j \varphi\\|_{L_{2}} \leq c(R)\bigl( \\|\varphi_{xxx}\\|_{H^{(j-3,0)}} + \\|\varphi\\|_{L_2}\bigr).$$ First consider the case $j=2$. For any $\psi\in L_2$ let $a_0(y)\equiv \displaystyle \int_0^R \psi(x,y)\,dx$, then $\\|a_0\\|_{L_2(0,L)} \leq c\\|\varphi\\|_{L_2}$. Let $\omega(x)\in C_0^\infty(0,R)$, $\\|\omega\\|_{L_2(0,R)}=1$. Define $\displaystyle\psi_0(x,y)\equiv \int_0^x \psi(z,y)\,dz -a_0(y)\omega(x)$, then $\\|\psi_0\\|_{L_2}, \\|\psi_{0x}\\|_{L_2}\leq c\\|\psi\\|_{L_2}$, $\psi_0\big\|_{x=0}=\psi_0\big\|_{x=R}=0$, $\psi = \psi_{0x} +a_0\omega'$. We have: $$\begin{gathered} \langle \varphi_{xx},\psi_{0x}\rangle = - \langle \varphi_{xxx},\psi_0\rangle \leq \\|\varphi_{xxx}\\|_{H^{(-1,0)}}\bigl(\\|\psi_0\\|_{L_2} +\\|\psi_{0x}\\|_{L_2}\bigr) \\ \leq c \\|\varphi_{xxx}\\|_{H^{(-1,0)}}\\|\psi\\|_{L_2},\end{gathered}$$ $$\langle \varphi_{xx}, a_0\omega'\rangle = \langle \varphi, a_0\omega'''\rangle \leq c\\|\varphi\\|_{L_2}\\|\psi\\|_{L_2}.$$ Therefore, $$\langle \varphi_{xx},\psi\rangle \leq c \bigl( \\|\varphi_{xxx}\\|_{H^{(-1,0)}} + \\|\varphi\\|_{L_2}\bigr)\\|\psi\\|_{L_2}$$ and for $j=2$ follows. Now let $j=1$. For $\psi\in L_2$ define $\displaystyle a_1(y)\equiv \int_0^R \psi_0(x,y)\,dx$, $\displaystyle \psi_1(x,y) \equiv \int_0^x \psi_0(z,y)\,dz -a_1(y)\omega(x)$. Then $\psi =\psi_{1xx}+a_0\omega' +a_1\omega''$ and similarly to the previous case $$\begin{gathered} \langle \varphi_{x},\psi_{1xx}\rangle = \langle \varphi_{xxx},\psi_1\rangle \leq \\|\varphi_{xxx}\\|_{H^{(-2,0)}}\bigl(\\|\psi_1\\|_{L_2} +\\|\psi_{1x}\\|_{L_2}+\\|\psi_{1xx}\\|_{L_2}\bigr) \\ \leq c \\|\varphi_{xxx}\\|_{H^{(-2,0)}}\\|\psi\\|_{L_2},\end{gathered}$$ $$\langle \varphi_{x}, a_0\omega'+a_1\omega''\rangle =- \langle \varphi, a_0\omega''+a_1\omega'''\rangle \leq c\\|\varphi\\|_{L_2}\\|\psi\\|_{L_2}.$$ Therefore, $$\langle \varphi_{x},\psi\rangle \leq c \bigl( \\|\varphi_{xxx}\\|_{H^{(-2,0)}} + \\|\varphi\\|_{L_2}\bigr)\\|\psi\\|_{L_2},$$ which finishes the proof. The paper is organized as follows. Auxiliary linear problems are considered in Section \[S2\]. Section \[S3\] is devoted to the well-posedness results for the original problems. Decay of solutions is studied in Section \[S4\] and boundary controllability in Section \[S5\]. Auxiliary linear problems {#S2} ========================= Consider a linear equation $$\label{2.1} u_t+bu_x+u_{xxx}+u_{xyy}=f(t,x,y).$$ For any interval $I\subset\mathbb R^x$ and $k$ introduce functional spaces $$\begin{gathered} Y_k((0,T)\times I\times (0,L)) =\{u(t,x,y): \partial_t^j u\in C([0,T];\widetilde H^{k-3j}(I\times (0,L)),\quad\text{if}\ j\leq k/3,\\ \partial_x^n u\in C_b(\overline{I};\widetilde H^{(k-n+1)/3,k-n+1}(B_T)),\quad \text{if}\ n\leq k+1\}\end{gathered}$$ (here and further the lower index ’b" means a bounded map), $$\begin{gathered} M_k((0,T)\times I\times (0,L))=\{f(t,x,y): \partial_t^j f\in L_2(0,T;\widetilde H^{k-3j}(I\times(0,L)),\\ \text{if}\ j\leq j_0=[(k+1)/3]\}.\end{gathered}$$ Let $\widetilde \Phi_0(x,y) \equiv u_0(x,y)$ and for $j\geq 1$ $$\widetilde\Phi_j(x,y) \equiv \partial^{j-1}_t f(0,x,y)- (b\partial_x+\partial_x^3+\partial_x\partial_y^2)\widetilde\Phi_{j-1}(x,y).$$ Solutions to an initial-boundary value problem in a domain $\Pi_T=(0,T)\times \Sigma$ with the initial profile for $(x,y)\in \Sigma$ and boundary conditions for $(t,x)\in (0,T)\times\mathbb R$ for equation can be constructed in a form (see [@F17]) $$\label{2.2} u(t,x,y)=S(t,x,y;u_0)+K(t,x,y;f),$$ where potentials $S$ and $K$ are given by formulas $$\label{2.3} \begin{gathered} S(t,x,y;u_0) \equiv \sum_{l=1}^{+\infty} \frac{1}{2\pi} \int_{\mathbb R}\, e^{it(\xi^3-b\xi+\lambda_l\xi)} e^{i\xi x}\widehat{u}_0(\xi,l)\,d\xi \psi_l(y),\\ K(t,x,y;f)\equiv\int^t_0 S(t-\tau,x,y;f(\tau,\cdot,\cdot))\,d\tau, \end{gathered}$$ where the functions $\widehat{u}_0(\xi,l)$ are defined similarly to . \[L2.1\] If $u_0\in \widetilde H^k(\Sigma)$, $f\in M_k(\Pi_T)$ for some $T>0$ and $k\geq 0$, then a unique solution $u(t,x,y)\in Y_k(\Pi_T)$ to problem , , exists and for any $t_0\in (0,T]$ $$\label{2.4} \begin{aligned} &\\|u\\|_{Y_k(\Pi_{t_0})}\\ &\leq c(T,k,b)\Bigl(\\|u_0\\|_{\widetilde H^k(\Sigma)}+ t_0^{1/6}\\|f\\|_{M_k(\Pi_{t_0})} +\sum_{j=0}^{j_0-1} \\|\partial_t^j f\big\|_{t=0}\\|_{\widetilde H^{k-3(j+1)}(\Sigma)}\Bigr). \end{aligned}$$ First of all note that uniqueness of solutions to the considered problem in the space $L_2(\Pi_T)$ (in fact, in a more wide class) was established in [@BF13]. Next, note that $$\label{2.5} \partial_t^j S(t,x,y;u_0)+\partial_t^j K(t,x,y;f)= S(t,x,y;\widetilde\Phi_j)+ K(t,x,y;\partial_t^j f).$$ Then the corresponding estimates on $\partial_t^j u$ in the norm $C([0,t_0];\widetilde H^{k-3j}(\Sigma))$ by $\\|\widetilde\Phi_j\\|_{\widetilde H^{k-3j}(\Sigma)}$ and $\\|\partial_t^j f\\|_{L_1(0,t_0;\widetilde H^{k-3j}(\Sigma))}$ easily follow. In turn, $$\label{2.6} \\|\widetilde\Phi_j\\|_{\widetilde H^{k-3j}(\Sigma)} \leq c(k,b)\Bigl( \\|u_0\\|_{\widetilde H^k(\Sigma)} + \sum_{m=0}^{j-1} \\|\partial_t^m f\big\|_{t=0}\\|_{\widetilde H^{k-3(m+1)}(\Sigma)}\Bigr).$$ It was proved in [@F17] that for $s\in [0,3]$ $$\label{2.7} \\|u\\|_{C_b(\mathbb{R};\widetilde H^{s/3,s}(B_{t_0}))} \leq c(T,b)\Bigl(\\|u_0\\|_{\widetilde H^{s-1}(\Sigma)}+t_0^{1/2-s/6} \\|f\\|_{L_2(0,t_0;\widetilde H^{s-1}(\Sigma))}\Bigr).$$ Applying – for $j=[(k+1-n-l)/3]\leq j_0$, $s=k+1-n-l-3j\in [0,3)$, we derive that $$\begin{gathered} \label{2.8} \\|\partial_t^j\partial_x^n\partial_y^l u\\|_{C_b(\mathbb{R};\widetilde H^{s/3,s}(B_{t_0}))} \leq \\|S(\cdot,\cdot,\cdot;\partial^n_x\partial^l_y\widetilde\Phi_j)\\|_{C_b(\mathbb{R};\widetilde H^{s/3,s}(B_{t_0}))} \\ +\\|K(\cdot,\cdot,\cdot;\partial^n_x\partial^l_y\partial_t^j f)\\|_{C_b(\mathbb{R};\widetilde H^{s/3,s}(B_{t_0}))} \leq c(T,k,b)\Bigl(\\|u_0\\|_{\widetilde H^k(\Sigma)} \\+ \sum_{m=0}^{j-1} \\|\partial_t^m f\big\|_{t=0}\\|_{\widetilde H^{k-3(m+1)}(\Sigma)} + t_0^{1/2-s/6}\\|\partial_t^j f\\|_{L_2(0,t_0;\widetilde H^{k-3j})}\Bigr). \end{gathered}$$ Finally, it is suffice to note that the minimal value $1/6$ for the degree $(1/2-s/6)$ in is achieved if $k+1-n-l=3j+2$. Next, consider an initial-boundary value problem in a domain $\Pi_T^-=(0,T)\times \Sigma_-$, $\Sigma_-=\mathbb R_-\times (0,L) =\{(x,y): x<0, 0<y<L\}$, for equation with initial condition for $(x,y)\in \Sigma_-$, boundary conditions for $(t,x)\in (0,T)\times \mathbb R_-$ and $$\label{2.9} u(t,0,y)=\nu_0(t,y),\quad u_x(t,0,y)=\nu_1(t,y), \quad (t,y)\in B_T.$$ Weak solutions to this problem are understood similarly to Definition \[D1.1\] with obvious changes, moreover, due to the absence of nonlinearity one can take solutions from the space $L_2(\Pi_T^-)$. \[L2.2\] A generalized solution to problem , , , is unique in the space $L_2(\Pi_T^-)$. According to [@F17] the backward problem in $\Pi_T^-$ for equation with boundary conditions $u\big\|_{t=T}=0$, $u\big\|_{x=0}=0$ and for $f\in C_0^\infty(\Pi_T^-)$ has a solution $u\in C([0,T];\widetilde H^3(\Sigma_-))$, $u_t\in C([0,T];L_2(\Sigma_-))$, therefore, the desired result is obtained via the standard Hölmgren’s argument. \[L2.3\] Let $u_0\equiv 0$, $\nu_0, \nu_1\in C_0^\infty(B_+)$, $f\equiv 0$. Then there exists a solution $u(t,x,y)$ to problem , , , such that $\partial^j_t u \in C_b(\overline{\mathbb R}_+^t; \widetilde H^n(\Sigma_-))$ for any $j$ and $n$. Let $v(t,x,y)\equiv u(t,x,y) -\nu_0(t,y)\eta(x+1)-\nu_1(t,y)x\eta(x+1)$, then the original problem is equivalent to the problem of , , , type for the function $v$ with homogeneous initial-boundary conditions and $f\equiv -\nu_{0t}\eta-\nu_{1t}x\eta-b\nu_0\eta'-b\nu_1(x\eta)' -\nu_0\eta'''-\nu_1(x\eta)''' -\nu_{0yy}\eta'-\nu_{1yy}(x\eta)'$. Let $\{\varphi_j(x): j=1,2,\dots\}$ be a set of linearly independent functions complete in the space $\{\varphi \in H^3(\mathbb R_-): \varphi(0)=0\}$. We use the Galerkin method and seek an approximate solution in a form $v_k(t,x,y)= \sum\limits_{j,l=1}^k c_{kjl}(t) \varphi_j(x)\psi_l(y)$ (remind that $\psi_l$ are the orthonormal in $L_2(0,L)$ eigenfunctions for the operator $(-\psi'')$ on the segment $[0,L]$ with corresponding boundary conditions) via conditions for $i,m=1,\dots,k$, $t\in [0,T]$ $$\label{2.10} \iint_{\Sigma_-} \bigl(v_{kt}\varphi_i(x)\psi_m(y) -v_k(b\varphi_i'\psi_m+\varphi'''_i\psi_m +\varphi'_i\psi''_m)\bigr)\,dxdy - \iint_{\Sigma_-} f\varphi_i\psi_m\,dxdy=0,$$ $c_{kjl}(0)=0$. In particular, $v_k\big\|_{t=0}=0$. Moreover, putting in $t=0$, multiplying by $c'_{kim}(0)$ and summing with respect to $i,m$, we obtain that $v_{kt}\big\|_{t=0}=0$. Next, differentiating $j$ times with respect to $t$ we derive that $$\begin{gathered} \label{2.11} \iint_{\Sigma_-} \bigl(\partial_t^{j+1}v_{k}\varphi_i\psi_m -\partial_t^j v_k(b\varphi_i'\psi_m+\varphi'''_i\psi_m +\varphi'_i\psi''_m)\bigr)\,dxdy \\- \iint_{\Sigma_-} \partial_t^j f\varphi_i\psi_m\,dxdy=0.\end{gathered}$$ Then by induction with respect to $j$ we find that $\partial_t^j v_k\big\|_{t=0} =0$ for all $j$. Since $\psi_m^{(2n)}(y)=(-\lambda_m)^n\psi_m(y)$ it follows from and that for all $j$ and $n$ $$\begin{gathered} \label{2.12} \iint_{\Sigma_-} \bigl(\partial_t^{j+1}\partial_y^{n} v_{k}\varphi_i\psi_m^{(n)} - \partial_t^j\partial_y^{n} v_k(b\varphi'_m\psi_m^{(n)}+ \varphi'''_i\psi_m^{(n)}+\varphi'_i\psi_m^{(n+2)})\bigr)\,dxdy \\- \iint_{\Sigma_-} \partial_t^j\partial_y^{n}f\varphi_i\psi_m^{(n)}\,dxdy=0.\end{gathered}$$ Multiplying by $2c^{(j)}_{kim}(t)$ and summing with respect to $i,m$, we find that $$\label{2.13} \frac d{dt} \iint_{\Sigma_-} (\partial_t^j\partial_y^n v_k)^2\,dxdy +\int_0^L (\partial_t^j\partial_y^n v_{kx})^2\big\|_{x=0}\,dy = 2\iint_{\Sigma_-} \partial_t^j\partial_y^nf \partial_t^j\partial_y^nv_k\,dxdy,$$ and, therefore, for all $j$ and $n$ $$\label{2.14} \\|\partial_t^jv_k\\|_{L_\infty(\mathbb R_+^t;\widetilde H^{(0,n)}(\Sigma_-))} \leq \\|\partial_t^j f\\|_{L_1(\mathbb R_+^t;\widetilde H^{(0,n)}(\Sigma_-))}.$$ Estimate provide existence of a weak solution $v(t,x,y)$ to the considered problem such that $\partial_t^j v \in C_b(\overline{\mathbb R}_+^t;\widetilde H^{(0,n)}(\Sigma_-))\ \forall n, j$ in the following sense: for any $T>0$ and a function $\phi\in L_2(0,T;\widetilde H^2(\Sigma_-))$, such that $\phi_t, \phi_{xxx}, \phi_{xyy}\in L_2(\Pi_T^-)$, $\phi\big\|_{t=T}=0$, $\phi\big\|_{x=0}=0$, the following equality holds: $$\label{2.15} \iiint_{\Pi_T^-}\Bigl[v(\phi_t+b\phi_x+\phi_{xxx}+\phi_{xyy}) +f\phi\Bigr]\,dxdydt =0.$$ Note, that the traces of the function $v$ satisfy zero condition and condition . Moreover, it follows from that $\partial_t^j\partial_y^n v_{xxx} \in C_b(\overline{\mathbb R}_+^t;H^{(-1,0)}(\Sigma_-))\ \forall n,j$, therefore, $\partial_t^j\partial_y^n v_x \in C_b(\overline{\mathbb R}_+^t;L_2(\Sigma_-))\ \forall n,j$ (see [@F17]) and one more application of yields that $\partial_t^j v_{xxx} \in C_b(\overline{\mathbb R}_-^t;\widetilde H^{(0,n)}(\Sigma_-))\ \forall n,$, the function $v$ satisfies the corresponding equation a.e. in $\Pi_T^+$ and its traces satisfy zero conditions . Finally, with the use of induction with respect to $m$ one can find that $\partial^j_t\partial_x^{3m} v \in C_b(\overline{\mathbb R}_+^t; \widetilde H^{(0,n)})$ for all $m,j,n$. In what follows, we need some properties of solutions to an algebraic equation $$\label{2.16} z^3 +az+p=0, \qquad a\in \mathbb R,\quad p=\varepsilon+i\theta \in \mathbb C.$$ For $\varepsilon>0$ we denote by $z_1(p,a)$ and $z_2(p,a)$ two roots of this equation with positive real parts (the rest root has the negative real part). Let $$\label{2.17} r_j(\theta,a) = \lim\limits_{\varepsilon\to +0} z_j(\varepsilon+i\theta),\quad j=1, 2.$$ The values $r_j(\theta,a)$ are roots of the equation $$\label{2.18} r^3+ar+i\theta=0$$ and $\Re r_j\geq 0$, $j=1$ and $2$. Moreover, it can be shown with the use of the Cardano formula, that for certain positive constants $c_0$, $c_1$ and all $\theta$ and $a$ $$\begin{gathered} \label{2.19} \|r_j(\theta,a)\| \leq c_1(\|\theta\|^{1/3}+\|a\|^{1/2}),\quad j=1,2,\\ \label{2.20} \|r_1(\theta,a)-r_2(\theta,a)\| \geq c_0(\|\theta\|^{1/3}+\|a\|^{1/2})\end{gathered}$$ (for more details see, for example, [@F07-2]). Now introduce special solutions of equation for $f\equiv 0$ of “boundary potential” type. \[D2.1\] Let $\nu \in\widetilde{\EuScript S}(\overline{B})$. Define for $x\leq 0$ $$\begin{gathered} \label{2.21} J_0(t,x,y;\nu) \equiv \sum\limits_{l=1}^{+\infty} \EuScript F^{-1}_t \Bigl[\frac{r_1 e^{r_2x}-r_2e^{r_1x}}{r_1-r_2}\widehat\nu(\theta,l)\Bigr](t) \psi_l(y),\\ \label{2.22} J_1(t,x,y;\nu) \equiv \sum\limits_{l=1}^{+\infty} \EuScript F^{-1}_t \Bigl[\frac{e^{r_1x}-e^{r_2x}}{r_1-r_2}\widehat\nu(\theta,l)\Bigr](t) \psi_l(y),\end{gathered}$$ where $\widehat\nu(\theta,l)$ is given by formula and $r_j=r_j(\theta,b-\lambda_l)$ – by formula . \[L2.4\] For any $s\in\mathbb R$ the notion of the function $J_0(t,x,y;\nu)$ can be extended by continuity in the space $C_b(\overline{\mathbb R}^x_-;\widetilde H^{s/3,s}(B))$ to any function $\nu\in \widetilde H^{s/3,s}(B)$. Moreover, for any $n$ $$\label{2.23} \\|\partial_x^n J_0(\cdot,\cdot,\cdot;\nu)\\|_{ C_b(\overline{\mathbb R}_-^x;\widetilde H^{(s-n)/3.s-n}(B))} \leq c(n,b)\\|\nu\\|_{\widetilde H^{s/3.s}(B)}$$ and $J_0\big\|_{x=0}=\nu$, $J_{0x}\big\|_{x=0}=0$. Since $$\partial_x^n \widehat J_0(\theta,x,l;\nu) = \frac{r_1r_2^n e^{r_2x}-r_2r_1^ne^{r_1x}}{r_1-r_2}\widehat\nu(\theta,l)$$ and $\Re (r_jx)\leq 0$ the assertion of the lemma follows from , . \[L2.5\] For any $s\in\mathbb R$ and $R>0$ the notion of the function $J_1(t,x,y;\nu)$ can be extended by continuity in the space $C([-R,0];\widetilde H^{s/3,s}(B))$ to any function $\nu\in \widetilde H^{s/3,s}(B)$. Moreover, $$\label{2.24} \\|x^{-1}J_1(\cdot,\cdot,\cdot;\nu)\\|_{ C_b(\overline{\mathbb R}_-^x;\widetilde H^{s/3.s}(B))} \leq \\|\nu\\|_{\widetilde H^{s/3.s}(B)},$$ for any $n\geq 1$ $$\label{2.25} \\|\partial_x^n J_1(\cdot,\cdot,\cdot;\nu)\\|_{ C_b(\overline{\mathbb R}_-^x;\widetilde H^{(s-n+1)/3.s-n+1}(B))} \leq c(n,b)\\|\nu\\|_{\widetilde H^{s/3.s}(B)}$$ and $J_1\big\|_{x=0}=0$, $J_{1x}\big\|_{x=0}=\nu$. Since $$\partial_x^n \widehat J_1(\theta,x,l;\nu) = \frac{r_1^n e^{r_1x}-r_2^ne^{r_2x}}{r_1-r_2}\widehat\nu(\theta,l)$$ and $\Re (r_jx)\leq 0$ (in particular, $\|\widehat J_1(\theta,x,l;\nu)\|\leq \|x\widehat\nu(\theta,l)\|$) the assertion of the lemma follows from , . \[R2.1\] In the most important for us case $s\geq 0$ the values $\widehat\nu(\theta,l)$ can be defined directly as limits in $L_2(B)$, for example, of integrals $\displaystyle \int_{-T}^T\!\int_0^L e^{-i\theta t}\psi_l(y)\nu(t,y)\,dtdy$, $T\to +\infty$. Then the functions $J_0(t,x,y;\nu)$ and $J_1(t,x,y;\nu)$ can be equivalently defined simply by formulas , . \[L2.6\] If $\nu\in \widetilde H^{(s+1)/3.s+1}(B)$ for certain $s\geq 0$, then for any $j\leq s/3$ there exists $\partial_t^j J_0(t,x,y;\nu)\in C_b(\mathbb R^t;\widetilde H^{s-3j}(\Sigma_-))$ and uniformly with respect to $t\in\mathbb R$ $$\label{2.26} \\|\partial_t^j J_0(t,\cdot,\cdot;\nu)\\|_{\widetilde H^{s-3j}(\Sigma_-)} \leq c(b,s,L) \\|\nu\\|_{\widetilde H^{(s+1)/3.s+1}(B)}.$$ If $\nu\in \widetilde H^{s/3.s}(B)$ for certain $s\geq 0$, then for any $j\leq s/3$ there exists $\partial_t^j J_1(t,x,y;\nu)\in C_b(\mathbb R^t;\widetilde H^{s-3j}(\Sigma_-))$ and uniformly with respect to $t\in\mathbb R$ $$\label{2.27} \\|\partial_t^j J_1(t,\cdot,\cdot;\nu)\\|_{\widetilde H^{s-3j}(\Sigma_-)} \leq c(b,s,L) \\|\nu\\|_{\widetilde H^{s/3.s}(B)}.$$ The proof is based on the following inequality, established in [@F07-2]: let $$I(t,x) \equiv \int_{\mathbb R} e^{i\theta t} e^{r_j(\theta,a)x}w(\theta)\,d\theta,$$ where $r_j(\theta,a)$, $j=1$ and $2$, are the roots of equation , defined in . Then there exists a positive constant $c$, such that uniformly with respect to $t\in\mathbb R$ $$\label{2.28} \\|I(t,\cdot)\\|_{L_2(\mathbb R_-)} \leq c\\|(\|\theta\|^{1/3} +\|a\|^{1/2})w(\theta)\\|_{L_2(\mathbb R)}.$$ Now let $$J(t,x,y)\equiv \sum\limits_{l=1}^{+\infty} \int_{\mathbb R} e^{i\theta t} e^{r_j(\theta,b-\lambda_l)x} w(\theta,l)\,d\theta \,\psi_l^{(m)}(y).$$ Then it follows from that uniformly with respect to $t\in\mathbb R$ since the system $\{\psi_l^{(m)}\}$ is also orthogonal in $L_2(0,L)$ and $\\|\psi_l^{(m)}\\|_{L_2(0,L)}\leq c(l/L)^m$ $$\begin{gathered} \label{2.29} \\|J(t,\cdot,\cdot)\\|_{L_2(\Sigma_-)} = \Bigl(\sum\limits_{l=1}^{+\infty} \Bigl\\|\int_{\mathbb R} e^{i\theta t} e^{r_j(\theta,b-\lambda_l)x} w(\theta,l)\,d\theta \Bigr\\|_{L_2(\mathbb R_-^x)}^2\Bigl\\|\psi^{(m)}_l\Bigr\\|_{L_2(0,L)}^2\Bigr)^{1/2} \\ \leq c(m,L)\Bigl(\sum\limits_{l=1}^{+\infty} \left\\|\bigl(\|\theta\|^{1/3}+\|b-\lambda_l\|^{1/2}\bigr)w(\theta,l)\right\\|_{L_2(\mathbb R^\theta)}^2 l^{2m}\Bigr)^{1/2}.\end{gathered}$$ Without loss of generality one can assume that $\nu \in\widetilde{\EuScript S}(\overline{B})$. Let $s$ be integer. Then for $3j+n+m=s$ $$\label{2.30} \partial_t^j \partial_x^n \partial_y^m J_0(t,x,y;\nu) = \sum\limits_{l=1}^{+\infty} \frac1{2\pi} \int_{\mathbb R} (i\theta)^j \frac{r_1r_2^n e^{r_2x}-r_2r_1^ne^{r_1x}}{r_1-r_2}\widehat\nu(\theta,l)\,d\theta\, \psi_l^{(m)}(y)$$ and inequalities , and yield that $$\begin{gathered} \\|\partial_t^j \partial_x^n \partial_y^m J_0(t,\cdot,\cdot;\nu)\\|_{L_2(\Sigma_-)} \leq c\Bigl(\sum\limits_{l=1}^{+\infty} \left\\|\bigl(\|\theta\|^{2/3}+l^2\bigr)^{(3j+n+m+1)/2}\widehat\nu(\theta,l)\right\\|_{L_2(\mathbb R^\theta)}^2\Bigr)^{1/2} \\ = c\\|\nu\\|_{\widetilde H^{(s+1)/3.s+1}(B)}.\end{gathered}$$ Similarly, $$\label{2.31} \partial_t^j \partial_x^n \partial_y^m J_1(t,x,y;\nu) = \sum\limits_{l=1}^{+\infty} \frac1{2\pi} \int_{\mathbb R} (i\theta)^j \frac{r_1^n e^{r_1x}-r_2^ne^{r_2x}}{r_1-r_2}\widehat\nu(\theta,l)\,d\theta\, \psi_l^{(m)}(y)$$ and, therefore, $$\begin{gathered} \\|\partial_t^j \partial_x^n \partial_y^m J_1(t,\cdot,\cdot;\nu)\\|_{L_2(\Sigma_-)} \leq c\Bigl(\sum\limits_{l=1}^{+\infty} \left\\|\bigl(\|\theta\|^{2/3}+l^2\bigr)^{(3j+n+m)/2}\widehat\nu(\theta,l)\right\\|_{L_2(\mathbb R^\theta)}^2\Bigr)^{1/2} \\ = c\\|\nu\\|_{\widetilde H^{s/3.s}(B)}.\end{gathered}$$ Finally, use interpolation. \[L2.7\] Let $\nu\in \widetilde H^{s/3,s}(B)$, then for any $T>0$ $$\label{2.32} \\|J_1(\cdot,\cdot,\cdot;\nu)\\|_{ C_b(\overline{\mathbb R}_-^x;\widetilde H^{(s+1)/3.s+1}(B_T))} \leq c(T,b,s,L)\\|\nu\\|_{\widetilde H^{s/3.s}(B)}.$$ Without loss of generality one can assume that $\nu \in\widetilde{\EuScript S}(\overline{B})$. There exists $l_0$ such that for $l>l_0$and all $\theta$ and there exists $\theta_0\geq 1$ such that for $\|\theta\|\geq \theta_0$ and all $l$ $$\label{2.33} \|r_1(\theta,b-\lambda_l)-r_2(\theta,b-\lambda_l)\|\geq c_0(\|\theta\|^{1/3}+l).$$ Divide $\nu$ into two parts: $$\nu_0(t,y)\equiv \sum\limits_{l=1}^{l_0} \EuScript F^{-1}_t\bigr[ \widehat \nu(\theta,l)\eta(\theta_0+1-\|\theta\|)\bigr](x)\psi_l(y),\quad \nu_{1}(t,y)\equiv \nu(t,y)-\nu_0(t,y).$$ For $\nu_0$ inequality yields, that for any $j$ and $m$ $$\begin{gathered} \\|\partial_t^j\partial_y^m J_1(\cdot,\cdot,\cdot;\nu_0)\\|_{C_b(\overline{\mathbb R}_-^x;L_2(B_T))} \leq T^{1/2}\sup\limits_{t\in [0,T]} \\|\partial_t^j J_1(t,\cdot,\cdot;\nu_0)\\|_{\widetilde H^{m+1}(\Sigma_-)} \\ \leq c(T,b,j,m,L)\\|\nu_0\\|_{\widetilde H^{(3j+m+1)/3,3j+m+1}(B)} \leq c(T,b,s,j,m,L) \\|\nu\\|_{\widetilde H^{s/3.s}(B)}.\end{gathered}$$ For $\nu_1$ by virtue of $$\begin{gathered} \\|J_1(\cdot,\cdot,\cdot;\nu_0)\\|_{ C_b(\overline{\mathbb R}_-^x;\widetilde H^{(s+1)/3.s+1}(B))} \\ \leq c\Bigl(\sum\limits_{l=1}^{+\infty} \\|\frac{(\|\theta\|^{2/3}+l^2)^{(s+1)/2}}{\|\theta\|^{1/3}+l} \widehat\nu_1(\theta,l)\\|^2_{L_2(\mathbb R^\theta)}\Bigr)^{1/2} \leq c_1(b,s) \\|\nu\\|_{\widetilde H^{s/3.s}(B)}.\end{gathered}$$ \[L2.8\] Let $\nu_0\in \widetilde H^{1/3,1}(B)$, $\nu_1\in L_2(B)$ and $\nu_0(t,y)=\nu_1(t,y)=0$ for $t<0$, then the function $u(t,x,y)\equiv J_0(t,x,y;\nu_0)+J_1(0,x,y;\nu_1)$ for any $T>0$ is a weak solution from the space$Y_0(\Pi_T^-)$ to problem (for $f\equiv 0$), (for $u_0\equiv 0$), , . First let $\nu_0,\nu_1\in C_0^\infty(B_+)$. Consider the smooth solution $u(t,x,y)$ to the considered problem constructed in Lemma \[L2.3\]. For any $p=\varepsilon+i\theta$, where $\varepsilon>0$, define the Laplace–Fourier transform-coefficients $$\widetilde u(p,x,l) \equiv \int_{\mathbb R_+}\!\! \int_0^L e^{-pt} \psi_l(y) u(t,x,y)\,dydt.$$ The function $\widetilde u(p,x,l)$ solves a problem $$\begin{gathered} p\widetilde u(p,x,l)+b\widetilde u_x(p,x,l)+\widetilde u_{xxx}(p,x,l)-\lambda_l \widetilde u_x(p,x,l)=0,\\ \widetilde u(p,0,l)=\widetilde \nu_0(p,l)\equiv \int_{\mathbb R_+}\!\! \int_0^L e^{-pt}\psi_l(y)\nu_0(t,y)\,dydt, \quad \widetilde u_x(p,0,l)=\widetilde \nu_1(p,l),\end{gathered}$$ whence, since $\widetilde u(p,x,l)\to 0$ as $x\to -\infty$, it follows, that $$\widetilde u(p,x,l)=\frac {z_1e^{z_2x}-z_2e^{z_1x}} {z_1-z_2} \widetilde \nu_0(p,l) + \frac {e^{z_1x}-e^{z_2x}} {z_1-z_2} \widetilde \nu_1(p,l).$$ where $z_j=z_j(p,b-\lambda_l)$ are defined in for $a=b-\lambda_l$. Using the formula of inversion of the Laplace transform we find, that the Fourier coefficients of the function $u(t,x,\cdot)$ are the following: $$\widehat u(t,x,l) = e^{\varepsilon t} \EuScript F_t^{-1}\left[\frac {z_1e^{z_2x}-z_2e^{z_1x}} {z_1-z_2}\widetilde \nu_0(\varepsilon+i\theta,l) + \frac {e^{z_1x}-e^{z_2x}} {z_1-z_2} \widetilde \nu_1(\varepsilon+i\theta,l)\right](t)$$ and, therefore, $$\begin{gathered} u(t,x,y) \\= \sum\limits_{l=1}^{+\infty} e^{\varepsilon t} \EuScript F_t^{-1}\left[\frac {z_1e^{z_2x}-z_2e^{z_1x}} {z_1-z_2}\widetilde \nu_0(\varepsilon+i\theta,l) +\frac {e^{z_1x}-e^{z_2x}} {z_1-z_2} \widetilde \nu_1(\varepsilon+i\theta,l)\right](t)\psi_l(y).\end{gathered}$$ Passing to the limit as $\varepsilon\to+0$, we derive that $u(t,x,y) \equiv J_0(t,x,y;\nu_0)+J_1(0,x,y;\nu_1)$. In the general case approximate the function $\mu$ by smooth ones, pass to the limit on the basis of estimates , , , for $s=0$, for $s=1$ and use the uniqueness result. \[L2.9\] Let $u_0\in \widetilde H^k(\Sigma_-)$, $\nu_0\in \widetilde H^{(k+1)/3,k+1}(B_T)$, $\nu_1\in \widetilde H^{k/3,k}(B_T)$, $f\in M_k(\Pi_T^-)$ for certain $T>0$, $k\geq 0$. Assume also that $\partial_t^j \nu_0(0,y) \equiv \widetilde \Phi_j(0,y)$ for $j<k/3$, $\partial_t^j \nu_1(0,y) \equiv \widetilde\Phi_{jx}(0,y)$ for $j<(k-1)/3$. Then there exists a unique solution $u(t,x,y)\in Y_k(\Pi_T^-)$ to problem , , , and for any $t_0\in (0,T]$ $$\begin{gathered} \label{2.34} \\|u\\|_{Y_k(\Pi_{t_0}^-)} \leq c(T,k,b,L)\Bigl(\\|u_0\\|_{\widetilde H^k(\Sigma_-)}+ \\|\nu_0\\|_{\widetilde H^{(k+1)/3,k+1}(B_T)}+ \\|\nu_1\\|_{\widetilde H^{k/3,k}(B_T)} \\+ t_0^{1/6}\\|f\\|_{M_k(\Pi_{t_0}^-)} +\sum_{j=0}^{j_0-1} \\|\partial_t^j f\big\|_{t=0}\\|_{\widetilde H^{k-3(j+1)}(\Sigma_-)}\Bigr), \quad j_0=[(k+1)/3].\end{gathered}$$ Extend the functions $u_0$ and $f$ to the whole real axis with respect to $x$ in the classes $\widetilde H^k(\Sigma)$ and $M_k(\Pi_T)$ respectively and consider the solution $U(t,x,y)$ to the initial value problem , , in the class $Y_k(\Pi_T)$ given by Lemma \[L2.1\]. Note that $$\widetilde \nu_0\equiv \nu_0-U\|_{x=0}\in \widetilde H^{(k+1)/3,k+1}(B_T),\quad \widetilde \nu_1\equiv \nu_1-U_x\|_{x=0}\in \widetilde H^{k/3,k}(B_T),$$ and by virtue of the compatibility conditions $\partial_t^j\widetilde \nu_0\big\|_{t=0}=0$ for $j<k/3$, $\partial_t^j\widetilde \nu_1\big\|_{t=0}=0$ for $j<(k-1)/3$, so the functions $\widetilde\nu_0$, $\widetilde\nu_1$ can be extended in the same spaces to the whole strip $B$, such that $\widetilde\nu_0(t,y)=\widetilde\nu_1(t,y)=0$ for $t<0$. Then Lemmas \[L2.1\]–\[L2.8\] for the function $$u(t,x,y)\equiv U(t,x,y)+J_0(t,x,y;\widetilde\nu_0)+J_1(t,x,y;\widetilde\nu_1)$$ provide the desired result. Now consider the problem in $Q_T$. \[L2.10\] Let $u_0\in \widetilde H^k$, $\mu_0,\nu_0\in \widetilde H^{(k+1)/3,k+1}(B_T)$, $\nu_1\in \widetilde H^{k/3,k}(B_T)$, $f\in M_k(Q_T)$ for certain $T>0$, $k\geq 0$. Assume also that $\partial_t^j \mu_0(0,y) \equiv \widetilde \Phi_j(0,y)$, $\partial_t^j \nu_0(0,y) \equiv \widetilde \Phi_j(R,y)$ for $j<k/3$, $\partial_t^j \nu_1(R,y) \equiv \widetilde\Phi_{jx}(R,y)$ for $j<(k-1)/3$. Then there exists a unique solution $u(t,x,y)\in Y_k(Q_T)$ to problem , – and for any $t_0\in (0,T]$ $$\begin{gathered} \label{2.35} \\|u\\|_{Y_k(Q_{t_0})} \leq c(T,k,b,R,L)\Bigl(\\|u_0\\|_{\widetilde H^k}+ \\|\mu_0\\|_{\widetilde H^{(k+1)/3,k+1}(B_T)}+ \\|\nu_0\\|_{\widetilde H^{(k+1)/3,k+1}(B_T)} \\+ \\|\nu_1\\|_{\widetilde H^{k/3,k}(B_T)} + t_0^{1/6}\\|f\\|_{M_k(Q_{t_0})} +\sum_{j=0}^{j_0-1} \\|\partial_t^j f\big\|_{t=0}\\|_{\widetilde H^{k-3(j+1)}}\Bigr).\end{gathered}$$ Solutions to the considered problem (similarly to the corresponding problem in [@F08]) are constructed in the form $$\label{2.36} u(t,x,y)=w(t,x,y)+v(t,x,y),$$ where $w(t,x,y)$ is a solution to an initial-boundary value problem in $\widetilde\Pi^-_{T}=(0,T)\times \widetilde \Sigma_{-}$, $\widetilde \Sigma_- = (-\infty,R)\times (0,L)$ for equation with initial and boundary conditions for $(x,y)\in \widetilde\Sigma_-$, for $(t,x)\in (0,T)\times (-\infty,R)$ and (where $x=0$ is substituted by $x=R$) in the class $Y_k(\widetilde \Pi^-_{T})$. Then according to ($u_0$ and $f$ are extended to $x<0$ in a appropriate way) $$\begin{gathered} \label{2.37} \\|w\\|_{Y_k(\widetilde \Pi_{t_0}^-)} \leq c(T,k,b,L)\Bigl(\\|u_0\\|_{\widetilde H^k}+ \\|\nu_0\\|_{\widetilde H^{(k+1)/3,k+1}(B_T)}+ \\|\nu_1\\|_{\widetilde H^{k/3,k}(B_T)} \\+ t_0^{1/6}\\|f\\|_{M_k(\widetilde Q_{t_0})} +\sum_{j=0}^{j_0-1} \\|\partial_t^j f\big\|_{t=0}\\|_{\widetilde H^{k-3(j+1)}}\Bigr).\end{gathered}$$ Moreover, $$\widetilde\mu_0(t,y)\equiv \mu_0(t,y)-w(t,0,y)\in H^{(k+1)/3,k+1}(B_T),$$ by virtue of the compatibility conditions on the line $(0,0,y)$ $\partial_t^j\widetilde \mu_0(0,y)\equiv 0$ for $j<k/3$ and $$\begin{gathered} \label{2.38} \\|\widetilde\mu_0\\|_{\widetilde H^{(k+1)/3,k+1}(B_T)} \leq c(T,k,b)\Bigl(\\|u_0\\|_{\widetilde H^k}+ \\|\mu_0\\|_{\widetilde H^{(k+1)/3,k+1}(B_T)}\\+ \\|\nu_0\\|_{\widetilde H^{(k+1)/3,k+1}(B_T)} + \\|\nu_1\\|_{\widetilde H^{k/3,k}(B_T)} + t_0^{1/6}\\|f\\|_{M_k(Q_{t_0})} +\sum_{j=0}^{j_0-1} \\|\partial_t^j f\big\|_{t=0}\\|_{\widetilde H^{k-3(j+1)}}\Bigr).\end{gathered}$$ In particular, the function $\widetilde\mu_0$ can be considered as extended in the same class to the whole strip $B$ such that $\widetilde\mu_0(t,y) = 0$ for $t<0$. Consider in $Q_T$ a problem for the function $v$: $$\begin{gathered} \label{2.39} v_t+bv_x+v_{xxx}+v_{xyy}=0, \\ \label{2.40} v\big\|_{t=0}=0, \quad v\big\|_{x=0}=\widetilde\mu_0, \quad v\big\|_{x=R}=v_x\big\|_{x=R}=0\end{gathered}$$ also with corresponding boundary conditions . In order to construct a solution to this problem we consider for $x\geq 0$ the boundary potential $J(t,x,y;\mu)$ for an arbitrary function $\mu\in \widetilde H^{(k+1)/3,k+1}(B)$, $\mu(t,y)=0$ for $t<0$. Such a potential was introduced in [@F17] as a solution to an initial-boundary value problem in $\Pi_T^+ =(0,T)\times \Sigma_+$, $\Sigma_+=\mathbb R_+ \times (0,L)$, for equation in the case $f\equiv 0$ with zero initial condition for $(x,y)\in \Sigma_+$, boundary condition for $(t,x)\in (0,T)\times\mathbb R_+$ and boundary condition $$\label{2.41} u(t,0,y)=\mu(t,y),\quad (t,y)\in B_T.$$ According to [@F17] the function $J$ is infinitely differentiable for $x>0$ and for any $\delta\in (0,T]$ $$\begin{gathered} \label{2.42} \\|J(\cdot,R,\cdot;\mu)\\|_{\widetilde H^{(k+1)/3,k+1}(B_\delta)}+ \\|\partial_x J(\cdot,R,\cdot;\mu)\\|_{\widetilde H^{k/3,k}(B_\delta)} \\ \leq c(T,k,b,R,L)\delta^{1/2}\\|\mu\\|_{L_2(B_\delta)}.\end{gathered}$$ Moreover, $\partial_t^j J(0,R,y;\mu)=\partial_t^j J_x(0,R,y;\mu)\equiv 0$ for all $j$. Consider in the domain $\widetilde \Pi_{\delta}^-$ the problem of , , , (for $x=R$) type, where $u_0\equiv 0$, $f\equiv 0$, $\nu_0\equiv -J(\cdot,R,\cdot;\mu)$, $\nu_1\equiv -\partial_x J(\cdot,R,\cdot;\mu)$. A solution to this problem $V\in Y_k(\widetilde \Pi_\delta^-)$ exists and, in particular, $$\label{2.43} \begin{aligned} &\\|V(\cdot,0,\cdot)\\|_{\widetilde H^{(k+1)/3,k+1}(B_\delta)} \\ &\leq c(T,k,b,L) \left(\\|J(\cdot,R,\cdot;\mu)\\|_{\widetilde H^{(k+1)/3,k+1}(B_\delta)}+ \\|\partial_x J(\cdot,R,\cdot;\mu)\\|_{\widetilde H^{k/3,k}(B_\delta)}\right). \end{aligned}$$ Moreover, it is obvious that $\partial_t^j V(0,0,y) \equiv 0$ if $j<k/3$. Consider a linear operator $\Gamma: \mu \mapsto V(\cdot,0,\cdot)$ in the space $\widetilde H^{(k+1)/3,k+1}(B_\delta)$, $\partial_t^j\mu(0,y)\equiv 0$ if $j<k/3$. For small $\delta=\delta(T,k,b,L)$ estimates and provide that the operator $(E+\Gamma)$ is invertible ($E$ is the identity operator) and setting $\mu\equiv (E+\Gamma)^{-1}\widetilde\mu_0$ we obtain the desired solution to problem , , $$v(t,x,y)\equiv J(t,x,y;\mu) + V(t,x,y),$$ where (also with the use of the corresponding estimate on $J$ from [@F17]) $$\label{2.44} \\|v\\|_{Y_k(Q_\delta)} \leq c(T,k,b,L)\\|\widetilde\mu_0\\|_{\widetilde H^{(k+1)/3,k+1}(B_T)}.$$ Thus, the solution $u(t,x,y)$ to problem , – in the domain $Q_\delta$ is constructed and according to (\[2.36\])–(\[2.38\]) and (\[2.44\]) is evaluated in the space $Y_k(Q_\delta)$ by the right part of (\[2.35\]). Moving step by step ($\delta$ is constant) we obtain the desired solution in the whole domain $Q_T$. Uniqueness of weak solutions to problem , – in $L_2(Q_T)$ succeeds from existence of smooth solutions to the adjoint problem $$\begin{gathered} \phi_t+b\phi_x+\phi_{xxx}+\phi_{xyy}=f\in C_0^\infty(Q_T), \\ \phi\big\|_{t=T}=0, \quad \phi\big\|_{x=0}=\phi_x\big\|_{x=0}=\phi\big\|_{x=R}=0\end{gathered}$$ and with the corresponding boundary conditions of type, which after simple change of variables transforms to the original one. \[R2.2\] In further lemmas of this section all intermediate argument is performed for smooth solutions constructed in Lemma \[L2.10\] with consequent pass to the limit on the basis of obtained estimates due to linearity of the problem. \[L2.11\] Let $u_0\in L_2$, $\mu_0=\nu_0\equiv 0$, $\nu_1\in L_2(B_T)$, $f\equiv f_0 +f_{1x}$, where $f_0\in L_1(0,T;L_2)$, $f_1\in L_2(Q_T)$. Then there exist a (unique) weak solution to problem , – from the space $X(Q_T)$ and a function $\mu_1\in L_2(B_T)$, such that for any function $\phi\in L_2(0,T;\widetilde H^2)$, $\phi_t, \phi_{xxx}, \phi_{xyy}\in L_2(Q_T)$, $\phi\big\|_{t=T}=0$, $\phi\big\|_{x=0}=\phi\big\|_{x=R} =0$, the following equality holds: $$\begin{gathered} \label{2.45} \iiint_{Q_T}\Bigl[u(\phi_t+b\phi_x+\phi_{xxx}+\phi_{xyy}) +f_0\varphi -f_1\phi_x \Bigr]\,dxdydt \\ +\iint u_0\phi\big\|_{t=0}\,dxdy + \iint_{B_T} \Bigl[\nu_1\phi_x\big\|_{x=R}-\mu_1\phi_x\big\|_{x=0}\Bigr]\,dydt =0.\end{gathered}$$ Moreover, for $t\in (0,T]$ $$\begin{gathered} \label{2.46} \\|u\\|_{X(Q_t)} +\\|\mu_1\\|_{L_2(B_t)}\leq c(T,b,R)\Bigl(\\|u_0\\|_{L_2} + \\|\nu_1\\|_{B_T} +\\|f_0\\|_{L_1(0,t;L_2)} \\ +\\|f_1\\|_{L_2(Q_{t})}\Bigr),\end{gathered}$$ and if either $\rho(x)\equiv 1$ or $\rho(x)\equiv 1+x$ $$\begin{gathered} \label{2.47} \iint u^2(t,x,y)\rho(x)\,dxdy + \int_0^t \!\! \iint (3u_x^2 +u_y^2 -bu^2)\rho'(x)\,dxdyd\tau + \rho(0)\iint_{B_t} \mu_1^2\,dyd\tau \\ = \iint u_0^2\rho(x)\,dxdy + \rho(R)\iint_{B_t} \nu_1^2\,dyd\tau + 2\int_0^t \!\! \iint f_0 u\rho(x)\,dxdyd\tau \\ -2\int_0^t \!\! \iint f_1\bigl(u\rho(x)\bigr)_x\,dxdyd\tau.\end{gathered}$$ Multiplying by $2u(t,x,y)\rho(x)$ and integrating over $\Omega$, we find that $$\begin{gathered} \label{2.48} \frac d{dt} \iint u^2\rho\,dxdy +\rho(0)\int_0^L u_{x}^2\big\|_{x=0}\,dy + \iint (3u_{x}+u_{y}^2 -bu^2)\rho' \,dxdy \\ = \rho(R)\int_0^L \nu_1^2\,dy + 2\iint f_0 u\rho\,dxdy -2\iint f_1(u\rho)_x\,dxdy.\end{gathered}$$ Note that $$\begin{gathered} \label{2.49} \Bigl\|\iint f_1(u\rho)_x\,dxdy\Bigr\| \leq c\\|f_1\\|_{L_2} \bigl\\|(\|u_{x}\|+\|u\|)\bigr\\|_{L_2}\\ \leq \varepsilon \iint \bigl(u_{x}^2+u^2\bigr)\,dxdy + c(\varepsilon)\\|f_1\\|_{L_2}^2,\end{gathered}$$ where $\varepsilon>0$ can be chosen arbitrarily small. Equality for $\rho\equiv 1+x$ and inequality imply that that for smooth solutions $$\label{2.50} \\|u\\|_{X(Q_T)} + \\|u_{x}\big\|_{x=0}\\|_{L_2(B_T)} \leq c.$$ The end of the proof is standard. \[R2.3\] The method of construction of weak solution in Lemma \[L2.13\] via closure ensures that $u\big\|_{x=0}=u\big\|_{x=R}=0$ in the trace sense (this fact can be also easily derived from equality , since $u_x \in L_2(Q_T)$). Moreover, if $f\in L_2(Q_T)$ then according to Lemma \[2.12\] $u\in Y_0(Q_T)$ and, in particular, $\mu_1\equiv u_x\big\|_{x=0}$. \[L2.12\] Let $u_0\in \widetilde H^{(0,1)}$, $\mu_0=\nu_0=\nu_1\equiv 0$, $f\in L_2(Q_T)$. Then for the unique weak solution $u(t,x,y)\in X(Q_T)$ to problem , – $u_y\in C([0,T];L_2)$, $\|Du_y\|\in L_2(Q_T)$ and for any $t\in (0,T]$ $$\begin{gathered} \label{2.51} \iint u_y^2(t,x,y)\,dxdy + \int_0^t\!\! \iint \|Du_y\|^2\,dxdyd\tau \\ \leq (1+R)\iint u_{0y}^2\,dxdy + b\int_0^t\!\! \iint u_y^2 \, dxdyd\tau - 2\int_0^t\!\! \iint (1+x) f u_{yy}\,dxdyd\tau.\end{gathered}$$ Multiply by $-2(1+x)u_{yy}(t,x,y)$ and integrate over $\Omega$, then $$\begin{gathered} \frac{d}{dt} \iint (1+x)u_y^2\,dxdy + \int_0^L u_{xy}^2\big\|_{x=0}\,dy + \iint (3u_{xy}^2+u_{yy}^2- bu_y^2)\,dxdy \\ = -2 \iint (1+x)fu_{yy}\,dxdy,\end{gathered}$$ whence the assertion of the lemma obviously follows. \[L2.13\] Let $u_0\in \widetilde H^2$, $u_0\big\|_{x=0}=u_0\big\|_{x=R}=u_{0x}\big\|_{x=R}\equiv 0$ and $u_{0xxx},u_{0xyy} \in L_2$, $\mu_0=\nu_0=\nu_1\equiv 0$, $f\in C([0,T];L_2)$, $f_t\in L_2(0,T;H^{-1})$. Then for the (unique) weak solution to problem , – from the space $X(Q_T)$ there exists $u_t\in X(Q_T)$, which is the weak solution to problem of , – type, where $f$ is substituted by $f_t$, $u_0$ – by $\bigl(f\big\|_{t=0}-bu_{0x} -u_{0xxx}- u_{0xyy}\bigr)$, $\mu_0=\nu_0=\nu_1\equiv 0$. The proof for the function $v\equiv u_t$ is similar to Lemma \[L2.11\]. \[L2.14\] Let the hypothesis of Lemma \[L2.13\] be satisfied and, in addition, $f\in L_1(0,T;\widetilde H^{(0,2)})$. Then there exists a (unique) solution to problem , – from the space $X^{2}(Q_T)$ and for any $t\in [0,T]$ $$\begin{gathered} \label{2.52} \\|u\\|^2_{X^2(Q_t)} \leq c(T,b,R)\Bigl(\\|u_{0yy}\\|^2_{L_2} + \\|f\\|^2_{C([0,t];L_2)}+ \\|u\\|^2_{C([0,t];L_2)} + \\|u_t\\|^2_{C([0,t];L_2)} \\+ \sup\limits_{\tau\in (0,t]}\Bigl\|\int_0^{\tau} \!\!\iint (1+x)f_{yy}u_{yy}\,dxdyds\Bigr\| \Bigr).\end{gathered}$$ For smooth solutions differentiating equality twice with respect to $y$, multiplying the obtained equality by $2u_{yy}(t,x,y)\rho(x)$, $\rho(x)\equiv (1+x)$, and integrating over $\Omega$ we derive, that $$\begin{gathered} \label{2.53} \frac d{dt} \iint u_{yy}^2\rho\,dxdy +\int_0^L u_{xyy}^2\big\|_{x=0}\,dy + \iint (3u_{xyy}^2+u_{yyy}^2-bu^2_{yy})\,dxdy \\ = 2\iint f_{yy}u_{yy}\rho\,dxdy,\end{gathered}$$ whence obviously follows that $$\label{2.54} \\|u_{yy}\\|_{X(Q_T)} \leq c.$$ Hence, for the weak solution also $u_{yy}\in X(Q_T)$. Lemmas \[L2.11\] and \[L2.13\] provide, that $u,u_t \in X(Q_T)$. Write equality in the form $$\label{2.55} u_{xxx} = f-u_t -bu_x -u_{xyy}.$$ Then, inequality for $j=2$ and yield that $$\begin{gathered} \label{2.56} \\|u_{xx}\\|_{L_2} \leq c(R)\bigl(\\|u_{xxx}\\|_{H^{(-1,0)}} + \\|u\\|_{L_2}\bigr) \\ \leq c(b,R)\bigl(\\|f\\|_{L_2} + \\|u_t\\|_{L_2} + \\|u_{yy}\\|_{L_2} + \\|u\\|_{L_2}\bigr).\end{gathered}$$ Since $$\iint u_{xy}^2\,dxdy = \iint u_{xx}u_{yy}\,dxdy,$$ estimates and yield that $u\in C([0,T];\widetilde H^{2})$ and $$\label{2.57} \\|u(t,\cdot,\cdot)\\|_{\widetilde H^{2}} \leq c\bigl( \\|f\\|_{L_2} + \\|u_t\\|_{L_2} + \\|u_{yy}\\|_{L_2} + \\|u\\|_{L_2}\bigr).$$ Next, $$\iint u_{xxy}^2\,dxdy = \iint u_{xxx}u_{xyy}\,dxdy +\int_0^L (u_{xyy}u_{xx})\big\|_{x=0}\,dy$$ and inequality provides that $$\label{2.58} \iint u_{xxy}^2\,dxdy \leq \iint (u_{xxx}^2+u_{xyy}^2)\,dxdy + \int_0^L u^2_{xyy}\big\|_{x=0}\,dy + c\iint u_{xx}^2\,dxdy.$$ From equality we derive, that $$\label{2.59} \iint u_{xxx}^2\,dxdy \leq c\iint (f^2+u_t^2+b^2u_x^2+u_{xyy}^2)\,dxdy,$$ and combining , – finish the proof. \[L2.15\] Let the hypothesis of Lemma \[L2.13\] be satisfied and, in addition, $u_0\in \widetilde H^3$, $f\in L_2(0,T;\widetilde H^{(0,2)})$. Then there exists a (unique) solution to problem , – from the space $X^{3}(Q_T)$ and for any $t\in (0,T]$ $$\begin{gathered} \label{2.60} \\|u\\|_{X^{3}(Q_t)} \leq c(T,b,R,L)\bigl(\\|u_0\\|_{\widetilde H^{3}} + \\|f\\|_{C([0,t];L_2)} + \\|f\\|_{L_2(0,t;\widetilde H^{(0,2)})} \\ +\\|f_{t}\\|_{L_2(0,t;H^{-1})}\bigr).\end{gathered}$$ First of all note that hypotheses of Lemmas \[L2.11\] (for $f_1\equiv 0$), \[L2.13\] and \[L2.14\] are satisfied. Therefore, taking into account also Remark \[R2.3\] we derive for smooth solutions that $$\label{2.61} \\|u\\|_{X^2(Q_T)} + \\|u_x\big\|_{x=0}\\|_{L_2(B_T)} + \\|u_t\\|_{X(Q_T)} + \\|u_{tx}\big\|_{x=0}\\|_{L_2(B_T)} \leq c.$$ Next, differentiating equality twice with respect to $y$, multiplying the obtained equality by $-2u_{yyyy}(t,x,y)\rho(x)$, $\rho(x)\equiv (1+x)$ and integrating over $\Omega$ we derive similarly to that $$\begin{gathered} \label{2.62} \frac d{dt} \iint u_{yyy}^2\rho\,dxdy +\int_0^L u_{xyyy}^2\big\|_{x=0}\,dy + \iint (3u_{xyyy}+u_{yyyy}^2-bu^2_{yyy})\,dxdy \\ = - 2\iint f_{yy}u_{yyyy}\rho\,dxdy.\end{gathered}$$ Here $$\Bigl\|2\iint f_{yy}u_{yyyy}\rho\,dxdy \Bigr\| \leq \varepsilon \iint u_{yyyy}^2\,dxdy + \frac {(1+R)^2}\varepsilon\iint f_{yy}^2 \,dxdy,$$ where $\varepsilon>0$ can be chosen arbitrarily small, and equality yields that $$\label{2.63} \\|u_{yyy}\\|_{X(Q_T)} + \\|u_{xyyy}\big\|_{x=0}\\|_{L_2(B_T)} \leq c.$$ Again apply equality . Then it follows from that we have the suitable estimate on $u_{xxxy}$ in the space $L_2(Q_T)$. Similarly to $$\iint u_{xxyy}^2\,dxdy \leq \iint (u_{xxxy}^2+u_{xyyy}^2)\,dxdy + \int_0^L u^2_{xyyy}\big\|_{x=0}\,dy + c\iint u_{xxy}^2\,dxdy,$$ whence follows the suitable estimate on $u_{xxyy}$ in $L_2(Q_T)$ and, as a result, on $u_y$ in $L_2(0,T;\widetilde H^3)$. One more application of yields the estimate on $u_{xxxx}$ in $L_2(Q_T)$. Therefore, $$\label{2.64} \\|u\\|_{L_2(0,T;\widetilde H^4)} \leq c.$$ Consider the extensions of the functions $u$ and $f$ for $y\in (L,2L]$ and $y\in [-L,0)$ in the case a) by the even reflections through $y=L$ and $y=0$, in the case b) – by the odd ones, in the case c) – by the corresponding combination of these methods, in the case d) – by the periodic extension. Then the functions $u$ and $f$ remain smooth in the more wide domain $[0,T]\times [0,R] \times [-L,2L]$, and equality also remains valid. Let $\eta_L(y) \equiv \eta(1+y/L)\eta(2-y/L)$, $\widetilde u(t,x,y) \equiv u(t,x,y)\eta_L(y)$, $\widetilde f(t,x,y) \equiv f(t,x,y)\eta_L(y)$. Now we apply the inequality (see, e.g. [@LM]) for the domain $\widetilde\Omega=(0,R)\times \mathbb R^y$ $$\\|g\\|_{H^2(\widetilde\Omega)}\leq c\bigl(\\|\Delta g\\|_{L_2(\widetilde\Omega)}+ \\|g\big\|_{\partial\widetilde\Omega}\\|_{H^{3/2}(\mathbb R)}+ \\|g\\|_{H^1(\widetilde\Omega)}\bigr)$$ for the function $g\equiv \widetilde u_x$. Note that $g\big\|_{x=R}=0$ and $$\Delta_{x,y} \widetilde u_x = \widetilde f -\widetilde u_t -b\widetilde u_x +2u_{xy}\eta_L'+u_x\eta_L''.$$ It follows from (\[2.61\]) that $$\\|\Delta_{x,y} \widetilde u_x\\|_{C([0,T];L_2(\widetilde\Omega))} \leq c.$$ Moreover, by virtue of (\[2.61\]), (\[2.63\]) and embedding $H^2(\widetilde\Omega)\subset H^{3/2}(\{{x=0}\}\times \mathbb R^y)$ (see [@LM]) $$\begin{gathered} \\|u_x\big\|_{x=0}\\|_{C([0,T];H^{3/2}(\mathbb R))} \leq \\|u_{0x}\big\|_{x=0}\\|_{H^{3/2}(\mathbb R)}\\ +2\\|u_{tx}\big\|_{x=0}\\|_{L_2((0,T)\times \mathbb R)}^{1/2} \\|u_x\big\|_{x=0}\\|_{L_2(0,T;H^3(\mathbb R))}^{1/2} \leq c.\end{gathered}$$ Therefore, $$\label{2.65} \\|u_x\\|_{C([0,T];H^2)} \leq c.$$ Estimates , – provide the desired result. At the end of this section consider the particular case of problem , – in $Q_T$ for $\mu_0=\nu_0=\nu_1\equiv 0$, $f\equiv 0$. Denote its solution by $Pu_0$, then it succeeds from Lemma \[L2.10\] that the operator $P$ is linear and bounded from $L_2$ to $Y_0(Q_T)$. Moreover, it easily follows from that $$\label{2.66} \bigl\\|\partial_x(Pu_0)\big\|_{x=0}\bigr\\|_{L_2(B_T)} \leq \\|u_0\\|_{L_2}.$$ For the controllability purposes we need the following observability result. \[L2.16\] If condition holds, then there exists a constant $c=c(T,b,R,L)>0$, such that $$\label{2.67} \\|u_0\\|_{L_2} \leq c \bigl\\|\partial_x(Pu_0)\big\|_{x=0}\bigr\\|_{L_2(B_T)}.$$ In the smooth case multiplying by $2(T-t)u(t,x,y)$ and integrating over $Q_T$ we find, that $$\iiint_{Q_T} u^2\,dxdydt -T\iint u_0^2\,dxdy +\iint_{B_T} (T-t)u_x^2\big\|_{x=0}\,dydt =0,$$ whence follows, that $$\label{2.68} \iint u_0^2\,dxdy \leq \frac 1T \iiint_{Q_T} (Pu_0)^2\,dxdydt + \iint_{B_T} \bigl(\partial_x(Pu_0)\big\|_{x=0}\bigr)^2\,dydt.$$ By continuity this estimate can be extended to any $u_0\in L_2$. Now assume, that inequality is not true. Then there exists a sequence $\{u_{0n}\in L_2\}_{n\in\mathbb N}$ such that $$\label{2.69} \\|u_{0n}\\|_{L_2}=1\quad \forall\ n,\qquad \lim\limits_{n\to+\infty} \bigl\\|\partial_x(Pu_{0n})\big\|_{x=0}\bigr\\|_{L_2(B_T)}=0.$$ It follows from that the sequence $\{Pu_{0n}\}$ is bounded in $L_2(0,T;H^1)$. Moreover, equality provides that the sequence $\{\partial_t Pu_{0n}\}$ is bounded in $L_1(0,T;H^{-2})$ and the standard argument provides that $\{Pu_{0n}\}$ is precompact in $L_2(Q_T)$. Extract the subsequence $n'$, such that $\{Pu_{0n'}\}$ converges in $L_2(Q_T)$. It follows from , that $\{u_{0n'}\}$ converges in $L_2$ to a certain function $\widetilde u_0\in L_2$. Continuity of the operator $P$ and the second property yield, that $P\widetilde u_0\in Y_0(Q_T)$ verifies $\partial_x(P\widetilde u_{0})\big\|_{x=0}=0$. In particular, according to for any function $\phi\in L_2(0,T;\widetilde H^2)$, $\phi_t, \phi_{xxx}, \phi_{xyy}\in L_2(Q_T)$, $\phi\big\|_{t=T}=0$, $\phi\big\|_{x=0}=\phi\big\|_{x=R} =0$, the following equality holds: $$\label{2.70} \iiint_{Q_T}P\widetilde u_0(\phi_t+b\phi_x+\phi_{xxx}+\phi_{xyy})\,dxdydt +\iint \widetilde u_0\phi\big\|_{t=0}\,dxdy=0.$$ For any natural $l$ let $$\label{2.71} v_l(t,x)\equiv \int_0^L (P\widetilde u_0)(t,x,y)\psi_l(y)\,dy,\quad v_{0l}(x) \equiv \int_0^L \widetilde u_0(x,y)\psi_l(y)\,dy.$$ Let $\vartheta(t,x)$ be an arbitrary function, such that $\vartheta\in L_2(0,T; H^3(0,R)\cap H_0^1(0,R))$, $\vartheta_t\in L_2((0,T)\times (0,R))$, $\vartheta\big\|_{t=T}\equiv 0$. Choose $\phi(t,x,y)\equiv \vartheta(t,x)\psi_l(y)$, then it follows from , , that $$\label{2.72} \iint_{(0,T)\times (0,R)} v_l\bigl(\vartheta_t+ (b-\lambda_l)\vartheta_x +\vartheta_{xxx}\bigr)\,dxdt + \int_0^R v_{0l}\vartheta\big\|_{t=0}\,dx =0.$$ It means, that the function $v_l \in C([0,T];L_2(0,R))$, $v_{lx} \in C([0,R];L_2(0,T))$ is a weak solution in the rectangle $(0,T)\times (0,R)$ to an initial-boundary value problem $$\begin{gathered} \label{2.73} v_t +(b-\lambda_l)v_x +v_{xxx}=0,\\ \label{2.74} v\big\|_{t=0}= v_{0l},\quad v\big\|_{x=0}=v_x\big\|_{x=0}=v\big\|_{x=R}=v_x\big\|_{x=R}=0.\end{gathered}$$ But the obvious generalization of results from [@R] (in that paper the case of the equation $v_t+v_x+v_{xxx}=0$ was considered) shows that under condition (if $b-\lambda_l\leq 0$ there are no restrictions on $R$) $v_{0l}\equiv 0$ and, therefore, $\widetilde u_0\equiv 0$, which contradicts the fact, that $\\|\widetilde u_0\\|_{L_2}=1$. Existence of solutions {#S3} ====================== Consider an auxiliary equation $$\label{3.1} u_t+bu_x+u_{xxx}+u_{xyy}+(g(u))_x+(\psi(t,x,y)u)_x=f(t,x,y).$$ The notion of a weak solution to problem , – is similar to Definition \[D1.1\]. \[L3.1\] Let $g\in C^1(\mathbb R)$, $g(0)=0$, $\|g'(u)\|\leq c\ \forall u\in\mathbb R$, $\psi\in L_2(0,T;L_\infty)$, $u_0\in L_2$, $f\in L_1(0,T;L_2)$, $\mu_0=\nu_0\equiv 0$, $\nu_1\in L_2(B_T)$. Then problem , – has a unique weak solution $u\in X(Q_T)$. We apply the contraction principle. For $t_0\in(0,T]$ define a mapping $\Lambda$ on $X(Q_{t_0})$ as follows: $u=\Lambda v\in X(Q_{t_0})$ is a weak solution to a linear problem $$\label{3.2} u_t+bu_x+u_{xxx}+u_{xyy} =f-(g(v))_x -(\psi v)_x$$ in $Q_{t_0}$ with initial and boundary conditions –. Since $$\begin{gathered} \\|g(v)\\|_{L_2(Q_{t_0})}\leq c\\|v\\|_{C([0,t_0];L_2)}<\infty, \\ \\|\psi v\\|_{L_2(Q_{t_0})}\leq c\\|\psi\\|_{L_2(0,t_0;L_\infty)} \\|v\\|_{C([0,t_0];L_2)}<\infty,\end{gathered}$$ Lemma \[L2.11\] provides that the mapping $\Lambda$ exists. Moreover, for functions $v,\widetilde{v}\in X(Q_{t_0})$ $$\\|g(v)-g(\widetilde{v})\\|_{L_2(Q_{t_0})}\leq c\\|v-\widetilde{v}\\|_{L_2(Q_{t_0})} \leq ct_0^{1/2}\\|v-\widetilde{v}\\|_{C([0,t_0];L_2)},$$ $$\\|\psi(v-\widetilde{v})\\|_{L_2(Q_{t_0})} \leq c\\|\psi\\|_{L_2(0,t_0;L_\infty)} \\|v-\widetilde{v}\\|_{C([0,t_0];L_2)}.$$ As a result, according to inequality $$\\|\Lambda v-\Lambda\widetilde{v}\\|_{X(Q_{t_0})}\leq c(T)\omega(t_0)\\|v-\widetilde{v}\\|_{X(Q_{t_0})},$$ where $\omega(t_0)\to 0$ as $t_0\to +0$ and $\omega$ depends on the properties of continuity of the primitive of the function $\\|\psi(t,\cdot,\cdot)\\|_{L_\infty}^2$ on $[0,T]$. Since the constant in the right side of this inequality is uniform with respect to $u_0$ and $f$, one can construct the solution on the whole time segment $[0,T]$ by the standard argument. Now we pass to the results of existence in Theorem \[T1.1\]. First of all we make zero the boundary data in for the function $u$ itself. Let $$\label{3.3} \psi(t,x,y)\equiv J(t,x,y;\mu_0)\eta(3/2-2x/R) + J(-t,R-x,y;\nu_{0-})\eta(2x/R-1/2),$$ where $\nu_{0-}(t,y)\equiv \nu_0(-t,y)$, the functions $\mu_0$ and $\nu_0$ are extended to the whole strip $B$ in the class $H^{s/3,s}(B)$, such that $\mu_0\equiv 0$ for $t<-1$, $\nu_0\equiv 0$ for $t>T+1$ and the function $J(t,x,y;\mu)$ is the aforementioned in the proof of Lemma \[L2.10\] solution to an initial-boundary value problem in $\Pi_T^+ =(0,T)\times \Sigma_+$ for equation in the case $f\equiv 0$ with zero initial condition for $(x,y)\in \Sigma_+$, boundary condition for $(t,x)\in (0,T)\times\mathbb R_+$ and boundary condition , introduced in [@F17]. Then the results of [@F17] provide, that $$\label{3.4} \left\{ \begin{aligned} &\widetilde\psi \equiv\psi_t +b\psi_x+\psi_{xxx}+\psi_{xyy} \in C^\infty(\overline{Q}_T), \\ &\psi\in Y_0(Q_T)\cap L_2(0,T;W^1_\infty),\\ &\psi\big\|_{x=0}=\mu_0, \quad \psi\big\|_{x=R}=\nu_0, \quad \\|\psi_x\big\|_{x=R}\\|_{L_2(B_T)}\leq c\\|\nu_0\\|_{H^{1/3,1}(B_T)}. \end{aligned} \right.$$ Consider a function $$\label{3.5} U(t,x,y) \equiv u(t,x,y) -\psi(t,x,y).$$ Then $u\in X(Q_T)$ is a weak solution to problem – iff $U\in X(Q_T)$ is a weak solution to an initial-boundary value problem in $Q_T$ for an equation $$\label{3.6} U_t+bU_x+U_{xxx}+U_{xyy}+UU_x+(\psi U)_x=F\equiv f-\widetilde\psi- \psi\psi_x,$$ with initial and boundary conditions $$\label{3.7} U\big\|_{t=0} = U_0\equiv u_0 -\psi\big\|_{t=0},\quad U\big\|_{x=0}= U\big\|_{x=R} =0,\quad U_x\big\|_{x=R}=V_1\equiv \nu_1-\psi_x\big\|_{x=R}$$ and the same boundary conditions on $(0,T)\times (0,R)$ as . Note also that the functions $U_0$, $F$, $V_1$ satisfy the same assumptions as the corresponding functions $u_0$, $f$, $\nu_1$ in the hypothesis of the theorem. For $h\in (0,1]$ consider a set of initial-boundary value problems in $Q_T$ for an equation $$\label{3.8} U_t+bU_x+U_{xxx}+U_{xyy}+\left(g_h(U)\right)_x+(\psi U)_x = F$$ with boundary conditions and . $$g_h(u)\equiv\int_0^u\Bigl[\theta\eta(2-h\|\theta\|)+\frac{2\operatorname{sign}\theta}{h}\eta(h\|\theta\|-1)\Bigr]\,d\theta.$$ Note that $g_h(u)= u^2/2$ if $\|u\|\leq 1/h$, $\|g'_h(u)\|\leq 2/h\ \forall u\in\mathbb R$ and $\|g'_h(u)\|\leq 2\|u\|$ uniformly with respect to $h$. According to Lemma \[L3.1\], there exists a unique solution to this problem $U_h\in X(Q_T)$. Next, establish appropriate estimates for functions $U_h$ uniform with respect to $h$ (we drop the index $h$ in intermediate steps for simplicity). First, note that $g'(U)U_x, \psi U_x, \psi_x U, F \in L_1(0,T;L_2)$ and so the hypothesis of Lemma \[L2.11\] is satisfied (for $f_1\equiv 0$). Write down the analogue of equality for $\rho\equiv 1$, then: $$\label{3.9} \iint U^2 \,dxdy \leq \iint U_0^2\,dxdy +\int_0^t \!\!\iint \bigl(2F- 2(g(U))_x-\psi_x U \bigr)U\,dxdyd\tau.$$ Since $$\label{3.10} (g(U))_xU=\partial_x\Bigl(\int_0^U g'(\theta)\theta \,d\theta\Bigr)$$ we derive that $$\label{3.11} \iint (g(U))_x U \,dxdy = 0.$$ Therefore, since $\psi_x\in L_2(0,T;L_\infty)$ uniformly with respect to $h$ $$\label{3.12} \\|u_h\\|_{C([0,T];L_2)} \leq c.$$ Next, equalities and provide that for $\rho(x)\equiv (1+x)$ $$\begin{gathered} \label{3.13} \iint U^2\,dxdy + \int_0^t\!\! \iint (3U_x^2 +U_y^2)\,dxdyd\tau \leq (1+R)\iint U_0^2\,dxdy \\ +b \int_0^t \!\!\iint U^2\, dxdyd\tau +(1+R)\iint_{B_t} V_1^2\,dyd\tau +2\int_0^t \!\!\iint FU\rho\,dxdyd\tau \\ +\int_0^t \!\!\iint (\psi-\psi_x\rho)U^2\,dxdyd\tau +2\int_0^t \!\!\iint \Bigl(\int_0^U g'(\theta)\theta \,d\theta \Bigr) \,dxdyd\tau.\end{gathered}$$ Note that $$\label{3.14} \Bigl\|\int_0^U g'(\theta)\theta \,d\theta \Bigr\| \leq c\|U\|^3.$$ Applying interpolating inequality (here the exact value of the constant is indifferent), we obtain that $$\label{3.15} \iint \|U\|^3\,dxdy \leq c\iint U^2\,dxdy \Bigl(\iint \bigl(\|DU\|^2 +U^2\bigr)\,dxdy\Bigr)^{1/2}$$ Since the norm of the functions $u_h$ in the space $L_2$ is already estimated in , it follows from – that uniformly with respect to $h$ $$\label{3.16} \\|u_h\\|_{X(Q_T)} \leq c.$$ From equation itself, estimate and the well-known embedding $L_1\subset H^{-2}$, it follows that uniformly with respect to $h$ $$\label{3.17} \\|u_{ht}\\|_{L_1(0,T;H^{-3})}\leq c.$$ Estimates , by the standard argument provide existence of a weak solution to problem – $u\in L_\infty(0,T;L_2)\cap L_2(0,T;\widetilde H^1)$, as a limit of functions $u_h$ when $h\to +0$. Finally, since by virtue of (here the exact value of the constant is again indifferent) $$\begin{gathered} \label{3.18} \iiint_{Q_T} U^4\,dxdydt \leq c\int_0^T \\|U(t,\cdot,\cdot)\\|^2_{H^1} \\|U(t,\cdot,\cdot)\\|^2_{L_2}\,dt \\ \leq c\\|U\\|^2_{L_2(0,T;H^1)}\\|U\\|^2_{L_\infty(0,T;L_2)} <\infty\end{gathered}$$ and $$\label{3.19} \iiint_{Q_T} \psi^2 U^2\,dxdydt \leq \\|\psi\\|^2_{L_2(0,T;L_\infty)} \\|U\\|^2_{L_\infty(0,T;L_2)} <\infty,$$ it follows from Lemma \[L2.11\] (where $f_1\equiv U^2/2 +\psi U$), that after possible modification on a set of zero measure $U\in C([0,T];L_2)$. Result on uniqueness and continuous dependence of weak solutions succeeds from the following theorem. \[T3.1\] For any $T>0$ and $M>0$ there exist constant $c=c(T,M,b,R,L)$, such that for any two weak solutions $u(t,x,y)$ and $\widetilde u(t,x,y)$ to problem –, satisfying $\\|u\\|_{X(Q_T)}, \\|\widetilde u\\|_{X(Q_T)} \leq M$, with corresponding data $u_0, \widetilde u_0\in L_2$, $\mu_0, \widetilde\mu_0 ,\nu_0, \widetilde\nu_0 \in \widetilde H^{1/3,1}(B_T)$, $\nu_1,\widetilde\nu_1 \in L_2(B_T)$ $f, \widetilde f\in L_1(0,T;L_2)$ the following inequality holds: $$\begin{gathered} \label{3.20} \\|u -\widetilde u\\|_{X(Q_T)} \leq c\bigl( \\|u_0 - \widetilde u_0\\|_{L_2} + \\|\mu_0-\widetilde\mu_0\\|_{H^{1/3,1}(B_T)} \\+ \\|\nu_0-\widetilde\nu_0\\|_{H^{1/3,1}(B_T)} +\\|\nu_1-\widetilde\nu_1\\|_{L_2(B_T)}+ \\|f-\widetilde f\\|_{L_1(0,T;L_2)}\bigr).\end{gathered}$$ Let the function $\psi$ is defined by formula , the function $\widetilde\psi$ in a similar way for $\widetilde\mu_0$, $\widetilde\nu_0$ and $\Psi\equiv \psi-\widetilde\psi$. Then, in particular, $$\label{3.21} \\|\Psi\\|_{X(Q_T)} \leq c\bigl(\\|\mu_0-\widetilde\mu_0\\|_{H^{1/3,1}(B_T)} + \\|\nu_0-\widetilde\nu_0\\|_{H^{1/3,1}(B_T)}\bigr).$$ Let $U_0\equiv u_0-\widetilde u_0 - \Psi\big\|_{t=0}$, $F\equiv f-\widetilde f -(\Psi_t+b\Psi_x+\Psi_{xxx}+\Psi_{xyy})$, $V_1\equiv \nu_1-\widetilde\nu_1 -\Psi_x\big\|_{x=R}$, then $$\label{3.22} \\|U_0\\|_{L_2} \leq \\|u_0-\widetilde u_0\\|_{L_2} + c\bigl(\\|\mu_0-\widetilde\mu_0\\|_{H^{1/3,1}(B_T)} + \\|\nu_0-\widetilde\nu_0\\|_{H^{1/3,1}(B_T)}\bigr),$$ $$\begin{gathered} \label{3.23} \\|F\\|_{L_1(0,T;L_2)} \leq \\|f-\widetilde f\\|_{L_1(0,T;L_2)} \\ + c\bigl(\\|\mu_0-\widetilde\mu_0\\|_{H^{1/3,1}(B_T)} + \\|\nu_0-\widetilde\nu_0\\|_{H^{1/3,1}(B_T)}\bigr),\end{gathered}$$ $$\label{3.24} \\|V_1\\|_{L_2(B_T)} \leq \\|\nu_1-\widetilde\nu_1\\|_{L_2(B_T)} + c\bigl(\\|\mu_0-\widetilde\mu_0\\|_{H^{1/3,1}(B_T)} + \\|\nu_0-\widetilde\nu_0\\|_{H^{1/3,1}(B_T)}\bigr).$$ The function $U(t,x,y) \equiv u(t,x,y) -\widetilde u(t,x,y) -\Psi(t,x,y)$ is a weak solution to an initial-boundary value problem in $Q_T$ for an equation $$U_t+bU_x+U_{xxx}+U_{xyy} = F -(uu_x-\widetilde u \widetilde u_x)$$ with initial and boundary conditions , $$U\big\|_{t=0} =U_0,\qquad U\big\|_{x=0}= U\big\|_{x=R}=0, \qquad U_x\big\|_{x=R}=V_1.$$ Apply Lemma \[L2.11\] where $f_1\equiv -(u^2-\widetilde u^2)/2$. Note that similarly to $f_1\in L_2(Q_T)$. Therefore, we derive from that for $t\in (0,T]$ and $\rho(x)\equiv (1+x)$ $$\begin{gathered} \label{3.25} \iint U^2\,dxdy +\int_0^t\!\! \iint (3U_x^2+U_y^2)\,dxdyd\tau \leq (1+R)\iint U_0^2\,dxdy \\ +b\int_0^t\!\! \iint U^2\,dxdyd\tau + (1+R)\iint_{B_t} V_1^2\,dyd\tau + 2\int_0^t \!\!\iint FU\rho\, dxdyd\tau \\+ \int_0^t \!\!\iint (u^2-\widetilde u^2)(U\rho)_x\, dxdyd\tau\end{gathered}$$ Here $u^2-\widetilde u^2 = (u+\widetilde u)(U+\Psi)$ and by virtue of $$\begin{gathered} \iint \|u(U+\Psi)U_x\|\,dxdy \\ \leq c \Bigl(\iint u^4 \,dxdy \iint (U^4+\Psi^4) \,dxdy\Bigr)^{1/4} \Bigl(\iint U_x^2\,dxdy\Bigr)^{1/2} \\ \leq c_1 \\|u\\|^{1/2}_{H^1}\\|u\\|_{L_2}^{1/2} \Bigl[\Bigl(\iint \|DU\|^2\,dxdy\Bigr)^{3/4} \Bigl(\iint U^2\,dxdy\Bigr)^{1/4} + \iint U^2\,dxdy \\+ \Bigl(\iint \|DU\|^2\,dxdy\Bigr)^{1/2}\\|\Psi\\|^{1/2}_{H^1}\\|\Psi\\|_{L_2}^{1/2}\Bigr]\end{gathered}$$ and, therefore, $$\begin{gathered} \label{3.26} \int_0^t\!\! \iint \|u(U+\Psi)U_x\|\,dxdyd\tau \leq \varepsilon \int_0^t\!\! \iint \|DU\|^2\,dxdyd\tau + \int_0^t \\|\Psi\\|^2_{H^1}\,d\tau\\ + c(\varepsilon) \int_0^t \gamma(\tau) \iint (U^2+\Psi^2)\,dxdyd\tau,\end{gathered}$$ where $\gamma(t)\equiv 1+\\|u(t,\cdot,\cdot)\\|^2_{H^1}\\|u(t,\cdot,\cdot)\\|^2_{L_2}\in L_1(0,T)$ and $\varepsilon>0$ can be chosen arbitrarily small. Then estimates –, and inequality provide the desired result. Finally, consider regular solutions. \[L3.3\] Let $g(u)\equiv u^2/2$, $\mu_0=\nu_0=\nu_1\equiv 0$, the functions $u_0$ and $f$ satisfy the hypothesis of Theorem \[T1.2\], $\psi\in X^3(Q_T)$. Then problem , – has a unique solution $u\in X^3(Q_T)$. For $t_0\in(0,T]$, $v\in X^3(Q_{t_0})$ let $u=\Lambda v\in X^3(Q_{t_0})$ be a solution to a linear problem (for $g(v)\equiv v^2/2$), –. Apply Lemma \[L2.15\]. We have: $$\begin{gathered} \label{3.27} \\|vv_x+\psi v_x+\psi_x v\\|_{C[0,t_0];L_2)} \leq \\|u_0u_{0x}+\psi\big\|_{t=0}u_{0x}+\psi_x\big\|_{t=0}u_0\\|_{L_2} \\ + \\|(vv_x)_t+(\psi v)_{tx}\\|_{L_1(0,t_0;L_2)}\end{gathered}$$ and with the use of derive that $$\begin{gathered} \label{3.28} \\|u_0u_{0x}\\|_{L_2} \leq c\\|u_0\\|_{L_{\infty}}\\|u_{0x}\\|_{L_2} \leq c_1\\|u_0\\|_{\widetilde H^3}^2, \\ \label{3.29} \\|\psi\big\|_{t=0}u_{0x}+\psi_x\big\|_{t=0}u_0\\|_{L_2} \leq c\\|\psi\big\|_{t=0}\\|_{H^1}\\|u_0\\|_{W^1_\infty} \leq c\\|\psi\\|_{X^3(Q_T)} \\|u_0\\|_{\widetilde H^3};\end{gathered}$$ next, $$\label{3.30} \\|vv_{tx}\\|_{L_1(0,t_0;L_2)} \leq \int_0^{t_0} \\|v\\|_{L_\infty}\\|v_{tx}\\|_{L_2}\,dt \\ \leq ct_0^{1/2}\\|v\\|_{X^2(Q_{t_0})} \\|v\\|_{X^3(Q_{t_0})},$$ $$\label{3.31} \\|v_xv_{t}\\|_{L_1(0,t_0;L_2)} \leq \int_0^{t_0} \\|v_x\\|_{L_4}\\|v_t\\|_{L_4}\,dt \leq ct_0^{1/2}\\|v\\|_{X^2(Q_{t_0})} \\|v\\|_{X^3(Q_{t_0})}$$ and similarly $$\label{3.32} \\|(\psi v)_{tx}\\|_{L_1(0,t_0;L_2)} \leq ct_0^{1/2}\\|\psi\\|_{X^3(Q_{T})} \\|v\\|_{X^3(Q_{t_0})}.$$ Next, $$\label{3.33} \\|vv_{t}\\|_{L_2(Q_{t_0})} \leq \Bigl(\int_0^{t_0} \\|v\\|^2_{L_\infty}\\|v_{t}\\|^2_{L_2}\,dt\Bigr)^{1/2} \leq ct_0^{1/2}\\|v\\|_{X^2(Q_{t_0})} \\|v\\|_{X^3(Q_{t_0})},$$ $(vv_x)_{yy} = vv_{xyy} +2v_yv_{xy}+v_xv_{yy}$, where similarly to $$\label{3.34} \\|vv_{xyy}\\|_{L_2(Q_{t_0})} \leq ct_0^{1/2}\\|v\\|_{X^2(Q_{t_0})} \\|v\\|_{X^3(Q_{t_0})},$$ $$\label{3.35} \\|v_yv_{xy}\\|_{L_2(Q_{t_0})} \leq \Bigl(\int_0^{t_0} \\|v_y\\|^2_{L_4}\\|v_{xy}\\|^2_{L_4}\,dt \Bigr)^{1/2} \leq ct_0^{1/2}\\|v\\|_{X^2(Q_{t_0})} \\|v\\|_{X^3(Q_{t_0})}$$ and similar estimate holds for $v_xv_{yy}$. Finally, similarly to – $$\label{3.36} \\|(\psi v)_t\\|_{L_2(Q_{t_0})}+ \\|(\psi v)_{xyy}\\|_{L_2(Q_{t_0})} \\ \leq ct_0^{1/2}\\|\psi\\|_{X^3(Q_{T})} \\|v\\|_{X^3(Q_{t_0})}.$$ Moreover, the assumptions on the function $\psi$ ensure that the corresponding boundary conditions on the function $vv_x+(\psi v)_x$ are satisfied for $y=0$ and $y=L$. Therefore, the mapping $\Lambda$ exists and one can use estimate to derive inequalities $$\label{3.37} \\|\Lambda v\\|_{X^3(Q_{t_0})}\leq \widetilde c + ct_0^{1/2}\bigl(\\|\psi\\|_{X^3(Q_{T})}\\|v\\|_{X^3(Q_{t_0})} +\\|v\\|_{X^3(Q_{t_0})}^2\bigr),$$ $$\begin{gathered} \label{3.38} \\|\Lambda v - \Lambda\widetilde v\\|_{X^3(Q_{t_0})}\leq ct_0^{1/2} \Bigl(\\|\psi\\|_{X^3(Q_{T})}\\|v-\widetilde v\\|_{X^3(Q_{t_0})} \\ +\bigl(\\|v\\|_{X^3(Q_{t_0})}+ \\|\widetilde v\\|_{X^3(Q_{t_0})}\bigr) \\|v-\widetilde v\\|_{X^3(Q_{t_0})} \Bigr),\end{gathered}$$ where the constant $c$ depends on the parameters $T,b,R,L$ and the constant$\widetilde c$ also on the properties of functions $u_0$, $f$, $\psi$. Hence, existence of the unique solution to the considered problem in the space $X^3(Q_{t_0})$ on the time interval $[0,t_0]$, depending on $\\|u_0\\|_{\widetilde H^3}$, follows by the standard argument. Now establish the following a priori estimate: if $u\in X^3(Q_{T'})$ is a solution to the considered problem for some $T'\in (0,T]$, then $$\label{3.39} \\|u\\|_{X^3(Q_{T'})} \leq c,$$ where the constant $c$ depends on $T,b,R,L$ and the properties of the functions $u_0$, $f$, $\psi$ from the hypothesis of the present lemma. It is already known, that (see ) $$\label{3.40} \\|u\\|_{X(Q_{T'})} \leq c.$$ Apply Lemma \[L2.12\], then by virtue of for $\rho(x)\equiv 1+x$ $$\begin{gathered} \label{3.41} \iint u_y^2\,dxdy + \int_0^t\!\! \iint \|Du_y\|^2\,dxdyd\tau \leq (1+R)\iint u_{0y}^2\,dxdy \\ + b\int_0^t\!\! \iint u_y^2 \, dxdyd\tau - 2\int_0^t\!\! \iint \bigl(f-uu_x-(\psi u)_x\bigr) u_{yy}\rho\,dxdyd\tau.\end{gathered}$$ Here for arbitrary $\varepsilon>0$ $$\begin{gathered} \label{3.42} 2\int_0^t\!\!\iint uu_x u_{yy}\rho\,dxdyd\tau = \int_0^t\!\!\iint (u-u_x\rho)u_y^2\,dxdyd\tau \\ \leq c\int_0^t \Bigl(\iint (u_x^2+u^2)\,dxdy \iint u_y^4\,dxdy\Bigr)^{1/2}\,d\tau \leq \varepsilon\int_0^t\!\! \iint \bigl(\|Du_y\|^2+u_y^2\bigr)\,dxdyd\tau \\+ c(\varepsilon)\int_0^t \gamma(\tau)\iint u_y^2\,dxdy\,d\tau,\end{gathered}$$ where $\gamma\equiv\\|u(t,\cdot,\cdot)\\|^2_{\widetilde H^1}\in L_1(0,T')$, $$\begin{gathered} \label{3.43} 2\int_0^t\!\! \iint (\psi u)_x u_{yy}\rho\,dxdyd\tau \\ \leq \sup\limits_{t\in [0,T]} \\|\psi(t,\cdot,\cdot)\\|_{W^1_\infty} \int_0^t\!\!\Bigl(\iint u_{yy}^2\,dxdy \iint (u_x^2+u^2)\,dxdy\Bigr)^{1/2}\,d\tau\\ \leq \varepsilon \int_0^t\!\!\iint u_{yy}^2\,dxdyd\tau + c(\varepsilon)\\|\psi\\|^2_{X^3(Q_T)}\\|u\\|^2_{X(Q_{T'})}.\end{gathered}$$ Therefore, inequality yields that $$\label{3.44} \\|u_y\\|_{C([0,T'];L_2)} + \bigl\\|\|Du_y\|\bigr\\|_{L_2(Q_{T'})} \leq c.$$ Next, since the hypothesis of Lemma \[L2.13\] is fulfilled, write down the corresponding analogue of equality for the function $u_t$ and $\rho(x)\equiv 1+x$: $$\begin{gathered} \label{3.45} \iint u_t^2\,dxdy + \int_0^t \!\! \iint (3u_{tx}^2+u_y^2)\,dxdyd\tau \\ \leq (1+R)\iint \bigl(f-bu_x-u_{xxx}-u_{xyy}-uu_x-(\psi u)_x\bigr)^2\big\|_{t=0}\,dxdy +b\int_0^t \!\! \iint u_t^2\,dxdyd\tau \\ + 2\int_0^t \!\! \iint f_t u_t\rho\,dxdyd\tau +2\int_0^t \!\! \iint \bigl(uu_t+(\psi u)_t\bigr)(u_t\rho)_x\,dxdyd\tau.\end{gathered}$$ Here similarly to , for arbitrary $\varepsilon>0$ $$\begin{gathered} 2\int_0^t\!\! \iint uu_t(u_t\rho)_x\,dxdyd\tau = \int_0^t\!\! \iint (u-u_x\rho)u_t^2\,dxdyd\tau \\\leq \varepsilon\int_0^t\!\! \iint \bigl(\|Du_t\|^2+u_t^2\bigr)\,dxdyd\tau + c(\varepsilon)\int_0^t \gamma(\tau)\iint u^2_t\,dxdy\,d\tau,\end{gathered}$$ where $\gamma\equiv\\|u(t,\cdot,\cdot)\\|^2_{\widetilde H^1}\in L_1(0,T')$, $$\begin{gathered} 2\int_0^t\!\! \iint \psi_t u (u_t\rho)_x\,dxdyd\tau \\ \leq c\int_0^t\!\!\Bigl(\iint (u_{tx}^2+u_t^2)\,dxdy\Bigr)^{1/2} \Bigl(\iint \psi_t^4\,dxdy \iint u^4\,dxdy\Bigr)^{1/4}\,d\tau\\ \leq \varepsilon \int_0^t\!\!\iint (u_{tx}^2+u_t^2)\,dxdyd\tau +c(\varepsilon)\\|\psi_t\\|^2_{X(Q_T)}\\|u\\|^2_{X(Q_{T'})},\end{gathered}$$ and $$\begin{gathered} 2\int_0^t\!\! \iint \psi u_t (u_t\rho)_x\,dxdyd\tau = \int_0^t\!\! \iint(\psi-\psi_x\rho) u_t^2\,dxdyd\tau \\ \leq c\\|\psi\\|_{X^3(Q_T)} \int_0^t\!\!\iint u_t^2\,dxdyd\tau.\end{gathered}$$ Consequently, it follows from , that $$\label{3.46} \\|u_t\\|_{X(Q_{T'})} \leq c.$$ Now apply Lemma \[L2.14\], then inequality and estimates , and yield that for any $t\leq T'$ and$\rho(x)\equiv 1+x$ $$\begin{gathered} \label{3.47} \\|u\\|^2_{X^2(Q_t)} \leq c+ c\\|uu_x\\|^2_{C([0,t];L_2)}+c\\|(\psi u)_x\\|^2_{C([0,t];L_2)} \\ + c \sup\limits_{\tau\in (0,t]}\Bigl\|\int_0^\tau \!\! \iint \bigl(uu_x+(\psi u)_x\bigr)_{yy}u_{yy}\rho\,dxdyds\Bigr\|.\end{gathered}$$ Uniformly with respect to $t\in [0,T']$ for arbitrary $\varepsilon>0$ $$\begin{gathered} \\|uu_x\\|^2_{L_2} \leq c \\|u\\|^2_{\widetilde H^1}\\|u_x\\|^2_{L_4} \leq \varepsilon \bigl\\| \|Du_x\| \bigr\\|^2_{L_2} +c(\varepsilon)\bigl(\\|u_t\\|^6_{X(Q_{T'})}+\\|u\\|^6_{X(Q_{T'})}+1\bigr),\\ \\|(\psi u)_x\\|^2_{L_2} \leq \\|\psi\\|^2_{W^1_\infty} \\|u\\|^2_{\widetilde H^1} \leq c\bigl(\\|u_t\\|^2_{X(Q_{T'})}+ \\|u\\|^2_{X(Q_{T'})}\bigr);\end{gathered}$$ then, $$\iint (uu_x)_{yy}u_{yy}\rho\,dxdy = \frac12\iint (u_x\rho-u)u^2_{yy}\,dxdy + 2\iint u_yu_{xy}u_{yy}\rho\,dxdy,$$ where $$\begin{gathered} \int_0^t\!\!\iint \|u_yu_{xy}u_{yy}\|\,dxdyd\tau \leq \sup\limits_{t\in [0,T']}\iint u_y^2\,dxdy \int_0^t\!\!\Bigl(\iint (u_{xy}^4+u_{yy}^4)\,dxdy\Bigr)^{1/2}d\tau \\ \leq \varepsilon\int_0^t\!\! \iint \|D^3 u\|^2\,dxdyd\tau + c(\varepsilon)\bigl(\\|u_t\\|^2_{X(Q_{T'})}+ \\|u\\|^2_{X(Q_{T'})}\bigr)\int_0^t \iint \|D^2 u\|^2\,dxdyd\tau,\end{gathered}$$ $$\begin{gathered} \int_0^t\!\!\iint \|u-u_x\rho\|u_{yy}^2\,dxdyd\tau \leq c\sup\limits_{t\in [0,T']}\\|u\\|^2_{H^1} \int_0^t \Bigl(\iint u_{yy}^4\,dxdy\Bigr)^{1/2}\,d\tau \\ \leq \varepsilon\int_0^t\!\! \iint \|Du_{yy}\|^2\,dxdyd\tau + c(\varepsilon)\bigl(\\|u_t\\|^2_{X(Q_{T'})}+ \\|u\\|^2_{X(Q_{T'})}\bigr)\int_0^t \!\!\iint u_{yy}^2\,dxdy\,d\tau;\end{gathered}$$ finally, $(\psi u)_{xyy} = \psi_{xyy}u +2\psi_{xy}u_y +\psi_{yy}u_x +\psi_xu_{yy} +2\psi_yu_{xy} +\psi u_{xyy}$, where $$\begin{gathered} \int_0^t\!\! \iint \|\psi_{xyy}u u_{yy}\|\,dxdtd\tau \\ \leq \sup\limits_{t\in [0,T]}\\|\psi_{xyy}\\|_{L_2} \int_0^t \Bigl(\iint u^4\,dxdy \iint u_{yy}^4\,dxdy\Bigr)^{1/4}\,d\tau \\ \leq \varepsilon\int_0^t\!\!\iint \Bigl(\|D^3 u\|^2+\|D^2 u\|^2\bigr)\,dxdyd\tau + c(\varepsilon)\\|\psi\\|^2_{X^3(Q_T)}\\|u\\|^2_{X(Q_{T'})},\end{gathered}$$ $$\begin{gathered} \int_0^t\!\! \iint \|\psi_{xy}u_y u_{yy}\|\,dxdtd\tau \\ \leq \int_0^t \!\! \Bigl(\iint \psi_{xy}^4\,dxdy\Bigr)^{1/4} \Bigl(\iint u_y^2\,dxdy\Bigr)^{1/2} \Bigl(\iint u_{yy}^4\,dxdy\Bigr)^{1/4}\,d\tau \\ \leq \varepsilon\int_0^t\!\! \iint \Bigl(\|D^3 u\|^2+\|D^2 u\|^2\bigr)\,dxdyd\tau + c(\varepsilon)\\|\psi\\|^2_{X^3(Q_T)}\\|u\\|^2_{X(Q_{T'})}\end{gathered}$$ and similar estimate holds for the integral of $\psi_{yy}u_xu_{yy}$. The rest integrals are estimated in an obvious way. As a result, it follows from that $$\label{3.48} \\|u\\|_{X^2(Q_{T'})} \leq c.$$ Finally, apply Lemma \[L2.15\] on the basis of the already obtained estimates , , then inequality and estimates – applied to $v\equiv u$ provide similarly to that for any $t_0\in (0,T']$ $$\\|u\\|_{X^3(Q_{t_0})}\leq \widetilde c + ct_0^{1/2}\bigl(\\|\psi\\|_{X^3(Q_{T})}+\\|u\\|_{X^2(Q_{T'})}\bigr) \\|u\\|_{X^3(Q_{t_0})},$$ whence follows. Let $\psi\in Y^3(Q_T)\subset X^3(Q_T)$ be the solution to problem , – for $f\equiv 0$ (see Lemma \[L2.12\]). Introduce the function $U$ by formula and consider problem , , (here $\widetilde\psi\equiv 0$, $V_1\equiv 0$). Then the functions $\psi$, $F\sim f$ and $U_0\sim u_0$ satisfy the hypothesis of Lemma \[L3.3\] and the result is immediate. Large-time decay of small solutions {#S4} =================================== Consider the solution to problem – $u\in X(Q_T)\ \forall T$. Note that $u^2\in L_2(Q_T)$ (see, for example, ). Apply Lemma \[L2.11\], then equality for $f_1\equiv u^2/2$, $\rho\equiv 1$ and equality for $g(u)\equiv u^2/2$ yield similarly to , that $$\label{4.1} \\|u(t,\cdot,\cdot)\\|^2_{L_2} \leq \\|u_0\\|^2_{L_2}+\\|\nu_1\\|^2_{L_2(B_+)}\leq \epsilon_0^2\quad \forall\ t\geq 0.$$ Next, it follows from equality for $\rho\equiv 1+x$, that $$\begin{gathered} \label{4.2} \iint u^2\rho\,dxdy + \iint_{B_t} \mu_1^2\,dyd\tau + \int_0^t \!\!\iint (3u_x^2+u_y^2-bu^2)\,dxdyd\tau \\ = \iint u_0^2\rho\,dxdy + (1+R)\iint_{B_t} \nu_1^2\,dyd\tau+ \frac{2}3 \int_0^t\!\!\iint u^3\,dxdyd\tau.\end{gathered}$$ Since $u^3\in L_1(Q_T)$ equality provides the following inequality in a differential form: for a.e. $t>0$ $$\label{4.3} \frac{d}{dt}\iint u^2\rho\,dxdy + \iint (3u_x^2+u_y^2-bu^2)\,dxdy \leq (1+R)\int_0^L \nu_1^2\,dy+\frac{2}3 \iint u^3\,dxdy.$$ Next, we show that inequality implies the following one: $$\begin{gathered} \label{4.4} \frac{d}{dt}\iint u^2\rho\,dxdy +\frac{\varkappa}{1+R}\iint u^2\rho\,dxdy \\ +\delta\iint \Bigl[1-\frac1{\varepsilon_0}\\|u(t,\cdot,\cdot)\\|_{L_2}\Bigr](3u_x^2+u_y^2)\,dxdy \leq (1+R)\int_0^L \nu_1^2\,dy.\end{gathered}$$ where $\delta$, $\varkappa$ and $\epsilon_0$ are from the hypothesis of the theorem. First of all note, that in all cases inequality implies, that $$\label{4.5} \iint u_x^2\,dxdy \geq \frac {\pi^2}{R^2} \iint u^2\,dxdy.$$ Further consider different cases separately. In the cases b) and d) it follows from inequality , that $$\label{4.6} (1-\delta) \iint (3u_x^2+u_y^2)\,dxdy -b\iint u^2\,dxdy \geq \frac{\varkappa}{1+R}\iint u^2\rho\,dxdy.$$ Moreover, by virtue of and $$\begin{gathered} \label{4.7} \frac 23 \iint u^3\,dxdy \leq \frac {4R}{3\pi} \Bigl(\iint u_x^2\,dxdy\Bigr)^{3/4} \Bigl(\iint u_y^2\,dxdy\Bigr)^{1/4}\Bigl( \iint u^2\,dxdy\Bigr)^{1/2} \\ + \frac {4R^{3/2}}{3L^{1/2}\pi^{3/2}} \iint u_x^2\,dxdy\Bigl( \iint u^2\,dxdy\Bigr)^{1/2} \leq \frac{\delta}{\epsilon_0}\\|u(t,\cdot,\cdot)\\|_{L_2}\iint (3u_x^2+u_y^2)\,dxdy,\end{gathered}$$ and follows. In the case a) we also use an inequality $$\label{4.8} \iint u_y^2\, dxdy \geq \frac {\pi^2}{L^2} \iint u^2\,dxdy$$ and, therefore, obtain with the corresponding $\varkappa$. Then we can alternatively derive, that either similarly to $$\begin{gathered} \frac 23 \iint u^3\,dxdy \leq \frac {4R}{3\pi} \Bigl(\iint u_x^2\,dxdy\Bigr)^{3/4} \Bigl(\iint u_y^2\,dxdy\Bigr)^{1/4}\Bigl( \iint u^2\,dxdy\Bigr)^{1/2} \\ \leq \frac{4R}{3^{7/4}\pi}\\|u(t,\cdot,\cdot)\\|_{L_2}\iint (3u_x^2+u_y^2)\,dxdy,\end{gathered}$$ or $$\begin{gathered} \label{4.9} \frac 23 \iint u^3\,dxdy \leq \frac {4L}{3\pi} \Bigl(\iint u_x^2\,dxdy\Bigr)^{1/4} \Bigl(\iint u_y^2\,dxdy\Bigr)^{3/4}\Bigl( \iint u^2\,dxdy\Bigr)^{1/2} \\ \leq \frac{4L}{3^{5/4}\pi}\\|u(t,\cdot,\cdot)\\|_{L_2}\iint (3u_x^2+u_y^2)\,dxdy,\end{gathered}$$ whence follows. In the case c) inequality must be substituted by the following one: $$\iint u_y^2\, dxdy \geq \frac {\pi^2}{4L^2} \iint u^2\,dxdy.$$ Similar modification must be done in and in this case also follows. Inequalities and imply, that $$\frac{d}{dt}\iint u^2\rho\,dxdy +\frac{\varkappa}{1+R}\iint u^2\rho\,dxdy \leq (1+R)\int_0^L \nu_1^2\,dy,$$ whence easily succeeds. Boundary controllability {#S5} ======================== First establish the result on boundary controllability for the linear equation. \[T5.1\] Let condition be satisfied for any natural $l$, such that $\lambda_l <b$. Let $T>0$, $f\equiv 0$, $\mu_0=\nu_0\equiv 0$. Then for any $u_0, u_T \in L_2$ there exists a function $\nu_1\in L_2(B_T)$, such that there exists a unique solution $u\in Y_0(Q_T)$ to problem , –, satisfying . Assume first that $u_0\equiv 0$. In the case $\nu_1\in L_2(B_T)$, $u_0\equiv 0$, $\mu_0=\nu_0\equiv 0$, $f\equiv 0$ denote the solution $u\in Y_0(Q_T)$ to problem , – by $P_1\nu_1$. Then Lemma \[L2.10\] provides, that $P_1$ is the linear bounded operator from $L_2(B_T)$ to $Y_0(Q_T)$. Let $P_{1T}\nu_1 \equiv P_1\nu_1\big\|_{t=T}$, then $P_{1T}$ is the linear bounded operator from $L_2(B_T)$ to $L_2$. Consider also the backward problem in $Q_T$ $$\begin{gathered} \label{5.1} \phi_t+b\phi_x+\phi_{xxx}+\phi_{xyy}=0,\\ \label{5.2} \phi\big\|_{t=T}=\phi_0(x,y), \quad \phi\big\|_{x=0}=\phi_x\big\|_{x=0}=\phi\big\|_{x=R}=0\end{gathered}$$ with corresponding boundary conditions of type, which after change of variables $(t,x,y)\to (T-t,R-x,y)$ transforms to the corresponding problem of , – type. In particular, if we denote $\phi=\widetilde P\phi_0$, then $\widetilde P$ is the linear bounded operator from $L_2$ to $Y_0(Q_T)$. Moreover estimates , yield, that for $\Lambda\phi_0 \equiv \partial_x(\widetilde P\phi_0)\big\|_{x=R}$ $$\label{5.3} \\|\Lambda\phi_0\\|_{L_2(B_T)} \leq \\|\phi_0\\|_{L_2} \leq c \\|\Lambda\phi_0\\|_{L_2(B_T)}.$$ In the smooth case multiplying equation by $P_1\nu_1$ and integrating over $Q_T$ one can easily derive an equality $$\label{5.4} \iint P_{1T}\nu_1\cdot\phi_0\,dxdy = \iint_{B_T} \nu_1\cdot\Lambda\phi_0\,dydt.$$ By continuity this equality can be extended to the case $\nu_1\in L_2(B_T)$, $\phi_0\in L_2$. Let $A \equiv P_{1T}\circ\Lambda$, then according to and the aforementioned properties of the operator $P_{1T}$ the operator $A$ is bounded in $L_2$. Moreover, and provide, that $$(A\phi_0,\phi_0) = \iint (P_{1T}\circ\Lambda)\phi_0\cdot \phi_0\,dxdy = \iint_{B_T} (\Lambda\phi_0)^2\,dydt \geq \frac 1{c^2} \\|\phi_0\\|^2_{L_2}.$$ Application of Lax–Milgram theorem implies, that $A$ is invertible and $A^{-1}=\Lambda^{-1}\circ P_{1T}^{-1}$ is bounded in $L_2$. Let $$\label{5.5} \Gamma \equiv \Lambda\circ A^{-1} = P_{1T}^{-1}$$ (linear bounded operator from $L_2$ to $L_2(B_T)$), then $\nu_1\equiv \Gamma u_T$ and $u\equiv P_1\nu_1$ provide the desired solution in the case $u_0\equiv 0$. In the general case the solution is given by the formula $$\label{5.6} \nu_1\equiv \Gamma(u_T-Pu_0\big\|_{t=T}),\quad u\equiv Pu_0 +P_1\nu_1$$ (remind that $Pu_0$ is the solution to problem , – for $\mu_0=\nu_0=\nu_1\equiv 0$, $f\equiv 0$). Now we can prove Theorem \[T1.4\]. Consider first linear problem , –. Let $u_0\equiv 0$, $\mu_0=\nu_0=\nu_1\equiv 0$, $f\equiv f_{1x}$, $f_1\in L_2(Q_T)$. Let $P_2f_1\in X(Q_T)$ be the solution to this problem, existing by virtue of Lemma \[L2.11\]. In particular, estimate yields, that $P_2$ is the linear bounded operator from $L_2(Q_T)$ to $X(Q_T)$. Obviously, a solution $\nu_1\in L_2(B_T)$, $u\in X(Q_T)$ to the controllability problem $$\begin{gathered} u_t+bu_x+u_{xxx}+u_{xyy}=f_{1x},\quad f_1\in L_2(Q_T),\\ u\big\|_{t=0}=u_0\in L_2,\quad u\big\|_{t=T}=u_T\in L_2, \quad u\big\|_{x=0}=u\big\|_{x=R}=0,\quad u_x\big\|_{x=R}=\nu_1\end{gathered}$$ is given by the formula $$\label{5.7} \nu_1\equiv \Gamma\bigl(u_T-Pu_0\big\|_{t=T}-P_2f_1\big\|_{t=T}\bigr), \quad u\equiv Pu_0 +P_1\nu_1 +P_2f_1.$$ The solution to the original problem is constructed as a fixed point of the map $$\label{5.8} u=\Theta v\equiv Pu_0 + (P_1\circ\Gamma)\bigl(u_T-Pu_0\big\|_{t=T}+P_2(v^2/2)\big\|_{t=T}\bigr) -P_2(v^2/2),$$ defined on $X(Q_T)$. Similarly to $$\begin{gathered} \\|v^2\\|_{L_2(Q_T)} \leq c\\|v\\|^2_{X(Q_T)}, \\ \\|v^2-\widetilde v^2\\|_{L_2(Q_T)} \leq c\bigl(\\|v\\|_{X(Q_T)}+\\|\widetilde v\\|_{X(Q_T)}\bigr) \\|v-\widetilde v\\|_{X(Q_T)}.\end{gathered}$$ Therefore, $$\begin{gathered} \\|\Theta v\\|_{X(Q_T)} \leq c\bigl(\\|u_0\\|_{L_2} + \\|u_T\\|_{L_2} + \\|v\\|^2_{X(Q_T)}\bigr),\\ \\|\Theta v -\Theta \widetilde v\\|_{X(Q_T)} \leq c\bigl(\\|v\\|_{X(Q_T)}+\\|\widetilde v\\|_{X(Q_T)}\bigr) \\|v-\widetilde v\\|_{X(Q_T)}\end{gathered}$$ and the standard contraction argument provides the desired result. [99]{} E. S. Baykova and A. V. Faminskii, [On initial-boundary-value problems in a strip for the generalized two-dimensional Zakharov–Kuznetsov equation]{}, Adv. Differential Equ. [18]{} (2013), 663-686. H. A. Biagioni and F. Linares, Well-posedness for the modified Zakharov–Kuznetsov equation, Progr. Nonlinear Differential Equ. Appl. 54 (2003), 181–189. E. Bustamante, J. Jimenez and J. Mejia, The Zakharov–Kuznetsov equation in weighted Sobolev spaces, J. Math. Anal. Appl. 433 (2016), 149–175. G. G. Doronin and N. A. Larkin, [Stabilization of regular solutions for the Zakharov–Kuznetsov equation posed on bounded rectangles and on a strip]{}, Proc. Edinburgh Math. Soc. [58]{} (2015), 661–682. A. V. Faminskii, [The Cauchy problem for quasilinear equations of odd order]{}, Mat. Sb. [180]{} (1989), 1183–1210. English transl. in Math. USSR-Sb. [68]{} (1991), 31–59. A. V. Faminski, [The Cauchy problem for the Zakharov–Kuznetsov equation]{}, Differ. Uravn. [31]{} (1995), 1070–1081. English transl. in Differential Equ. [31]{} (1995), 1002–1012. A . V. Faminskii, [On the mixed problem for quasilinear equations of the third order]{}, J. Math. Sci. [110]{} (2002), 2476–2507. A. V. Faminskii, [On the nonlocal well-posedness of a mixed problem for the Zakharov–Kuznetsov equation]{}, J. Math. Sci. [147]{} (2007), 6524–6537. A. V. Faminskii, [Global well-posedness of two initial-boundary-value problems for the Korteweg–de Vries equation]{}, Differential Integral Equ. [20]{} (2007), 601–642. A. V. Faminskii, [Well-posed initial-boundary value problems for the Zakharov–Kuznetsov equation]{}, Electronic J. Differential Equ. No. 127 (2008), 1–23. A. V. Faminskii and I. Yu. Bashlykova, [Weak solutions to one initial-boundary value problem with three boundary conditions for quasilinear equations of the third order]{}, Ukrainian Math. Bull. [5]{} (2008), 83–98. A. V. Faminskii, [Weak solutions to initial-boundary-value problems for quasilinear evolution equations of an odd order]{}, Adv. Differential Equ. [17]{} (2012), 421–470. A. V. Faminskii, An initial-boundary value problem in a strip for two-dimensional Zakharov–Kuznetsov–Burgers equation, Nonlinear Analysis 116 (2015), 132–144. A. V. Faminskii, [An initial-boundary value problem in a strip for two-dimensional equations of Zakharov–Kuznetsov type]{}, Contemp. Math. [653]{} (2015), 137–162. A. V. Faminski, [Initial-boundary value problems in a half-strip for two-dimensional Zakharov–Kuznetsov equation]{}, arXiv: 1703.05660v1 \[math.AP\] 16 Mar 2017. L. G. Farah, F. Linares and A. Pastor, [A note on the 2D generalized Zakharov–Kuznetsov equation: local, global and scattering results]{}, J. Differential Equ. [253]{} (2012), 2558–2571. G. Fonseca and M. Panchón, Well-posedness for the two dimensional generalized Zakharov–Kuznetsov equation in anisotropic weighted Sobolev spaces, J. Math. Anal. Appl. 443 (2016), 566–584. A. Grünrock, On the generalized Zakharov–Kuznetsov equation at critical regularity, arXiv: 1509.09146v1 \[math.AP\] 30 Sep 2015. A. Grünrock and S. Herr, The Fourier restriction norm method for the Zakharov–Kuznetsov equation, Discrete Cont. Dyn. Syst. (A) 34 (2014), 2061–2068. D. Han-Kwan, From Vlasov–Poisson to Korteweg–de Vries and Zakharov–Kuznetsov, Comm. Math. Phys. 324 (2013), 961–993. O. A. Ladyzhenskaya, V. A. Solonnikov and N. N. Uraltseva, [Linear and quasilinear equations of parabolic type]{}, Trans. of Math. Monogr. [23]{}, American Math. Soc., Providence, R.I., 1968. D. Lannes, F. Linares and J.-C. Saut, [The Cauchy problem for the Euler-Poisson system and derivation of the Zakharov–Kuznetsov equation]{}, Progress Nonlinear Differential Equ. Appl. [84]{} (2013), 183–215. N. A. Larkin, [Exponential decay of the $H^1$-norm for the 2D Zakharov–Kuznetsov equation]{}, J. Math. Anal. Appl. [405]{} (2013), 326–335. N. A. Larkin, The 2D Zakharov–Kuznetsov–Burgers equation with variable dissipation on a strip, Electronic J. Differential Equ. (2015), no. 60, 1–20. N. A. Larkin, The 2D Zakharov–Kuznetsov–Burgers equation on a strip, Bol. Soc. Parana Mat. (3) 34 (2016), 151–172. N. A. Larkin and E. Tronco, [Regular solutions of the 2D Zakharov–Kuznetsov equation on a half-strip]{}, J. Differential Equ. [254]{} (2013), 81–101. F. Linares and A. Pastor, [Well-posedness for the two-dimensional modified Zakharov–Kuznetsov equation]{}, SIAM J. Math. Anal. [41]{} (2009), 1323–1339. F. Linares and A. Pastor, [Well-posedness for the 2D modified Zakharov–Kuznetsov equation]{}, J. Funct. Anal., [260]{} (2011), 1060–1085. F. Linares, A. Pastor and J.-C. Saut, [Well-posedness for the Zakharov–Kuznetsov equation in a cylinder and on the background of a KdV soliton]{}, Comm. Partial Differential Equ., [35]{} (2010), 1674–1689. J.-L. Lions and E. Magenes, Problèmes aux limites non homogènes et applications, Dunod, Paris, 1968. L. Molinet and D. Pilod, Bilinear Strichartz estimates for the Zakharov–Kuznetsov equation and applications, Ann. Inst. H. Poincare (C) Analyse Non Linéaire 32 (2015), 347–371. F. Ribaud and S. Vento, [A note on the Cauchy problem for the 2D generalized Zakharov–Kuznetsov equation]{}, C. R. Acad. Sci. Paris [350]{} (2012), 499–503. L. Rosier, [Exact boundary controllability for the Korteweg–de Vries equation on a bounded domain]{}, ESAIM: Control, Optimization Calculus Variations [2]{} (1997), 33–55. J.-C. Saut, [Sur quelques generalizations de l’equation de Korteweg–de Vries]{}, J. Math. Pures Appl., [58]{} (1979), 21–61. J.-C. Saut and R. Temam, [An initial boundary value problem for the Zakharov–Kuznetsov equation]{}, Adv. Differential Equ., [15]{} (2010), 1001–1031. J.-C. Saut, R. Temam and C. Wang, [An initial and boundary-value problem for the Zakharov–Kuznetsov equation in a bounded domain]{}, J. Math. Phys. [53]{} (2012), 115612. V.E. Zakharov and E.A. Kuznetsov, [On three-dimensional solutions]{}, Zhurnal Eksp. Teoret. Fiz., [66]{} (1974), 594–597. English transl. in Soviet Phys. JETP, [39]{} (1974), 285–288. [^1]: The publication was financially supported by the Ministry of Education and Science of the Russian Federation (the Agreement 02.A03.21.0008 and the Project 1.962.2017/PCh)	{ "pile_set_name": "ArXiv" }
--- abstract: 'It is well known that the distributions of hitting times in Markov chains are quite irregular, unless the limit as time tends to infinity is considered. We show that nevertheless for a typical finite irreducible Markov chain and for nondegenerate initial distributions the tails of the distributions of the hitting times for the states of a Markov chain can be ordered, i.e., they do not overlap after a certain finite moment of time. If one considers instead each state of a Markov chain as a source rather than a sink then again the states can generically be ordered according to their efficiency. The mechanisms underlying these two orderings are essentially different though. Our results can be used, e.g., for a choice of the initial distribution in numerical experiments with the fastest convergence to equilibrium/stationary distribution, for characterization of the elements of a dynamical network according to their ability to absorb and transmit the substance (“information”) that is circulated over the network, for determining optimal stopping moments (stopping signals/words) when dealing with sequences of symbols, etc.' address: 'School of Mathematics, Georgia Tech, Atlanta GA, 30332-0160, USA' author: - Yuri Bakhtin - Leonid Bunimovich bibliography: - 'mc.bib' title: The optimal sink and the best source in a Markov chain --- Introduction ============ Hitting and recurrence times are a classical subject in the theory of random processes. However, the relevant studies have always been concerned with averages (expectations) of hitting and recurrence times and relations between their distributions for a fixed state [@Kac:MR0022323],[@Galves-Schmitt:MR1483874],[@Hirata-Saussol-Vaienti:MR1736991],[@Lacroix:MR1952624],[@Abadi:MR2040782],[@Haydn-Vaienti:MR2018869],[@Haydn-Lacroix-Vaienti:MR2165587],[@Kupsa-Lacroix:MR2123204],[@Chazottes-Ugalde:MR2151722]. It is well-known that the distributions of recurrence times are quite regular for many random processes and dynamical systems [@Hirata-Saussol-Vaienti:MR1736991],[@Haydn-Vaienti:MR2018869],[@Haydn-Lacroix-Vaienti:MR2165587],[@Chazottes-Ugalde:MR2151722]. On the other hand, the distribution functions of the first hitting times are very irregular [@Galves-Schmitt:MR1483874],[@Lacroix:MR1952624],[@Abadi:MR2040782],[@Haydn-Lacroix-Vaienti:MR2165587],[@Kupsa-Lacroix:MR2123204],[@Chazottes-Ugalde:MR2151722],[@Keller-Liverani:MR2535206]. This seems to be natural because an ergodic process returns to any set of positive measure infinitely many times with probability 1, while the hitting event occurs only once. Therefore, our first result that for a typical irreducible Markov chain and for a typical initial distribution, the distribution tails of the first hitting times for the states of the chain can be ordered is quite surprising. This striking regularity property means that there is a finite moment of time $n_0$ such that the tails of the survival probabilities $P_i(n)$, $n\ge n_0$, $i=1,2,\ldots,N$, form an ordered set, i.e., $P_{\sigma_1}(n)<P_{\sigma_2}(n)<\ldots<P_{\sigma_N}(n)$ for all $n\ge n_0$, where $\sigma_i\in \{1,2,\ldots,N\}$ for all $i$. From this point of view, the state $\sigma_1$ is the most efficient sink (absorber of “information”) out of all the states of the Markov chain. The question of the choice of the best (worst) sink naturally arises in the theory of dynamical networks. A dynamical network is a dynamical system that is generated by individual dynamics of its elements (cells, power stations, neurons, etc.), the interactions between these elements and the structure of the graph of interactions (often called the topology of the network). These three characteristics determine the long term dynamics of a network [@Dynamical-networks:MR2335082]. Traditionally, the theory of dynamical systems deals with asymptotic in time ($t\to\infty$) properties. It has been found though recently [@Yurchenko] that it is also possible to effectively answer some natural questions on finite time dynamics. For instance, placing a hole in a proper place in the phase space of chaotic dynamical systems guarantees that survival probabilities for this hole for all times $n\ge n_0$ are smaller than for other holes of the same size (measure). However, it was recently discovered [@Yurchenko],[@Which-hole:MR2593912] that it is possible to make finite time predictions of dynamics even if there are no small/large parameters in equations which govern dynamics of a system. The results of the present paper (as the ones in [@Yurchenko],[@Which-hole:MR2593912]) are not only generally unexpected but often counterintuitive as well. For instance we provide the examples where the best sink or source is not the state with the maximal equilibrium/stationary probability. It is always tempting and important to try to characterize elements of networks by their ability to absorb and transmit “information”. By combining the ideas and approaches of [@Yurchenko],[@Dynamical-networks:MR2335082] it was shown in [@Which-hole:MR2593912] that, indeed, one can characterize the elements of networks by their ability to leak “information” out of the system. Thus the elements of networks could be characterized by their dynamical properties rather than by standard static characteristics like centrality, betweenness, etc., which are based only on the topology of a network, rather than on its dynamics. Typically, chaotic dynamical systems, even the most chaotic ones, have a fast decaying but still infinite memory. Therefore, the studies of statistical properties of dynamical systems always make use of results of the probability theory and, if needed, require to prove some modifications of the existing limit theorems, etc. It is a very natural approach because a memory in chaotic dynamical systems is (most often) infinite and such systems are approximated by random processes with a finite memory. However, even for such random processes standard approach is to analyze only their asymptotic in time properties. We show here, though, that some interesting finite time properties of random processes can also be rigorously studied. For instance, our results show that for hitting times one can find not only relations between their averages, but also between their distribution functions. It occurred that the infinite tails of these distribution never overlap after a finite moment of time that can be effectively computed. Our results also generalize those of [@Yurchenko],[@Which-hole:MR2593912] to an essentially larger class of dynamical systems. However, the question that we address in this paper seems to have never been considered even in the theory of Markov chains. Our results show that for hitting times one can find not only relations between their averages, but even between their distributions. Another problem considered in this paper is to find the most efficient source in a Markov chain. To the best of our knowledge, this problem has not been addressed before. It is also motivated by the dynamical networks where the following question is of utmost importance: which node (element of a network) one should apply a perturbation to, in order to achieve the strongest effect? We show that for a typical irreducible Markov chain, there also exists a hierarchy of its states with respect to the rate at which the initial perturbation converges to the stationary state. Thus one can find an optimal node to apply perturbation to in order to achieve the fastest relaxation. And again typically there exists a finite moment of time after which the states of a Markov chain form an ordered set with respect to their ability to transmit information to the entire chain (network) or to serve as sources. Generalizations to the case when a sink/source consists not of one but of several states of a Markov chain are straightforward. Another straightforward (although important for applications where one deals, e.g., with a network of chemical reactions, supply chains, etc) generalization deals with nonnegative (rather with transition probabilities) matrices and uses Perron-Frobenius theorem instead of the Markov theorem. The results of our paper could be used e.g. for choosing an appropriate (e.g. the fastest convergent) initial distribution in computer experiment, for choosing an appropriate sequence of stopping/observing times when dealing with the sequences of symbols and for dynamical characterization of the elements of networks. These finite time probabilistic predictions allowed to realize that some natural basic questions have never been addressed in the theory of stochastic processes and even for Markov chains. This gap should be filled in. Most efficient sink =================== It is intuitively clear that for most Markov chains some of the states are more important for the dynamics than the others. The goal of this section is to introduce and study a measure of importance of the states based on the escape rate through a state (or a family of states, since this generalization of our approach is straightforward). Let $P=(P_{ij})_{i,j=1}^N$ be the transition probability matrix of an irreducible Markov chain (see, e.g., [@Feller:MR0228020 Chapter XV]), on state space $\{1,\ldots,N\}$ for some $N\in{\mathbb{N}}$. Let us fix $k\in\{1,\ldots,N\}$ and stop our Markov chain as soon as it reaches state $k$. In other words, whenever the original Markov chain makes a transition to $k$, it gets killed, so that the state $k$ can be considered as a cemetery state for the Markov chain, or a hole through which the mass leaks out of the system. There are at least two equivalent ways one can describe the resulting dynamics with. One is to treat the new system as a new Markov chain with absorbing state $k$ and introduce the associated transition matrix $P^{(k)}$ by $$P^{(k)}_{ij}=\begin{cases} P_{ij},& i\ne k,\\ 1,& j=i=k,\\ 0,& j\ne i=k. \end{cases}$$ Another way is to introduce a matrix $Q^{(k)}=(Q_{ij}^{(k)})_{i,j\ne k}$ obtained from $P$ by crossing out its $k$-th row and column. The matrix $P^{(k)}$ is a stochastic matrix whereas $Q^{(k)}$ is strictly substochastic (or sub-Markov) since it does not account for the mass leaking out through the state $k$. We assume that even after the removal of an arbitrary state $k$ the system remains irreducible and aperiodic, i.e., for some $n_0=n_0(k)$ and all $n>n_0$, all entries of the matrix $(Q^{(k)})^n$ are positive. The aperiodicity assumption is standard, see, e.g., [@Feller:MR0228020 Section XV.9], and we make it to avoid unnecessary although straightforward technicalities. Let us denote the simplex of all probability distributions on $\{1,\ldots,N\}$ by $\Delta_N$. Suppose we are given the initial distribution $p=(p_1,\ldots,p_N)\in\Delta_N$. After $n$ steps, the distribution of the Markov chain with a hole at state $k$ is given by $p(P^{(k)})^n$. The irreducibility of $P$ implies that, as $n\to\infty$, this distribution converges to the one concentrated at $k$. This is the only stationary distribution, i.e., the only eigenvector corresponding to the leading eigenvalue 1 of the stochastic matrix $P^{(k)}$. The rate of convergence to this obvious equilibrium is characterized by the second largest eigenvalue, $\mu_k<1$. It is easy to see that the spectrum of $P^{(k)}$ coincides with that of $Q^{(k)}$ except for a simple eigenvalue 1. Therefore, $\mu_k$ is also the leading positive eigenvalue of matrix $Q^{(k)}$. In our setting, the classical Perron–Frobenius (PF) theorem guarantees that $\mu_k$ is simple and there is an associated eigenvector $q^{(k)}=(q^{(k)}_i)_{i\ne k}$ with all positive coordinates. We can choose $q^{(k)}$ so that besides the equality $$\label{eq:q-k-quasistationary} q^{(k)} Q^{(k)}=\mu_kq^{(k)},$$ it satisfies $$\sum_{i\ne k} q^{(k)}_i=1,$$ thus defining a probability distribution. Notice that is exactly the definition of a quasi-stationary distribution for the sub-Markov kernel $Q^{(k)}$. Since the matrix is sub-Markov, there is no stationary distribution, and the total mass of a vector $qQ^{(k)}$ may be less than 1 for a probability vector $q$. However, if we normalize the distribution to have total mass 1 after each step then we end up with the notion of quasi-stationary distributions defined by . This equation means that under the stationary distribution, the total mass that has not leaked through $k$ multiplies by $\mu_k<1$ at every step. Therefore, $\lambda_k=-\ln \mu_k$ can serve as the escape rate through $k$. It can happen that $\mu_k=0$, in this case, all mass escapes the system in finitely many steps, and we set $\lambda_k=\infty$. For $q=(q_i)_{i\ne k}$, we define $M_n^{(k)}(q)$ to be the survival probability, or the total mass remaining in the sub-Markov chain defined by $Q^{(k)}$ after $n$ steps: $$M^{(k)}_{n}(q)=\sum_{i\ne k} (q(Q^{(k)})^n)_i.$$ If $p=(p_1,\ldots,p_N)$, we define $p^{(k)}$ to be an $N-1$-dimensional vector $(p_i)_{i\ne k}$ and denote $$M^{(k)}_{n}(p)=M^{(k)}_{n}(p^{(k)}).$$ Since every nonzero vector with nonnegative components has a nontrivial positive component in the direction of the PF eigenvector, the following statement holds true. \[th:leak\_rate\] Let $\lambda_k<\infty$ for some $k\in\{1,\ldots,N\}$. Then for any $p\in\Delta_N$ with $p_k<1$, there are numbers $c_1(p),c_2(p)$ depending only on $p$ such that $$c_1(p)e^{-\lambda_k n}\le M^{(k)}_{n}(p)\le c_2(p)e^{-\lambda_k n}.$$ The next corollary compares leaking through different holes. 1. If $\lambda_i>\lambda_j$, then for any $p,q\in\Delta_N$ with $q_j<1$, there is $n_0=n_0(p,q)\in{\mathbb{N}}$ such that for all $n\ge n_0$, $$M^{(i)}_{n}(p)< M^{(j)}_{n}(q).$$ 2. Suppose $\sigma$ is a permutation on $\{1,\ldots,N\}$ such that $$\lambda_{\sigma_N}<\ldots<\lambda_{\sigma_1}<\infty.$$ Then for any family of distributions $(p(i)\in \Delta_N,\ i=1,\ldots,N)$ satisfying $p_i(i)<1$ for all $i$, there is $n_0=n_0(p(1),\ldots,p(N))\in{\mathbb{N}}$ such that for all $n\ge n_0$, $$M^{(\sigma_1)}_{n}(p(\sigma_1))< M^{(\sigma_2)}_{n}(p(\sigma_2))<\ldots< M^{(\sigma_N)}_{n}(p(\sigma_N)).$$ 3. Suppose the state $i\in 1,\ldots,N$ is such that $\lambda_i>\lambda_k$ for all $k\ne i$. For any $p\in\Delta_N$ and any $k\ne i$, if $p_k<1$ then there is a time $n_0=n_0(p)$ such that for all $n\ge n_0$, $$M_n^{i}(p)<M_n^k(p).$$ [[[Proof: ]{}]{}]{}The first part follows directly from Theorem \[th:leak\_rate\]. The other two parts are consequences of the first one.[[[ $\Box$ ]{}]{}]{} \[ex:1\]Not only the size (stationary probability) of a state of the Markov chain matters for the escape rate through that state. Consider a Markov chain with transition matrix $$\left( \begin{array}{ccc} 1/3 & 1/6 & 1/2 \\ 1/3 & 5/12 & 1/4 \\ 1/3 & 5/12 & 1/4 \end{array} \right).$$ The stationary distribution for this Markov chain is uniform, i.e., $\pi_i=1/3$, $i=1,2,3$. However, the leading eigenvalues in the reduced matrices $Q^{(i)}$, $i=1,2,3$ are different. Namely $\mu_1=2/3$, $\mu_2=(7+\sqrt{97})/24$, $\mu_3=(9+\sqrt{33})/24$. Therefore, the fastest escape is through the hole in the third state and the slowest one is through the hole in the second state. This example belongs to a more general class than the one considered in [@Yurchenko]. This Markov chain is generated, e.g., by a piecewise linear map $f:[0,1]\to[0,1]$ shown on Fig. \[fig:piecewise\]. States 1,2,3 correspond to intervals $[0,1/3)$, $[1/3,2/3)$, and $[2/3,1]$, respectively, and the stationary measure is Lebesgue measure. The Markov chain in the next example is also generated by a certain 1D expanding piecewise linear map. For the sake of brevity we do not present it here though. ![Piecewise linear map generating the Markov chain of Example \[ex:1\].[]{data-label="fig:piecewise"}](piecewise.eps){height="6cm"} It is possible that the escape is slower through a state with greater stationary probability (bigger “hole”). Consider a Markov chain with transition matrix $$\left( \begin{array}{ccc} 1/2 & 1/12 & 5/12 \\ 1/2 & 0 & 1/2 \\ 1/3 & 1/3 & 1/3 \end{array} \right).$$ The stationary distribution is given by the vector $(36/83,14/83,33/83)$. The largest eigenvalues of the matrices $Q^{(i)}$, $i=1,2,3$ equal $\mu_1=(1+\sqrt{7})/6$, $\mu_2=(5+\sqrt{21})/12$, $\mu_3=(3+\sqrt{15})/12$. Therefore, the escape through the third state is faster than through the first one, although the stationary probability (“size”) of the first state is larger than that of the third state. This example belongs to a more general class of systems than the one considered in [@Which-hole:MR2593912]. A generalization of Theorem 1 to the case of a non-stochastic but just non-negative matrix is straightforward since the Perron-Frobenius theorem is still applicable. It is easy and straightforward to address the situations where our assumption on irreducibility of the Markov chain after the removal of a vertex is violated. For example, one can consider the case where besides one strongly connected component $A$ satisfying our original set of assumptions there are several extra vertices connected to $A$ but unreachable from $A$. In this case, the rate of escape through any vertex of $A$ may depend on the initial distribution. In fact, the rate is determined by the minimum of the “internal” escape rate of $A$ through that vertex and the rates of escape to $A$ from the vertices outside of $A$ that support the initial distribution. Another also easily and directly analyzed situation appears when after a removal of a state the remaining states form several isolated subsets. Clearly in this case escape from each of these subsets should be treated separately (and absolutely analogously to the proof above). It is also easy to see that nothing else besides these two situations in case of reducibility of the resulting after removal of the state Markov chain can appear. Most efficient source ===================== In this section we classify the states of a Markov chain with respect to their efficiency in distributing the information or any perturbation over the entire state space. Here we assume that the Markov chain is just irreducible and aperiodic. Let the evolution be initiated at state $k\in\{1,\ldots,N\}$. Then for any step $n\ge 0$, the distribution of the Markov chain at time $n$ is given by $e_k P^n$ where $e_k$ is the $k$-th coordinate vector and $P^n$ is the $n$-step transition matrix. We can study the total variation distance between the distribution at time $n$ and $\pi=(\pi_k)_{k=1}^N$, the stationary distribution, the existence and uniqueness of which is guaranteed by the Perron–Frobenius Theorem: $$D_k(n)=\|e_kP^n-\pi\|_1,\quad k\in{1,\ldots,N},\ n\ge 0,$$ where $\|v\|_1=\sum_{i=1}^N\|v\|$ is the $L^1$ norm of $v$. Ideally, we would like to say that initial state $k_1$ allows for faster convergence to the stationary distribution than initial state $k_2$ if there is $n_0\in{\mathbb{N}}$ such that $D_{k_1}(n)< D_{k_2}(n)$ for all $n\ge n_0$. However, there are situations where this property holds due to the specific choice of $\|\cdot\|_1$ to measure distances, and will be destroyed if one replaces $\|\cdot\|_1$ with a different (equivalent) norm. So, we choose to work with a partial order on states that does not depend on the concrete choice of the norm in ${\mathbb{R}}^N$. We denote by ${\mathcal{N}}$ the set of all norms on ${\mathbb{R}}^N$. We say that a sequence of vectors $(v_n)_{n\in {\mathbb{N}}}$ in ${\mathbb{R}}^N$ dominates another sequence of vectors $(u_n)_{n\in {\mathbb{N}}}$, if for any $H\in{\mathcal{N}}$, there is a number $n_0=n_0(u,v,H)$ such that $$H(u_n)<H(v_n),\quad n\ge n_0.$$ Obviously, $(v_n)_{n\in {\mathbb{N}}}$ does not dominate $(u_n)_{n\in {\mathbb{N}}}$ iff there is $H\in{\mathcal{N}}$ and a sequence $(n_m)_{m\in{\mathbb{N}}}$ increasing to infinity such that $$H(u_{n_m})\ge H(v_{n_m}),\quad m\in{\mathbb{N}}.$$ We shall say initial state $k_1$ allows for faster convergence to the stationary distribution than initial state $k_2$, if $(e_{k_1}P^n-\pi)_{n\in{\mathbb{N}}}$ is dominated by $(e_{k_2}P^n-\pi)_{n\in{\mathbb{N}}}$. In general, we can say that an initial distribution $u$ allows for faster convergence to the stationary distribution than initial distribution $v$, if $(uP^n-\pi)_{n\in{\mathbb{N}}}$ is dominated by $(vP^n-\pi)_{n\in{\mathbb{N}}}$. This introduces a partial order on $\Delta_N$, and our goal is to give an equivalent definition of this partial order in terms of projections on vectors in a (real) canonical Jordan basis $(w_i)_{i=1}^N$ associated to $P$ (we refer to [@HSD-MR2144536] for the background on canonical forms). We assume that $w_N=\pi$, the stationary distribution for $P$, a positive eigenvector of $P$ with simple eigenvalue 1, a unique eigenvalue of $P$ equal to 1 in magnitude. To each $w_i$, $i=1,\ldots,N-1,$ we associate $\lambda_i$ with $\|\lambda_i\|<1$ and ${\mathop{\mathrm{Im}}}\lambda_i\ge 0$, the eigenvalue of the generalized eigenspace that $w_i$ belongs to (since complex eigenvalues come in conjugate pairs, we choose ${\mathop{\mathrm{Im}}}\lambda_i\ge 0$). If $\lambda_i\in{\mathbb{R}}$, then we define $$r_i=\min\{r\in{\mathbb{N}}:\ w_i(P-\lambda_i I)^r=0\}.$$ Recalling that for a nonreal eigenvalue $\lambda$, the canonical basis vectors are grouped in pairs, we can define $r_i=r_j$ analogously for a pair $(w_i,w_j)$ of canonical basis vectors corresponding to $\lambda_i\not\in{\mathbb{R}}$. In both cases, the numbers $r_i$ enumerate the generalized eigenvectors within one generalized eigenspace, and the pair $(\lambda_i,r_i)$ determines the rate of decay of $w_i$ under iterations of $P$, namely, $\lambda_i$ is the exponential rate of decay, and $r_i-1$ is the degree of the polynomial factor, see Lemma \[lm:exact-proj\] below. If $\mu\ge 0$ and $k\in{\mathbb{N}}$, we denote by $\Pi_{\mu,k}v$ the vector projection on the vector subspace spanned by all $w_i$ such that $\|\lambda_i\|=\mu$ and $r_i=k$ (the projection is taken along the span of all other vectors of the Jordan basis). If this subspace is empty, the projection is assumed to be 0. For two vectors $u,v\in {\mathbb{R}}^N$ we write $u<v$ if there is a real number $a$ with $\|a\|<1$ such that $u=av$. \[thm:ordering\_initial\_distributions\] Initial distribution $u$ allows for faster convergence to the stationary distribution than initial distribution $v$ if and only if there are $\mu_0\in(0,1)$ and $r_0\in{\mathbb{N}}$ such that the following conditions are satisfied: 1. If either (i) $ \mu\in(\mu_0,1)$, or (ii) $\mu=\mu_0$ and $r>r_0$, then $\Pi_{\mu,r}u=0$. 2. $ \Pi_{\mu_0,r_0} u < \Pi_{\mu_0,r_0} v.$ Intutively, it is natural to think of $\mu_0$ as of the second largest eigenvalue of $P$. However, the theorem holds true even in such a degenerate situation where the projections of both $u$ and $v$ on the eigenspace associated to the second largest eigenvalue vanish. \[cor: second\_eigenvalue\] Suppose that the image of the projection operator $\Pi_{\mu_0,r_0}$ is 1-dimensional (this is guaranteed if the second largest in magnitude eigenvalue of $P$ is real and simple). Let us denote $q_i=\|\Pi_{\mu_0,r_0}e_i\|$, $i=1,\ldots,N$. Suppose $\sigma$ is a permutation on $\{1,\ldots,N\}$ such that $$q_{\sigma_1}<\ldots<q_{\sigma_N}.$$ Then for any $N\in{\mathcal{N}}$ there is a number $n_0=n_0(H)$ such that for any $n>n_0$, $$H(e_{\sigma_1}P^n-\pi)<\ldots<H(e_{\sigma_N}P^n-\pi),$$ so that for any $i,j$ with $i<j$, the initial state $\sigma_i$ allows for faster convergence to the stationary distribution than the initial state $\sigma_j$. In particular, the initial state $\sigma_1$ allows for faster convergence than any other initial state. This hierarchy of states may fail to exist in the case where the dimension of the image of $\Pi_{\mu_0,r_0}$ is greater than one, e.g., where the second highest eigenvalue of $P$ is non-real, or where there are multiple Jordan blocks associated to $\mu_0$. Often, the best source state from the point of view of the hierarchy established in Corollary \[cor: second\_eigenvalue\] is the state with maximal stationary probability. However, this is not necessarily so, as the following example shows. Suppose the transition probability matrix is $$\left( \begin{array}{ccc} 1/8 & 5/8 & 1/4 \\ 3/8 & 9/16 & 1/16 \\ 1/24 & 1/12 & 7/8 \end{array} \right).$$ Then, there are three simple eigenvalues: $1$, $3/4$, and $-3/16$. Their respective eigenvectors are: $\pi=(1/6,1/3,1/2)$, $w_1=(-1/6,-1/3,1/2)$, and $w_2=(-16/3,13/3,1)$. Notice that the stationary probability is maximized by state $3$ since $\pi_3>\pi_2>\pi_1$. However, decomposing $$\begin{aligned} e_1&=\pi-\frac{11}{15}w_1-\frac{2}{15}w_2,\\ e_2&=\pi-\frac{17}{15}w_1-\frac{1}{15}w_2,\\ e_3&=\pi+1\cdot w_1+0\cdot w_2,\end{aligned}$$ comparing the projections on $w_1$, and noticing that $11/15<1<17/15$, we can use Theorem \[thm:ordering\_initial\_distributions\] to conclude that state 1 allows for faster convergence than the two other states. A generalization of Theorem 2 for non-negative (but non-stochastic) matrices is straightforward. Proof of Theorem \[thm:ordering\_initial\_distributions\] ========================================================= We begin with several elementary auxiliary statements. First, we recall formulas for powers of Jordan blocks. For a condition $A$, we use $${{\bf 1}}_A=\begin{cases} 1,& \text{if}\ A\ \text{\rm holds},\\ 0,&\text{\rm otherwise.} \end{cases}$$ \[lm:exact-proj\] 1. Let vectors $(w_{i_r})_{r=1}^m$ form a generalized eigenspace of $P$ with eigenvalue $\lambda\in{\mathbb{R}}$, i.e., $w_{i_r}P=\lambda w_{i_r}+{{\bf 1}}_{2\le r\le m}w_{i_{r-1}}$. Then $$w_{i_r}P^n=\sum_{k=1}^r \binom{n}{r-k} \lambda^{n-(r-k)}w_{i_k}.$$ 2. Let $\lambda=\mu e^{i\phi}$, where $\mu>0$ and $\phi\in(0,\pi)$, and vectors $(w_{i_r})_{r=1}^m, (w_{j_r})_{r=1}^m$ form a generalized eigenspace of $P$ with eigenvalue $\lambda$, i.e., for any $a,b\in{\mathbb{R}}$, $$\begin{aligned} (aw_{i_r}+bw_{j_r})P &=\mu (a \cos \phi-b\sin\phi) w_{i_r}+a{{\bf 1}}_{2\le r\le m}w_{i_{r-1}}\\ &+ \mu (a \sin \phi+b\cos\phi) w_{j_k}+b{{\bf 1}}_{2\le r\le m}w_{j_{r-1}}.\end{aligned}$$ Then $$\begin{gathered} (aw_{i_r}+bw_{j_r})P^n\\ =\sum_{k=1}^r {\binom{n}{r-k}}\mu^{n-(r-k)}\Bigl[(a \cos ((n-(r-k))\phi)-b\sin((n-(r-k))\phi)) w_{i_k}\\ +(a \sin((n-(r-k))\phi)+b\cos((n-(r-k))\phi)) w_{j_k}\Bigr].\end{gathered}$$ \[lm:finding\_norm\] Let $u,v\in{\mathbb{R}}^N$. If $u<v$, then $H(u)<H(v)$ for any $H\in{\mathcal{N}}$. If $u=v$, then $H(u)=H(v)$ for any $H\in{\mathcal{N}}$. If $u\ne v$ and $u\not<v$, then there is $H\in{\mathcal{N}}$ such that $H(u)>H(v)$. [[[Proof: ]{}]{}]{}First two statements of the lemma are trivial. It is sufficient to prove the third one for the case where $u$ and $v$ are not proportional to each other. To that end, let us take a linear bijection that sends vectors $(2,0,0,\ldots,0)$ and $(0,1,0,0,\ldots,0)$ to $u$ and $v$. The pushforward of the Euclidean norm under this map satisfies the desired property. [[[ $\Box$ ]{}]{}]{} \[lm:finding\_norm\_derivative\] Suppose $x,y\in{\mathbb{R}}^N$ and they are not multiples of each other. Then there is $H\in {\mathcal{N}}$, a neighborhood $U$ of $x$, and a constant $c>0$ such that for all $z\in U$, $ \frac{d}{d{\varepsilon}} H(z+{\varepsilon}y)\bigr\|_{{\varepsilon}=0} $ is well defined and exceeds $c$. [[[Proof: ]{}]{}]{}Let us take a linear bijection that sends vectors $(1,0,0,\ldots,0)$ and $(1,1,0,0,0,\ldots,0)$ to $x$ and $y$. The pushforward of the Euclidean norm under this map satisfies the desired property. [[[ $\Box$ ]{}]{}]{} [[[Proof of Theorem \[thm:ordering\_initial\_distributions\]: ]{}]{}]{}First, we notice that $\Pi_{1,1}u=\Pi_{1,1}v=\pi=w_N$. This follows from the following facts: $P$-iterates of $u$ and $v$ converge to $\pi$; $\Pi_{1,1}u$ and $\Pi_{1,1}v$ are invariant under $P$; $P$-iterates of all other projections decay exponentially. Suppose that there are $\mu_0$ and $r_0$ such that conditions 1 and 2 of the Theorem hold true. Decomposing $u$ and $v$ w.r.t. the canonical basis and using Lemma \[lm:exact-proj\], we immediately see that $u$ allows for faster convergence than $v$. Suppose now that $u$ allows for faster convergence than $v$. Let us choose $\mu_0$ and $r_0$ so that $\Pi_{\mu_0,r_0}v\ne 0$ and if either (i) $ \mu\in(\mu_0,1)$, or (ii) $\mu=\mu_0$ and $r>r_0$, then $\Pi_{\mu,r}v=0$. Lemma \[lm:exact-proj\] immediately implies now that condition 1 of the theorem is satisfied. To prove condition 2, let us assume that the opposite holds, i.e., $\Pi_{\mu_0,r_0} u\not< \Pi_{\mu_0,r_0} v$. First, we consider the case where $\Pi_{\mu_0,r_0} u \ne \Pi_{\mu_0,r_0} v$. Lemma \[lm:finding\_norm\] allows us to find a norm $H\in{\mathcal{N}}$ such that $H(\Pi_{\mu_0,r_0} u)> H(\Pi_{\mu_0,r_0} v)$. For any small neighborhoods $U$ of $\Pi_{\mu_0,r_0}u$ and $V$ of $\Pi_{\mu_0,r_0}v$ we can use Lemma \[lm:exact-proj\] to find an infinite sequence $n_m\to\infty$ such that $$\begin{aligned} \frac{\Pi_{\mu_0,r_0} u P^{n_m}} {\binom{n_m}{r_0 -1}\mu_0^{n_m}}\in U,\quad m\in{\mathbb{N}},\label{eq:return_u} \\ \frac{\Pi_{\mu_0,r_0} v P^{n_m}}{\binom{n_m}{r_0 -1}\mu_0^{n_m}}\in V,\quad m\in{\mathbb{N}}.\label{eq:return_v}\end{aligned}$$ This is trivially true with $n_m\equiv m$ if all the eigenvalues with magnitude $\mu_0$ are real and equal to $\mu_0$. If the arguments of some of these eigenvalues are not zero, then we can use the recurrence property of the shift on the multidimensional torus induced by these arguments. Using Lemma \[lm:exact-proj\] to compute the leading terms of $(u-\pi) P^{n_m}$ and $(v-\pi) P^{n_m}$, we see that $$\lim_{m\to\infty}\frac{(z-\pi) P^{n_m}-(\Pi_{\mu_0,r_0} z) P^{n_m}}{\binom{n_m}{r_0 -1}\mu_0^{n_m}}=0,\quad z=u,v. \label{eq:main_contrib_is_mu_0_k_0}$$ Therefore, $$\begin{aligned} \frac{(u-\pi) P^{n_m}} {\binom{n_m}{r_0 -1}\mu_0^{n_m}}\in U,\quad m\in{\mathbb{N}},\label{eq:return_u_1} \\ \frac{(v-\pi) P^{n_m}}{\binom{n_m}{r_0 -1}\mu_0^{n_m}}\in V,\quad m\in{\mathbb{N}}.\label{eq:return_v_1}\end{aligned}$$ so that choosing $U$ and $V$ disjoint and sufficiently small and using inequality $H(\Pi_{\mu_0,r_0} u)> H(\Pi_{\mu_0,r_0} v)$ along with the continuity of $H$, we conclude that $u$ does not allow for faster convergence than $v$. This contradicts our assumption and therefore it remains to consider $$\Pi_{\mu_0,r_0} u = \Pi_{\mu_0,r_0} v. \label{eq:equal_at_mu_0_k_0}$$ Faster convergence for $u$ is clearly impossible in the situation where $u=v$. Assuming $u\ne v$, we can find numbers $\mu_1$ and $r_1$ such that $\Pi_{\mu_1,r_1}u\ne \Pi_{\mu_1,r_1}v$ and if (i) $\mu_1<\mu<\mu_0$, or (ii) $\mu=\mu_1$ and $r>r_1$, then $\Pi_{\mu,r}u=\Pi_{\mu,r}v$. Let $U$,$H$, and $c$ be defined in Lemma \[lm:finding\_norm\_derivative\] applied to $x=\Pi_{\mu_0,k_0} u = \Pi_{\mu_0,k_0} v$, and $y=\Pi_{\mu_1,r_1} u-\Pi_{\mu_1,r_1} v$ which is not a multiple of $x$. Due to Lemma \[lm:exact-proj\], there is a sequence of numbers $n_m\to\infty$ and a sequence of vectors $u_m,v_m$ such that $$\begin{aligned} \label{eq:decomposition1} \frac{ (u-\pi) P^{n_m}}{ \binom{n_m}{r_0 -1}\mu_0^{n_m}}=z_m+u_m,\quad m\in{\mathbb{N}},\\ \label{eq:decomposition2} \frac{(v-\pi) P^{n_m}}{\binom{n_m}{r_0 -1}\mu_0^{n_m}}=z_m+v_m,\quad m\in{\mathbb{N}},\end{aligned}$$ where $z_m\in U$ for all $m$, and $$\begin{aligned} u_m=\frac{\binom{n_m}{r_1 -1}\mu_1^{n_m} \Pi_{\mu_1,r_1} u}{\binom{n_m}{r_0 -1}\mu_0^{n_m}} +o\left(\frac{\binom{n_m}{r_1 -1}\mu_1^{n_m}}{\binom{n_m}{r_0 -1}\mu_0^{n_m}}\right),\quad m\to\infty, \label{eq:decomposition3}\\ v_m=\frac{\binom{n_m}{r_1 -1}\mu_1^{n_m} \Pi_{\mu_1,r_1} v}{\binom{n_m}{r_0 -1}\mu_0^{n_m}}+o\left(\frac{\binom{n_m}{r_1 -1}\mu_1^{n_m}}{ \binom{n_m}{r_0 -1}\mu_0^{n_m}}\right),\quad m\to\infty, \label{eq:decomposition4}\end{aligned}$$ We can use relations and to derive $$\begin{gathered} H(z_m+u_m)-H(z_m+v_m)=\frac{\binom{n_m}{r_1 -1}\mu_1^{n_m}}{\binom{n_m}{r_0 -1}\mu_0^{n_m}}\cdot \frac{d}{d{\varepsilon}}H(z_m+{\varepsilon}(\Pi_{\mu_1,r_1} u - \Pi_{\mu_1,r_1} v))\bigr\|_{{\varepsilon}=0} \\+o\left(\frac{\binom{n_m}{r_1 -1}\mu_1^{n_m}}{\binom{n_m}{r_0 -1}\mu_0^{n_m}}\right),\quad m\to\infty.\end{gathered}$$ Since $z_m\in U$, Lemma \[lm:finding\_norm\_derivative\] allows us to conclude that the derivative in the r.h.s. of the last identity exceeds $c>0$. Therefore, relations and imply that $H((u-\pi)P^{n_m})>H((v-\pi)P^{n_m})$ for all $m$, which contradicts our assumption that $u$ allows for faster convergence than $v$. Therefore, is impossible and the proof of the necessity of conditions 1 and 2 of the theorem is complete. [[[ $\Box$ ]{}]{}]{} Concluding remarks ================== We have shown that the tails of the distributions of hitting times for different states of irreducible Markov chains and for typical initial distributions can be ordered. This means that there is a finite moment of time $n^*$ after which the tails of these distributions never intersect. This property allows to determine the optimal sink in a Markov chain or in a dynamical network. Our results hold for any nondegenerate initial distribution and in this respect they essentially generalize those in [@Which-hole:MR2593912], where only Lebesgue measure was considered. We also demonstrated that one can determine the best source in a Markov chain. Again it is a finite time result and the hierarchy of the Markov chain states emerges in their ability to serve as a source. For a network, this suggests the node or element one should apply a perturbation to, or inject information at, so that the perturbation spreads over the network and converges to the stationary distribution in the fastest way. Our results are also true (with obvious adjustments) if the matrix $P$ is nonnegative and not necessarily stochastic. Acknowledgments =============== Y.B. was partially supported by NSF CAREER grant DMS-0742424. L.B. was partially supported by NSF grant DMS-0900945.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Cascaded regression method is a fast and accurate method on finding 2D pose of objects in RGB images. It is able to find the accurate pose of objects in an image by a great number of corrections on the good initial guess of the pose of objects. This paper explains the algorithm and shows the result of two experiments carried by the researchers. The presented new method to quickly and accurately predict 3D positions of body joints from a single depth image, using no temporal information. We take an object recognition approach, designing an intermediate body parts representation that maps the difficult pose estimation problem into a simpler per-pixel classification problem. Our large and highly varied training dataset allows the classifier to estimate body parts invariant to pose, body shape, clothing. Finally, we generate confidence-scored 3D proposals of several body parts by re-projecting the classification result and finding local modes.' author: - Wenye He title: A Generic Regression Framework for Pose Recognition on Color and Depth Images --- Introduction {#introduction .unnumbered} ============ Detection and localization provide a helpful function in computer vision. Detection finds out whether an object is contained in the image, while localization tells people which specific place of the image an object is in. For example, there is an image of tree. Detection can tell us whether a bird is in the image and localization shows where the bird it is. Localization could answer whether the bird flies in the sky or stays on the tree. More specifically, localization enable us to know which groups of pixels represent a bird in the image. On the other hand, robust interactive human body tracking has different applications that include gaming, human computer interaction, security, telepresence and even the health care. The task has recent been simplified by the introduction of real time depth cameras. However even the best existing system has limitations. Until the launch of Kinect, none ran at interactive rates on consumer hardware while handling a full range of human body shapes and sizes undergoing general body motions. Detecting body parts from a single depth image is a challenging task from a small set of 3D position candidates for each skeletal joint [@ierf4]. They focused on pre-frame initialization and recovery is designed to complement any appropriate tackling algorithm In this paper, we demonstrate an algorithm to answer a question â€œIs an object $o$ with pose $\theta$ located in the image $I$. In their work, after making a raw guess of the pose of an object in a set of image, they use cascaded pose regression to detect which image has such the object and locate which region in the image contains the object. Pose-indexed features[@iref2] and its assumed weak invariance are used in the algorithm. Furthermore, random ferns[@iref3] regressors are applied in CPR. The system runs at 200 frames per second on consumer hardware. The evaluation shows high accuracy on both synthetic and real test sets, and investigates the effect of several training parameters. They achieved state of the art accuracy in our comparison with related work and demonstrate improved generalization over exact whole-skeleton nearest neighbour matching. Related Works {#related-works .unnumbered} ============= This paper uses features to achieve pose estimate. In computer vision, features have great functions. The popular use of features is recognition. Paul and Michael[@iref4] use rectangle features for object detection. Lowe[@iref5] present an approach to identify objects using higly distinctive and scale - invariant features, and has been applied for many applications [@tan13]. Features also help people in 3D model retrieval. Ryutarou et al.[@iref6] obtain 3D model using salient local visual features. In this paper, researchers apply pose-indexed features[@iref2] whose concept is proposed by Francois and Donald to estimate pose. Early work in pose estimation contains active contour models[@iref7], 2D range scans matching[@iref8], and active appearance models[@iref9]. In recent work, pose estimation is employed to reconstruct 3D shape[@iref10] and detect and localize object[@iref11]. \[[@iref25]\] use the marginal statistics of unlabelled data to improve pose estimation. \[[@iref41], [@shen13]\] proposed a local mixture of Gaussian Processes to regress human pose. Auto-context was used in \[[@iref40]\] to obtain a coarse body part labelling. \[[@iref42]\] track a hand clothed in a colored glove. Our system could be automatically inferring the colors of a virtual colored suit from a depth image and used for many RGB-D applications, such as virtual reality [@Su13]. Method {#method .unnumbered} ====== This section describes how Cascaded Pose Regression works. First of all, the researchers create an image model $G : \mathcal{O} \times \Theta \to \mathcal{I}$. The model show how an image $I \in \mathcal{I}$ is constructed from an object $o \in \mathcal{O}$ and pose $\theta\in\Theta$. Given that poase estimate is suppose to be unique in the model, $\mathcal{G}(o_1,\theta_1)=\mathcal{G}(o_2,\theta_2)$ iff $o_1=o_2$ and $\theta_1=\theta_2$. Moreover, the operator $\circ$ is designed for combination of two poses. The researchers write a formula for a new pose composed of $\theta_1$ and $\theta_2$, $\theta=\theta_1\circ\theta_1$. $\overline{\theta}$ is set as the inverse of $\overline{\theta}$, and $e$ as the identity element. A function to calculate relative error between two poses is formed, $d : \Theta \times \Theta \to \mathcal{R}$ where $d(\theta_1,\theta_2)$ depends on $\overline{\theta_1}\circ\theta_2$ or equivalently that $d(\theta_\delta\circ\theta_1,\theta_\delta\circ\theta_2)=d(\theta_1,\theta_2)$ for all $\theta_1,\theta_2,\theta_\delta\in\Theta$.\ Pose-indexed features and weak invariance are introduced in CPR. A pose-indexed features is a function, $h : \Theta\times\mathcal{I} \to \mathcal{R}$. $h$ is weakly invariant if $$h(\theta,G(o,e))=h(\theta_{\delta}\circ\theta,G(o,\theta_{\delta})),$$ where $\forall\theta,\theta_\delta\in\Theta$, or equivalently, $$h(\theta_1,G(o,\theta_2))=h(\theta_{\delta}\circ\theta_1,G(o,\theta_{\delta}\circ\theta_2)),$$ where $\forall\theta_1,\theta_2,\theta_\delta\in\Theta$. Each pose-indexed feature is composed of control point features and pose. Each control feature, $h_{p_1,p_2}$, is computed as the difference of two pixels, $p_1$ and $p_2$ at predefined image locations, so $h_{p_1,p_2}(I)=I(p_1)-I(p_2)$ where $I(p)$ is the grayscale value of image $I$ at location $p$. Therefore, taking pose $\theta$ into account, the researchers define an pose-indexed feature $h_{p_1,p_2}(\theta,I) = I(H_{\theta}p_1)-I(H_{\theta}p_2)$, where $H_\theta$ is an associated $3\times3$ homography matrix. Figure 1 shows pose-indexed features in mice and zebra fish.\ After introducing pose-indexed features and weakly invariance, the evaluation and training algorithm on CPR is demonstrated here and is shown in Figures 2 and Figure 3. The full formula to evaluate CPR is $\theta^t=\theta^{t-1}\circ R^t(h^t(\theta^{t-1},I))$, where $t=1...T$ and $R$ is a cascaded regressor trainned in Figure 3, given an input pose, initial pose, $\theta^0$ and an image $I$, outputting $\theta^T$. The full evaluating procedures is shown by testing images in Figure 4.\ The objective for training a new regressor is to reduce the difference between the true pose and the pose calculated by the previous pose-indexed features and previous regressor. The goal is to optimize the loss $\mathcal{L}=\sum^N_{i=1}d(\theta^T_i,\theta_i)$. In step 1 and 2 of figure 3, for each i, there is a pose in the zero phase, $\theta_i^0=\theta^0={\arg\!\min}_\theta\sum_i d(\theta,\theta_i)$. $\theta^0$ is the single pose estimate with the minimized training error and without depending on regressors. In every iteration, the researchers compute the pose-indexed features for all the training images set $I$ using the previous pose estimate $\theta_i^{t-1}$ corresponding to the image $I_i$. Then, they compute a new pose $\tilde{\theta}_i$ using the average of the poses in previous one phase, $\overline{\theta}_i^{t-1}$, and the true pose $\theta_i$. In Step 6, the researchers find a regressor $R^t$ by minimizing the loss. In the formula in Step 6, $R$ is a set of regressors, $R=(R^1,...,R^{t-1})$. Using the new regressor, the researchers get the new poses for all the images. In Step 8 and Step 9, the researchers calculate the error by comparing the losses in current phase and the previous one phase. The loop stop once the new iteration cannot decline the training error.\ Inspired by random fern classifier and random forests regressor, the researchers train a random fern regressor at each phase in the cascade. A fern regressor outputs $y_i\in\mathbb{R}$ by taking the featurs $x_i\in\mathbb{R}^F$ as input. Given that a Fern is generated by $S$ pairs of random features from all the features in the image and one pair of features can turn to be a value of 1 or 0, each $x_i$ can be viewed as a value range from 0 to $2^S-1$. $y$ is the prediction for the mean of $y_i$’s of the training examples matching the value of each $x_i$. This method can converge very quickly and accurately in predicting 3D positions of body joints from a single depth map, using no temporal information. CPR may fail to make a correct pose estimate due to some bad initial pose. To improve the performance of CPR, the researchers find the region that have more sucessful initial poses by running CPR $K$ times with a variety of random intial poses for every image. Experiments and results {#experiments-and-results .unnumbered} ======================= This section covers the experiments and results CPR on two datasets: Mice and Fish. Human annotations are used to make poses of mice and fish clearer, which are shown in Figure 5. For each mouse, its pose include the location, orientation, scale and aspect ratio of an ellipse fitting around For each fish, its pose is a 3-part model with location, orientation and scale of a central body part, and the angles of the tail and head with respect to the body. Figure 6 shows how the number of phases influence the error. The alogrithm converges after 512 stages for both datasets. Reference {#reference .unnumbered} ========= [1]{} DollÃ¡r, Piotr, Peter Welinder, and Donald Geman. “Cascaded pose regression.” Computer Vision and Pattern Recognition (CVPR), 2010 IEEE Conference on. IEEE, 2012. Fleuret, FranÃ§ois, and Pietro Perona. “Stationary features and cat detection.” Journal of Machine Learning Research 9.Nov (2008): 2549-2578. J. Shen and J. Yang, “Automatic human animation for non-humanoid 3d characters," International Conference on Computer-Aided Design and Computer Graphics (CAD/Graphics), pp. 220-221, 2015. Michael Jones, et al. “Fast keypoint recognition using random ferns.” IEEE transactions on pattern analysis and machine intelligence 32.3 (2012): 448-461. Viola, Paul, and Ozuysal, Mustafa. “Rapid object detection using a boosted cascade of simple features.” Computer Vision and Pattern Recognition, 2001. CVPR 2001. Proceedings of the 2001 IEEE Computer Society Conference on. Vol. 1. IEEE, 2001. Lowe, David G. “Distinctive image features from scale-invariant keypoints.” International journal of computer vision 60.2 (2008): 91-110. Ohbuchi, Ryutarou, et al. “Salient local visual features for shape-based 3D model retrieval.” Shape Modeling and Applications, 2012. IEEE International Conference on. IEEE, 2012. J. Shen and S. C. S. Cheung, “Layer Depth Denoising and Completion for Structured-Light RGB-D Cameras," IEEE Conference on Computer Vision and Pattern Recognition, pp. 1187-1194, 2013. Kass, Michael, Andrew Witkin, and Demetri Terzoopoulos. “Snakes: Active contour models.” International journal of computer vision 1.4 (2000): 321-331. Timothy F. “Robot pose estimation in unknown environments by matching 2d range scans.” Computer Vision and Pattern Recognition, 1994. Proceedings CVPR’94. IEEE Computer Society Conference on. IEEE, 2004. Cootes, A.W. Fitzgibbon., Gareth J. Edwards, and Christopher J. Taylor. “Active appearance models.” European conference on computer vision. Springer Berlin Heidelberg, 2002. J. Shen and W. Tan, “Image-based indoor place-finder using image to plane matching," IEEE International Conference on Multimedia and Expo, pp. 1-6, 2013. Ronsen Basri, A.W. Fitzgibbon, and R. Cipolla. The joint manifold model for semi-supervised multi-valued regression. In Proc. ICCV, 2005 Arie-Nachimson, Mica, and Ronen Basri. “Constructing implicit 3d shape models for pose estimation.” Computer Vision, 2009 IEEE 12th International Conference on. IEEE, 2012. J. Shen, P. C. Su, S. c. S. Cheung, J. Zhao, “Virtual mirror rendering with stationary rgb-d cameras and stored 3-d background," IEEE Transactions on Image Processing, vol. 22, no. 9, pp. 3433-3448, 2013. Ozuysal, Mustafa, Z. Tu, and Pascal Fua. “Pose estimation for category specific multiview object localization.” Computer Vision and Pattern Recognition. CVPR 2011 Vecent Lepettite. Auto-context and its application to high-level vision tasks. In Proc. CVPR, 2009.2 S. Urtasun. Local probabilistic regression for activityindependent human pose inference. In Proc. CVPR, 2006. 2 R. ZHANG and J. Popovi´c. Real-time hand-tracking with a color glove. In Proc. ACM SIGGRAPH, 2004	{ "pile_set_name": "ArXiv" }
--- address: \| ^\^Intel Corporation\ Istanbul, Turkey\ [email protected]\ ^$\dagger$^Department of Computer Engineering, Boğaziçi University\ Istanbul, Turkey\ {hakan.demir, arzucan.ozgur}@boun.edu.tr\ bibliography: - 'xample.bib' title: 'Named Entity Recognition on Twitter for Turkish using Semi-supervised Learning with Word Embeddings' --- Introduction ============ Microblogging environments, which allow users to post short messages, have gained increased popularity in the last decade. Twitter, which is one of the most popular microblogging platforms, has become an interesting platform for exchanging ideas, following recent developments and trends, or discussing any possible topic. Since Twitter has an enormously wide range of users with varying interests and sharing preferences, a significant amount of content is being created rapidly. Therefore, mining such platforms can extract valuable information. As a consequence, extracting information from Twitter has become a hot topic of research. For Twitter text mining, one popular research area is opinion mining or sentiment analysis, which is surely useful for companies or political parties to gather information about their services and products [@Kokciyan-2013]. Another popular research area is content analysis, or more specifically topic modeling, which is useful for text classification and filtering applications on Twitter [@Hong-2010]. Moreover, event monitoring and trend analysis are also other examples of useful application areas on microblog texts [@Kireyev-2009]. In order to build successful social media analysis applications, it is necessary to employ successful processing tools for Natural Language Processing (NLP) tasks such as Named Entity Recognition (NER). NER is a critical stage for various NLP applications including machine translation, question answering and opinion mining. The aim of NER is to classify and locate atomic elements in a given text into predefined categories like the names of the persons, locations, and organizations (PLOs). NER on well-written texts is accepted as a solved problem for well-studied languages like English. However, it still needs further work for morphologically rich languages like Turkish due to their complex structure and relatively scarce language processing tools and data sets [@Seker-2012]. In addition, most of the NER systems are designed for formal texts. The performance of such systems drops significantly when applied on informal texts. To illustrate, the state-of-the-art Turkish NER system has CoNLL F-score of 91.94% on news data, but the performance drops to F-score of 19.28% when this system is adopted to Twitter data [@Celikkaya-2013].\ There are several challenges for NER on tweets, which are also summarized in , due to the very short text length and informal structure of the language used. Missing proper grammar rules and punctuation, lack of capitalization and apostrophes, usage of hashtags, abbreviations, and slang words are some of those challenges. In Twitter, using contracted forms and metonymic expressions instead of full organization or location names is very common as well. The usage of non-diacritic characters and the limited annotated data bring additional challenges for processing Turkish tweets.\ Due to the dynamic language used in Twitter, heavy feature engineering is not feasible for Twitter NER. developed a semi-supervised approach for Turkish NER on formal (newswire) text using word embeddings obtained from unlabeled data. They obtained promising results without using any gazetteers and language dependent features. We adopted this approach for informal texts and evaluated it on Turkish tweets, where we achieved the state-of-the-art F-score performance. Our results show that using word embeddings for Twitter NER in Turkish can result in better F-score performance compared to using text normalization as a pre-processing step. In addition, utilizing in-domain word embeddings can be a promising approach for Twitter NER. Related Work ============ There are various important studies of NER on Twitter for English. presented a two-phase NER system for tweets, T-NER, using Conditional Random Fields (CRF) and including tweet-specific features. proposed a hybrid NER approach based on K-Nearest Neighbors and linear CRF. presented a factor graph-based method for NER on Twitter. described an unsupervised approach for tweets, called TwiNER. described an NLP pipeline for tweets, called TwitIE. Very recently, have shown the effectiveness of Brown clusters and word vectors on Twitter NER for English. For Turkish NER on formal texts, presented the first study with a Hidden Markov Model based approach. presented an automatic rule learning system. used CRF for Turkish NER, and proposed a hybrid approach. A CRF-based model by is the state-of-the-art Turkish NER system with CoNLL F-score of 91.94%, using gazetteers. achieved a similar F-score of 91.85%, without gazetteers and language dependent features, using a semi-supervised model with word embeddings. For Turkish NER on Twitter, presented the first study by adopting the CRF-based NER of with a text normalizer. adopted a multilingual rule-based NER by extending the resources for Turkish. adopted a rule-based approach for Turkish tweets, where diacritics-based expansion to lexical resources and relaxing the capitalization yielded an F-score of 48% with strict CoNLL-like metric. NER for Turkish Tweets using Semi-supervised Learning ===================================================== To build a NER model with a semi-supervised learning approach on Turkish tweets, we used a neural network based architecture consisting of unsupervised and supervised stages. Unsupervised Stage ------------------ In the unsupervised stage, our aim is to learn distributed word representations, or word embeddings, in continuous vector space where semantically similar words are expected to be close to each other. Word vectors trained on large unlabeled Turkish corpus can provide additional knowledge base for NER systems trained with limited amount of labeled data in the supervised stage. A word representation is usually a vector associated with each word, where each dimension represents a feature. The value of each dimension is defined to be representing the amount of activity for that specific feature. A distributed representation represents each word as a dense vector of continuous values. By having lower dimensional dense vectors, and by having real values at each dimension, distributed word representations are helpful to solve the sparsity problem. Distributed word representations are trained with a huge unlabeled corpus using unsupervised learning. If this unlabeled corpus is large enough, then we expect that the distributed word representations will capture the syntactic and semantic properties of each word and this will provide a mechanism to obtain similar representations for semantically and syntactically close words. Vector space distributed representations of words are helpful for learning algorithms to reach better results in many NLP tasks, since they provide a method for grouping similar words together. The idea of using distributed word representations in vector space is applied to statistical language modeling for the first time by using a neural network based approach with a significant success by . The approach is based on learning a distributed representation of each word, where each dimension of such a word embedding represents a hidden feature of this word and is used to capture the word’s semantic and grammatical properties. Later on, proposed to use distributed word representations together with the supervised neural networks and achieved state-of-the art results in different NLP tasks, including NER for English. We used the public tool, word2vec[^1], released by to obtain the word embeddings. Their neural network approach is similar to the feed-forward neural networks [@Bengio-2003; @Collobert-2011]. To be more precise, the previous words to the current word are encoded in the input layer and then projected to the projection layer with a shared projection matrix. After that, the projection is given to the non-linear hidden layer and then the output is given to softmax in order to receive a probability distribution over all the words in the vocabulary. However, as suggested by , removing the non-linear hidden layer and making the projection layer shared by all words is much faster, which allowed us to use a larger unlabeled corpus and obtain better word embeddings. ![Skip-gram model architecture to learn continuous vector representation of words in order to predict surrounding words [@Mikolov-2013].[]{data-label="fig.1"}](skipgram_v3.eps) Among the methods presented in , we used the continuous Skip-gram model to obtain semantic representations of Turkish words. The Skip-gram model uses the current word as an input to the projection layer with a log-linear classifier and attempts to predict the representation of neighboring words within a certain range. In the Skip-gram model architecture we used, we have chosen 200 as the dimension of the obtained word vectors. The range of surrounding words is chosen to be 5, so that we will predict the distributed representations of the previous 2 words and the next 2 words using the current word. Our vector size and range decisions are aligned with the choices made in the previous study for Turkish NER by . The Skip-gram model architecture we used is shown in Figure \[fig.1\]. Supervised Stage ---------------- In this stage, a comparably smaller amount of labeled data is used for training the final NER models. We used the publicly available neural network implementation by [^2], which actually follows the study by , where a regularized averaged multiclass perceptron is used. Note that although non-local features are proven to be useful for the NER task on formal text types such as news articles, their usage and benefit is questionable for informal and short text types. Due to the fact that each tweet is treated as a single document with only 140 characters, it is difficult to make use of non-local features such as context aggregation and prediction history for the NER task on tweets. On the other hand, local features are mostly related to the previous and next tokens of the current token. With this motivation, we explored both local and non-local features but observed that we achieve better results without non-local features. As a result, to construct our NER model on tweets, we used the following local features: - [Context: All tokens in the current window of size two.]{} - [Capitalization: Boolean feature indicating whether the first character of a token is upper-case or not. This feature is generated for all the tokens in the current window.]{} - [Previous tags: Named entity tag predictions of the previous two tokens.]{} - [Word type information: Type information of tokens in the current window, i.e. all-capitalized, is-capitalized, all-digits, contains-apostrophe, and is-alphanumeric.]{} - [Token prefixes: First characters with length three and four, if exists, of current token.]{} - [Token suffixes: Last characters with length one to four, if exists, of current token.]{} - [Word embeddings: Vector representations of words in the current window.]{} In addition to tailoring the features used by for tweets, there are other Twitter-specific aspects of our NER system such as using word embeddings trained on an unlabeled tweet corpus, applying normalization on labeled tweets, and extracting Twitter-specific keywords like hashtags, mentions, smileys, and URLs from both labeled and unlabeled Turkish tweets. For text normalization as a pre-processing step of our system, we used the Turkish normalization interface[^3] developed for social media text with ill formed word detection and candidate word generation [@Torunoglu-2014]. Along with the features used, the representation scheme for named entities is also important in terms of performance for a NER system. Two popular such encoding schemes are BIO and BILOU. The BIO scheme identifies the Beginning, the Inside and the Outside of the named entities, whereas the BILOU scheme identifies the Beginning, the Inside and the Last tokens of multi-token named entities, plus the Outside if it is not a named entity and the Unit length if the entity has single token. Since it is shown by that BILOU representation scheme significantly outperforms the BIO encoding scheme, we make use of BILOU encoding for tagging named entities in our study. Furthermore, we applied normalization to numerical expressions as described in , which helps to achieve a degree of abstraction to numerical expressions. Data Sets ========= Unlabeled Data -------------- In the unsupervised stage, we used two types of unlabeled data to obtain Turkish word embeddings. The first one is a Turkish news-web corpus containing 423M words and 491M tokens, namely the BOUN Web Corpus[^4] [@Sak-2008; @Sak-2011]. The second one is composed of 21M Turkish tweets with 241M words and 293M tokens, where we combined 1M tweets from TS TweetS[^5] by and 20M Turkish Tweets[^6] by Bolat and Amasyal[i]{}. We applied tokenization on both Turkish news-web corpus and Turkish tweets corpus using the publicly available Zemberek[^7] tool developed for Turkish. We have also applied lower-casing on both corpora in order to limit the number of unique words. Since our combined tweets corpus is composed of Twitter-specific texts, we applied what we call Twitter processing where we replaced mentions, hashtags, smileys and URLs with certain keywords. Labeled Data ------------ In the supervised stage, we used two types of labeled data to train and test our NER models. The first one is Turkish news data annotated with ENAMEX-type named entities, or PLOs [@Tur-2003]. It includes 14481 person, 9409 location, and 9034 organization names in the training partition of 450K words. This data set is popularly used for performance evaluation of NER systems for Turkish, including the ones presented by , by and by . The second type of labeled data is annotated Turkish tweets, where we used two different sets. The first set, TwitterDS-1, has around 5K tweets with 54K tokens and 1336 annotated PLOs [@Celikkaya-2013]. The second set, TwitterDS-2[^8], which is publicly available, has 2320 tweets with around 21K tokens and 980 PLOs in total [@Kucuk-2014-1]. The counts for each of the ENAMEX-type named entities for these Turkish Twitter data sets are provided in Table \[tab.0\]. -------------------------- ------------------ ------------------ Twitter DS-1* Twitter DS-2 (TwtDS-1) (TwtDS-2) Data Size (\#tokens) 54K 21K Person 676 457 Location 241 282 Organization 419 241 Total PLOs 1336 980 -------------------------- ------------------ ------------------ : Number of PLOs in Turkish Twitter data sets.[]{data-label="tab.0"} Experiments and Results ======================= We designed a number of experimental settings to investigate their effects on Turkish Twitter NER. These settings are as follows: the text type of annotated data used for training, the text type of unlabeled data used to learn the word embeddings, using the capitalization feature or not, and applying text normalization. We evaluated all models on ENAMEX types with the CoNLL metric and reported phrase-level overall F-score performance results. To be more precise, the F-score values presented in Table \[tab.1\], Table \[tab.2\] and Table \[tab.3\] are micro-averaged over the classes using the strict metric. NER Models Trained on News -------------------------- Most of our NER models are trained on annotated Turkish news data by and tested on tweets, due to the limited amount of annotated Turkish tweets. -- ----- --------- --------- ----------- Web Twt W+T ON 36.55 35.14 38.11 OFF 38.52 31.57 40.18 ON 41.82 41.54 OFF 40.50 39.31 41.04 ON 40.53 40.44 41.86 OFF 41.63 36.43 44.00 ON 45.74 46.27 OFF 44.17 44.91 45.27 ON 53.14 47.72 54.01 OFF 54.09 48.15 55.45 ON 55.20 52.23 OFF 54.75 49.43 56.12 -- ----- --------- --------- ----------- : Phrase-level overall F-score performance results of the NER models trained on news.[]{data-label="tab.1"} In addition to using TwitterDS-1 and TwitterDS-2 as test sets, we detected 291 completely non-Turkish tweets out of 5040 in TwitterDS-1 and filtered them out using the isTurkish[^9] tool [@Sahin-2013] to obtain TwitterDS-1\_FT. We also used the normalized versions of these data sets. As shown in Table \[tab.1\], turning off the capitalization feature is better when text normalization is not applied (bold entries), but the best results are achieved when normalization is applied and the capitalization feature is used (underlined bold entries). To observe the effects of the type of the source text used to learn the word embeddings, we have three models as Web, Twt, and Web+Twt where we used the Turkish web corpus, tweet corpus, and their combination respectively to learn the word embeddings. Including in-domain data from a relatively smaller tweet corpus together with a larger web corpus yields in better Twitter NER performance. ### Word Embeddings versus Text Normalization We examined the effects of word embeddings on the performance of our NER models, and compared them to the improvements achieved by applying normalization on Turkish tweets. The baseline NER model is built by using the features explained in section 3.2, except the capitalization and word embeddings features. Using word embeddings obtained with unsupervised learning from a large corpus of web articles and tweets results in better NER performance than applying a Twitter-specific text normalizer, as shown in Table \[tab.2\]. This is crucial since Turkish text normalization for unstructured data is a challenging task and requires successful morphological analysis, whereas extracting word embeddings for any language or domain is much easier, yet more effective. ----------------------- ----------- -------------- ----------- Twt Twt Twt DS-1 DS-1\_FT DS-2 Baseline(BL) 22.16 25.98 35.16 BL+Norm 33.05 39.23 37.17 BL+WordE 40.18 44.00 55.45 BL+WordE+Norm 41.04 45.27 56.12 Baseline(BL)+Cap 27.16 30.21 37.32 BL+Cap+Norm 36.70 40.78 42.18 BL+Cap+WordE 38.11 41.86 54.01 BL+Cap+WordE+Norm 42.79 46.61 56.79 ----------------------- ----------- -------------- ----------- : Phrase-level overall F-score performance results to compare word embeddings and normalization.[]{data-label="tab.2"} --------------------- -------------------- ----------- ---------- ------------------ --------------------- ------------- --------------- Phrase-level Gazet Norm Cap Other (Overall) Çelikkaya et al., Turkish News (2013) (Tür et al., 2003) Küçük et al., relaxed & extended TwtDS-1 36.11 (2014) gazetteer TwtDS-2 42.68 Küçük and diacritics expanded TwtDS-1 38.01 Steinberger, (2014) gazetteer TwtDS-2 48.13 word embeddings + filter non-Turkish word embeddings TwtDS-2 56.79 Turkish Tweets word embeddings + (TwtDS-2) filter non-Turkish --------------------- -------------------- ----------- ---------- ------------------ --------------------- ------------- --------------- NER Models Trained on Tweets ---------------------------- Although an ideal Turkish NER model for Twitter should be trained on similar informal texts, all previous Turkish Twitter NER systems are trained on news data due to the limited amount of annotated Turkish tweets. We also experimented training NER models on relatively smaller labeled Twitter data with 10-fold cross-validation. Our best phrase-level F-score of 46.61% achieved on TwitterDS-1\_FT is increased to 48.96% when trained on the much smaller tweets data, TwitterDS-2, instead of news data. Comparison with the State-of-the-art ------------------------------------ The best F-scores of the previously published Turkish Twitter NER systems [@Celikkaya-2013; @Kucuk-2014-1; @Kucuk-2014-2] as well as our proposed NER system are shown in Table \[tab.3\]. We used the same training set with the first system [@Celikkaya-2013] in our study, but the second NER system [@Kucuk-2014-1] uses a different multilingual news data and the third system [@Kucuk-2014-2], which is rule based, does not have a training phase at all. All of these previous NER systems use gazetteer lists for named entities, which are manually constructed and highly language-dependent, whereas our system does not. Note that there is no publicly available gazetteer lists in Turkish. achieved the state-of-the-art performance results for Turkish Twitter NER with their best model settings (shown in italic). These settings are namely using gazetteers list, with capitalization feature turned off, and with no normalization, together by expanding their gazetteer lists of named entities with diacritics variations. Our proposed system outperforms the state-of-the-art results on both Turkish Twitter data sets, even without using gazetteers (shown in bold). We achieved our best performance results with Turkish word embeddings obtained from our Web+Tweets corpus, when we apply normalization on tweets and keep the capitalization as a feature. Conclusion ========== We adopted a neural networks based semi-supervised approach using word embeddings for the NER task on Turkish tweets. At the first stage, we attained distributed representations of words by employing a fast unsupervised learning method on a large unlabeled corpus. At the second stage, we exploited these word embeddings together with language independent features in order to train our neural network on labeled data. We compared our results on two different Turkish Twitter data sets with the state-of-the-art NER systems proposed for Twitter data in Turkish and showed that our system outperforms the state-of-the-art results on both data sets. Our results also show that using word embeddings from an unlabeled corpus can lead to better performance than applying Twitter-specific text normalization. We discussed the promising benefits of using in-domain data to learn word embeddings at the unsupervised stage as well. Since the only language dependent part of our Turkish Twitter NER system is text normalization, and since even without text normalization it outperforms the previous state-of-the-art results, we believe that our approach can be adapted to other morphologically rich languages. Our Turkish Twitter NER system, namely TTNER, is publicly available[^10]. We believe that there is still room for improvement for NLP tasks on Turkish social media data. As a future work, we aim to construct a much larger in-domain resource, i.e., unlabeled Turkish tweets corpus, and investigate the full benefits of attaining word embeddings from in-domain data on Twitter NER. Acknowledgements ================ This research is partially supported by Boğaziçi University Research Fund Grant Number 11170. We would also like to thank The Scientific and Technological Research Council of Turkey (TÜBİTAK), The Science Fellowships and Grant Programmes Department (BİDEB) for providing financial support with 2210 National Scholarship Programme for MSc Students. Bibliographical References {#main:ref} ========================== [^1]: https://code.google.com/p/word2vec/ [^2]: http://cogcomp.cs.illinois.edu/Data/ACL2010\_NER\_Experim ents.php [^3]: http://tools.nlp.itu.edu.tr/Normalization [^4]: http://79.123.177.209/ hasim/langres/BounWebCorpus.tgz [^5]: http://tscorpus.com/en [^6]: http://www.kemik.yildiz.edu.tr/data/File/20milyontweet.rar [^7]: https://github.com/ahmetaa/zemberek-nlp [^8]: http://optima.jrc.it/Resources/2014\_JRC\_Twitter\_TR\_NER-dataset.zip [^9]: http://tools.nlp.itu.edu.tr/IsTurkish [^10]: http://tabilab.cmpe.boun.edu.tr/projects/ttner/	{ "pile_set_name": "ArXiv" }
--- abstract: \| Recent work in CT image reconstruction has seen increasing interest in the use of total variation (TV) and related penalties to regularize problems involving reconstruction from undersampled or incomplete data. Superiorization is a recently proposed heuristic which provides an automatic procedure to “superiorize” an iterative image reconstruction algorithm with respect to a chosen objective function, such as TV. Under certain conditions, the superiorized algorithm is guaranteed to find a solution that is as satisfactory as any found by the original algorithm with respect to satisfying the constraints of the problem; this solution is also expected to be superior with respect to the chosen objective. Most work on superiorization has used reconstruction algorithms which assume a linear measurement model, which in the case of CT corresponds to data generated from a monoenergetic X-ray beam. Many CT systems generate X-rays from a polyenergetic spectrum, however, in which the measured data represent an integral of object attenuation over all energies in the spectrum. This inconsistency with the linear model produces the well-known beam hardening artifacts, which impair analysis of CT images. In this work we superiorize an iterative algorithm for reconstruction from polyenergetic data, using both TV and an anisotropic TV (ATV) penalty. We apply the superiorized algorithm in numerical phantom experiments modeling both sparse-view and limited-angle scenarios. In our experiments, the superiorized algorithm successfully finds solutions which are as constraints-compatible as those found by the original algorithm, with significantly reduced TV and ATV values. The superiorized algorithm thus produces images with greatly reduced sparse-view and limited angle artifacts, which are also largely free of the beam hardening artifacts that would be present if a superiorized version of a monoenergetic algorithm were used. address: - '$^1$ Division of Engineering and Mathematics, University of Washington Bothell, Bothell, WA 98011' - '$^2$ Department of Mathematics, Oregon State University, Corvallis, OR 97331' author: - 'T. Humphries$^1$, J. Winn$^1$, and A. Faridani$^2$' title: 'Superiorized algorithm for reconstruction of CT images from sparse-view and limited-angle polyenergetic data' --- [Keywords]{}: computed tomography, superiorization, total variation, beam hardening Introduction ============ Superiorization [@CDH10; @HGDC12] is an optimization heuristic in which an iterative algorithm for solving an inverse problem is [superiorized]{} by perturbing the solution within each iteration, in order to improve it with respect to some objective function. The goal of the superiorized algorithm is to produce a solution that satisfies the constraints of the inverse problem to the same extent as any solution produced by the original algorithm, which is also superior with respect to this objective function. The authors prove in [@HGDC12] that this can be achieved provided that the original algorithm is [strongly perturbation resilient]{}, i.e., satisfies certain mathematical conditions that we describe later. Concurrently, a great deal of recent work in computed tomography (CT) has focused on incorporating total variation (TV) minimization into iterative reconstruction. One application that has attracted a great deal of interest is sparse-view imaging, in which the image is reconstructed from fewer views than would be acquired in a typical scan. Sparse-view imaging is motivated primarily by the desire to reduce dose to the patient [@BSHS10; @MCKL12], as well as by results from the field of compressive sensing, which state that signals that are sparse in some sense can be recovered from significantly fewer measurements than classical sampling theory dictates. A major result in compressive sensing is that under certain conditions on the measurement model (e.g. that it satisfies a restricted isometry property, or RIP), finding the sparsest signal that satisfies the measured data is equivalent to the much more tractable problem of finding the signal with the smallest $\ell_1$ norm. Early papers developing these results include [@CRT06a; @CRT06b; @D06b], and comprehensive treatments of the topic can be found, for example, in [@EK12; @FR13]. In CT imaging, it may be reasonable to assume that the image is piecewise constant with a relatively small number of edges, meaning that its discrete gradient is sparse. Total variation is simply the $\ell_1$ norm of the vector of discrete gradient magnitudes, and so TV minimization can therefore be used to find the image with the sparsest gradient which satisfies the measurements [@NW13b]. It must be noted, however, that the system matrices encountered in CT imaging have not been proven to satisfy necessary conditions, such as the RIP, for these theoretical results to be applicable [@JSP13; @JS15]. Superiorization does not claim to find a constraints-compatible solution that is [optimal]{} with respect to a chosen objective (e.g. sparsity), but rather a solution that is simply [superior]{} to the solution found by the original algorithm. The original algorithm is typically an iterative algorithm for solving large systems of linear equations, such as the algebraic reconstruction technique (ART) or one of its variants. The main appeal of superiorization is that it provides a straightforward method for modifying iterative algorithms in order to produce superior solutions, as well as theoretical guarantees that this is possible under certain circumstances. Total variation is a common choice of objective function, especially for studies on sparse-view CT; some examples of algorithms which have been applied in this context are ASD-POCS [@SP08; @BSHS10; @HBRSP12; @BYBH14], and methods presented in [@HD08; @CDH10; @HGDC12]. Other approaches for incorporating TV minimization into image reconstruction are also possible, and have been demonstrated in papers such as [@TNC09; @CWZS10; @DVL11; @RBFK11; @SJP12]. Aside from TV, algorithms have been superiorized using other objectives such as the $\ell_1$ norm of the Haar transform [@GHD11] and maximum entropy [@DHC09], though with less successful results. An anisotropic variant of TV was used to superiorize simultaneous ART (SART) in [@CJLW13] for the case of limited-angle CT, in order to exploit the directional nature of artifacts that occur in this case. Superiorization has also been applied to problems in proton computed tomography [@PSCR10], positron emission tomography [@GH14], and radiation therapy [@DCSG15]. The goal of this paper is to present and validate a superiorized algorithm for reconstruction of CT images from polyenergetic data. Algorithms for CT image reconstruction which use a linear measurement model (such as ART) implicitly assume that measurements are generated from a monoenergetic X-ray beam. Many clinical CT systems generate X-rays from a polyenergetic spectrum, which produces measured data that is a nonlinear function of the object attenuation. The inconsistency with the linear model produces the well-known beam hardening artifacts, such as cupping and streaking [@BD76]. These artifacts were first noted in early images of the brain [@JS78], and have also been shown to impede diagnosis in renal and cardiac imaging [@BHLB07; @RRDR10; @KGALL10]. Many techniques for addressing beam hardening have been developed over the years, including corrections applied to the measured data or the reconstructed image [@JS78; @H79; @HT83; @JR97; @KMPK10], as well as iterative reconstruction algorithms which incorporate X-ray polychromaticity directly into the imaging model [@HMDJ00; @DNDMS01; @EF02; @VVDB11; @LS14b]. The algorithm we choose to superiorize is polyenergetic SART (pSART) [@LS14b], which uses interpolation to model the attenuation map of the object at all energies as a function of the attenuation at a reference energy. We apply the superiorized algorithm in numerical phantom experiments modeling sparse-view and limited-angle reconstruction scenarios, using both TV and the anisotropic variant described in [@CJLW13]. While we are not able to prove that pSART is perturbation resilient – and hence that a superiorized version is guaranteed to produce solutions that are as constraint-compatible as the original algorithm – our superiorized version of pSART does produce solutions of this type in these numerical experiments. Methodology {#S:methods} =========== We first describe the general framework for superiorization of an iterative method. This is followed by a discussion of the mathematical model used for both monoenergetic and polyenergetic CT data, a description of the pSART method, and finally our description of the superiorized version of pSART. Superiorization --------------- The framework for superiorization is described in detail in [@HGDC12]; here we summarize the main points from that paper. As an illustrative example, we consider the problem of finding a non-negative solution to a linear system of equations $A {\bi{x}}= {\bi{b}}$ with $I$ equations for $J$ unknowns. The domain of the problem is the set of all vectors with $J$ non-negative components, $\Omega = R^J_+$. A [problem set]{}, $\mathbb{T}$, is the set of all possible matrices $A \in \mathbb{R}^{I \times J}$ and right-hand sides ${\bi{b}}\in \mathbb{R}^I$ which describe an inverse problem of this type; a [problem]{}, $T \in \mathbb{T}$, refers to the constraints imposed by a specific matrix and right-hand side. Since it is often impossible to satisfy all constraints of a problem exactly, we define a [proximity function]{}, $Pr$, which assigns to every problem $T$ an operator $Pr_T: \Omega \to \mathbb{R}_+$ that quantifies the extent to which a solution satisfies the constraints of the problem. A simple choice is the residual: $$Pr_T({\bi{x}}) = {\left\Vert{\bi{b}}- A {\bi{x}}\right\Vert}_2. \label{E:res_mono}$$ We then define an [$\varepsilon$-compatible]{} solution to the problem as any solution such that $Pr_T({\bi{x}}) \leq \varepsilon$, where $\varepsilon \geq 0$. Together, $\mathbb{T}$ and $Pr$ form a [problem structure]{}. An [algorithm]{}, $\mathbf{R}$, for a given problem structure assigns an operator $\mathbf{R}_T: \Delta \to \Omega$ to every problem $T$. The set $\Delta$ is such that $\Omega \subseteq \Delta \subseteq \mathbb{R}^J$; for example, $\mathbf{R}_T$ could take as input vectors in $\mathbb{R}^J$ with negative components, even if the desired output has only non-negative components. From any initial point ${\bi{x}}^{(0)}$, this operator produces a sequence of iterates, $$\left( (\mathbf{R}_T)^k {\bi{x}}^{(0)} \right)_{k=0}^\infty = {\bi{x}}^{(0)}, \mathbf{R}_T ( {\bi{x}}^{(0)} ), \mathbf{R}_T (\mathbf{R}_T ({\bi{x}}^{(0)})), \dots.$$ The algorithm $\mathbf{R}$ is said to be [strongly perturbation resilient]{} if the following two conditions are satisfied for every problem $T \in \mathbb{T}$: 1. There exists an $\varepsilon > 0$ such that $\mathbf{R}_T$ eventually produces an $\varepsilon$-compatible solution, from all starting points ${\bi{x}}^{(0)} \in \Omega$, and 2. For all $\varepsilon \in \mathbb{R}_+$ for which $\mathbf{R}_T$ is able to produce an $\varepsilon$-compatible solution, the modified iteration $${\bi{x}}^{(k+1)} = \mathbf{R}_T \left({\bi{x}}^{(k)} + \beta_k {\bi{v}}^{(k)} \right) \label{E:perturb}$$ produces a solution that is $\varepsilon'$-compatible, for any $\varepsilon' > \varepsilon$. Here $(\beta_k)^\infty_{k=0}$ represents a summable sequence of non-negative real numbers, and $({\bi{v}}^{(k)})^\infty_{k=0}$ is a bounded sequence of vectors ${\bi{v}}^{(k)} \in \mathbb{R}^J$. Essentially these conditions state that even if the solutions generated by $\mathbf{R}_T$ are perturbed after every iteration (subject to some restrictions on the size of these perturbations), one is able to obtain a solution that is equally as constraints-compatible as that obtained by the unperturbed algorithm. The authors of [@HGDC12] give some sufficient conditions on $\mathbf{R}$ and $Pr$ for the algorithm to be strongly perturbation resilient. The essential idea of superiorization is that the perturbation directions ${\bi{v}}^{(k)}$ can be chosen as nonascending directions of an objective function, $\phi : \Delta \to \mathbb{R}$. Thus, any strongly perturbation resilient algorithm can be “steered” according to this objective without jeopardizing its convergence towards an $\varepsilon$-compatible solution. We discuss how this is achieved practically in Section \[S:pSARTsup\]. Mathematical model ------------------ For the sake of simplicity we restrict our discussion to the two-dimensional case. Let $\mu({\bi{y}}, E): \mathbb{R}^2 \times \mathbb{R} \to \mathbb{R}$ denote the object attenuation as a function of position ${\bi{y}}$ and energy $E$. For a monoenergetic X-ray beam with energy $E_0$, the measurement along a line $\ell$ is given by $$\begin{aligned} I_\ell = I_0 \exp \left( -\int_\ell \mu({\bi{y}}, E_0) \: d {\bi{y}}\right), \textrm{ or} \label{E:mono1} \\ -\ln \left( I_\ell / I_0 \right) = \int_\ell \mu({\bi{y}}, E_0) \: d {\bi{y}}\label{E:mono2}\end{aligned}$$ where $I_\ell$ is the measured intensity and $I_0$ is the initial intensity of the beam. Suppose that $\mu({\bi{y}}, E_0)$ is discretized as an $n \times n$ pixel image, and we collect a total of $I$ measurements. Letting $J=n^2$, the data can be represented as a linear system of equations $${\bi{b}}= A {\bi{x}}, \label{E:mono3}$$ where ${\bi{x}}\in \mathbb{R}^{J}_+$ is a vector representing the image of $\mu({\bi{y}}, E_0)$, ${\bi{b}}\in \mathbb{R}^I_+$ is the log-transformed projection data (left side of ), and $A \in \mathbb{R}^{I \times J}$ is the system matrix whose $(i,j)$th element is the length of intersection of the $i$th line with the $j$th pixel of ${\bi{x}}$. If the X-ray beam is polyenergetic, becomes $$I_\ell = \int S(E) \exp \left( -\int_\ell \mu({\bi{y}}, E) \: d {\bi{y}}\right) d E,$$ where $S(E)$ represents the spectrum of the beam as a function of energy. For the purpose of image reconstruction, both the line integral and the integral with respect to energy are discretized, giving $$I_\ell = \sum_{h=1}^{N_h} S_h \exp \left( -a_i {\bi{x}}_h \right),\label{E:poly}$$ where $a_i$ is the row of $A$ corresponding to measuring along line $\ell$, $N_h$ is the number of quadrature points (i.e., energies) used in the discretization, $S_h$ is the quadrature weight corresponding to energy $E_h$, and ${\bi{x}}_h$ is the discretized image of $\mu({\bi{y}}, E_h)$. Without additional modeling, the image reconstruction problem is not tractable as it requires reconstructing the images ${\bi{x}}_h$ at $N_h$ different energy levels, from a single set of tomographic data. Thus, techniques for reconstruction of images from polyenergetic data must typically include a model for how $\mu({\bi{y}}, E)$ varies with energy. We discuss the model used by pSART in the next section. Polyenergetic SART ------------------ We first describe the simultaneous algebraic reconstruction technique (SART) [@AK84], an iterative algorithm for image reconstruction based on the linear model . Following [@CE02], we define diagonal matrices $D \in \mathbb{R}^{J \times J}$ and $M \in \mathbb{R}^{I \times I}$ with $$\begin{aligned} D_{jj} &= \frac{1}{\zeta_j}, ~~\zeta_j = \sum_{k=1}^m \|a_{kj}\|, j = 1 \dots J \\ M_{ii} &= \frac{1}{\eta_i}, ~~\eta_i = \sum_{k=1}^n \|a_{ik}\|, i = 1 \dots I. \label{E:diagmatrices} \end{aligned}$$ In other words, $\zeta_j$ is the 1-norm of the $j$th column of $A$, and $\eta_i$ is the 1-norm of the $i$th row. Furthermore, we organize the measurements corresponding to each angular view into $N_w$ subsets, denoted by $s(w)$ for $w = 1 \dots N_w$. Let $A_{s(w)}$ denote the matrix obtained by extracting only the rows of $A$ corresponding to $s(w)$, and similarly for $D_{s(w)}$, $M_{s(w)}$ and the vector of measurements, ${\bi{b}}_{s(w)}$. Then, a block-iterative algorithm for solving the problem $T$ described by is given by $$\begin{aligned} \mathbf{R}_T({\bi{x}}) = \mathbf{Q} \mathbf{B}_{N_w} \dots \mathbf{B}_2 \mathbf{B}_1 ({\bi{x}}), \label{E:SART1}\end{aligned}$$ where $$\begin{aligned} \mathbf{B}_w ({\bi{x}}) = {\bi{x}}- D_{s(w)} (A_{s(w)})^T M_{s(w)}\left[ A_{s(w)} {\bi{x}}- {\bi{b}}_{s(w)} \right],\label{E:SART2}\end{aligned}$$ and $$\begin{aligned} (\mathbf{Q} {\bi{x}})_j = \max\{0, x_j\},~~ j \in [1, J]. \label{E:SART3}\end{aligned}$$ Here $\mathbf{B}_w$ is an iterative step aiming to satisfy the constraints imposed by the $w$th subset of measurements, and $\mathbf{Q}$ is an operator that sets negative values of ${\bi{x}}$ to zero. The subsets can be chosen in several ways. For a parallel-beam geometry consisting of $N_v$ angular views of the object, the case where $N_w = N_v$ (i.e., one view per subset) corresponds to the classical version of SART [@AK84], while the case where $N_w=1$ (all views are processed simultaneously) is typically referred to as SIRT (simultaneous iterative reconstruction technique). In this paper we use an ordered subsets approach [@HL94], where each subset consists of $N_v / N_w$ equally spaced views around the object. Polyenergetic SART (pSART) [@LS14b] replaces the monoenergetic forward model used by SART with a log-transformed polyenergetic forward projection operator $\mathcal{P}: \mathbb{R}_+^{J} \to \mathbb{R}^I_+$, given by: $$[\mathcal{P} ({\bi{x}})]_i = - \ln \left[ \sum_{h=1}^{N_h} S_h \exp \left( - a_i {\boldsymbol \mu}({\bi{x}}, E_h) \right) \right] \biggl/ \sum_{h=1}^{N_h} S_h. \label{E:project_poly}$$ The image to be reconstructed, ${\bi{x}}$, is the attenuation map at some reference energy $E_0$. The function ${\boldsymbol \mu}: \mathbb{R}_+^J \times \mathbb{R} \to \mathbb{R}_+^J$ maps ${\bi{x}}$ to ${\bi{x}}_h$ (cf. ) using linear interpolation between tabulated attenuation curves for basis materials such as air, fat, soft tissue and bone. In particular, if $x_j$ denotes the linear attenuation coefficient (LAC) of pixel $j$ at $E_0$, then the attenuation coefficient of that pixel at all other energies is computed as $$\mu(x_j, E) = \frac{ [\mu_{m+1}(E_0) - x_j] \mu_{m}(E) + [x_j - \mu_{m}(E_0) ] \mu_{m+1}(E)}{\mu_{m+1}(E_0) - \mu_{m}(E_0)}, \label{E:mu_interp}$$ where $\mu_m(E)$ and $\mu_{m+1}(E)$ are the energy-dependent LAC functions for the two basis materials with LAC values adjacent to $x_j$ at the reference energy. So, if a pixel’s intensity is halfway betwen the LAC for fat and soft tissue at the reference energy, for example, then it is assumed that its value is halfway between the LAC for fat and soft tissue at all other energies as well. The pSART algorithm can be expressed in the same way as algorithm , but $\mathbf{B}_w$ changed to $$\begin{aligned} \mathbf{B}_w ({\bi{x}}) = {\bi{x}}- D_{s(w)} (A_{s(w)})^T M_{s(w)}\left[ \mathcal{P}_{s(w)}( {\bi{x}}) - {\bi{b}}_{s(w)} \right],\label{E:pSART1}\end{aligned}$$ where $\mathcal{P}_{s(w)}$ is the same as but using only rows of $A_{s(w)}$ rather than $A$, and the elements of ${\bi{b}}$ are given by $$b_\ell = -\ln \left( I_\ell \biggl/ \sum_{h=1}^{N_h} S_h \right).\label{E:pSART2}$$ Finally, the proximity function is the residual of the log-transformed data (cf. ): $$Pr_T({\bi{x}}) = {\left\Vert{\bi{b}}- \mathcal{P}({\bi{x}})\right\Vert}_2, \label{E:res_poly}$$ The pSART algorithm is a nonlinear fixed point iteration, whose convergence properties are not well-established. Our previous work in [@H15] shows that, in constrast to SART, convergence of the algorithm cannot be established in general. Numerical and physical phantom experiments in [@LS14b; @H15], however, indicate that it works well in practice, i.e., with realistic system matrices and measurements. We have chosen to superiorize pSART in this paper due to its similarity to SART and other algorithms that have been successfully superiorized. Superiorized pSART {#S:pSARTsup} ------------------ We consider superiorization with respect to two objective functions, total variation (TV) and anisotropic TV (ATV). Suppose that ${\bi{x}}$ is represented as a two-dimensional image, with $x_{m,n}$ denoting the pixel in the $m$th row and $n$th column of ${\bi{x}}$. We define a discrete, local gradient operator by $$\nabla {\bi{x}}_{m,n} = \left[ x_{m+1,n}-x_{m,n},~x_{m,n+1}-x_{m,n} \right]^T,$$ and can then define the TV as $$\begin{aligned} \phi_{TV}({\bi{x}}) &= \sum_{m,n} {\left\Vert \nabla {\bi{x}}_{m,n} \right\Vert}_2 \nonumber \\ &= \sum_{m,n} \sqrt{ (x_{m+1,n}-x_{m,n})^2 + (x_{m,n+1}-x_{m,n})^2 + \epsilon^2},\label{E:TV1}\end{aligned}$$ where $\epsilon$ is a small parameter which is added to ensure that the function is differentiable everywhere. The partial derivatives of $\phi_{TV}$ with respect to ${\bi{x}}_{m,n}$ can be computed straightforwardly, and form the gradient vector $\nabla \phi_{TV} ({\bi{x}}) \in \mathbb{R}^J$. The anisotropic TV (ATV) variant considered here was proposed in [@CJLW13] for limited-angle reconstruction. It is defined as $$\phi_{ATV}({\bi{x}}) = \sum_{i=1}^{N_\alpha} \omega_i \sum_{m,n} {\left\Vert \nabla_{\alpha_i} {\bi{x}}_{m,n}\right\Vert}_2, \label{E:ATV1}$$ where $$\nabla_{\alpha} {\bi{x}}_{m,n} = \left( \nabla {\bi{x}}_{m,n} \cdot {\bi{e}}_{\alpha} \right) {\bi{e}}_\alpha \label{E:ATV2}$$ and ${\bi{e}}_\alpha = [\cos \alpha,~\sin \alpha]^T$ is a unit vector which has angle $\alpha$ with the positive $x$-axis. The ATV is computed by selecting $N_\alpha$ directions ${\bi{e}}_\alpha$ and projecting the discrete, local gradient at every point onto the corresponding unit vector. The sum of the 2-norms of those vectors gives the total variation in the direction of ${\bi{e}}_\alpha$, and the total ATV is a weighted sum of these directional total variations. The weights, $\omega_i$, should sum to one and can be chosen to give higher weight along certain directions. In particular, reconstruction from limited-angle data produces artifacts whose directions are determined by the missing data (see Section \[S:SVLA\]), and so the $\omega_i$ can be chosen to have stronger weighting orthogonal to those directions. A heuristic for determining $\omega_i$ based on the missing angles, which we use in this paper, is described in [@CJLW13]. As in , we include a small parameter $\epsilon$ when computing ${\left\Vert \nabla_{\alpha_i} {\bi{x}}_{m,n}\right\Vert}_2$ to ensure differentiability, and form $\nabla \phi_{ATV} ({\bi{x}})$ from the partial derivatives with respect to ${\bi{x}}_{m,n}$. The superiorized version of pSART is presented in . The inputs to the algorithm are: a starting point ${\bi{x}}^{(0)}$, parameter $N$ specifying the number of inner loop iterations to perform, parameter $\gamma \in (0, 1)$ controlling the step size for minimizing $\phi$, and target constraint-compatibility value $\varepsilon_{target}$ indicating when to terminate the algorithm. The objective function $\phi$ is either $\phi_{TV}$ or $\phi_{ATV}$ in our experiments. The operators $\mathbf{B}_w$ are as specified in and require as additional input the system matrix $A$, log-transformed measurements ${\bi{b}}$, number of projection subsets $N_w$, and the quadrature weights $S_h$, reference energy $E_0$, and basis material attenuation curves used in and . $\mathbf{Q}$ is the non-negativity operator . $\ell \gets -1$ ${\bi{x}}^{(k,0)} \gets {\bi{x}}^{(k)}$ ${\bi{v}}^{(k,n)} \gets -\nabla \phi ({\bi{x}}^{(k,n)}) / \left( {\left\Vert \nabla \phi({\bi{x}}^{(k,n)}\right\Vert} + \delta \right)$ $\ell \gets \ell+1$ $\beta_{k,n} \gets \gamma^\ell$ ${\bi{z}}\gets {\bi{x}}^{(k,n)} + \beta_{k,n} {\bi{v}}^{(k,n)}$ ${\bi{x}}^{(k,n+1)} \gets {\bi{z}}$ ${\bi{x}}^{(k+1)} \gets \mathbf{Q} \mathbf{B}_{N_w} \dots \mathbf{B}_2 \mathbf{B}_1 ({\bi{x}}^{(k,N)})$ ${\bi{x}}^{(k+1)}$ The inner loop over $n$ (lines 4 to 12) is the key component that is added to pSART to superiorize it. On line 5 we set ${\bi{v}}^{(k,n)}$ to be the negative, normalized gradient at the point ${\bi{x}}^{(k,n)}$. The small parameter $\delta$ is included to ensure no division by zero. Lines 6 to 12 form another loop in which ${\bi{x}}^{(k,n)}$ is perturbed in the direction of ${\bi{v}}^{(k,n)}$ by a decreasing step size until a point ${\bi{z}}$ is found such that $\phi({\bi{z}}) \leq \phi({\bi{x}}^{(k)})$. One can prove inductively that since ${\bi{v}}^{(k,n)}$ is a nonascending vector of $\phi$ at ${\bi{x}}^{(k,n)}$, this loop must eventually terminate [@HGDC12]. This point ${\bi{z}}$ is then used in the next iteration of the inner loop over $n$. Following exactly $N$ repetitions of this loop, the resulting point is passed as input to pSART (line 13). We then check to see whether the solution is $\varepsilon$-compatible for some pre-specified value $\varepsilon_{target}$, and return the solution if so. The stopping criterion for the algorithm is based on producing an $\varepsilon_{target}$-compatible solution, so one must ensure that it is possible to achieve such a solution if the algorithm is ever to terminate. The value of $\varepsilon_{target}$ can be obtained by taking the value of $\varepsilon$ corresponding to a solution produced by the unsuperiorized version of the algorithm; if the algorithm is strongly perturbation resilient, then the superiorized version will be able to find an equally compatible solution [@HGDC12]. It is not clear, however, that pSART is actually a strongly perturbation resilient algorithm. Some sufficient conditions for strong perturbation resilience are given in [@HGDC12], which include that the algorithm is convergent for all problems; we have shown previously that this is not the case for pSART [@H15]. In the numerical experiments of the next section, however, our superiorized version of pSART is able to produce solutions that are equally constraints-compatible to those found by pSART. The superiorized pSART algorithm is similar to one previously presented by the authors in [@HF15], which we denoted as pSART-iTV. This algorithm also combined pSART with TV minimization, using a technique based on the improved TV (iTV) method of [@RBFK11]. The technique used by iTV is similar to superiorization, but the size of the descent steps during the TV minimization process is controlled using a different heuristic. In this paper we have chosen to use a superiorized version of pSART following the technique of [@HGDC12]. We will refer to the superiorized algorithm as pSART-TV or pSART-ATV, depending on the choice of the objective function $\phi$. We note that algorithms combining polyenergetic modeling and minimization of secondary objectives (such as TV) have been considered previously in some studies. The polyenergetic iterative method of [@EF02] includes an edge-preserving Huber penalty which is similar to TV, though the primary purpose of this penalty was to denoise the image, and the paper did not consider sparse-view or limited-angle data. In [@BB15], the authors use a compressive sensing approach to reconstruct sparse-view polyenergetic data, but used constraints based on wavelet sparsity and matrix factorization rather than TV. Finally, an algorithm for material decomposition from spectral CT data presented in [@BSSP16] incorporates polyenergetic modeling within each energy window as well as a TV minimization constraint. In this paper we do not consider the case of spectral CT, in which photon-counting detectors allow for simultaneous acquisition of several data sets in different energy windows, but rather the more traditional model where the data represent an integral over the entire range of X-ray energies. Beam-hardening artifacts are likely to be more severe in this case as the attenuation coefficients of materials vary more over a larger energy range. Numerical Experiments {#S:results} ===================== In this section we demonstrate the effectiveness of the superiorized pSART algorithm using numerical phantom experiments. As our numerical experiments are concerned with sparse-view and limited-angle data, we first define what is meant by these two terms. Sparse-view and limited-angle data {#S:SVLA} ---------------------------------- Recall that the set of all line integrals through a function $f: \mathbb{R}^2 \to \mathbb{R}$ is given by the Radon transform: $$\begin{aligned} \mathcal{R} f (\theta,s)= \int_{-\infty}^\infty f \left(s \boldsymbol \xi(\theta) + t \boldsymbol \xi^{\perp} (\theta) \right) \: dt, \label{E:radon} \\ \qquad \textrm{where } \boldsymbol \xi(\theta) = (\cos \theta, \sin \theta)^T, \: \boldsymbol \xi^\perp(\theta) = (-\sin \theta, \cos \theta)^T. \nonumber\end{aligned}$$ This transform is frequently used to model X-ray data. Assuming $f$ is compactly supported on the unit disc, it can be recovered exactly from $\mathcal{R}f$ if the latter is known for all $\theta \in [-\pi/2, \pi/2)$ and $s \in [-1, 1]$. In practice one only measures X-ray data for a discrete set of values $\theta_j$ and $s_l$; for a parallel-beam geometry with equally-spaced views and rays measured over a 180$^\circ$ arc, these are given by: $$\begin{aligned} \theta_j = \pi(j-1)/p, \qquad & j = 1,2, \dots p, \label{E:sampling1}\\ s_l = l/q, & l = -q,-q+1, \dots q \nonumber.$$ This discrete sampling of $\mathcal{R}f$ can be used to reconstruct an image of $f$ accurately, provided that the rate of sampling is adequate. In particular, if $f$ is $b$-band limited (i.e., its Fourier transform vanishes for frequencies higher than $b$), then it is sufficient to take $p \geq b$ and $q \geq b/ \pi$ [@N86] (p. 71). Furthermore, to reconstruct an $n \times n$ pixel image of $f$, one can simply take $p = \lceil n \pi / 2 \rceil$ and $q = \lceil n/2 \rceil$, since features with frequency higher than $n\pi/2$ cannot be represented at that resolution. This corresponds to acquiring $p$ views of the object equally spaced over $[-90^\circ,90^\circ)$ – or equivalently, $[0^\circ, 180^\circ)$ – with $n+1$ measurements per view[^1]. This choice of $p$ is conservative, as the true value of $b$ may be less than $n\pi / 2$. We can then define sparse-view imaging as the case where $p \ll n \pi/2$, and limited-angle tomography as the case where there exist indices $1 \leq j_a < j_b \leq p$ such that either $$\begin{aligned} j = j_a, j_a + 1, \dots j_b, \textrm{ or} \label{E:limitedangle} \label{E:sampling2} \\ j = 1, 2, \dots j_a, j_b, j_b +1, \dots p, \nonumber \end{aligned}$$ in , i.e., where some portion of the angular range is missing. shows some characteristic artifacts that arise in these two cases. For this 200$\times$200 pixel phantom, we have $n\pi/2 = 314$, which we round up to the more convenient value of $p=360$ to define complete data. The sparse-view reconstruction (centre right image) uses $p=36$, producing high-frequency artifacts characteristic of this case. The limited-angle reconstruction (right image) uses $p=360$ but $j = 1, \dots 60, 181, \dots 3600$, corresponding to 240 views acquired over $[0^\circ,120^\circ)$. In this case the artifacts consist of lost edges and streaks whose orientations are determined by the missing values of $\theta$. Specifically, one sees that edges are lost corresponding to the missing range of angles $\theta \in \left[ \pi / 6, \pi / 2\right)$, when there are no measurements acquired along lines tangent to those edges. These artifacts are characteristic of limited-angle reconstructions [@Q93; @FQ13]. ![Numerical phantom experiment showing artifacts characteristic of sparse-view and limited-angle reconstructions. All reconstructions are generated from monoenergetic data using SART. Left: Discretized phantom at resolution of $200 \times 200$ pixels. Centre left: reconstruction from complete data consisting of 360 equally spaced views on $[0^\circ,180^\circ)$. Centre right: sparse-view reconstruction from 36 equally spaced views on $[0^\circ,180^\circ)$. Right: limited-angle reconstruction from 240 equally spaced views on $[0^\circ,120^\circ)$. []{data-label="F:artifacts"}](fig2.pdf){width="\linewidth"} Mathematical head phantom ------------------------- Our first set of numerical experiments use the FORBILD head phantom [@YNDW12]. The phantom at a reference energy of 70 keV is shown in Fig. \[F:FORBILD\]. The intensity of the ROIs are all within $\pm 1\%$ of the background soft tissue intensity of 0.203 cm$^{-1}$, so all images are displayed on a narrow greyscale window of $[0.195, 0.215]$ cm$^{-1}$. The attenuation coefficient of the bone (white regions) at 70 keV is 0.495 cm$^{-1}$. The phantom size is $800 \times 800$ pixels with a pixel width of 0.375 mm. We generated several data sets from this phantom. The first set consisted of “consistent” projection data generated from the discrete phantom, using the same forward model assumed by pSART , . Thus the true phantom represented the exact solution to the nonlinear system of equations . This situation is unrealistic (an example of the “inverse crime”) but is useful for examining the performance of the algorithm in the ideal case. The 130 kVp spectrum is shown in and was generated using the Michigan Image Reconstruction Toolbox (MIRT) [@Fessler]. We refer to this consistent data set as Set C. We also generated three inconsistent data sets to test the performance of the approach under more realistic circumstances. For all three data sets, projections were generated using analytically computed line integrals through the FORBILD phantom, and Poisson-distributed noisy measurements were generated from the analytic data. Additionally, we introduced error into the spectrum used for reconstruction, by computing an approximation using the composite trapezoid rule. The parameters for these three data sets were: - (Set I-1) 130 kVp spectrum with noise proportional to $\displaystyle \int I_0(E) \: dE = 4 \times 10^6$. - (Set I-2) 130 kVp spectrum with $\displaystyle \int I_0(E) \: dE = 1 \times 10^6$. - (Set I-3) 80 kVp spectrum (see ) with $\displaystyle \int I_0(E) \: dE = 4 \times 10^6$. Sets I-2 and I-3 are more challenging to accurately reconstruct than Set I-1, due to the higher level of noise in both datasets and lower average energy of the spectrum in Set I-3, which exacerbates beam hardening artifacts. The use of a low energy spectrum of this type is common in dual-energy CT imaging. ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![Left: FORBILD phantom at reference energy of 70 keV. Centre left: Normalized 130 kVp and 80 kVp spectra used to generate polyenergetic data. Centre right: Image reconstructed using from soft tissue corrected data using SART. Right: image reconstructed from soft tissue corrected data using TV-superiorized SART.[]{data-label="F:FORBILD"}](fig3a.pdf "fig:"){width="0.2\linewidth"} ![Left: FORBILD phantom at reference energy of 70 keV. Centre left: Normalized 130 kVp and 80 kVp spectra used to generate polyenergetic data. Centre right: Image reconstructed using from soft tissue corrected data using SART. Right: image reconstructed from soft tissue corrected data using TV-superiorized SART.[]{data-label="F:FORBILD"}](fig3b.pdf "fig:"){height="0.2\linewidth"} ![Left: FORBILD phantom at reference energy of 70 keV. Centre left: Normalized 130 kVp and 80 kVp spectra used to generate polyenergetic data. Centre right: Image reconstructed using from soft tissue corrected data using SART. Right: image reconstructed from soft tissue corrected data using TV-superiorized SART.[]{data-label="F:FORBILD"}](fig3c.pdf "fig:"){width="0.2\linewidth"} ![Left: FORBILD phantom at reference energy of 70 keV. Centre left: Normalized 130 kVp and 80 kVp spectra used to generate polyenergetic data. Centre right: Image reconstructed using from soft tissue corrected data using SART. Right: image reconstructed from soft tissue corrected data using TV-superiorized SART.[]{data-label="F:FORBILD"}](fig3d.pdf "fig:"){width="0.2\linewidth"} ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- As discussed in the previous section, reconstructing an $800 \times 800$ pixel image requires (conservatively) $400 \pi \approx 1250$ views acquired over 180$^\circ$. For convenience, we took $p=1440$ to obtain an integer multiple of 360, and defined this choice along with $j = 1, 2, \dots p$ to be the “complete” data set for both the consistent and inconsistent data. Our sparse-view and limited-angle experiments were then obtained by reducing $p$ and/or taking a subset of $j$ values as described in . Before presenting the results of our experiments, we provide some motivation for superiorizing a polyenergetic reconstruction algorithm such as pSART. One possible approach to reconstructing an image from sparse-view or limited-angle polyenergetic data would be to generate soft tissue corrected data and then use a superiorized version of a monoenergetic algorithm, such as SART. Soft tissue correction [@JS78; @HT83] is a simple procedure in which monoenergetic data is simulated from polyenergetic data using the following steps: 1. Solve $\displaystyle I_\ell = \sum_{h=1}^{N_h} S_h \exp \left(-\mu_s(E_h)T_{\ell} \right)$ for $T_\ell$, for all measurements $\ell$. 2. Let $m_\ell = I_0 \exp \left(-\mu_s(E_0\right) T_\ell)$. Here $\mu_s(E)$ denotes the attenuation coefficient of soft tissue at energy $E$. The first step consists of solving a nonlinear equation for each measurement to determine the equivalent length, $T_\ell$, of soft tissue through which ray $\ell$ would have to pass to generate measurement $I_\ell$. The second step then generates the corresponding monoenergetic data at energy $E_0$, which can be reconstructed with a conventional algorithm, such as SART. While this technique is effective in lessening cupping artifacts, it does not remove streaking artifacts caused by high-attenuation material such as bone or contrast agents. shows that this is true also for a TV-superiorized version of SART. The image in the centre right shows the FORBILD phantom reconstructed from 288 soft tissue corrected views using SART, while the image on the right shows a reconstruction from the same data using TV-superiorized SART. While the TV minimization is effective in removing high-frequency artifacts caused by the sparse-view data, it does not remove the streaking artifacts caused by beam hardening. While techniques similar to soft tissue correction can be extended to account for artifacts of this nature [@JS78; @JR97], they typically require image segmentation and do not account for mixtures of different tissue types within pixels. By superiorizing pSART we avoid both of these issues. Sparse-view experiments ----------------------- In addition to the complete data set with $p=1440$, we reconstructed data using $p = 720$, 360, 288, 144 and 72. Images were reconstructed using both pSART and pSART-TV, beginning with an initial guess of ${\bi{x}}^{(0)} = \mathbf{0}$ in all cases, using a reference energy of 70 keV for data sets C, I-1 and I-2, and 50 keV for data set I-3, due to the lower average energy of the spectrum. Air, soft tissue, and bone were used as the basis materials, with the energy-dependent attenuation curves obtained from the MIRT. The number of subsets, $N_w$, was chosen such that every subset consisted of 12 views; this corresponds to $N_w = 120$ for $p=1440$, down to $N_w = 6$ for $p=72$. pSART-TV was run with $\gamma = 0.999$ and $N = 20$ as the counter for the TV minimization loop. As discussed in Section \[S:pSARTsup\], the parameter $\varepsilon_{target}$ should correspond to a solution produced by the unsuperiorized algorithm, pSART. This leads to the question of when pSART should be terminated to generate this solution. In our experiments we chose to terminate the algorithm after a fixed number of iterations were run. This was the approach used in [@LS14b] and has the advantage of being simple, although the total number of iterations to run must be determined ad-hoc. When using inconsistent data we ran fewer iterations of pSART since it exhibits semiconvergence behaviour similar to SART and other methods (see e.g. [@EHN14]). Semiconvergence refers to the degradation of image quality after a certain number of iterations, as the algorithm begins fitting noise and other inconsistencies in the data. In both the consistent and inconsistent data case the total number of iterations to run was increased as $p$ decreased, since the cost per iteration decreases proportionately to the number of views. This ensured that the same amount of computation (in terms of forward projection operations) was done for all values of $p$. Some representative images are shown in . In the consistent data case (top row), the image reconstructed by pSART-TV using $p=288$ (fourth column) is virtually indistinguishable from the image obtained with $p=1440$ (second column), though there is a slight artifact on the left side of the phantom opposite the ear insert. This artifact becomes more pronounced when $p$ is reduced further to 144 views (sixth column), with additional artifacts also arising. Due to the polyenergetic forward model, the images are virtually free of any beam hardening artifacts (cf. ), although the streak caused by the ear insert is faintly visible. For data sets I-1, I-2 and I-3, the image quality is diminished due to the various sources of inconsistency. The reconstructed intensity values from data set I-1 (second row) are too low due to the mismatch in the spectrum, and the large streak caused by the ear insert is somewhat more prominent than for the consistent data. The addition of noise also causes the images to have a patchier appearance, and makes it more difficult to discern the three small low-contrast ROIs near the edge of the skull, especially as $p$ decreases. When the noise level increases in Set I-2, this problem is exacerbated and some of the low-contrast features are not apparent in the image. Finally, in the images reconstructed from Set I-3 (which are displayed using a greyscale window of $[0.235, 0.255]$ cm$^{-1}$ due to the lower reference energy), the noise level is not as high as in Set 1-B, but the beam hardening artifacts are somewhat more noticeable, as one might expect from a lower energy spectrum. ![Images of FORBILD phantom reconstructed from sparse-view data using pSART and pSART-TV. From top to bottom, rows correspond to images reconstructed from data sets C, I-1, I-2 and I-3, respectively. First, third and fifth columns show images reconstructed with pSART, while second, fourth and sixth columns show corresponding images reconstructed with pSART-TV. Number of views used for each column pair is indicated at the bottom.[]{data-label="F:sparseview_images"}](fig4a_new.pdf "fig:"){width="\linewidth"}\ ![Images of FORBILD phantom reconstructed from sparse-view data using pSART and pSART-TV. From top to bottom, rows correspond to images reconstructed from data sets C, I-1, I-2 and I-3, respectively. First, third and fifth columns show images reconstructed with pSART, while second, fourth and sixth columns show corresponding images reconstructed with pSART-TV. Number of views used for each column pair is indicated at the bottom.[]{data-label="F:sparseview_images"}](fig4b_new.pdf "fig:"){width="\linewidth"} ------------ -- -- $p = 1440$ ------------ -- -- While pSART-TV did not produce high quality images in all cases, the key claim of superiorization is simply that the superiorized algorithm should be able to find solutions that are as constraints-compatible as those found pSART, with lower TV. This was the case in every experiment that was run. Plots of relevant quantities are shown in . The leftmost plot shows the $\varepsilon$ values obtained by both pSART and pSART-TV in every simulation, as a function of the number of views. The much smaller value of $\varepsilon$ for Set C (black curve) can be attributed to the fact that the data were consistent, and also that pSART was run for more iterations than in the inconsistent data cases. Additionally, since $\varepsilon$ is not normalized (see ), the values also decrease as fewer views are included. Importantly, the middle plot shows that the images found by pSART-TV had substantially smaller TV than the solutions found by pSART, as desired. One can also see that the TV values of the images reconstructed using pSART-TV were fairly consistent as $p$ decreased, and relatively close to the true TV values of the phantom at 70 keV (for the first three data sets) or 50 keV (for Set I-3). Finally, the rightmost plot shows the total number of iterations required by pSART and pSART-TV to achieve the same $\varepsilon$ value in each experiment. Unsurprisingly, pSART-TV required more iterations than pSART in nearly every case, as the perturbation introduced at every iteration competes with the goal of fitting the data (i.e. reducing $\varepsilon$). This was especially true for the noisy data sets. ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![Plots of important quantities from sparse-view experiments. Left plot: Values of $\varepsilon$ obtained by pSART and pSART-TV for each simulation. Middle plot: TV values for pSART reconstructions plotted against TV values for pSART-TV reconstructions.Black dashed line is the line $y=x$, blue and green dashed lines represent the true TV of the phantom at 70 keV and 50 keV, respectively. Leftmost points correspond to $p=1440$ in all cases, with $p$ decreasing as one follows the line. Right plot: number of iterations required by each algorithm for each simulation. Leftmost points correspond to $p=1440$, as before.[]{data-label="F:sparseview_plots"}](fig5a.pdf "fig:"){height="0.2\linewidth"} ![Plots of important quantities from sparse-view experiments. Left plot: Values of $\varepsilon$ obtained by pSART and pSART-TV for each simulation. Middle plot: TV values for pSART reconstructions plotted against TV values for pSART-TV reconstructions.Black dashed line is the line $y=x$, blue and green dashed lines represent the true TV of the phantom at 70 keV and 50 keV, respectively. Leftmost points correspond to $p=1440$ in all cases, with $p$ decreasing as one follows the line. Right plot: number of iterations required by each algorithm for each simulation. Leftmost points correspond to $p=1440$, as before.[]{data-label="F:sparseview_plots"}](fig5b.pdf "fig:"){height="0.2\linewidth"} ![Plots of important quantities from sparse-view experiments. Left plot: Values of $\varepsilon$ obtained by pSART and pSART-TV for each simulation. Middle plot: TV values for pSART reconstructions plotted against TV values for pSART-TV reconstructions.Black dashed line is the line $y=x$, blue and green dashed lines represent the true TV of the phantom at 70 keV and 50 keV, respectively. Leftmost points correspond to $p=1440$ in all cases, with $p$ decreasing as one follows the line. Right plot: number of iterations required by each algorithm for each simulation. Leftmost points correspond to $p=1440$, as before.[]{data-label="F:sparseview_plots"}](fig5c.pdf "fig:"){height="0.2\linewidth"} ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Limited-angle experiments ------------------------- To conduct limited-angle experiments, we reconstructed images from data acquired over angular extents of 165$^\circ$ and 150$^\circ$, corresponding to $1320$ and $1200$ total views, respectively. In all cases we used $p=1440$; for limited angle data this parameter is no longer equal to the total number of views, but rather defines the spacing between consecutive views as $180^\circ / p$, or $0.2^\circ$ in this case. For each angular extent, two different acquisition arcs were considered; one consisting of views acquired between $0^\circ$ and $180^\circ$, and the another consisting of views acquired between $90^\circ$ and $270^\circ$. For example, the two $165^\circ$ acquisitions consisted of views acquired on $[97.5^\circ, 262.5^\circ)$ and $[7.5^\circ, 172.5^\circ)$. With respect to Equations and , these two cases correspond to $j = 61, 62, \dots 1380$ and $j = 1, 2, \dots 660, 781, 782, \dots 1440$, respectively. For ease of reference, however, we will denote each acquisition arc by the starting angle (first view acquired) and the angular extent. All images were reconstructed using both pSART and pSART-ATV with an initial guess of ${\bi{x}}^{(0)} = \mathbf{0}$. For the ATV calculation , we took $N_\alpha=4$ and $\alpha=\{0^\circ, 45^\circ, 90^\circ, 135^\circ\}$. As in the sparse-view case, the number of subsets $N_w$ was chosen such that every subset consisted of 12 views. Because the artifacts introduced by the limited-angle problem were, in general, more severe than those encountered in the sparse-view problem, we altered the pSART-ATV parameters to $\gamma = 0.9999$ and $N = 60$, which we found was more effective in lessening artifacts. As in the sparse-view experiments, we terminated pSART after a fixed number of iterations, increasing that number as the angular extent decreased, and running fewer iterations for the inconsistent data cases. Representative images of the limited-angle reconstructions are shown in . The images obtained from the $165^\circ$ arcs and consistent data (top row, second and fourth columns) are generally accurate reconstructions of the phantom; some artifacts characteristic of limited-angle reconstructions are evident, but they are greatly reduced compared to the corresponding images reconstructed with pSART. When the starting view is $7.5^\circ$, the missing views correspond to the horizontal axis (and nearby angles) and so the streaks and lost edges occur along directions tangent to those missing views. When the starting view is $97.5^\circ$, the missing directions are close to the vertical axis and so the streaks occur along those directions, and are generally less severe since the phantom has fewer sharp edges in the vertical direction. When the angular extent is further reduced to $150^\circ$ (sixth and eighth columns), the artifacts become more severe and pSART-ATV is not able to remove them entirely. Artifacts caused by beam hardening, however, are only faintly visible (if at all). ![Images of FORBILD phantom reconstructed from limited angle data by pSART and pSART-ATV. Layout of rows and columns is the same as .[]{data-label="F:limitedangleimages"}](fig6a_new.pdf "fig:"){width="\linewidth"}\ ![Images of FORBILD phantom reconstructed from limited angle data by pSART and pSART-ATV. Layout of rows and columns is the same as .[]{data-label="F:limitedangleimages"}](fig6b_new.pdf "fig:"){width="\linewidth"} --------------------- --------------------- ---------------------- --------------------- Extent: 165$^\circ$ Extent: 165$^\circ$ Extent: 150$^\circ$ Extent: 150$^\circ$ Start: $97.5^\circ$ Start: $7.5^\circ$ Start: $105.0^\circ$ Start: $15.0^\circ$ --------------------- --------------------- ---------------------- --------------------- Reconstructions from ithe inconsistent data sets are shown in the second, third and fourth rows. As in the sparse-view case, there is a loss of image quality due to the sources of inconsistency in the data; nonetheless, the superiorized algorithm remains effective at removing limited-angle artifacts. These artifacts appear to be no more severe than they were in the consistent data case, with the exception of the images in the sixth column (starting view of $105^\circ$). This can be attributed primarily to the fact that the image reconstructed from consistent data (top row) was produced from 128 iterations of pSART-ATV, while the images reconstructed from inconsistent data (second to fourth row) were produced from between 43–54 iterations. We discuss the relation between the stopping criterion for pSART and the number of iterations run by pSART-ATV later in this section. Aside from this, we note that as before, data sets I-2 and I-3 were more challenging to reconstruct than data set I-1; the images are visibly noisier, and the beam hardening artifact is more severe with data set I-3. shows a plot of relevant quantities from these experiments. As in the sparse-view simulations, pSART-ATV was able to find a solution with equal $\varepsilon$ value to pSART in every experiment, and with substantially lower ATV. In this case, the values of $\varepsilon$ obtained across different experiments (upper left plot) were fairly consistent, due in large part to the fact that the total number of measurements used to calculate the residual did not vary as much as in the sparse-view experiments. The plots of the ATV values (upper middle and right plots) indicate that the ATV values of the images reconstructed by pSART-ATV were substantially lower than those of the images reconstructed by pSART. For the three data sets using a 130 kVp spectrum, the ATV values of the reconstructed images were also fairly close to the average ATV of the true phantom at 70 keV (the ATV varies slightly across experiments as the weighting factors $\omega_i$ in depend on the angular extent and starting angle). For data set I-3, the average ATV of the true phantom was roughly 5730, somewhat lower than the reconstructed values. ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ![Plots of important quantities from limited-angle experiments. Upper left plot: Values of $\varepsilon$ obtained by pSART and pSART-ATV for each simulation. Points marked with $+$ correspond to images in first and second columns of , moving across that figure as one follows the line. Upper middle plot: TV values for pSART reconstructions plotted against TV values for pSART-ATV reconstructions. Blue dashed line represents the average true ATV of the phantom at 70 keV. Upper right plot: equivalent plot for data set I-3, shown on different axes. Bottom middle and left figures: Number of iterations required by each algorithm for each simulation.[]{data-label="F:limitedangle_plots"}](fig7a.pdf "fig:"){height="0.2\linewidth"} ![Plots of important quantities from limited-angle experiments. Upper left plot: Values of $\varepsilon$ obtained by pSART and pSART-ATV for each simulation. Points marked with $+$ correspond to images in first and second columns of , moving across that figure as one follows the line. Upper middle plot: TV values for pSART reconstructions plotted against TV values for pSART-ATV reconstructions. Blue dashed line represents the average true ATV of the phantom at 70 keV. Upper right plot: equivalent plot for data set I-3, shown on different axes. Bottom middle and left figures: Number of iterations required by each algorithm for each simulation.[]{data-label="F:limitedangle_plots"}](fig7b.pdf "fig:"){height="0.2\linewidth"} ![Plots of important quantities from limited-angle experiments. Upper left plot: Values of $\varepsilon$ obtained by pSART and pSART-ATV for each simulation. Points marked with $+$ correspond to images in first and second columns of , moving across that figure as one follows the line. Upper middle plot: TV values for pSART reconstructions plotted against TV values for pSART-ATV reconstructions. Blue dashed line represents the average true ATV of the phantom at 70 keV. Upper right plot: equivalent plot for data set I-3, shown on different axes. Bottom middle and left figures: Number of iterations required by each algorithm for each simulation.[]{data-label="F:limitedangle_plots"}](fig7c.pdf "fig:"){width="0.2\linewidth"} \[10pt\] ![Plots of important quantities from limited-angle experiments. Upper left plot: Values of $\varepsilon$ obtained by pSART and pSART-ATV for each simulation. Points marked with $+$ correspond to images in first and second columns of , moving across that figure as one follows the line. Upper middle plot: TV values for pSART reconstructions plotted against TV values for pSART-ATV reconstructions. Blue dashed line represents the average true ATV of the phantom at 70 keV. Upper right plot: equivalent plot for data set I-3, shown on different axes. Bottom middle and left figures: Number of iterations required by each algorithm for each simulation.[]{data-label="F:limitedangle_plots"}](fig7d.pdf "fig:"){height="0.2\linewidth"} ![Plots of important quantities from limited-angle experiments. Upper left plot: Values of $\varepsilon$ obtained by pSART and pSART-ATV for each simulation. Points marked with $+$ correspond to images in first and second columns of , moving across that figure as one follows the line. Upper middle plot: TV values for pSART reconstructions plotted against TV values for pSART-ATV reconstructions. Blue dashed line represents the average true ATV of the phantom at 70 keV. Upper right plot: equivalent plot for data set I-3, shown on different axes. Bottom middle and left figures: Number of iterations required by each algorithm for each simulation.[]{data-label="F:limitedangle_plots"}](fig7e.pdf "fig:"){height="0.2\linewidth"} ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ The iterations required by pSART and pSART-ATV to obtain the same $\varepsilon$ value are shown in the bottom two plots. The “N” shape of the plot for all four data set indicates that the reconstructions from data acquired over the $0^\circ$ to $180^\circ$ arcs (fourth and eighth columns of ) generally required many more iterations of pSART-ATV to obtain an equivalent $\varepsilon$ value than the reconstructions from corresponding data acquired over the $90^\circ$ to $270^\circ$ arc (second and sixth columns). This is in spite of the fact that the $\varepsilon$ value in both cases was virtually the same. This phenomenon is difficult to explain, but is likely attributable to the differing nature of the artifacts that arose depending on which subset of angular data was missing. In general, the relationship between the value of $\varepsilon$ obtained by pSART and the number of iterations that pSART-ATV would require to obtain this value was difficult to predict. As desired, however, pSART-ATV was eventually able to obtain a $\varepsilon$-compatible solution in every case. Anatomical phantom ------------------ As a more clinically realistic test case, we performed a second set of experiments using the XCAT anatomical phantom [@SSMGT10]. The phantom size was $512 \times 512$ pixels with a pixel width of 0.75 mm. The simulation modeled a transaxial slice through the torso, including iodinated contrast agent in the bloodstream. The inclusion of iodine requires a modification to the pSART algorithm to determine whether to interpolate between soft tissue and iodine, or soft tissue and bone, within image pixels with LAC values higher than soft tissue [@LS14b]. This is accomplished by a user-provided mask indicating which pixels are likely to contain bone, under the assumption that this can be determined [a priori]{}. Simulated data were generated using XCAT’s analytic projection algorithm [@SMBFT08]. The same 80 kVp spectrum was used as in the FORBILD phantom experiments, with noise proportional to $\int I_0(E) = 2 \times 10^5$. To model a realistic clinical acquisition, a fan-beam geometry was simulated with source to iso-centre distance of 80 cm, source-to-detector distance of 160 cm, and fan beam half-angle of 17.5$^\circ$. Since the fan-beam geometry is not symmetric over 180$^\circ$, we collected 900 views over a 360$^\circ$ rotation arc. The fan-beam system matrices used for the reconstruction algorithms were generated using the MIRT. An image of the phantom is shown in , along with an image reconstructed by SART using soft tissue corrected data. While the beam hardening artifacts are not as apparent in this image as they were for the FORBILD phantom (due in part to the wider greyscale used to display the image), a subtraction of the soft tissue corrected SART reconstruction from the reconstruction obtained from the same dataset using pSART indicates that these artifacts are present in the former. In particular, there are streaking artifacts caused by bony structures such as the ribs and sternum, and the $\mu$ values of the body tissue and contrast agent tend to be underestimated. ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![XCAT anatomical phantom used for experiments with fan-beam data. Left: True phantom at reference energy of 50 keV. Phantom is displayed on a greyscale window of $[0.1, 0.35]$ cm$^{-1}$ to emphasize soft tissue contrast and artifacts in reconstructed images. Centre: image reconstructed using SART from water-corrected data. Right: Difference between image reconstructed using pSART (leftmost image in ) and image reconstructed using SART. Greyscale window is $[-0.04, 0.10]$ cm$^{-1}$.[]{data-label="F:xcat_phantom"}](fig8a.pdf "fig:"){width="0.25\linewidth"} ![XCAT anatomical phantom used for experiments with fan-beam data. Left: True phantom at reference energy of 50 keV. Phantom is displayed on a greyscale window of $[0.1, 0.35]$ cm$^{-1}$ to emphasize soft tissue contrast and artifacts in reconstructed images. Centre: image reconstructed using SART from water-corrected data. Right: Difference between image reconstructed using pSART (leftmost image in ) and image reconstructed using SART. Greyscale window is $[-0.04, 0.10]$ cm$^{-1}$.[]{data-label="F:xcat_phantom"}](fig8b.pdf "fig:"){width="0.25\linewidth"} ![XCAT anatomical phantom used for experiments with fan-beam data. Left: True phantom at reference energy of 50 keV. Phantom is displayed on a greyscale window of $[0.1, 0.35]$ cm$^{-1}$ to emphasize soft tissue contrast and artifacts in reconstructed images. Centre: image reconstructed using SART from water-corrected data. Right: Difference between image reconstructed using pSART (leftmost image in ) and image reconstructed using SART. Greyscale window is $[-0.04, 0.10]$ cm$^{-1}$.[]{data-label="F:xcat_phantom"}](fig8c.pdf "fig:"){width="0.25\linewidth"} ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- For the sparse-view experiments, we reconstructed the phantom from 900, 225, 180, 90, and 45 views taken over $360^\circ$. As before, the number of iterations for which pSART was run scaled with the number of views, from 9 iterations for the 900-view dataset up to 180 iterations for the 45-view data set. The number of subsets was chosen such that each subset consisted of 15 views in all cases. pSART-TV was run with $\gamma = 0.9995$ and $N=60$. ![Images of XCAT phantom reconstructed from sparse-view data using pSART and pSART-TV. First, third and fifth images are reconstructed with pSART, while second, fourth and sixth images are corresponding reconstructions with pSART-TV.[]{data-label="F:sparseview_images_XCAT"}](fig9_new.pdf){width="\linewidth"} ----------- -- -- $p = 900$ ----------- -- -- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![Plots of important quantities from sparse-view experiments with XCAT phantom. Left plot: Values of $\varepsilon$ obtained by pSART and pSART-TV. Middle plot: TV values for pSART reconstructions plotted against TV values for pSART-TV reconstructions. Black dashed line is the line $y=x$, blue dashed line represents true TV of phantom at 50 keV. Leftmost point corresponds to $p=900$. Right plot: number of iterations required by each algorithm for each simulation. Leftmost points correspond to $p=900$.[]{data-label="F:sparseview_plots_XCAT"}](fig10a.pdf "fig:"){height="0.2\linewidth"} ![Plots of important quantities from sparse-view experiments with XCAT phantom. Left plot: Values of $\varepsilon$ obtained by pSART and pSART-TV. Middle plot: TV values for pSART reconstructions plotted against TV values for pSART-TV reconstructions. Black dashed line is the line $y=x$, blue dashed line represents true TV of phantom at 50 keV. Leftmost point corresponds to $p=900$. Right plot: number of iterations required by each algorithm for each simulation. Leftmost points correspond to $p=900$.[]{data-label="F:sparseview_plots_XCAT"}](fig10b.pdf "fig:"){height="0.2\linewidth"} ![Plots of important quantities from sparse-view experiments with XCAT phantom. Left plot: Values of $\varepsilon$ obtained by pSART and pSART-TV. Middle plot: TV values for pSART reconstructions plotted against TV values for pSART-TV reconstructions. Black dashed line is the line $y=x$, blue dashed line represents true TV of phantom at 50 keV. Leftmost point corresponds to $p=900$. Right plot: number of iterations required by each algorithm for each simulation. Leftmost points correspond to $p=900$.[]{data-label="F:sparseview_plots_XCAT"}](fig10c.pdf "fig:"){height="0.2\linewidth"} --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Some representative images are shown in . With 900 views, the effect of pSART-TV is primarily to reduce some of the noise in the image. As the number of views is reduced, undersampling artifacts become more prevalent in the reconstructions generated using pSART. The superiorized algorithm is successful in reducing these artifacts, although they are not entirely removed. shows plots of the $\varepsilon$ values, TV values and number of iterations required for each of the reconstructions. The trends are consistent with the experiments conducted with the FORBILD phantom; pSART-TV found a solution that was as constraints-compatible as the image reconstructed by pSART, with significantly lower TV values. As before, obtaining this solution required running many more iterations of the algorithm, in general. For the limited-angle scenarios, we were not able to use pSART-ATV in this experiment, since the weights used in the computation of the ATV are calculated based on a parallel-beam geometry [@CJLW13]. Instead, we used pSART-TV; while [@CJLW13] indicates that an ATV penalty is more effective in removing limited angle artifacts, TV is also able to reduce these artifacts considerably. As in the previous experiments, we simulated two acquisitions over 165$^\circ$ and two over 150$^\circ$, with different starting points. Other parameters to the reconstruction algorithms were the same as in the sparse-view experiments. ![Images of XCAT phantom reconstructed from limited angle data using pSART and pSART-TV. First, third, fifth and seventh images are reconstructed with pSART; second, fourth, sixth and eighth images are corresponding images reconstructed with pSART-TV.[]{data-label="F:limitedangle_images_XCAT"}](fig11_new.pdf){width="\linewidth"} --------------------- --------------------- ---------------------- --------------------- Extent: 165$^\circ$ Extent: 165$^\circ$ Extent: 150$^\circ$ Extent: 150$^\circ$ Start: $97.5^\circ$ Start: $7.5^\circ$ Start: $105.0^\circ$ Start: $15.0^\circ$ --------------------- --------------------- ---------------------- --------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![Plots of important quantities from limited angle experiments with XCAT phantom. Plotted quantities are the same as for . Leftmost point corresponds to first and second columns of in all cases, moving across that figure as one follows the line.[]{data-label="F:limitedangle_plots_XCAT"}](fig12a.pdf "fig:"){height="0.2\linewidth"} ![Plots of important quantities from limited angle experiments with XCAT phantom. Plotted quantities are the same as for . Leftmost point corresponds to first and second columns of in all cases, moving across that figure as one follows the line.[]{data-label="F:limitedangle_plots_XCAT"}](fig12b.pdf "fig:"){height="0.2\linewidth"} ![Plots of important quantities from limited angle experiments with XCAT phantom. Plotted quantities are the same as for . Leftmost point corresponds to first and second columns of in all cases, moving across that figure as one follows the line.[]{data-label="F:limitedangle_plots_XCAT"}](fig12c.pdf "fig:"){height="0.2\linewidth"} ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Representative images and plots of important quantities are shown in Figures \[F:limitedangle\_images\_XCAT\] and \[F:limitedangle\_plots\_XCAT\], respectively. It is apparent from the reconstructed images that pSART-TV is effective in reducing the limited angle artifacts that are present in the corresponding images reconstructed by pSART. In the two experiments with 165$^\circ$ angular extents, the artifacts are largely eliminated aside from a few in the vicinity of the sternum and ribs. When the angular extent is further reduced to 150$^\circ$, the severity of the artifacts increases, and pSART-TV is less effective in removing them. As the plots in Figure \[F:limitedangle\_images\_XCAT\] indicate, however, the superiorized algorithm is again able to find a $\varepsilon$-compatible solution with substantially lower TV in every case. Conclusions {#S:conclusions} =========== In this paper we present a superiorized algorithm for reconstruction of CT images from sparse-view and limited-angle polyenergetic data. The primary contribution of the paper is the application of the superiorization methodology to an iterative algorithm for reconstruction of images from polyenergetic data, whereas previously it has been applied to algorithms which assume a monoenergetic model. The polyenergetic reconstruction algorithm is pSART [@LS14b], and the chosen objective functions for superiorization are total variation (TV) and anisotropic TV (ATV) [@CJLW13], which have been successfully applied in the past to reconstruction of sparse-view and limited-angle monoenergetic data. Our superiorized algorithm significantly reduces artifacts that arise from sparse-view and limited-angle data, as well as the beam hardening artifacts that would be present if a superiorized version of a monoenergetic algorithm were used. A sufficient condition to guarantee success of a superiorized algorithm is for the algorithm to be strongly perturbation resilient. We are not able to establish that pSART is a strongly perturbation resilient algorithm, as the conditions under which the algorithm converges are not well understood [@H15]. In our numerical experiments, however, the superiorized pSART algorithm is successful in every instance. It produces solutions that are as constraints-compatible as those produced by pSART, with TV or ATV values that are typically 30 – 60% lower. As the authors note in [@HGDC12], strong perturbation resilience is a sufficient condition for the success of superiorization, but not a necessary one. Thus, the inability to prove whether or not an algorithm is strongly perturbation resilient does not preclude the use of superiorization to obtain solutions that are improved with respect to the chosen objective. Whether the image is superior with respect to the chosen objective is, of course, a separate question from whether the image quality is improved; this depends on how effective the objective function is at penalizing undesirable image characteristics. In our experiments we find that TV and ATV are effective at eliminating artifacts caused by sparse-view and limited-angle data, up to a point. For the FORBILD phantom, we were able to obtain reconstructions that were largely free of artifacts from 165$^\circ$ of data in limited angle experiments, and roughly one quarter to one fifth the angular sampling rate that (conservatively) would be required to reconstruct the phantom at a resolution of $800 \times 800$ pixels, in our sparse-view experiments. Results of a second experiment using the XCAT anatomical phantom with a simulated fan-beam acquisition gave similar results, showing that the approach is not dependent on a parallel-beam geometry. While we were successful in superiorizing the pSART algorithm, the pSART-TV and pSART-ATV algorithms are not without drawbacks. A primary deficiency is the increased computation time required to find an $\varepsilon$-compatible solution. In addition to the extra cost per iteration to reduce the objective function value, it may take many more iterations than the original algorithm to find such a solution. This was particularly true in the limited-angle experiments, where it often required more than twice as many iterations to find an $\varepsilon$-compatible solution (with consistent data) and sometimes more than ten times as many iterations (with inconsistent data). This may be due to the choice of the parameters $\gamma$ and $N$ in these experiments, which we found needed to be larger than in the sparse-view experiments to effectively remove artifacts. Additionally, because the pSART algorithm is an algebraic technique, it does not model noise as accurately as a statistical technique. Superiorizing a statistical polyenergetic algorithm (e.g. [@DNDMS01]) and using a likelihood-based proximity function, as was done in [@GH14], could provide better results in the case of noisy data. References {#references .unnumbered} ========== [xx]{} Andersen, A. H. Kak, A. C. 1984 . Simultaneous algebraic reconstruction technique ([SART]{}): a superior implementation of the [ART]{} algorithm, [Ultrasonic Imaging]{} [ 6]{}(1): 81–94. Barber, R. F., Sidky, E. Y., Schmidt, T. G. Pan, X. 2016. An algorithm for constrained one-step inversion of spectral ct data, [Phys. Med. Biol.]{} [61]{}(10): 3784. Bian, J., Siewerdsen, J. H., Han, X., Sidky, E. Y., Prince, J. L., Pelizzari, C. A. Pan, X. 2010. Evaluation of sparse-view reconstruction from flat-panel-detector cone-beam [CT]{}, [Phys. Med. Biol.]{} [55]{}(22): 6575. Bian, J., Yang, K., Boone, J. M., Han, X., Sidky, E. Y. Pan, X. 2014. Investigation of iterative image reconstruction in low-dose breast [CT]{}, [Phys. Med. Biol.]{} [59]{}(11): 2659. Birnbaum, B. A., Hindman, N., Lee, J. Babb, J. S. 2007. Renal cyst pseudoenhancement: Influence of multidetector [CT]{} reconstruction algorithm and scanner type in phantom model, [Radiology]{} [244]{}(3): 767–775. Blumensath, T. Boardman, R. 2015. Non-convexly constrained image reconstruction from nonlinear tomographic [X-ray]{} measurements, [Philosophical Transactions of the Royal Society of London A: Mathematical, Physical and Engineering Sciences]{} [373]{}(2043): 20140393. Brooks, R. A. Di Chiro, G. 1976 . Beam hardening in [X-ray]{} reconstructive tomography, [Phys. Med. Biol.]{} [21]{}(3): 390–398. Cand[è]{}s, E. J., Romberg, J. Tao, T. 2006a. Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information, [IEEE Trans. Inf. Theory]{} [ 52]{}(2): 489–509. Cand[è]{}s, E., Romberg, J. K. Tao, T. 2006b. Stable signal recovery from incomplete and inaccurate measurements, [Communications on pure and applied mathematics]{} [59]{}(8): 1207–1223. Censor, Y., Davidi, R. Herman, G. T. 2010. Perturbation resilience and superiorization of iterative algorithms, [Inverse Problems]{} [26]{}: 1–12. Censor, Y. Elfving, T. 2002. Block-iterative algorithms with diagonally scaled oblique projections for the linear feasibility problem, [[SIAM]{} Journal on Matrix Analysis and Applications]{} [24]{}(1): 40–58. Chen, Z., Jin, X., Li, L. Wang, G. 2013. A limited-angle [CT]{} reconstruction method based on anisotropic [TV]{} minimization, [Phys. Med. Biol.]{} [58]{}(7): 2119. Choi, K., Wang, J., Zhu, L., Suh, T.-S., Boyd, S. Xing, L. 2010. Compressed sensing based cone-beam computed tomography reconstruction with a first-order method, [Med. Phys.]{} [37]{}(9): 5113–5125. Davidi, R., Censor, Y., Schulte, R. W., Geneser, S. Xing, L. 2015. Feasibility-seeking and superiorization algorithms applied to inverse treatment planning in radiation therapy, [Contemporary Mathematics]{} [ 636]{}: 83–92. Davidi, R., Herman, G. Censor, Y. 2009. Perturbation-resilient block-iterative projection methods with application to image reconstruction from projections, [International Transactions in Operational Research]{} [16]{}(4): 505–524. De Man, B., Nuyts, J., Dupont, P., Marchal, G. Suetens, P. 2001. An iterative maximum-likelihood polychromatic algorithm for [CT]{}, [IEEE Trans. Med. Imag.]{} [20]{}(10): 999–1008. Defrise, M., Vanhove, C. Liu, X. 2011. An algorithm for total variation regularization in high-dimensional linear problems, [Inverse Problems]{} [27]{}(6): 065002. Donoho, D. L. 2006. Compressed sensing, [IEEE Trans. Inf. Theory]{} [ 52]{}(4): 1289–1306. Elbakri, I. A. Fessler, J. A. 2002. Statistical image reconstruction for polyenergetic [X-ray]{} computed tomography, [IEEE Trans. Med. Imag.]{} [21]{}(2): 89–99. Eldar, Y. C. Kutyniok, G. 2012 . , Cambridge University Press. Elfving, T., Hansen, P. C. Nikazad, T. 2014. Semi-convergence properties of [Kaczmarz]{}â€™s method, [Inverse Problems]{} [30]{}(5): 055007. Fessler, J. Accessed 2016.. Michigan image reconstruction toolbox. http://web.eecs.umich.edu/$\sim$fessler/code/. Foucart, S. Rauhut, H. 2013. , Vol. 1, Springer. Frikel, J. Quinto, E. T. 2013. Characterization and reduction of artifacts in limited angle tomography, [Inverse Problems]{} [29]{}(12): 125007. Gardu[ñ]{}o, E., Herman, G. Davidi, R. 2011. Reconstruction from a few projections by â„“1-minimization of the haar transform, [Inverse problems]{} [27]{}(5): 055006. Gardu[ñ]{}o, E. Herman, G. T. 2014. Superiorization of the [ML-EM]{} algorithm, [IEEE Transactions on Nuclear Science]{} [61]{}(1): 162–172. Han, X., Bian, J., Ritman, E. L., Sidky, E. Y. Pan, X. 2012. Optimization-based reconstruction of sparse images from few-view projections, [Phys. Med. Biol.]{} [57]{}(16): 5245. Herman, G. T. 1979. Correction for beam hardening in computed tomography, [Phys. Med. Biol.]{} [24]{}(1): 81–106. Herman, G. T. Davidi, R. 2008. Image reconstruction from a small number of projections, [Inverse Problems]{} [24]{}(4): 045011. Herman, G. T., Garduño, E., Davidi, R. Censor, Y. 2012. Superiorization: An optimization heuristic for medical physics, [ Med. Phys.]{} [39]{}: 5532–5546. Herman, G. Trivedi, S. 1983. A comparative study of two post-reconstruction beam hardening correction methods, [IEEE Trans. Med. Imag.]{} [MI-2]{}(3): 128–135. Hsieh, J., Molthen, R., Dawson, C. Johnson, R. 2000. An iterative approach to the beam hardening correction in cone beam [CT]{}, [Med. Phys.]{} [27]{}(1): 23–29. Hudson, H. M. Larkin, R. S. 1994. Accelerated image reconstruction using ordered subsets of projection data, [IEEE Transactions on Medical Imaging]{} [13]{}(4): 601–609. Humphries, T. 2015. Technical note: Convergence analysis of a polyenergetic [SART]{} algorithm, [Medical Physics]{} [42]{}(7): 1407–1404. Humphries, T. Faridani, A. 2015 . Reconstruction of sparse-view polyenergetic [CT]{} data using total variation minimization, [2015 IEEE Nuclear Science Symposium Conference Record]{}. J[ø]{}rgensen, J. S. Sidky, E. 2015. How little data is enough? [Phase]{}-diagram analysis of sparsity-regularized [X-ray]{} computed tomography, [Philosophical Transactions of the Royal Society of London A: Mathematical, Physical and Engineering Sciences]{} [373]{}(2043): 20140387. J[ø]{}rgensen, J. S., Sidky, E. Y. Pan, X. 2013. Quantifying admissible undersampling for sparsity-exploiting iterative image reconstruction in [X-ray]{} [CT]{}, [IEEE Trans. Med. Imag.]{} [32]{}(2): 460–473. Joseph, P. M. Spital, R. D. 1978. A method for correcting bone induced artifacts in computed tomography, [Journal of Computer Assisted Tomography]{} [ 2]{}(1): 100–108. Joseph, P. Ruth, C. 1997. A method for simultaneous correction of spectrum hardening in [CT]{} images containing both bone and iodine, [Med. Phys.]{} [24]{}(10): 1629–1634 Kitagawa, K., George, R. T., Arbab-Zadeh, A., Lima, J. A. Lardo, A. 2010. Characterization and correction of beam-hardening artifacts during dynamic volume [CT]{} assessment of myocardial perfusion, [Radiology]{} [ 256]{}(1): 111–118. Kyriakou, Y., Meyer, E., Prell, D. Kachelrie[ß]{}, M. 2010. Empirical beam hardening correction ([EBHC]{}) for [CT]{}, [Med. Phys.]{} [37]{}(10): 5179–5187. Lin, Y. Samei, E. 2014. An efficient polyenergetic [SART]{} ([pSART]{}) reconstruction algorithm for quantitative myocardial [CT]{} perfusion, [Med. Phys.]{} [ 41]{}(2): 021911–1 – 021911–14. McCollough, C. H., Chen, G. H., Kalender, W., Leng, S., Samei, E., Taguchi, K., Wang, G., Yu, L. Pettigrew, R. I. 2012. Achieving routine submillisievert [CT]{} scanning: report from the summit on management of radiation dose in [CT]{}, [Radiology]{} [ 264]{}(2): 567–580. Natterer, F. 1986. , Springer. Needell, D. Ward, R. 2013. Stable image reconstruction using total variation minimization, [ SIAM Journal on Imaging Sciences]{} [6]{}(2): 1035–1058. Penfold, S., Schulte, R. W., Censor, Y. Rosenfeld, A. B. 2010. Total variation superiorization schemes in proton computed tomography image reconstruction, [Medical physics]{} [37]{}(11): 5887–5895. Quinto, E. T. 1993. Singularities of the [X-ray]{} transform and limited data tomography in [$R^2$]{} and [$R^3$]{}, [[SIAM]{} Journal on Mathematical Analysis]{} [ 24]{}(5): 1215–1225. Ritschl, L., Bergner, F., Fleischmann, C. Kachelrie[ß]{}, M. 2011. Improved total variation-based [CT]{} image reconstruction applied to clinical data, [Phys. Med. Biol.]{} [56]{}(6): 1545. Rodr[í]{}guez-Granillo, G. A., Rosales, M. A., Degrossi, E. Rodriguez, A. E. 2010. Signal density of left ventricular myocardial segments and impact of beam hardening artifact: implications for myocardial perfusion assessment by multidetector [CT]{} coronary angiography, [Int. J. Cardiovasc. Imag.]{} [ 26]{}(3): 345–354. Segars, W., Mahesh, M., Beck, T. J., Frey, E. C. Tsui, B. 2008. Realistic [CT]{} simulation using the [4D XCAT]{} phantom, [Medical physics]{} [35]{}(8): 3800–3808. Segars, W., Sturgeon, G., Mendonca, S., Grimes, J. Tsui, B. 2010. phantom for multimodality imaging research, [Med. Phys.]{} [37]{}(9): 4902–4915. Sidky, E. Y., J[ø]{}rgensen, J. H. Pan, X. 2012. Convex optimization problem prototyping for image reconstruction in computed tomography with the [Chambolle–Pock]{} algorithm, [Physics in Medicine and Biology]{} [57]{}(10): 3065. Sidky, E. Y. Pan, X. 2008. Image reconstruction in circular cone-beam computed tomography by constrained, total-variation minimization, [Phys. Med. Biol.]{} [ 53]{}(17): 4777. Tang, J., Nett, B. E. Chen, G.-H. 2009. Performance comparison between total variation ([TV]{})-based compressed sensing and statistical iterative reconstruction algorithms, [ Phys. Med. Biol.]{} [54]{}(19): 5781. Van Gompel, G., Van Slambrouck, K., Defrise, M., Batenburg, K., de Mey, J., Sijbers, J. Nuyts, J. 2011. Iterative correction of beam hardening artifacts in [CT]{}, [Med. Phys.]{} [38]{}(7): S36–S49. Yu, Z., Noo, F., Dennerlein, F., Wunderlich, A., Lauritsch, G. Hornegger, J. 2012. Simulation tools for two-dimensional experiments in [X-ray]{} computed tomography using the [FORBILD]{} head phantom, [Phys. Med. Biol.]{} [ 57]{}(13): N237–N252. [^1]: To distinguish between the two, we will use radians to refer to the parameter $\theta$ in , and degrees when referring to views of the object, which are acquired orthogonally to $\theta$.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The use of hyperasymptotics and the Weniger transformation has been proposed, in a joint fashion, for decoding the divergent asymptotic series generated by the steepest descent on a wide class of saddle-point integrals [evaluated across Stokes sets]{} \[R. Borghi, Phys. Rev. E [78,]{} 026703 (2008)\]. In the present sequel, the full development of the H-WT up to the second order in H is derived. Numerical experiments, carried out on several classes of saddle-point integrals, including the swallowtail diffraction catastrophe, show the effectiveness of the 2nd-level H-WT, in particular when the integrals are evaluated beyond the asymptotic realm.' author: - Riccardo Borghi title: 'Joint use of the Weniger transformation and hyperasymptotics for accurate asymptotic evaluations of a class of saddle-point integrals. II. Higher-order transformations' --- Introduction {#intro} ============ The present paper is a sequel of a previous work[@borghiPRE-08] concerning the evaluation of saddle-point integrals of the form $$\mathcal{I}(k)=\displaystyle\int_{\mathcal{C}}\, g(s)\,\exp[-k\,f(s)]\,\mathrm{d}s, \label{sd.1}$$ where $\mathcal{C}$ is a suitable integration path in the complex $s$-plane, $g(s)$ and $f(s)$ are functions which, for simplicity, will be assumed to be nonsingular, and $k$ will be intended as a “large” (in modulus) complex parameter. As is well known, the numerical evaluation of integrals of the kind in Eq. (\[sd.1\]) is customarily required for solving several classes of physical problems, occurring in optics, quantum mechanics, statistical physics, fluid mechanics, and so on. In optics, the evaluation of several diffraction integrals is customarily carried out asymptotically by identifying the parameter $k$ as the wavenumber of the radiation. In quantum mechanics, the same role is played by the inverse of the Planck’s constant, while in fluid mechanics by the Reynold’s number[@berryLectures-89]. In the stationary phase treatment of diffraction integrals the values of the associated complex wavefield are asymptotically evaluated by taking the contributions coming from the stationary points of $f(s)$, each of them associated to a “ray" in the corresponding geometrical picture. Of particular importance is the birth and the death, as the spatial parameters ruling the “phase" function $\mathrm{i} f(s)$ vary, of “evanescent" rays across sets of codimension 1, named “Stokes sets" [@wrightJPA-80; @berryNL-90]. The $\delta$-, or Weniger, transformation[@wenigerCPR-89; @wenigerJMP-04; @calicetiArXiv07] (WT for short), is particularly efficient for resumming the factorial divergent asymptotic series well away from Stokes sets, as well as sets where two or more saddles are symmetrically placed in the complex singulant space[@endnote31]. [ Unfortunately, as with other resummation techniques[@wenigerCPR-89; @brezinski], the WT fails to perform across Stokes sets. The reason of such a failure stems to the extreme “specialization” of the transformation itself, which requires, for a successful resummation, an alternating sign pattern of the sequence of the single terms of the series[@jentschuraCPC-99]. Several methods have been conceived for resumming nonalternating, slowly convergent or divergent, sequences [@jentschuraPRD-00; @calicetiArXiv07], some of them being based on the serial combination of various resummation techniques[@jentschuraCPC-99; @aksenovCPC-03]. For the class of saddle-point integrals in Eq. (\[sd.1\]), the marriage between hyperasymptotics[@berryPRSA-90; @berryPRSA-91] (H for short) and the WT [@borghiPRE-08], generating the so-called H-WT (which stands for hyperasymptotic-Weniger transformation), allows the WT to successfully operate also across Stokes sets. ]{} Basically, the H-WT consists in the sequential application, to the integral in Eq. (\[sd.1\]), of a classical hyperasymptotic treatment, as described in Ref. [@berryPRSA-91], followed by the action of the WT on all asymptotic divergent series generated by H. In particular, the results obtained have shown how the 1st-order H-WT, for which only the first-stage of H is anticipated to the WT, is able to provide relative errors several orders of magnitude smaller than those achievable via the use of full hyperasymptotic treatments and with considerably lighter computational complexity and effort. A key aspect is that, differently from H, the first truncation operated on the starting asymptotic series has not to be an optimal, in the sense of superasymptotics (i.e., at the least term) one, but rather the corresponding truncation order, say $N$, must be used as a free parameter for the subsequent application of the WT. A question which was not addressed in Ref. [@borghiPRE-08], but mentioned only in the last sentence, is whether WT and H can be combined to higher orders in H, and if so, how the accuracy improves with order. The present paper is aimed at giving a first answer to such a question. We shall limit our analysis only to the second stage of H. Further increasing of the H-WT order would be achievable along the same guidelines outlined here. On the other hand, it should be noted how, on increasing the order of H, the number of asymptotic series associated to the corresponding remainder that have to be resummed exponentially grows for topologies involving more than two saddles and, at the same time, the number of free parameters (i.e., the truncation orders at each H step) linearly increases. Accordingly, from a mere computational viewpoint, it is mandatory to find a compromise between the H-WT order and the computational effort. Some limits of the 1st-order H-WT have already been emphasized in Ref. [@borghiPRE-08], where asymptotic evaluations of saddle-point integrals for “small" values of the asymptotic parameter were considered. In such cases, in fact, to increase the parameter $N$ of the 1st-order H-WT does not necessarily lead to an improvement of the reached accuracy, but often results in the opposite, i.e., a worsening of it. It is just the above scenario that we are interested in when the 2nd-level H-WT will be developed. Numerical experiments will be carried out on the class of saddle-point integrals already considered in the numerical sections of the first paper [@borghiPRE-08]. Moreover, asymptotic evaluations of the so-called swallowtail diffraction catastrophe [@berryPIO-80] will be proposed as a new numerical experiment. The swallowtail function is defined via Eq. (\[sd.1\]) with $f(s)$ being a 5th-order polynomial with respect to $s$, thus involving a four saddle network. We will present a study of the accuracy achievable via H-WT asymptotic evaluations of the swallowtail diffraction catastrophe for points placed at the Stokes set, following the prescriptions by Berry and Howls[@berryNL-90]. In doing so, we will find that the corresponding asymptotic expanding coefficients can be expressed in closed-form terms. [For any practical implementation of the H-WT, a key role is played by the numerical evaluation of the corresponding hyperterminants[@berryPRSA-90; @berryPRSA-91] which are defined through suitable multiple integrals. For the lowest-order hyperterminant the exact analytical expression is available from literature[@berryPRSA-90], but unfortunately this is not true for higher-order hyperterminants, included those involved in the 2nd-level H-WT. In the present paper we solve the problem of the 2nd-level hyperterminant exact evaluation for a particular, but very important, choice of the hyperterminant parameters, which often occurs in the implementation of H for evaluating a wide class of saddle-point integrals. Up to our knowledge, this is a new result which also provides an interesting connection of such hypeterminants to the Meijer-G functions [@GR]. Moreover, although the closed-form evaluation of 2nd-level hypeterminants for arbitrary choices of their parameters seems to remain an open problem, in the present paper we find a semi-analytical representation which turns out to be suitable for numerical calculations via standard integration packages. ]{} Similarly as done in Ref. [@borghiPRE-08], one of our aims is to keep the paper reasonably self-consistent. Accordingly, in the next section a brief review of H, up to the 2nd-level, is given. As far as the WT is concerned, we believe that what is contained in Ref. [@borghiPRE-08], together with the extensive bibliography, should be enough also for a nonexpert reader. For this reason, we do not repeat it in the present paper. Resuming Hyperasymptotics {#hyperAsymptotics} ========================= Preliminaries and notations {#notations} --------------------------- For simplicity, we shall refer to the asymptotic evaluation of saddle-point integrals of the type in Eq. (\[sd.1\]) where the set of saddle points of $f(s)$ will be denoted $\mathcal{S}$ and the integration path $\mathcal{C}$ will be thought of as the union of a finite number of steepest descent arcs each of them, say $\mathcal{C}_n$, passing through the contributive saddle point $s_n$, which will be supposed to be a simple one. Accordingly, the quantity $\mathcal{I}(k)$ can generally be written as $$\mathcal{I}(k)=\displaystyle\int_{\mathcal{C}}\,g(s)\,\exp[-k\,f(s)]\,\mathrm{d}s= \displaystyle\sum_{n \in \mathcal{S}'}\,\mathcal{I}^{(n)}(k), \label{sdReview.5}$$ where $\mathcal{S}'$ denotes the subset of $\mathcal{S}$ containing all the contributive saddles, and $$\mathcal{I}^{(n)}(k)= \displaystyle\int_{\mathcal{C}_n}\, g(s)\,\exp[-k f(s)]\,\mathrm{d}s. \label{sdReview.5.1}$$ The last integral can be written as[@berryPRSA-91] $$\mathcal{I}_n(k)=k^{-1/2}\,\exp(-kf_n)\,T^{(n)}(k), \label{sdReview.3}$$ where $f_n=f(s_n)$, and where $T^{(n)}(k)$ can formally be written through the following asymptotic series expansion: $$T^{(n)}(k)=\displaystyle\sum_{r=0}^{\infty}\,k^{-r}\,T^{(n)}_r, \label{sdReview.4}$$ the expanding coefficients $T^{(n)}_r$ being expressed via the integral representation [@berryPRSA-91] $$T^{(n)}_r= \displaystyle\frac{(r-1/2)!}{2\pi\mathrm{i}}\, \oint_{n}\,\displaystyle\frac{g(s)}{[f(s)-f_n]^{r+1/2}}\,\mathrm{d}s, \label{sdReview.5.0}$$ where the subscript $n$ denotes a small positive loop around the saddle $s_n$. Development of H up to the second stage {#H} --------------------------------------- H starts by writing Eq. (\[sdReview.4\]) in the form $$T^{(n)}(k)=\displaystyle\sum_{r=0}^{N-1}\,k^{-r}\,T^{(n)}_r+ R^{(n)}(k,N), \label{sdReview.5bis}$$ where $N$ represents a positive integer and $R^{(n)}(k,N)=\sum_{r=N}^\infty\,k^{-r}\,T^{(n)}_r$, denotes the corresponding remainder which, due to the diverging character of the asymptotic series, turns out to be a diverging quantity too. H is based on a couple of fundamental results, found via a nontrivial analysis in Ref. [@berryPRSA-91]. The first is that the value of the expanding coefficients $T^{(n)}_r$ at the saddle $s_n$ is strictly related to the values of the expanding coefficients $T^{(m)}_r$ at all those saddles, say $\{s_m\}$, which are adjacent to $s_n$, via the following formal [resurgence]{} relation[@berryPRSA-91]: $$T^{(n)}_r=\displaystyle\frac 1{2\pi\mathrm{i}}\, \displaystyle\sum_{m\in \mathcal{A}_n}\,(-1)^{\gamma_{nm}}\, \displaystyle\sum_{l=0}^\infty\, \displaystyle\frac{(r-l-1)!}{F^{r-l}_{nm}}\,T^{(m)}_l, \label{resurgence.1}$$ where $\mathcal{A}_n$ denotes the set containing the indexes pertinent to all saddles adjacent to $s_n$, the quantities $F_{nm}$, called singulants, are defined by $$F_{nm}=f_m-f_n, \label{singulant}$$ and the binary quantities $\gamma_{nm} \in \{0,1\}$ are obtained through a topological rule[@berryPRSA-91]. The other fundamental tool of H is the following integral representation of the remainder $R^{(n)}(k,N)$[@berryPRSA-91]: $$\begin{array}{l} R^{(n)}(k,N)=\displaystyle\frac 1{2\pi\mathrm{i}}\, \displaystyle\sum_{m \in \mathcal{A}_n}\,\displaystyle\frac{(-1)^{\gamma_{nm}}}{(kF_{nm})^N}\\ \\ \times \displaystyle\int_0^\infty\, \mathrm{d}v\,\displaystyle\frac{v^{N-1}\,\exp(-v)}{1-\displaystyle\frac v{kF_{nm}}}\, T^{(m)}\left(\displaystyle\frac v{F_{nm}}\right). \end{array} \label{remainder.2}$$ Equations (\[sdReview.4\])-(\[remainder.2\]) allow hyperasymptotic expansions for the saddle integral in Eq. (\[sdReview.5.1\]) to be built up in principle to any order[@berryPRSA-91]. For instance, the direct substitution of Eq. (\[sdReview.4\]) into Eq. (\[remainder.2\]) leads to the 1st-stage hyperasymptotic expansion [@borghiPRE-08], $$\begin{array}{lcl} T^{(n)}(k) = \displaystyle\sum_{r=0}^{N-1}\,k^{-r}\,T^{(n)}_r\\ \\ + \displaystyle\frac {(-1)^N}{2\pi\mathrm{i}}\, \displaystyle\sum_{m \in \mathcal{A}_n}\,(-1)^{\gamma_{nm}}\\ \\ \,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\times \displaystyle\sum_{r=0}^{\infty}\, (-1)^r\,k^{-r}\,T^{(m)}_r\,K^{(1)}_{N-r}(-kF_{nm}), \end{array} \label{Istage.complete}$$ where the function $K^{(1)}_{n}(\beta)$, called hyperterminant of order 1[@berryPRSA-90; @berryPRSA-91], is defined through the integral $$\begin{array}{lcl} K^{(1)}_{n}(\beta)= \displaystyle\frac 1{\beta^n}\, \displaystyle\int_0^\infty\, \mathrm{d}v\,\displaystyle\frac{v^{n-1}\,\exp(-v)}{1+\displaystyle\frac v\beta}, \end{array} \label{remainder.3.1}$$ where, in order for it to converge, $n>0$. Moreover, it can be shown that[@borghiPRE-08] $$K^{(1)}_{n}(\beta)= \exp(\beta)\,\displaystyle\frac{E_n(\beta)}{\beta^{n-1}}\,(n-1)!+(-1)^{n-1}\,\mathrm{i}\pi \epsilon\,\exp(\beta), \label{remainder.3.1.1}$$ where $E_n(\cdot)$ denotes the exponential integral function [@GR], while $\epsilon$ equals 1 if $\beta<0$ and zero otherwise. The presence of the term containing $\epsilon$ has to be ascribed to the evaluation of the integral in Eq. (\[remainder.3.1\]), when $\beta<0$, in the Cauchy principal-value sense. Equation (\[Istage.complete\]) represents the first hyperasymptotic stage, at which the divergence of the asymptotic series in Eq. (\[sdReview.4\]) is led back to the presence of adjacent saddles [@berryPRSA-91]. Furthermore, the asymptotic series in Eq. (\[Istage.complete\]) are only formal, since for $r>N$ the terminant $K^{(1)}_{N-r}$ diverges. In Ref. [@borghiPRE-08] Eq. (\[Istage.complete\]) was taken as the starting point for introducing the H-WT. In particular, instead of using the WT directly on the single terms of the series in Eq. (\[sdReview.4\]), it is employed for resumming the asymptotic series associated to all saddles $s_m$, with $m\in\mathcal{A}_n$, which appear in Eq. (\[Istage.complete\]). Of course, due to the fact that $r \le N$ in Eq. (\[Istage.complete\]), it is mandatory that $N$ be left as a free parameter, in order for the WT to be able in decoding the above asymptotic series. The 2nd-level H can be derived by truncating each of the asymptotic series in Eq. (\[Istage.complete\]) at an order, say $M$, and by generating, for each adjacent saddle $s_m$, with $m\in \mathcal{A}_n$, a list of asymptotic series associated to all saddles, say $s_h$, such that $h \in \mathcal{A}_m$. In Appendix \[IIstage\], only for the reader convenience, the derivation of the 2nd-level hyperasymptotic expansion of the integral in Eq. (\[sd.1\]) is briefly recalled according to the formalism of Ref. [@berryPRSA-91]. In particular, it is found that $$\begin{array}{lcl} T^{(n)}(k)=\displaystyle\sum_{r=0}^{N-1}\,k^{-r}\,T^{(n)}_r\\ \\+ \displaystyle\frac {(-1)^N}{2\pi\mathrm{i}}\, \displaystyle\sum_{m \in \mathcal{A}_n}\,(-1)^{\gamma_{nm}}\\ \\ \,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\times \displaystyle\sum_{r=0}^{M-1}\, (-1)^r\,k^{-r}\,T^{(m)}_r\,K^{(1)}_{N-r}(-kF_{nm})\\ \\ + \displaystyle\frac{(-1)^{N+M}}{(2\pi\mathrm{i})^2}\, \displaystyle\sum_{m \in \mathcal{A}_n} \displaystyle\sum_{h \in \mathcal{A}_m}\,(-1)^{\gamma_{nm}+\gamma_{mh}}\,\\ \\ \,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\times \displaystyle\sum_{r=0}^\infty\,k^{-r}T^{(h)}_r\, K^{(2)}_{M-r,N-M}\left(-kF_{nm};-\displaystyle\frac{F_{mh}}{F_{nm}}\right), \end{array} \label{IIstage.complete}$$ where $K^{(2)}_{n,m}(\beta;\gamma)$, the hyperterminant of order 2, is now defined through the double integral $$\begin{array}{lcl} K^{(2)}_{n,m}(\beta;\gamma) = \displaystyle\frac 1{\beta^{n+m}}\,\\ \\ \times \displaystyle\int_0^\infty\, \displaystyle\int_0^\infty\, \mathrm{d}u\,\mathrm{d}v\, \exp(-u-\gamma\,v)\, \displaystyle\frac{u^{m-1} v^{n-1}}{\left(1+\displaystyle\frac u\beta\right)\,\left(1+\displaystyle\frac v u \right)}. \end{array} \label{terminant2}$$ Similarly as we done in Ref. [@borghiPRE-08], Eq. (\[IIstage.complete\]) can be used to give estimates of $T^{(n)}(k)$, as functions of the two (free) parameters $N$ and $M$, by resumming, via the WT, all asymptotic series generated at the second stage of H which are inside the double sum with respect to $h$ and $m$. On the evaluation of $K^{(2)}_{n,m}(\beta;\gamma)$ {#K2} ================================================== The numerical evaluation of the hyperterminants represents a step of fundamental importance for any practical implementation of the H-WT algorithm. Unfortunately, differently from the lowest-order H-WT, for which the corresponding hyperterminants are achievable via the closed-form expression in Eq. (\[remainder.3.1.1\]), there are no analytical expressions available for higher-order hyperterminants. In a series of important papers, Olde Daalhuis [@daalhiusJCAM-96; @daalhiusJCAM-98] addressed the general problem of the hyperterminants evaluation, up to arbitrary precisions, through the use of convergent series representations based on hypergeometric functions. However, for the particular case of the 2nd-level hyperterminant, it seems that some new, at least up to our knowledge, results could be established. From Eq. (\[terminant2\]) where, in order for it to converge, $n>0$ and $m>0$, the hyperterminant can be written as $$\begin{array}{lcl} K^{(2)}_{n,m}(\beta;\gamma) = \displaystyle\frac 1{\beta^{n+m}}\,\\ &&\\ \times \displaystyle\int_0^\infty\, \displaystyle\int_0^\infty\, \mathrm{d}u\,\mathrm{d}v\, \exp(-u-\gamma\,v)\, \displaystyle\frac{u^{m} v^{n-1}}{\left(1+\displaystyle\frac u\beta\right)\,(u+v)}, \end{array} \label{K2.0.1}$$ and, by formally expanding the factor $1/(1+u/\beta)$ as a geometric series, after some algebra takes the form $$\begin{array}{lcl} K^{(2)}_{n,m}(\beta;\gamma) = \displaystyle\frac {(-1)^{m}}{\beta^{n}} \displaystyle\sum_{k=m}^\infty\, \left(-\displaystyle\frac 1\beta\right)^k\,\\ \\ \times \displaystyle\int_0^\infty\, \displaystyle\int_0^\infty\, \mathrm{d}u\,\mathrm{d}v\, \exp(-u-\gamma\,v)\, \displaystyle\frac{u^{k} v^{n-1}}{u+v}=\\ \\ =(-1)^m\, \displaystyle\frac{(n-1)!}{(\beta\gamma)^{n}}\\ \\ \times \displaystyle\sum_{k=m}^\infty\, \left(-\displaystyle\frac 1\beta\right)^k\, \displaystyle\frac{k!}{k+n}\, F\left(n,1;k+n+1;1-\displaystyle\frac{1}{\gamma}\right), \end{array} \label{K2.0.2}$$ where $F(\cdot,\cdot;\cdot;\cdot)$ denotes the hypergeometric function[@GR]. Although the series in Eq. (\[K2.0.2\]) is divergent, it can be decoded via Borel summation, i.e., by replacing the term $k!$ by its integral representation, i.e., $$k!=\displaystyle\int_0^\infty\,\mathrm{d}t\, \exp(-t)\,t^k, \label{K2.0.3}$$ which, once substituted into Eq. (\[K2.0.2\]), leads to $$\begin{array}{lcl} K^{(2)}_{n,m}(\beta;\gamma) = (-1)^m\, \displaystyle\frac{(n-1)!}{(\beta\gamma)^{n}}\, \displaystyle\int_{0}^{\infty}\, \mathrm{d}t\, \exp(-t)\\ \\ \times \displaystyle\sum_{k=m}^{\infty}\, \left(-\displaystyle\frac t\beta\right)^k\, \displaystyle\frac{1}{k+n}\, F\left(n,1;k+n+1;1-\displaystyle\frac{1}{\gamma}\right). \end{array} \label{K2.0.3.1}$$ Although, as we shall in a moment, it is possible to express the series inside the last equation through a closed form, it is better to carry out the evaluations for the case $\gamma=1$ and $\gamma \ne 1$ separately. On letting into Eq. (\[K2.0.3.1\]) $\gamma=1$ we have $$\begin{array}{l} K^{(2)}_{n,m}(\beta;1) = (-1)^m\,\displaystyle\frac{(n-1)!}{\beta^{n}}\\ \\ \times \displaystyle\int_{0}^{\infty}\, \mathrm{d}t\, \exp(-t)\, \displaystyle\sum_{k=m}^{\infty}\, \left(-\displaystyle\frac t\beta\right)^k\, \displaystyle\frac{1}{k+n}=\\ \\ = \displaystyle\frac{(n-1)!}{(m+n)\beta^{n-1}}\\ \\ \times \displaystyle\int_0^\infty\, \mathrm{d}t\, \exp(-t)\, \left(\displaystyle\frac t\beta \right)^m\, F\left(m+n,1;m+n+1;-\displaystyle\frac t\beta\right). \end{array} \label{K2.0.3.2}$$ The integral in Eq. (\[K2.0.3.2\]) can be evaluated by using the representation of the hypergeometric function given by formula 9.34.7 of Ref. [@GR]. In particular, it turns out that $$\begin{array}{l} K^{(2)}_{n,m}(\beta;1)= \displaystyle\frac{(n-1)!}{\beta^{n-1}}\,\\ \\ \times \displaystyle\int_0^\infty\, \mathrm{d}t\, \exp(- t)\, \left(\displaystyle\frac t\beta \right)^{m+1}\, G^{12}_{22} \left(\displaystyle\frac t\beta\left\|\begin{array}{l} -n-m,\,-1\\-1,\,-n-m-1\end{array}\right.\right), \end{array} \label{hyper2.10.3}$$ where $G^{mn}_{pq}(\cdot)$ denotes the Meijer function[@GR]. Finally, by using formulas 9.31.5, 7.813.1, and 9.31.2 of [@GR], after some algebra it is found that $$\begin{array}{lcl} K^{(2)}_{n,m}(\beta;1)& = &(n-1)!\, G^{31}_{23}\left(\beta\left\|\begin{array}{c} 1-n-m,1\\1-n,1-n-m,0\end{array}\right.\right). \end{array} \label{terminant2.1}$$ Equation (\[terminant2.1\]) represents one of the main results of the present paper. As we shall see in the numerical section, in applying the 2nd-level H-WT the evaluation of the hyperterminants $K^{(2)}_{n,m}(\beta;\gamma)$ is often required for $\gamma=1$. This happens whenever the contributive saddle $s_{n}$ turns out to be adjacent to itself after two hyperasymptotic stages, i.e., when $h=n$ into Eq. (\[IIstage.complete\])[@berryPRSA-91]. For $\gamma \ne 1$, the series inside the integral in Eq. (\[K2.0.2\]) can still be expressed through a closed form, although the subsequent integral unfortunately not. However, a semi-analytical expression, which turns out to be suitable for being evaluated via standard numerical integration packages can be derived. In doing this, Eq. (\[K2.0.2\]) is first rewritten as $$\begin{array}{lcl} K^{(2)}_{n,m}(\beta;\gamma) = (-1)^{m}\,\displaystyle\frac{(n-1)!}{(\beta\gamma)^{n}}\\ \\ \times \left[ \mathcal{S}- \displaystyle\sum_{k=0}^{m-1}\, \left(-\displaystyle\frac 1\beta\right)^k\, \displaystyle\frac{k!}{k+n}\, F\left(n,1;k+n+1;\displaystyle\frac{\gamma-1}{\gamma}\right) \right], \end{array} \label{gn1.1}$$ where $$\begin{array}{lcl} \mathcal{S}= \displaystyle\sum_{k=0}^\infty\, \left(-\displaystyle\frac 1\beta\right)^k\, \displaystyle\frac{k!}{k+n}\, F\left(n,1;k+n+1;\displaystyle\frac{\gamma-1}{\gamma}\right), \end{array} \label{gn1.2}$$ so that the task is to evaluate the series in Eq. (\[gn1.2\]) for $\gamma \ne 1$. On substituting from Eq. (\[K2.0.3\]) into Eq. (\[gn1.2\]) we have $$\begin{array}{lcl} \mathcal{S}= \displaystyle\int_{0}^{\infty}\,\mathrm{d}t\, \exp(-t)\,\\ \\ \times \displaystyle\sum_{k=0}^\infty\, \left(-\displaystyle\frac t\beta\right)^k\, \displaystyle\frac{F\left(n,1;k+n+1;1-\displaystyle\frac{1}{\gamma}\right)}{k+n}=\\ \\ =\gamma\,\displaystyle\int_{0}^{\infty}\,\mathrm{d}t\, \exp(-t)\,\\ \\ \times \displaystyle\sum_{k=0}^\infty\, \left(-\displaystyle\frac t\beta\right)^k\, \displaystyle\frac{F(1+k,1;n+1+k;1-\gamma)}{k+n}, \end{array} \label{gn1.3}$$ where use has been made of the relation\[see Ref. [@PrudnikovIII] p. 347\] $$\begin{array}{lcl} F\left(n,1;k+n+1;1-\displaystyle\frac{1}{\gamma}\right) =\gamma\,F(1+k,1;n+1+k;1-\gamma). \end{array} \label{gn1.4}$$ Finally, on writing Eq. (\[gn1.3\]) as $$\begin{array}{lcl} \mathcal{S}=\displaystyle\frac\gamma n\, \displaystyle\int_{0}^{\infty}\,\mathrm{d}t\,\exp(-t)\,\\ \\ \times \displaystyle\sum_{k=0}^\infty\, \displaystyle\frac{(1)_{k}}{k!}\, \displaystyle\frac{(n)_{k}}{(n+1)_{k}}\, \left(-\displaystyle\frac t\beta\right)^k\, F(1+k,1;n+1+k;1-\gamma), \end{array} \label{gn1.5}$$ where $(\cdot)_{k}$ denotes the Pochhammer symbol, formula 6.7.1.8 of Ref. [@PrudnikovIII] gives at once $$\begin{array}{lcl} \mathcal{S}=\displaystyle\frac\gamma n\, \displaystyle\int_{0}^{\infty}\,\mathrm{d}t\, \displaystyle\frac{\exp(-t)}{1+\displaystyle\frac t\beta}\, F\left(1,1;n+1;1-\displaystyle\frac\gamma{1+\displaystyle\frac t\beta}\right). \end{array} \label{gn1.6}$$ Notice that, when $\mathrm{Re}[\beta]>0$, the function $S$ can also be evaluated through the alternative form $$\begin{array}{lcl} \mathcal{S}=\displaystyle\frac{\beta\gamma}n\, \displaystyle\int_{0}^{1}\,\mathrm{d}p\, \displaystyle\frac {\exp\left[-\beta\left(\displaystyle\frac 1p-1\right)\right]}{p}\, F(1,1;n+1;1-\gamma p). \end{array} \label{gn1.7}$$ Although it seems that the above expressions cannot be further simplified, the numerical evaluation of the function $S$ can be done with high accuracies by using standard integration packages. Finally, it should be stressed that, for $\beta<0$, the evaluation of the double integral defining $K^{(2)}_{n,m}(\beta;\gamma)$ has to be done, with respect to the $v$-variable, in the Cauchy principal value sense, in order to overcome the singularity placed at $v=-\beta$. This, in turn, implies that an extra term must be added to the result. In Appendix \[appC\] such term is analytically evaluated starting from the definition in Eq. (\[terminant2\]), and turns out to be $$\begin{array}{l} \mathrm{i}\pi\,(-1)^{n+m-1}(n-1)!\displaystyle\frac{\exp[\beta(1-\gamma)]} {(-\beta)^{n-1}} \,E_{n}(-\beta\gamma). \end{array} \label{extraTerm}$$\ All subsequent numerical experiments will be done within the Mathematica language. Numerical experiments {#numericalExperiments} ===================== Evaluation of the Airy function across the Stokes line {#airy} ------------------------------------------------------ [ Consider the evaluation of the Airy function, defined as $$\mathrm{Ai}(x)=\displaystyle\frac1{2\pi}\, \displaystyle\int_{\mathcal{C}}\, \exp\left[\mathrm{i}\left(\displaystyle\frac{s^3}3+xs\right)\right]\, \mathrm{d}s, \label{airy.1}$$ which is of the form given in Eq. (\[sd.1\]) with $g(s)=1/(2\pi)$, $f(s)=-\mathrm{i} (s^3/3+xs)$, and $k=1$. The detailed analysis of the saddle topology, as well as the expanding coefficients $T^{(n)}_r$ has been summarized in Ref. [@borghiPRE-08], so that here it will not be given again. We only recall that the two saddles are $s_1= (-x)^{1/2}$ and $s_2=-s_1$, and that $\mathcal{A}_{{1\atop 2}}=\{{2\atop 1}\}$, $\gamma_{12}=0$, $\gamma_{21}=1$. We focus our attention on the evaluation of the Airy function across the Stokes line[@boyd-99], i.e., for $\arg\{x\}=2\pi/3$, in order to compare the performances of the 2nd-level H-WT with respect those displayed by the 1st-order H-WT in the same situation. More precisely, we write the argument of the Airy function as $x=(3/4\times F)^{2/3}\,\exp(\mathrm{i}2\pi/3)$, where $F$ is a real positive parameter, whose value coincides with the singulant $F_{12}$. The study of the asymptotic evaluation of the Airy function across its Stokes line has played a pivotal role in the development of several asymptotic technique, mainly in light of the relative simplicity of the involved saddle topology. Such a simplicity could help in grasping, whereas possible, some conceptual aspects related to the use of the H-WT. Differently from what done in Ref. [@borghiPRE-08], where the relative error values were displayed via the use of tables, in the present paper we are going to resort to graphical visualizations, due to the presence of the two “free” parameters $N$ and $M$. In the first experiment, whose results are shown in Figure \[FigErrorAiryF16\], the Airy function is evaluated for $F=16$. Note that the same experiment was carried out in Ref. [@berryPRSA-90] via the use of H. The values of the relative error, obtained through the 1st-level H-WT, are shown, as black dots, versus the values of $N$, reported on the abscissa axis. For each value of $N$, the values of the relative error obtained via the 2nd-level H-WT, with $M \in [3,N-1]$, are also plotted and, for the sake of clarity, are joined with lines of different color each of them, which corresponds to a different value of $N$, departing from $N$ itself. This can be noted from the figure, where it is immediately seen how the higher the $N$, the longer the corresponding coloured “leg” is. From a first look to the figure, it appears that the relative error, obtained with both the 1st- and the 2nd-level H-WT, is lower bounded. We shall find that all subsequent numerical experiments present the same characteristic. As a first remark, it should be noting that the improvement of the estimate accuracy induced by the 2nd-level H-WT with respect to that obtained via the 1st-level one of the same order appears, at least for values of $N$ not too large, not to adequately refund the unavoidable increase of the computational complexity required by the application of the 2nd-level transformation. For example, it is seen from Fig. \[FigErrorAiryF16\] that a relative error of the order of $10^{-18}$, achieved through the 2nd-level H-WT with $N=11$ and $M=8$, would be reached via a 1st-order H-WT by letting $N=13$, but with a considerable saving of computational effort. The above example clearly suggests the use of the 2nd-level H-WT only in those cases where the best accuracy attainable via the 1st-level H-WT turns out to be not adequate. This happens, for instance, when the integral is attempted to be evaluated beyond the asymptotic realm. To put into evidence this aspect, Figure \[FigErrorAiryF14To2\] shows the same as in Fig. \[FigErrorAiryF16\], but for [a decreasing sequence of values of $F$, namely]{} 14 (a), 10 (b), 6 (c), and 2 (d). In particular, in Fig. \[FigErrorAiryF14To2\](d), where the Airy function argument is located at a distance $\simeq 1$ from the origin of the complex plane, the 1st-level H-WT provides a best error of the order of $10^{-4}$, achieved for $N=4$. Higher accuracies are not allowed because the information gained at the first H stage turns out to be no longer sufficient to generate WT-resummable sequences. The 2nd-level H-WT, on the other hand, provides a best error of the order of $10^{-10}$, which is attained for $(N,M)=(15,11)$. Some intuitive insights about the resummation process associated to the H-WT could be grasped by noting that a lower bounded error is an intrinsic imprint of superasymptotic and hyperasymptotic resummations, whereas it is not generally featured by the application of the WT to alternating factorial divergent series[@wenigerCPR-89]. Accordingly, one should be inclined to think that such an error behavior could be ascribed to the presence of the “regularization" step operated by H on the raw input data. [ Speaking within a more general context, this should be somewhat related to the possible presence on nonanalytic, nonperturbative correction terms which cannot be grasped simply by resummation processes, but rather require the use of “generalized nonanalytic expansions" [@calicetiArXiv07]. ]{}\ In a second experiment concerning the Airy function, the asymptotic parameter $F$ is let running within the interval $[2,4]$ and, for each value of $F$, an exhaustive search of the optimal values of the truncations $N$ and $(N,M)$, which minimize the 1st- and 2nd-level relative errors, respectively, is operated. The results are shown in Fig. \[FigOptimalErrorAiry\], where the optimal relative errors obtained via the 1st- (open circles) and the 2nd-level (dots) H-WT are shown as functions of $F$. The values of the optimal truncation $N$ for the 1st-level H-WT are also reported, versus $F$, in Fig. \[FigOptimalNAiry\], while those of $N$ and $M$, for the 2nd-level H-WT, in Fig. \[FigOptimalNMAiry\](a) and (b), respectively. We will come back later on the above results.\ In concluding the present section, however, we want to provide a table of explicit values obtained through the use of the 2nd-level H-WT. We choose to evaluate the Airy function for $F=2$, for which the optimal setting of the truncation parameters turns out to be $(N,M)=(15,11)$. The preliminary step is the evaluation, through a simple WT, of the contribution, to the Airy integral, coming from the saddle $s_2$. The result is shown in Table \[table.1\]. Furthermore, the subsequent action of the 2nd-level H-WT on the saddle $s_1$ is shown in Table \[table.2\], where the complete estimates of the Airy function provided are reported together with the corresponding values of the truncation $M \in [3,14]$, with $N=15$. ]{} Instanton integral {#instanton} ------------------ The second numerical experiment concerns the evaluation of the instanton integral $$\mathcal{N}(k)=k^{1/2}\,\displaystyle\int_{-\infty}^{+\infty}\, \exp[-k (s^2-1)^2]\,\mathrm{d}s, \label{inst.1}$$ with $k>0$, [already considered in Ref. [@berryPW-93] as a simplified prototype for the modeling of instanton tunneling between symmetric double wells. It was shown in Ref. [@borghiPRE-08] that the integral in Eq. (\[inst.1\]) can be written as[@borghiPRE-08] ]{} $$\mathcal{N}(k)=2\,k^{1/2}\, \mathrm{Re}\left\{\mathcal{I}(k)\right\}, \label{inst.2}$$ where, by referring to Eq. (\[sd.1\]), $g(s)=1$, $f(s)=(s^2-1)^2$, and where $\mathcal{C}$ is the steepest descent path connecting the points $-\mathrm{i}\infty$ and $+\infty$ via the lines $\mathrm{Im}\{s\}\le 0$ and $\mathrm{Re}\{s\}\ge 0$. The complete saddle topology, as well as the expressions of the expanding coefficients associated to all saddles have been described in Ref. [@borghiPRE-08]. In particular, there are three saddles, $s_1=-1$, $s_2=0$, and $s_3=1$, with $\mathcal{A}_{1}=\{2\}$, $\mathcal{A}_{2}=\{1,3\}$, and $\mathcal{A}_{3}=\{2\}$. [The saddles involved in the evaluation of $\mathcal{I}(k)$ are $s_2$ and $s_3$, but only the latter requires a H-WT treatment, since the associated singulant is $F_{32}=1 >0$, and the corresponding asymptotic series turns out to be nonalternating.]{} Furthermore, $\gamma_{12}= 1$, $\gamma_{21}= \gamma_{23}= 0$, and $\gamma_{32}=1$, while we recall that the integral in Eq. (\[inst.1\]) can be expressed in closed form via $$\mathcal{N}(k)= \displaystyle\frac{\pi\,\sqrt k}{2}\,\exp(-k/2)\, \left[ I_{-1/4}\left(\displaystyle\frac k2\right) + I_{1/4}\left(\displaystyle\frac k2\right) \right], \label{inst.4}$$ where $I_n(\cdot)$ denotes the $n$th-order modified Bessel function of the first kind. The first experiment concerns the evaluation of $\mathcal{N}(1/2)$ via the 1st- and the 2nd-level H-WT. In Fig. \[FigInstantonek1ov2\] it is seen how the 2nd-level relative error is bounded, with a minimum value of the order of $10^{-3}$, achieved for $(N, M)=(6,5)$. [On the opposite, the 1st-level H-WT turns out to be completely inadequate to provide a reasonably accurate estimate of the function, due to the very low value of $k$.]{} The searching for optimal values has also been carried out in the present case, but using, as the varying asymptotic parameter, $k \in [1/2,3]$, i.e., where the 1st-order H-WT displays the worst results in terms of accuracy, as shown in Fig. 2a of Ref. [@borghiPRE-08]. The error values are shown in Fig. \[FigOptimalErrorInstanton\], versus $k$, while the optimal settings of $N$ and of $(N,M)$ are plotted, against $k$, in Fig. \[FigOptimalNInstanton\] and \[FigOptimalNMInstanton\], respectively. [It is worth now comparing the results pertinent to the Airy and the $\mathcal{N}(k)$ functions. What we are going to show can seem at first sight somewhat surprising, but gives a possible first hint toward the understanding of the H-WT mechanisms. For simplicity we shall refer to the 1st-level transformation, but the results will apply also to higher-order levels. In Fig. \[FigComparisonAiryInst\], the values of the relative error obtained for the Airy function (dots) and for the instanton function (solid curve) are plotted, versus $N$, when the values of the parameter $F$ and the parameter $k$ are numerically equal. In particular, figure (a) corresponds to $k=F=3$, (b) to 7, (c) to 12, and (d) to 20. It is clearly seen that the behavior of the relative error follows basically the same law. To give a possible explanation of this, in Fig. \[FigCOmpAiryInst\] a pictorial representation of the complete saddle network and the complex integration path involved in the evaluation of the Airy (a) and of the instanton (b) functions is plotted. In both pictures, the black dot denotes the saddle for which the H-WT is required. Although the two saddle distributions are clearly different, they present some common features that, together with Eq. (\[Istage.complete\]), are enough to justify what happens in Fig. \[FigComparisonAiryInst\]. Each of the “black” saddles is adjacent to a single saddle. For the Airy function $s_{1}$ is adjacent to $s_{2}$, while for the instanton function $s_{3}$ is adjacent to $s_{2}$. The values of the corresponding singulants are $F$ and 1, respectively. The use of the resurgence relation in Eq. (\[resurgence.1\]) now gives, for the two “black” saddles, $$\begin{array}{lcr} T^{(1)}_r \propto \displaystyle\frac{(r-1)!}{F^r}, % && % T^{(2)}_r \propto \displaystyle\frac{(r-1)!}{(-F)^r},\\ \end{array} \label{airy1st}$$ for the Airy function and $$\begin{array}{lcr} T^{(3)}_r \propto \displaystyle\frac{(r-1)!}{k^r}, % && % T^{(2)}_r \propto \displaystyle\frac{(r-1)!}{(-k)^r},\\ \end{array} \label{inst1st}$$ for the instanton function. From the above equation it is seen that the behavior of the expanding coefficients follows the same asymptotic law as soon as $F=k$. At the same time, however, the above equality guarantees that also the asymptotic laws for the expanding coefficients corresponding to the adjacent saddles is identical. In fact, for the Airy function the saddle adjacent to $s_{2}$ is $s_{1}$ itself, with a singulant value of $-F$. As far as the instanton function is concerned, the saddles adjacent to $s_{2}$ are $s_{1}$ and $s_{3}$, but for both of them the singulant values are -1. Accordingly, the use of Eq. (\[resurgence.1\]), together with the condition $F=k$, provides again an equivalence between the asymptotic laws of $T^{(2)}_{r}$ for the Airy function and $T^{(2)}_{r}$ for the instanton function. Finally, on using Eq. (\[Istage.complete\]) it is not difficult to convince that the retrieving process is the same for the two functions at the 1st-level[@endnote30]. Leaving a deeper understanding of this phenomenon to future investigations, it is here worthwhile to point out how an immediate consequence of the above described “topological equivalence" could be the restriction of the study of the H-WT retrieving performances to a few classes of prototype test cases. ]{} Swallowtail diffraction catastrophe {#sw} ----------------------------------- As a last numerical experiment we consider asymptotic evaluations of the so-called swallowtail diffraction catastrophe [@berryPIO-80; @connorJPA-84; @nye; @nyePRSA-07], which is defined through the following integral: $$S(x,y,z)= \displaystyle\int_{\mathcal{C}}\, \exp\left[\mathrm{i}\left( \displaystyle\frac{s^5}5+x\,\displaystyle\frac{s^3}3+y\,\displaystyle\frac{s^2}2+zs\right)\right]\, \mathrm{d}s, \label{sw.1}$$ which is of the form given in Eq. (\[sd.1\]) with $g(s)=1$, $f(s)=-\mathrm{i} (s^5/5+x s^3/3+y s^2/2+zs)$, and $k=1$. The integration path $\mathcal{C}$ can be thought of as the union of steepest descent paths approaching, for $\|s\|\gg 1$, the directions $\varphi=(2n+1/2)\pi/5$, with $n=0,1,...,4$. Although a systematic treatment of the swallowtail asymptotics, along the general classical rules recalled in Sec. \[notations\], can be derived by paralleling the analysis carried out, for the Pearcey function, in Ref. [@berryPRSA-91], up to our knowledge it is not present in the current literature. As shown in appendix \[swT\], all expanding coefficients $T^{(n)}_r$ are given by $$\begin{array}{l} T^{(n)}_r= \displaystyle\frac {(5\mathrm{i})^{r+1/2}\,(r-1/2)!} {(10s^{3}_{n}+5s_{n}x+5y/2)^{5r/3+1/2}}\, B^{(r+1/2)}_{2r}(\alpha,\beta), \end{array} \label{generatingC.2}$$ where $$\begin{array}{l} \alpha=\displaystyle\frac{5s_{n}}{(10s^{3}_{n}+5s_{n}x+5y/2)^{1/3}},\\ \\ \beta= \displaystyle\frac {10 s^{2}_{n} + {5x}/{3}}{(10s^{3}_{n}+5s_{n}x+5y/2)^{2/3}}, \end{array} \label{sw.8.1.1.1.1.1.1.1.1}$$ and where the polynomials $B^{(\lambda)}_{n}(u,v)$ are defined via the generating function formula $$\begin{array}{l} \displaystyle\sum_{n=0}^\infty\, {t^n}\,B^{(\lambda)}_n(u,v)= \displaystyle\frac 1{(t^3+u t^2+v t+1)^\lambda}. \end{array} \label{generatingC.3}$$ It is not difficult to prove that Eq. (\[generatingC.3\]) allows the numerical evaluation of the polynomials $B^{(\lambda)}_n(u,v)$ to be efficiently performed via the use of the following recurrence rule, whose derivation is outlined in Appendix \[recurrenceRule\]: $$\begin{array}{l} nB^{(\lambda)}_n=-(n-3+3\lambda)\,B^{(\lambda)}_{n-3}-u\,(n-2+2\lambda)\,\,B^{(\lambda)}_{n-2}\\ \\-v\,(n-1+\lambda)\,B^{(\lambda)}_{n-1}, \end{array} \label{generatingC.4}$$ with the triggering values $B^{(\lambda)}_0(u,v)=1$, $B^{(\lambda)}_1(u,v)=-\lambda v$, and $B^{(\lambda)}_2(u,v)=-u\lambda+v^2\lambda(\lambda+1)/2$. The numerical experiments we are going to illustrate concern asymptotic evaluations of $S(x,y,z)$ at points belonging to the corresponding Stokes set, [which has been estensively studied in Ref. [@berryNL-90] (see, in particular, Fig. 3 of this reference)]{}. Accordingly, the triplets $(x,y,z)$ have been chosen following the prescriptions given in Ref. [@berryNL-90], in order to investigate points at the intersection between the Stokes surface and the plane $x=0$, along the branch corresponding to $y>0$ and $z>0$. This leads to triplets of the form $(x,y,z)=(0, \kappa^{3/2}, \kappa^2\times\,0.23012\ldots)$, with $\kappa$ being a positive parameter[@berryNL-90]. We start by considering the case $\kappa=2$. The saddle topology is constituted by four saddles, which are listed below together with the corresponding list of adjacent ones: $$\begin{array}{lcl} s_1=0.8062\ldots- \mathrm{i}\,1.2357\ldots , & & \mathcal{A}_1=\{4\},\\ &&\\ s_2=s^_1, & & \mathcal{A}_2=\{4\},\\ &&\\ s_3=-1.2828\ldots, & & \mathcal{A}_3=\{4\},\\ &&\\ s_4=-0.3296\ldots, & & \mathcal{A}_4=\{1,2,3\}, \end{array} \label{sw.1.1}$$ with the orientation matrix being $$\{\gamma_{nm}\}= \left[ \begin{array}{cccc} \cdot & \cdot & \cdot & 1 \\ \cdot & \cdot & \cdot & 0 \\ \cdot & \cdot & \cdot & 0 \\ 0 & 1 & 1 & \cdot \end{array} \right], \label{sw.1.2}$$ [where the values of $\gamma_{nm}$ which are not relevant for the present experiment have been replaced by dots.]{} [Three of the four saddles, namely $s_{2}$, $s_{3}$ and $s_{4}$, do contribute to the integral. In particular,]{} the integration path consists in the union of three steepest-descent arcs, the first connecting the points $\infty\,\exp(\mathrm{i}9\pi/10)$ and $\infty\,\exp(\mathrm{i}13\pi/10)$ passing through $s_3$, the second connecting the point $\infty\,\exp(\mathrm{i}13\pi/10)$ to the saddle $s_2$, passing through $s_4$, and the third connecting the saddle $s_2$ to the point $\infty\,\exp(\mathrm{i}\pi/10)$. Accordingly, the Stokes phenomenon occurs via the so-called saddle connection* between saddles $s_4$ and $s_2$[@berryNL-90]. [A pictorial representation of the topology above described is given, for the reader convenience, in Fig. \[FigSwPaths\].]{} [Before showing the numerical results about the performances of the 1st- and the 2nd-level H-WT, it is worth giving some details about the way the saddle topology of the swallowtail integral influences the retrieving capabilities of the WT. We refer, in particular, to the contribution, to the swallowtail integral, of $s_{3}$ and $s_{4}$. As far as the former is concerned, since there is only one adjacent saddle, $s_{4}$, due to the fact that the corresponding singulant $F_{34}=\mathrm{i}\,0.602168\ldots$ is purely imaginary, it turns out that $T^{(3)}_{r} \propto (-\mathrm{i})^{r}\,(r-1)!/\|F_{3,4}\|^{r}$, thus allowing the WT to operate the resummation. Similar considerations can be done for the contributive saddle $s_{2}$. The situation is somewhat different for the saddle $s_{4}$, which is connected to $s_{2}$. In fact, from the above described topology, it turns out that the adjacent saddle $s_{3}$ is dominant, i.e., presents the minimum value of $\|F_{4,m}\|$, with $m \in \mathcal{A}_{4}$. Accordingly, one should conclude also for $T^{(4)}_{r}$ an asymptotic “factorial divided by power” law, similar to that corresponding to $T^{(3)}_{r}$ and, due to the fact that $F_{4,3}=-\mathrm{i}\,0.602168\ldots$, one should expect the WT to be able in resumming the corresponding asymptotic series. But this, on the contrary, does not happen, because of the presence of the other two, nondominant, saddles $s_{1}$ and $s_{2}$, symmetrically placed in the complex singulant space. This can be explained by expliciting Eq. (\[resurgence.1\]) as $$T^{(4)}_r\approx \displaystyle\frac {(r-1)!}{2\pi\mathrm{i}}\, \left[ -\displaystyle\frac{T^{(3)}_{0}}{F^{r}_{43}}+ \left( \displaystyle\frac{T^{(1)}_{0}}{F^{r}_{41}}- \displaystyle\frac{T^{(2)}_{0}}{F^{r}_{42}} \right) \right], \label{resurgence.1Bis}$$ where $T^{(2)}_{0}=\mathrm{i}[T^{(1)}_{0}]^{}$. This example shows how the divergent asymptotic series generated on Stokes sets do not necessarily display a strictly nonalternating sign pattern as, for example, happened for the Airy function, but rather how the asymptotic behavior of their single terms can display more complex patterns, depending on the whole saddle topology. Figure \[FigSWk2\] shows the relative errors obtained through the 1st- and the 2nd-level H-WTs in the case of swallowtail evaluation across the Stokes set defined above, when $\kappa=2$. As for Figs. \[FigErrorAiryF16\], \[FigErrorAiryF14To2\], and \[FigInstantonek1ov2\], the errors are plotted versus the parameters $N$ and $M$. In this case their optimal values turn out to be $N=4$ (for the 1st-level) and $(N,M)=(7,6)$ (for the 2nd-level), with corresponding error values of $2\cdot10^{-3}$ and $5\cdot10^{-5}$, respectively, evaluated with respect to the “exact" value of $S(0,2.8284\ldots,0.9205\ldots)$, obtained via the method recently proposed in Ref. [@borghiJOSAA-08bis]. Finally, an experiment about optimal resummation of swallowtail functions has been carried out, by using $\kappa \in [2,4]$ as the parameter representative of the “asymptoticity" features. The error values are shown, versus $\kappa$, in Fig. \[FigOptimalErrorsw\], while the optimal settings of $N$ and of $(N,M)$ are plotted, against $\kappa$, in Fig. \[FigOptimalNsw\] and \[FigOptimalNMsw\], respectively. ]{} Conclusions =========== [The H-WT was introduced in Ref. [@borghiPRE-08] as a powerful and easily implementable refinement of H aimed at allowing the WT to successfully decode those divergent asymptotic series generated through the application of the steepest descent method to saddle point integrals evaluated across Stokes sets, for which their single terms do not display a strictly alternating sign pattern. ]{} The scheme proposed in [@borghiPRE-08] employed the WT only on the asymptotic series generated by the first-stage hyperasymptotic treatment of the corresponding diverging remainder. In the present sequel we reported on the possibility of combining H and the WT to higher orders in H. In particular, the full development of the 2nd-level H-WT has been detailed within the classical framework of H. The results obtained from the application of the 2nd-level H-WT, also in comparison to those obtained via the 1st-level one, on the different types of saddle-point integrals considered, showed how the increase of complexity and computational effort required by the new transformation be adequately refunded, in terms of accuracy of the estimate, particularly when the integrals are evaluated for values of their parameters which are beyond the asymptotic regime, whereas H turns out to be inapplicable and the 1st-order H-WT unavoidably lacks precision. At the same time, however, it should be noted how, for “ordinary" asymptotic evaluations, at least in the cases considered in the present work, the performances of the 1st- and the 2nd-level H-WTs seem to be comparable in terms of the estimate accuracy, against a considerably difference in the computational efforts required by the two transformations. [Although the H-WT has been developed, here and in Ref. [@borghiPRE-08], with reference to the evaluation of saddle-point integrals of the type in Eq. (\[sd.1\]), we believe it could be useful also in dealing with problems of different nature like, for instance, the hyperasymptotic treatment of a wide class of linear and nonlinear ordinary differential equations, which has been recently considered [@berryJPA-99; @daalhuisPRSA-05; @daalhuisPRSA-05b]. The semi-analytical algorithms proposed for the numerical evaluation of the 2nd-level hyperterminants would reveal useful in this perspective.]{} Before concluding, it is worth pointing out, as an important open problem, the need for an a priori* algorithm for estimating the values of $N$ and $M$ that lead to optimal results. Especially in cases where it is not convenient (or possible) to evaluate the original function, such an algorithm would certainly be of great help to a typical user. Unfortunately, differently from H, for which the optimal settings of the hyperseries truncations are directly extracted from the singulant values[@berryPRSA-90; @berryPRSA-91], at present it does not seem possible to provide similar information for the H-WT. The difficulties in giving practical guidelines for the choice of $N$ and $(N,M)$ can also be appreciated from the results presented in Sec. \[numericalExperiments\] and especially from Figs. \[FigOptimalNAiry\], \[FigOptimalNMAiry\], \[FigOptimalNInstanton\], \[FigOptimalNMInstanton\], \[FigOptimalNsw\], and \[FigOptimalNMsw\], where it seems quite difficult to obtain general rules for the optimal settings of them. [A possible hint, grasped from a quantitative comparison between the results obtained for the Airy and the instanton functions, seems to be given by the strong connection between the H-WT retrieving performances and the saddle topology associated to the integral under consideration. What we found is that different saddle networks can share a sort of “topological equivalence" property, which is related to the set of the saddles adjacent to that under consideration and to the values of the relevant singulants. If two networks turn out to be equivalent at a certain hyperasymptotic level, this would result in the same computational effort, in terms of relative error, as far as the corresponding H-WT retrieved estimates are concerned. This, in particular, would imply that the study of the H-WT retrieving performances could be, in principle, carried out only on a restricted class of prototype functions. ]{} I would like to thank all anonymous reviewers for their constructive criticisms and suggestions. I am also grateful to Turi Maria Spinozzi for his invaluable help during the preparation of the present work. Second-stage hyperasymptotics {#IIstage} ============================= The starting point is again the integral representation of the remainder $R^{(n)}(k,N)$ given in Eq. (\[remainder.2\]), where we let $$T^{(m)}\left(\displaystyle\frac v{F_{nm}}\right)= \displaystyle\sum_{r=0}^{M-1}\,\displaystyle\frac{v^{-r}}{F^{-r}_{nm}}\,T^{(m)}_r+ \displaystyle\sum_{r=M}^{\infty}\,\displaystyle\frac{v^{-r}}{F^{-r}_{nm}}\,T^{(m)}_r, \label{IIstage.1}$$ with $M$ being a new truncation order. Substitution of Eq. (\[IIstage.1\]) into Eq. (\[remainder.2\]) gives $$\begin{array}{lcl} R^{(n)}(k,N) = \displaystyle\frac {(-1)^N}{2\pi\mathrm{i}}\, \displaystyle\sum_{m \in \mathcal{A}_n}\,(-1)^{\gamma_{nm}}\,\\ \\ \times \displaystyle\sum_{r=0}^{M-1}\, (-1)^r\,k^{-r}\,T^{(m)}_r\,K^{(1)}_{N-r}(-kF_{nm}),\\ \\+ \displaystyle\frac {1}{2\pi\mathrm{i}}\, \displaystyle\sum_{m \in \mathcal{A}_n}\,(-1)^{\gamma_{nm}}\\ \\ \times \displaystyle\sum_{r=M}^{\infty}\,\displaystyle\frac{(-1)^{N-r}}{(-kF_{nm})^{N-r}}\,k^{-r}\,T^{(m)}_r\, \displaystyle\int_0^\infty\, \mathrm{d}u\, \exp(-u)\, \displaystyle\frac{u^{N-r-1}}{1-\displaystyle\frac{u}{kF_{nm}}}. \end{array} \label{IIstage.2}$$ On using resurgence, i.e., Eq. (\[resurgence.1\]), on $T^{(m)}_r$, namely $$\begin{array}{lcl} T^{(m)}_r= \displaystyle\frac {1}{2\pi\mathrm{i}}\, \displaystyle\sum_{h \in \mathcal{A}_m}\,(-1)^{\gamma_{mh}}%\\ \displaystyle\sum_{l=0}^\infty\, \displaystyle\frac{(r-l-1)!}{F^{r-l}_{mh}}\,T^{(h)}_l, \end{array} \label{IIstage.3}$$ and $$\begin{array}{lcl} (r-l-1)!= \displaystyle\int_0^\infty\, \mathrm{d}v\, \exp(-v)\,v^{r-l-1}, \end{array} \label{IIstage.3.0.1}$$ after straightforward algebra the second term in Eq. (\[IIstage.2\]) becomes $$\begin{array}{lcl} \left(\displaystyle\frac {1}{2\pi\mathrm{i}}\right)^2\, \displaystyle\sum_{m \in \mathcal{A}_n} \displaystyle\sum_{h \in \mathcal{A}_m}\,(-1)^{\gamma_{nm}+\gamma_{mh}}\, \displaystyle\sum_{l=0}^\infty\,k^{-l}T^{(h)}_l\\ &&\\ \times \displaystyle\sum_{r=M}^\infty\, \displaystyle\frac 1{(kF_{nm})^{N-r}}\, \displaystyle\frac 1{(kF_{mh})^{r-l}}\,\\ \\ \times \displaystyle\int_0^\infty\, \displaystyle\int_0^\infty\, \displaystyle\frac{\mathrm{d}u}{u}\, \displaystyle\frac{\mathrm{d}v}{v}\, \exp(-u-v)\,\displaystyle\frac{u^{N-r}\,v^{r-l}}{1-\displaystyle\frac u{kF_{nm}}}, \end{array} \label{IIstage.3.1}$$ or, by interchanging the integral with the $r$-sum, $$\begin{array}{lcl} \left(\displaystyle\frac {1}{2\pi\mathrm{i}}\right)^2\, \displaystyle\sum_{m \in \mathcal{A}_n} \displaystyle\sum_{h \in \mathcal{A}_m}\,(-1)^{\gamma_{nm}+\gamma_{mh}}\, \displaystyle\sum_{l=0}^\infty\,k^{-l}T^{(h)}_l\, \displaystyle\frac{(kF_{mh})^{l}}{(kF_{nm})^{N}}\\ &&\\ \times \displaystyle\int_0^\infty\, \displaystyle\int_0^\infty\, \displaystyle\frac{\mathrm{d}u}{u}\, \displaystyle\frac{\mathrm{d}v}{v}\, \exp(-u-v)\,\displaystyle\frac{u^{N}\,v^{-l}}{1-\displaystyle\frac u{kF_{nm}}}\,\\ \\ \times \displaystyle\sum_{r=M}^\infty\, \left(\displaystyle\frac vu\,\displaystyle\frac{F_{nm}}{F_{mh}}\right)^{r} =\\ &&\\ \left(\displaystyle\frac {1}{2\pi\mathrm{i}}\right)^2\, \displaystyle\sum_{m \in \mathcal{A}_n} \displaystyle\sum_{h \in \mathcal{A}_m}\,(-1)^{\gamma_{nm}+\gamma_{mh}}\, \displaystyle\sum_{l=0}^\infty\,k^{-l}T^{(h)}_l\, \displaystyle\frac{(kF_{mh})^{l}}{(kF_{nm})^{N}}\\ &&\\ \times \displaystyle\int_0^\infty\, \displaystyle\int_0^\infty\, \displaystyle\frac{\mathrm{d}u}{u}\, \displaystyle\frac{\mathrm{d}v}{v}\, \exp(-u-v)\,\displaystyle\frac{u^{N}\,v^{-l}}{1-\displaystyle\frac u{kF_{nm}}}\, \displaystyle\frac{\left(\displaystyle\frac vu\,\displaystyle\frac{F_{nm}}{F_{mh}}\right)^{M}}{1-\left(\displaystyle\frac vu\,\displaystyle\frac{F_{nm}}{F_{mh}}\right)}. \end{array} \label{IIstage.3.2}$$ Eventually, on making the substitution $-v F_{nm}/F_{mh} \to v$, after some algebra the quantity in Eq. (\[IIstage.3.2\]) can be written as $$\begin{array}{lcl} \displaystyle\frac{(-1)^{N+M}}{(2\pi\mathrm{i})^2}\, \displaystyle\sum_{m \in \mathcal{A}_n} \displaystyle\sum_{h \in \mathcal{A}_m}\,(-1)^{\gamma_{nm}+\gamma_{mh}}\,\\ \\ \times \displaystyle\sum_{l=0}^\infty\,k^{-l}T^{(h)}_l K^{(2)}_{M-l,N-M}\left(-kF_{nm};-\displaystyle\frac{F_{mh}}{F_{nm}}\right), \end{array} \label{IIstage.4}$$ where the 2nd-level hyperterminant $K^{(2)}_{n,m}(\beta;\gamma)$, defined by $$\begin{array}{lcl} K^{(2)}_{n,m}(\beta;\gamma) = \displaystyle\frac 1{\beta^{n+m}}\,\\ \\ \times \displaystyle\int_0^\infty\, \displaystyle\int_0^\infty\, \mathrm{d}u\,\mathrm{d}v\, \exp(-u-\gamma\,v)\, \displaystyle\frac{u^{m-1} v^{n-1}}{\left(1+\displaystyle\frac u\beta\right)\,\left(1+\displaystyle\frac v u \right)}, \end{array} \label{IIstage.6}$$ has been introduced. Accordingly, the complete expression of the 2nd-level hyperasymptotic expansion is $$\begin{array}{lcl} T^{(n)}(k)=\displaystyle\sum_{r=0}^{N-1}\,k^{-r}\,T^{(n)}_r\\ \\+ \displaystyle\frac {(-1)^N}{2\pi\mathrm{i}}\, \displaystyle\sum_{m \in \mathcal{A}_n}\,(-1)^{\gamma_{nm}}\\ \\ \,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\times \displaystyle\sum_{r=0}^{M-1}\, (-1)^r\,k^{-r}\,T^{(m)}_r\,K^{(1)}_{N-r}(-kF_{nm})\\ \\ + \displaystyle\frac{(-1)^{N+M}}{(2\pi\mathrm{i})^2}\, \displaystyle\sum_{m \in \mathcal{A}_n} \displaystyle\sum_{h \in \mathcal{A}_m}\,(-1)^{\gamma_{nm}+\gamma_{mh}}\,\\ \\ \,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\times \displaystyle\sum_{r=0}^\infty\,k^{-r}T^{(h)}_r\, K^{(2)}_{M-r,N-M}\left(-kF_{nm};-\displaystyle\frac{F_{mh}}{F_{nm}}\right), \end{array} \label{IIstage.complete.bis}$$ which coincides with Eq. (\[IIstage.complete\]). Derivation of the extra term for $\beta <0$ {#appC} =========================================== [ We first recast the double integral in Eq. (\[terminant2\]) as follows: $$\begin{array}{l} K^{(2)}_{n,m}(\beta;\gamma) = \displaystyle\frac 1{\beta^{n+m-1}}\, \displaystyle\int_0^\infty\, \mathrm{d}v\,\exp(-\gamma v)\,v^{n-1}\,\\ \\ \times \displaystyle\int_0^\infty\, \mathrm{d}u\,\exp(-u)\, \displaystyle\frac{u^{m}}{(u+\beta)(u+v)}, \end{array} \label{betaNegativo.1}$$ which, if $\beta$ were not negative, would take the form $$\begin{array}{l} K^{(2)}_{n,m}(\beta;\gamma) = \displaystyle\frac {m!}{\beta^{n+m-1}}\, \displaystyle\int_0^\infty\, \mathrm{d}v\,\exp(-\gamma v)\,v^{n-1}\\ \\ \times \left[\displaystyle\frac{\exp(v)E_{m+1}(v)-\exp(\beta)E_{m+1}(\beta)}{\beta-v}\right], \end{array} \label{betaNegativo.2}$$ where the integrand turns out to be continuous even for real, positive values of $\beta$, since $$\begin{array}{l} \displaystyle\lim_{v\to\beta}\, \displaystyle\frac{\exp(v)E_{m+1}(v)-\exp(\beta)E_{m+1}(\beta)}{\beta-v}=\\ \\ = \displaystyle\frac{1-\exp(\beta)\beta^{m}(m+\beta)\Gamma(-m,\beta)}{\beta}. \end{array} \label{betaNegativo.2.1}$$ If $\beta <0$, Eq. (\[betaNegativo.1\]) must be written as $$\begin{array}{l} K^{(2)}_{n,m}(\beta;\gamma) = \displaystyle\frac 1{\beta^{n+m-1}}\, \displaystyle\int_0^\infty\, \mathrm{d}v\,\exp(-\gamma v)\,v^{n-1}\,\\ \\ \times \mathcal{P}\displaystyle\int_0^\infty\, \mathrm{d}u\,\exp(-u)\, \displaystyle\frac{u^{m}}{(u+\beta)(u+v)}, \end{array} \label{betaNegativo.4}$$ where $\mathcal{P}\displaystyle\int_{0}^{\infty}\,\ldots$ denotes the principal part operator. On evaluating the integral in Eq. (\[betaNegativo.4\]) we have $$\begin{array}{l} \mathcal{P}\displaystyle\int_0^\infty\, \mathrm{d}u\,\exp(-u)\, \displaystyle\frac{u^{m}}{(u+\beta)(u+v)}=\\ \\ = {m!}\, \left[\displaystyle\frac{\exp(v)E_{m+1}(v)-\exp(\beta)E_{m+1}(\beta)} {\beta-v}\right]\\ \\- \mathrm{i}\pi\exp(\beta)\,\displaystyle\frac{(-\beta)^{m}}{\beta-v}. \end{array} \label{betaNegativo.5}$$ Finally, the direct substitution of Eq. (\[betaNegativo.5\]) into Eq. (\[betaNegativo.4\]) leads to the closed-form of the extra term as $$\begin{array}{l} \displaystyle\frac {\mathrm{i}\pi\exp(\beta)(-\beta)^{n}}{\beta^{n+m-1}}\, \displaystyle\int_0^\infty\, \mathrm{d}v\,\exp(-\gamma v)\, \displaystyle\frac{v^{n-1}}{v-\beta}=\\ \\ =\mathrm{i}\pi\,(-1)^{n+m-1}(n-1)!\displaystyle\frac{\exp[\beta(1-\gamma)]} {(-\beta)^{n-1}} \,E_{n}(-\beta\gamma). \end{array} \label{betaNegativo.6.1}$$ ]{} Asymptotic coefficients $T^{(n)}_r$ for the swallowtail function {#swT} ================================================================ The swallowtail function in Eq. (\[sd.1\]) corresponds to let $g=1$, $k=1$, and $f(s)=-\mathrm{i}(s^{5}/5+xs^{3}/3+ys^{2}/2+zs)$ in Eq. (\[sw.1\]). The (four) saddle points $\{s_n\}$ are solutions of the algebraic equation $$s^4+x\,s^{2}+y\,s+z=0, \label{sw.1.1.1.1.1}$$ which can be exactly solved by using, for instance, Cardano’s formula. The evaluation of the expanding coefficients $T^{(n)}_r$ requires to evaluate the integral in Eq. (\[sdReview.5.0\]). The first step is to expand $f(s)-f_n$ around $s_n$, i.e., $$\begin{array}{lll} f(s)-f_n= \\ \\= -\mathrm{i}\left[\displaystyle\frac{s^{5}-s_n^{5}}5+ x\,\displaystyle\frac{s^{3}-s_n^{3}}3+ y\,\displaystyle\frac{s^{2}-s_n^{2}}2+ z\,({s-s_n})\right]=\\ \\ =\displaystyle\frac{(s-s_n)^2}{5\mathrm{i}}\,P(s), \end{array} \label{sw.3}$$ where $$P(s)= s^{3}+ 2s_{n}s^{2}+ \left(3s^{2}_{n}+\displaystyle\frac 53x\right)\,s+ \left(4s^{3}_{n}+\displaystyle\frac {10 s_{n}}3x+\displaystyle\frac 52y\right). \label{sw.3.1}$$ Substitution into Eq. (\[sdReview.5.0\]) gives $$\begin{array}{l} T^{(n)}_r = \\ \\= \displaystyle\frac{(r-1/2)!\,\,\,(5\,\mathrm{i})^{r+1/2}}{2\pi\mathrm{i}}\, \displaystyle\oint_{n}\,\displaystyle\frac{\mathrm{d}s} {(s-s_n)^{2r+1}\,[P(s)]^{r+1/2}}=\\ \\ =\displaystyle\frac{(r-1/2)!\,\,\,(5\,\mathrm{i})^{r+1/2}\,}{(2r)!}\, \left\{ \displaystyle\frac{\mathrm{d}^{2r}}{\mathrm{d} s^{2r}}\, \displaystyle\frac 1{[P(s)]^{r+1/2}} \right\}_{s=s_n}. \end{array} \label{sw.5}$$ Equation (\[sw.5\]) can be further simplified by letting $u=s-s_n$ and then by changing $u$ into $t=u/(10s^{3}_{n}+5s_{n}x+5y/2)^{1/3}$, which yields $$\begin{array}{l} T^{(n)}_r= \displaystyle\frac {(5\mathrm{i})^{r+1/2}\,\,\,(r-1/2)!} {(10s^{3}_{n}+5s_{n}x+5y/2)^{5r/3+1/2}}\\ \\ \times \displaystyle\frac{1}{ (2r)!}\, \left[\displaystyle\frac{\mathrm{d}^{2r}}{\mathrm{d} t^{2r}}\, \displaystyle\frac{1}{(t^3+\alpha t^2+\beta t +1)^{r+1/2}}, \right]_{t=0}, \end{array} \label{generatingC.1.5.1}$$ where $$\begin{array}{l} \alpha=\displaystyle\frac{5s_{n}}{(10s^{3}_{n}+5s_{n}x+5y/2)^{1/3}},\\ \\ \beta= \displaystyle\frac {10 s^{2}_{n} + {5x}/{3}}{(10s^{3}_{n}+5s_{n}x+5y/2)^{2/3}}. \end{array} \label{sw.8.1}$$ Note that Eq. (\[generatingC.1.5.1\]) can also be given the alternative following form: $$\begin{array}{l} T^{(n)}_r= \displaystyle\frac {(5\mathrm{i})^{r+1/2}\,(r-1/2)!} {(10s^{3}_{n}+5s_{n}x+5y/2)^{5r/3+1/2}}\, B^{(r+1/2)}_{2r}(\alpha,\beta), \end{array} \label{generatingC.2.1}$$ where the function $B^{(n)}_{\lambda}(u,v)$ is defined through $$\begin{array}{l} \displaystyle\sum_{n=0}^\infty\, {t^n}\,B^{(\lambda)}_n(u,v)= \displaystyle\frac 1{(t^3+u t^2+v t+1)^\lambda}, \end{array} \label{generatingC.3.1}$$ which coincides with Eq. (\[generatingC.3\]). Derivation of the recurrence rule in Eq. (\[generatingC.4\]) {#recurrenceRule} ============================================================ The starting point is the definition of $B^{(\lambda)}_n(u,v)$ through the generating function in Eq. (\[generatingC.3.1\]) which, once derived with respect to $t$, gives $$\begin{array}{l} \displaystyle\sum_{n=0}^\infty\, t^{n}\,n\,B^{(\lambda)}_n(u,v)=-\lambda \displaystyle\frac {3t^3+2ut^2+vt}{(t^3+u t^2+v t+1)^{\lambda+1}}, \end{array} \label{recurrenceRule.1}$$ and, by taking Eq. (\[generatingC.3.1\]) into account once again, leads to $$\begin{array}{l} (3\lambda t^3+2\lambda ut^2+\lambda vt)\, \displaystyle\sum_{n=0}^\infty\, t^{n}\,B^{(\lambda)}_n(u,v)\\ \\+ (t^3+u t^2+v t+1)\,\displaystyle\sum_{n=0}^\infty\, t^{n}\,n\,B^{(\lambda)}_n(u,v)=0. \end{array} \label{recurrenceRule.2}$$ By operating the products term by term, by rearranging the series indices, after long but straightforward algebra Eq. (\[generatingC.4\]) follows. [00]{} R. Borghi, “Joint use of the Weniger transformation and hyperasymptotics for accurate asymptotic evaluations of a class of saddle-point integrals," Phys. Rev. E [78,]{} 026703 (2008). M. Born and E. Wolf, [Principles of Optics]{}, 7th ed. (Cambridge University Press, Cambridge, 1999). M. V. Berry, “Some quantum-to-classical asymptotics," in Les Houches Lecture Series LII, eds. M.-J. Giannoni, A. Voros, and J. Zinn-Justin, Chaos and Quantum Physics, Elsevier North-Holland, Amsterdam, 251-304 (1991). F. J. Wright, “The Stokes set of the cusp diffraction Catastrophe," J. Phys. A 13, 2913-2928 (1980). M. V. Berry and C. J. Howls, “Stokes surfaces of diffraction catastrophes with codimension three," Nonlinearity 3, 281-291 (1990). E. J. Weniger, “Nonlinear sequence transformations for the acceleration of convergence and the summation of divergent series," Comput. Phys. Rep. [10,]{} 189-371 (1989). E. J. Weniger, “Mathematical properties of a new Levin-type sequence transformation introduced by Cizek, Zamastil, and Skala. I. Algebraic theory", J. Math. Phys. [45,]{} 1209-1246 (2004). For an updated review about methods for decoding diverging series, see for instance E. Caliceti, M. Meyer-Hermann, P. Ribeca, A. Surzhykov, and U. D. Jentschura, “From Useful Algorithms for Slowly Convergent Series to Physical Predictions Based on Divergent Perturbative Expansions," Phys. Rep. 446, 1-96 (2007). arXiv:0707.1596v1. C. Brezinski and M. Redivo Zaglia, Extrapolation Methods (North-Holland, Amsterdam, 1991). U. D. Jentschura, P. J. Mohr, G. Soff, and E. J. Weniger, “Convergence acceleration via combined nonlinear-condensation transformations," Comput. Phys. Comm. 116, 28-54 (1999). D. Jentschura, “Resummation of nonalternating divergent perturbative expansions," Phys. Rev. D 62, 076001 (2000). S. V. Aksenov, M. A. Savageau, U. D. Jentschura, J. Becher, G. Soff, and P. J. Mohr, “Application of the combined nonlinear-condensation transformation to problems in statistical analysis and theoretical physics," Comput. Phys. Comm. 150, 1-20 (2003). M. V. Berry and C. J. Howls, “Hyperasymptotics," Proc. R. Soc. A 430, 653-668 (1990). M. V. Berry and C. J. Howls, “Hyperasymptotics for integrals with saddles," Proc. R. Soc. A 434, 657-675 (1991). M. V. Berry and C. Upstill, “Catastrophe optics: morphologies of caustics and their diffraction patterns," Prog. Opt. XVIII, 257-346 (1980). I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and Products, 6th ed., A. Jeffrey and D. Zwillinger, eds. (Academic, 2000). A. B. Olde Daalhuis, “Hyperterminants I," J. Comput. Appl. Math. 76, 255-264 (1996). A. B. Olde Daalhuis, “Hyperterminants II," J. Comput. Appl. Math. 89, 87-95 (1998). A. P. Prudnikov, Yu. A. Brychkov, and O. I. Marichev, Integrals and Series. Vol. III (Gordon Breach Publisher, New York, 1986). J. P. Boyd, “The Devil’s Invention: Asymptotic, Superasymptotic and Hyperasymptotic Series," Acta Applicandae Mathematicae 56, 1-98 (1999). M. V. Berry and C. J. Howls, “Infinity interpreted," Phys. World 6, 35 - 39, (1993). It could be proved that the above “equivalence" between the Airy and the instanton functions keeps its validity also for the 2nd-level H-WT. J. N. L. Connor, P. R. Curtis, and D. Farrelly, “The uniform asymptotic swallowtail approximation: practical methods for oscillating integrals with four coalescing saddle points,” J. Phys. A: Math. Gen. [17,]{} 281-310 (1984). J. F. Nye, [Natural Focusing and Fine Structures of Light]{}, (Institute of Physics Publishing, Bristol, 1999). J. F. Nye, “Dislocation lines in the swallowtail diffraction catastrophe," Proc. Roy. Soc. A [463,]{} 343-355 (2007). R. Borghi, “On the numerical evaluation of cuspoid diffraction catastrophes,” J. Opt. Soc. Am. A 25, 1682-1690 (2008). A. B. Olde Daalhuis, “Hyperasymptotics for nonlinear ODEs. I. A Riccati equation," Proc. R. Soc. Lond. A 461, 2503-2520 (2005). A. B. Olde Daalhuis, “Hyperasymptotics for nonlinear ODEs II. The first Painlevé equation and a second-order Riccati equation," Proc. R. Soc. Lond. A 461, 3005-3021 (2005). M. V. Berry and A. Shelankov, “The Aharonov-Bohm wave and the Cornu spiral," J. Phys. A 32, L447-L455 (1999). A. Sidi, “A new method for deriving Padé approximants for some hypergeometric functions," J. Comput. Appl. Math. 7, 37-40 (1981). List of tables {#list-of-tables .unnumbered} ============== WT order Contribution of saddle $s_2$ ---------- ---------------------------------- 2 15166515+4009986036 $\mathrm{i}$ 3 3913417+0495099 $\mathrm{i}$ 4 968041+4053805 $\mathrm{i}$ 5 9073+30429 $\mathrm{i}$ 6 29942+4613 $\mathrm{i}$ 7 0901+279 $\mathrm{i}$ 8 8+66 $\mathrm{i}$ 9 8+62 $\mathrm{i}$ 10 + $\mathrm{i}$ 11 + $\mathrm{i}$ ... ... : [Values, provided by the WT, of the contribution coming from the saddle $s_2$ in the evaluation of the Airy function across its Stokes set, when $F=2$. ]{} []{data-label="table.1"} M H-WT estimate ------- ------------------------------------------ 3 906105793-0.4032083434 $\mathrm{i}$ 4 61079570-44207316 $\mathrm{i}$ 5 8070334-9375922 $\mathrm{i}$ 6 955837-155117 $\mathrm{i}$ 7 34126-4847 $\mathrm{i}$ 8 6207-275 $\mathrm{i}$ 9 5835-060 $\mathrm{i}$ 10 5939-20 $\mathrm{i}$ 11 51-5 $\mathrm{i}$ 12 244-296 $\mathrm{i}$ 13 563466-928581 $\mathrm{i}$ 14 472-428 $\mathrm{i}$ exact 0.7015826047 - 0.3518080182 $\mathrm{i}$ : [Estimates, provided by the 2nd-level H-WT, of the Airy function across its Stokes set, when $F=2$. Note that the truncation parameter $N$ has been fixed to 15 (corresponding, from Fig. \[FigOptimalNMAiry\], to the optimal setting). ]{} []{data-label="table.2"} List of figure captions {#list-of-figure-captions .unnumbered} ======================= Fig. 1 - R. Borghi Fig. 2 - R. Borghi Fig. 3 - R. Borghi Fig. 4 - R. Borghi Fig. 5 - R. Borghi Fig. 6 - R. Borghi Fig. 7 - R. Borghi Fig. 8 - R. Borghi Fig. 9 - R. Borghi Fig. 10 - R. Borghi Fig. 11 - R. Borghi Fig. 12 - R. Borghi Fig. 13 - R. Borghi Fig. 14 - R. Borghi Fig. 15 - R. Borghi Fig. 16 - R. Borghi	{ "pile_set_name": "ArXiv" }
--- abstract: 'We present a new experimental set-up that creates a shear flow with zero mean advection velocity achieved by counterbalancing the nonzero streamwise pressure gradient by moving boundaries, which generates plane Couette-Poiseuille flow. We carry out the first experimental results in the transitional regime for this flow. Using flow visualization we characterize the subcritical transition to turbulence in Couette-Poiseuille flow and show the existence of turbulent spots generated by a permanent perturbation. Due to the zero mean advection velocity of the base profile, these turbulent structures are nearly stationary. We distinguish two regions of the turbulent spot: the active, turbulent core, which is characterized by waviness of the streaks similar to traveling waves, and the surrounding region, which includes in addition the weak undisturbed streaks and oblique waves at the laminar-turbulent interface. We also study the dependence of the size of these two regions on Reynolds number. Finally, we show that the traveling waves move in the downstream (Poiseuille) direction.' author: - 'L. Klotz' - 'G. Lemoult' - 'I. Frontczak' - 'L.S. Tuckerman' - 'J.E. Wesfreid' title: 'Couette-Poiseuille flow experiment with zero mean advection velocity: Subcritical transition to turbulence' --- Introduction {#intro} ============ The transition to turbulence in wall bounded shear flows is characterized by the presence of localized turbulent regions containing coherent structures in the form of streamwise streaks [@schmid_stability_2001; @landahl_a_1980]. These have been observed in different classical confined shear flows such as boundary layers [@gaster_experimental_1975; @cantwell_structure_1978], water tables [@emmons_laminar-turbulent_1951], and pipe [@hof_turbulence_2005; @mullin_experimental_2011], channel [@lemoult_turbulent_2013] and plane Couette flows [@tillmark_experiments_1992; @bottin_discontinuous_1998]. These turbulent structures are advected downstream with a speed that is approximately proportional to the bulk velocity. When this bulk velocity is non-zero, a very long test section is required to retain the turbulent spots for an appreciable time interval. Another difficulty is that these turbulent structures must be tracked as they move downstream. However, it is possible to cancel the mean flow velocity, as has been realized by pioneering experiments in plane Couette flow [@tillmark_experimental_1991; @daviaud_subcritical_1992]. In these experimental set-ups, the base flow is induced by imposing opposite velocities at each wall of the test section, which generates a linear profile with zero mean velocity. If a turbulent spot is generated under such conditions, it remains stationary in the laboratory framework and there is no time limit on the observation of its evolution. The great advantage of a zero mean velocity has motivated us to construct a facility which is a generalization of the plane Couette experimental set-up. We combine the effect of one moving wall (which introduces a Couette component) and a streamwise pressure gradient due to the backflow generated by imposing zero mean flux rate (responsible for a Poiseuille component). The resulting base flow is a plane Couette-Poiseuille flow with zero mean advection velocity, shown in Fig. \[fig:Scheme1\]. To our knowledge, this is the first experimental investigation of subcritical transition to turbulence in plane Couette-Poiseuille flow. There are a number of theoretical results concerning Couette-Poiseuille flow. The linear stability analysis of this flow (necessarily two dimensional in the streamwise-cross-channel directions, due to Squire’s theorem) was carried out [@potter_stability_1966; @reynolds_finite-amplitude_1967; @cowley_stability_1985; @hains_stability_1967; @drazin_hydrodynamic_1981; @balakumar_finite-amplitude_1997; @ozgen_heat_2006; @savenkov_features_2010], showing that when the Couette component is increased, the linear instability threshold shifts to higher values and the critical wave number decreases with respect to that of pure plane Poiseuille flow (see Fig. 4 in Ref. ). Even a relatively small component of Couette flow is sufficient to completely stabilize plane Poiseuille flow [@drazin_hydrodynamic_1981]. In this case, the linear instability threshold is infinite as it is for pure plane Couette flow. Specifically, it has been proved in Ref. that when the velocity of the Couette component exceeds $70\%$ of the center velocity of the Poiseuille component, the flow becomes stable to infinitesimal perturbations for all finite values of Reynolds number. This is in agreement with other results [@reynolds_finite-amplitude_1967; @cowley_stability_1985] (note however, a slight difference of the coefficient values characterising the contribution of Couette and Poiseuille components reported in Ref. ). Weakly nonlinear stability analysis was used to prove that, while it is stable to infinitesimal disturbances, Couette-Poiseuille flow is unstable to finite amplitude perturbations [@reynolds_finite-amplitude_1967; @cowley_stability_1985; @balakumar_finite-amplitude_1997; @zhuk_asymptotic_2006]. The only fully nonlinear (but still two-dimensional) numerical study of transition to turbulence [@ehrenstein_two-dimensional_2008] used Poiseuille-Couette homotopy to continue a streamwise-localized finite-amplitude solution from Poiseuille to Couette flow. Another two dimensional study used a weakly nonlinear approach to investigate the time evolution of localized solutions in mariginally stable Couette-Poiseuille flow in the framework of the Ginzburg-Landau equation [@jennings_when_1999]. However, it is now believed that two dimensional evolution is not dynamically relevant for subcritical transition to three dimensional turbulence in shear flows. One of the features of subcritical transition to turbulence and three dimensional flow with streamwise or quasi-streamwise elongated streaks is transient linear growth. Its origin is the nonorthogonality of the linearized Navier-Stokes operator and the fact that streaks are the structures most amplified by this process [@butler_threedimensional_1992]. Investigation of transient growth in Couette-Poiseuille flow has shown that adding even a small Couette component (introduced by a moving wall) to a Poiseuille flow (driven by a pressure gradient) significantly increases the nonmodal growth of the energy [@bergstrom_nonmodal_2005]. The flow thus becomes more sensitive to perturbations. As transient growth usually governs the dynamics of flow at early stages of transition to turbulence, one would expect Couette-Poiseuille flow to be less stable than pure Poiseuille flow. ![Schematic representation of plane Couette-Poiseuille flow, which is a combination of Couette (in red) and Poiseuille (in blue) flows. The Couette flow is forced by the upper moving wall, whereas the Poiseuille flow is induced by streamwise pressure gradient. Dash-dotted gray line in the right subfigure separates the two inner regions of the flow: the upper one with $U>0$ dominated by the Couette component (high-shear region) and the lower one with $U<0$ dominated by the Poiseuille component (low-shear region).[]{data-label="fig:Scheme1"}](Fig1 "fig:")\ Recall that linear transient growth cannot explain why turbulence does not decay for sufficiently high Reynolds number. A nonlinear cyclic process has been proposed that makes the turbulence self-sustained [@waleffe_self-sustaining_1997]. Instability of the streaks, manifested by their sinusoidal streamwise waviness [@duriez_self-sustaining_2009], has been found to be necessary to maintain the turbulence. It has been shown quantitatively that the self-sustaining process is relevant to the evolution of turbulent spots in channel flow [@lemoult_turbulent_2014]. We note that fully developed turbulence in Couette-Poiseuille flow has been studied numerically [@pirozzoli_large-scale_2011; @bernardini_statistics_2011; @gretler_calculation_1997; @kuroda_direct_1995], experimentally [@huey_plane_1974; @stanislas_experimental_1992; @telbany_velocity_1980; @telbany_turbulence_1981; @thurlow_experimental_2000; @nakabayashi_similarity_2004] and theoretically [@lund_asymptotic_1980; @wei_scaling_2007]. Specifically, the flow with zero net flux was investigated in Ref. . We complete our survey of Couette-Poiseuille flow by stating that a similar kind of flow profile appears in a long lid-driven cavity [@FLM1] or between two horizontal coaxial cylinders where the gap is partially filled with water (as in the case of Taylor-Dean instability in circular Couette-Poiseuille flow investigated in Ref. . Despite the studies cited above, plane Couette-Poiseuille flow has received relatively little attention until now, especially in the transitional regime. Having a relatively large test section and a base flow with zero mean advection velocity, we have been able to gain more insight into the dynamics of intermittent turbulent structures in shear flows. Until now, the only experimental attempts to generate stationary turbulent structures have been in plane Couette flow, which by definition has no streamwise pressure gradient. We present for the first time nearly stationary turbulent structures in a flow with non-zero streamwise pressure gradient, which represents a wide class of flows with practical relevance. Specifically, we report the first observations of turbulent spots localized in both the streamwise and spanwise directions. Another result of our research is the observation of the macro-organization of turbulence to form oblique turbulent bands in plane Couette-Poiseuille flow, as has been observed for Taylor-Couette [@coles_transition_1965], Taylor-Dean [@mutabazi_spatiotemporal_1990], plane Couette [@prigent_large-scale_2002; @prigent_long-wavelength_2003; @barkley_mean_2007; @duguet_formation_2010; @philip_temporal_2011] and plane Poiseuille flow [@tuckerman_turbulent-laminar_2014; @tsukahara_experimental_2014]. The article is organized as follows: in section \[sec:2\] we describe our new experimental set-up. Next, in section \[sec:3\], we present a general characterization of our installation, including the natural transition to turbulence due to intrinsic noise of the facility. In section \[sec:4\] we characterize the forced transition to turbulence which we triggered by applying a steady, continuous disturbance into the test section. Finally, in section \[sec:5\] we discuss our results. Description of the experimental set-up {#sec:2} ====================================== ![Sketch of: a) the plane Couette experimental set-up [@tillmark_experiments_1992; @daviaud_subcritical_1992]; b) our new facility to investigate plane Couette-Poiseuille flow. The upper moving wall induces the Couette flow to the right, which in turns generates the streamwise pressure gradient inducing Poiseuille flow to the left (compare with Fig. \[fig:Scheme1\]). The Roman numerals correspond to those of Fig. \[fig:ExSetUp\]. Red arrow in the inset marks the lower layer of the plastic belt.[]{data-label="fig:1"}](Fig2a "fig:")\ ![Sketch of: a) the plane Couette experimental set-up [@tillmark_experiments_1992; @daviaud_subcritical_1992]; b) our new facility to investigate plane Couette-Poiseuille flow. The upper moving wall induces the Couette flow to the right, which in turns generates the streamwise pressure gradient inducing Poiseuille flow to the left (compare with Fig. \[fig:Scheme1\]). The Roman numerals correspond to those of Fig. \[fig:ExSetUp\]. Red arrow in the inset marks the lower layer of the plastic belt.[]{data-label="fig:1"}](Fig2b "fig:")\ ![Perspective view of the new experimental set-up with cross-section at the $y_=0$ plane.[]{data-label="fig:ExSetUp"}](Fig3 "fig:")\ ![Configuration we use to: a) perform flow visualizations. The source of conventional, incoherent light is placed at the top and the camera at the side of the test section; b) perform 2D PIV measurements.[]{data-label="fig:5"}](Fig4a) ![Configuration we use to: a) perform flow visualizations. The source of conventional, incoherent light is placed at the top and the camera at the side of the test section; b) perform 2D PIV measurements.[]{data-label="fig:5"}](Fig4b) ![a) Instantaneous snapshot of laminar flow with parabolic streamwise velocity profiles superposed on it. The velocity vectors are measured with the PIV technique by cross-correlating the particles seen in the photo. The stationary wall and moving belt are located at the bottom ($y_=-1$) and top ($y_=1$) respectively. Near the moving belt, we are not able to measure the velocity with PIV; b) representation of streamwise velocity profile $U_(y_)$ (blue crosses) as a function of wall-normal direction normalized with the belt speed. This example corresponds to the central velocity profile in a). We also show a quadratic interpolation of the measured velocity profile (solid red line), which can be observed to fit the data. The blue solid line at the top here and in a) represents the instantaneous $y_$ position of the moving belt; the green dashed line marks the last point which can be measured with PIV. The interpolating function representing the velocity profile and the blue line cross very close to the $(U_=U/U_{\rm belt}=1,y_=1)$ point, which is precisely the position of the moving belt obtained from image processing.[]{data-label="fig:ExPiv"}](Fig5 "fig:")\ ![Time series of deviations of the position of the moving belt. The ordinate represents the deviation from its time-averaged position ($y_=1$). Black and red lines represent the belt position obtained with image treatment ($y_{\rm belt,img}$) and with interpolation ($y_{\rm belt,int}$) respectively. Two sharp peaks on the black lines at $t_=400$ and $t_=900$ are artifacts corresponding to the passage of the adhesive tape joining the two ends of the belt and do not represent the real film displacement. The vertical lines with numbers represent the instants at which instantaneous velocity profiles will be shown (see Fig. \[fig:InterpolationProfiles\]).[]{data-label="fig:TimeSeries"}](Fig6 "fig:")\ ![a) Instantaneous streamwise velocity profiles obtained from interpolation of the measured streamwise velocity. The numbers corresponds to the instants marked by vertical lines in Fig. \[fig:TimeSeries\]; b) time-averaged velocity profiles $U_(y_)$ over one full period of the belt motion for different Reynolds numbers.[]{data-label="fig:InterpolationProfiles"}](Fig7 "fig:")\ First, we denote by $x,y,z$ the streamwise, wall-normal and spanwise directions respectively. The center of the coordinate system is placed in the center of the test section. Our new installation is a generalization of the classical plane Couette experimental set-up (see Fig. \[fig:1\]a and Ref. ). As shown in Fig. \[fig:1\]b we use looped plastic belt to impose the speed on one wall of the test section, while the other wall remains stationary. The moving wall drives the Couette flow toward the right side (red velocity profile on the left of Fig. \[fig:Scheme1\]), which in turn increases the pressure in tank 2. This positive streamwise pressure gradient induces the reverse Poiseuille flow (blue velocity profile in the center of Fig. \[fig:Scheme1\]). The resulting plane Couette-Poiseuille flow (black velocity profile on the right of Fig. \[fig:Scheme1\] and inset of Fig. \[fig:1\]b) is a superposition of these two contributions. It has zero mean advection velocity $\int_{-1}^{1} U(y)dy = 0$. We will also sometimes refer hereafter to the high-shear/low-shear regions close to the moving/stationary wall as the Couette/Poiseuille regions respectively. In Fig. \[fig:ExSetUp\] we present a perspective view with a cross-section at the midgap plane to show in detail the side of the test section where the moving belt is placed. There are two lines of guiding cylinders (marked as I and II, see also Fig. \[fig:1\]b) which guide both layers of the plastic belt into the test section. We can regulate the $y$ position of both ends of these cylinders. There is one additional cylinder (III in Fig. \[fig:1\], \[fig:ExSetUp\]) to keep the plastic belt tight. Its position (just after the motorized cylinder) was chosen carefully to provide us the best stabilization of the $z$ position of the belt when it moves. All the cylinders we use are provided by Interoll with the exception of the motorized cylinder, which requires a large diameter (to diminish the slip between the cylinder and the moving belt) and a slightly tapered shape at both ends (to better control the $z$ position of the belt when it moves). For these reasons we have manufactured it on a three dimensional printer. Four additional external steel beams reinforce the test section and diminish the deflection of the side walls of the test section due to the hydrostatic pressure of the water. The experimental set-up is mounted on a heavy granite table which provides mechanical stability and thermal inertia. The motorized cylinder is driven by a servo-motor produced by Yaskawa Electrics ($100$ W) with a gear reduction of 1:26. We use glass plates (of $8$ mm thickness) as side walls, a Plexiglas beam as the upper wall and a transparent plastic belt made of Mylar(of $175$ $\mu$m thickness), granting optical access to the test section. The gap between the glass walls of the test section is $2h_1 = 14$ mm. However, one can observe in schematic Fig. \[fig:1\]b that there are two layers of the plastic belt in the vicinity of the upper wall, which bounds the effective gap of the test section. We measure it with an optical method as $2h = 11.5$ mm. Hereafter we use the phrase moving belt to refer only to the lower (inner) layer of plastic film (indicated by the red arrow in the inset of Fig. \[fig:1\]b). The streamwise and spanwise dimensions of the test section are $2000$ mm and $540$ mm respectively. However, the width of the plastic belt is $520$ mm, which is our effective spanwise dimension. The belt is positioned slightly asymmetrically in the spanwise direction. Taking all of this into consideration, the aspect ratios of the test section in the streamwise/spanwise directions are $L_x/h = 347.8$ and $L_z/h = 90.4$ respectively. Spatial coordinates and time are nondimensionalized with the effective half gap $h$ and $h/U_{\rm belt}$ respectively and are marked by $$ subscript. The velocity profile normalized by the belt speed is $$U_=U/U_{\rm belt}=\frac{3}{4} ({y_}^2-1)+\frac{1}{2}(y_* +1), ~~{\rm where}~~~ y_* = y/h \in (-1,1).$$ The Reynolds number is based on effective half gap $h$ (following the convention from plane Couette and plane Poiseuille flows) and the speed of the moving wall, $U_{\rm belt}$, namely ${\mbox{\textrm{Re}}}=U_{\rm belt} h/\nu$. We explore the range between ${\mbox{\textrm{Re}}}=160$ and ${\mbox{\textrm{Re}}}=780$. We photograph the flow visualizations with a Nikon D200camera (3800$\times$2800 pixels matrix) and Nikkor$f=35$ mm lens. Its optical axis is collinear with the $y$ axis and the source of the white light is at the top of the test section (Fig. \[fig:5\]a). In addition, we also acquire supplementary video of flow visualization with video camera Canon 3CCD XM2 Pal (720$\times$520 pixels), which will be described in details at the end of the section \[sec:4\]. To perform the PIV measurements we use a Phantom MIRO M120camera (1920x1600 pixels) with Nikkor$f=85$ mm lens and a Darvin Duo laser (double-headed, maximum output $80$ W, wavelength $527$ nm) in the configuration presented in Fig. \[fig:5\]b. We acquire a sequence of either single or double frame snapshots, which are then post-processed with Dantec Dynamic Studio4.2 software. We use a high concentration of small seeding particles made of Polyamid with a diameter of $5\mu$m. We use a rectangular 256 $\times$ 16 pixel interrogation window in the $x$ and $y$ directions with a 50% overlap. This unconventional choice is justified by the dominant streamwise velocity component, which implies that the streamwise pixel displacement is an order of magnitude larger than that in the wall-normal direction. The rectangular window enables us to increase the signal-to-noise ratio, keeping a high spatial resolution in the $y$ wall-normal direction. With this procedure we measure instantaneously three velocity profiles with $0.3h$ spacing in $x$ and with 100 points across the gap. When the camera is placed on top of the test section (with its optical axis aligned along the $z$ axis), we can measure the streamwise velocity with PIV only in the vicinity of the upper wall of the test section due to the high concentration of seeding particles. In order to measure the streamwise velocity profiles at different spanwise locations, we put the camera on the side of the test section (with its optical axis inclined at $45^{\circ}$ with respect to the $z$ axis, see Fig. \[fig:5\]b). We also use a Scheimpflug mount to record a well-focused image despite the inclination of the camera, as well as a water prism to reduce the optical distortions due to the difference in refractive indices of water and air. In order to study the laminar base flow, we measure the instantaneous streamwise velocity for different Reynolds numbers (${\mbox{\textrm{Re}}}\in (250,340,430,510)$) in the central part of the test section. We acquire a single image sequence and correlate two consecutive images. We set a high enough frequency (from $19$ Hz to $42$ Hz, depending on Reynolds number) to retain the time correlation between two snapshots. We need to record about 2800 images on the 3 GB internal memory of the Phantom camera to cover one period of the belt motion. For this reason we use part of the camera matrix (512 pixels in $x$ and 896 pixels in $y$). This procedure provides us with the best possible temporal resolution for a given spatial resolution (directly related to the size of the camera matrix) and for a given measurement time. In Fig. \[fig:ExPiv\]a we present one example of an instantaneous PIV vector field acquired for ${\mbox{\textrm{Re}}}=510$, which shows three similar velocity profiles within the measurement area. We plot the central profile in $x$ in Fig. \[fig:ExPiv\]b. Fig. \[fig:ExPiv\]a shows that the width of the optical image of the belt is about 1 mm, which is more than its actual thickness. This is a consequence of the inclination of the optical axis of the camera with respect to the laser sheet of finite width. Indeed, the light coming from the laser sheet of finite width is reflected by the belt and then registered on the camera matrix as a thick line. There are additional contributions from the defocusing and scattering of the laser light in the vicinity of the moving belt. In the region above the dashed green line we are not able to measure the velocity with our PIV technique, as it produces many spurious vectors. We define the center of this thick line as the instantaneous position of the moving belt ($y_{* \rm belt,img}$) and we determine it for each image using edge detection techniques. In Fig. \[fig:ExPiv\] we mark this instantaneous belt position by a solid blue line and we assign the value $y_=1$ to this location. In Fig. \[fig:TimeSeries\] we present as a solid black line a time series of the deviations of the moving belt position from its time-averaged location. The actual belt position changes smoothly in time. As in the measured data in Fig. \[fig:ExPiv\]b there is a maximum and due to the fact that we expect the laminar Couette-Poiseuille flow to be a quadratic function of $y_$, we interpolate the measured velocity points using a quadratic polynomial of the form $U_(y_)=\sigma_1(y_^2-1) + \sigma_2(y_+1)$ (red line in Fig. \[fig:ExPiv\]b). The interpolation fits the data very well. Then we estimate the $y_{\rm belt,int}$ position at which the interpolation function reaches the known value of belt speed ($U(y_{\rm belt,int})=U_{\rm belt}$). In Fig. \[fig:ExPiv\]b the interpolated streamwise velocity profile (red line) and the measured belt position (thick blue line) cross very close to the point $(1,1)$ as expected, which confirms the validity of our interpolation curve. We also compare the time evolution of the two wall positions obtained by these two methods ($y_{\rm belt,int}$ as the red and $y_{\rm belt,img}$ as the black line). The belt position predicted by interpolation matches very well the real position of the moving belt, with a deviation of less than 0.1 mm. We note that this is the first time that a detailed study of the gap variation is performed for this type of experiment with moving walls (including plane Couette facilities). In order to check whether the base flow is affected by temporal fluctuations of the moving belt position we plot in Fig. \[fig:InterpolationProfiles\]a instantaneous interpolations of the streamwise velocity profiles for six different instants, which are marked in Fig. \[fig:TimeSeries\] by dashed vertical lines and numbers from 1 to 6. These velocity profiles are virtually the same, which proves that the base flow does not depend in a significant way on the phase of the belt motion. We also calculate the time averaged velocity profiles $<U(y_)>_t$ for different Reynolds numbers (Fig. \[fig:InterpolationProfiles\]b). They collapse onto a single curve after being normalized with the belt speed. Finally, we calculate the mean advection velocity of the time averaged velocity profile ($U_{\rm avg}=\frac{1}{2h}\int_{-1}^1 <U(y_)>_tdy$), which does not exceed $0.03 U_{\rm belt}$ in the central part of the test section. Having determined the fluctuations of the belt position (Fig. \[fig:TimeSeries\]), as well as the variation of the velocity profiles in time (Fig. \[fig:InterpolationProfiles\]a) and for a given Reynolds number (Fig. \[fig:InterpolationProfiles\]b), we can estimate the total error of the local Reynolds number (for a given $z$ position) as lower than 5%. The variation of the effective gap for different $z_$ locations is lower than $\pm 0.5$ mm. The spatial variation of the temperature in the test section does not exceed $0.2^{\circ}$C. The resulting error related to the fluid viscosity is less than $0.6\%$. We estimate the global variation of the Reynolds number for different $z_$ locations as lower than $7.6\%$. The cross-flow component of the base flow (in the spanwise direction) is lower than $2.0\%$. In all subsequent figures the direction of motion of the plastic belt is toward the right. Characterization of the natural transition to turbulence triggered by the intrinsic noise of the installation {#sec:3} ============================================================================================================= ![Flow visualizations for different Reynolds numbers: a) uniform laminar flow (${\mbox{\textrm{Re}}}=330$); b) featureless turbulent region in the entire test section (${\mbox{\textrm{Re}}}=780$).[]{data-label="fig:2"}](Fig8){width="\linewidth"} ![Flow visualizations of a localized turbulent spot surrounded by laminar flow in plane Couette-Poiseuille flow (${\mbox{\textrm{Re}}}=530$). Sequence of pictures shows the slow advection of the turbulent structure toward the right (with time interval of $96$ advection units).[]{data-label="fig:4"}](Fig9 "fig:"){width="\linewidth"}\ ![Example of macro-organization of turbulent spots in the form of oblique turbulent bands (${\mbox{\textrm{Re}}}=670$).[]{data-label="fig:3"}](Fig10 "fig:"){width="\linewidth"}\ ![Spatio-temporal diagram of: a) measured streamwise velocity, $U_(t,y_)$, for the laminar flow (${\mbox{\textrm{Re}}}=480$). The black solid line represents the time-averaged profile; b) measured streamwise velocity, $U(t,y_)$, for the intermittent flow (${\mbox{\textrm{Re}}}=520$); c) streamwise velocity fluctuations $u_'(t,y_)=(U(t,y_)-U_{\rm base}(y_))/U_{\rm belt}$. We consider the time-averaged streamwise velocity profile in the range $t_ \in (20-120)$ as the base flow without perturbations. The pattern at $t_* \in (800, 1400)$ is a signature of the unsteady, wavy structure of the turbulent spot. The black profiles have been calculated by time-averaging the instantaneous velocity profiles within the ranges marked by white dashed lines.[]{data-label="fig:STturb"}](Fig11 "fig:")\ We perform flow visualizations to characterize qualitatively the flow in the test section. For this purpose we use reflective aluminium flakes (STAPA IL HYDROLAN 2154 55900/G produced by ECKART) of typical diameter $d_{al}\in(30$ $\mu$m$,80$ $\mu$m$)$, which are dispersed in water. These tracers enable the detection of three dimensional vortical structures in the flow, through spatial fluctuations of reflected light intensity. In contrast, the light distribution in laminar regions is nearly uniform and featureless. In this way, turbulent regions can be distinguished. The pictures presented here are taken with a Nikon camera with a 3800$\times$ 2800 pixel matrix, with a pixel pitch equal to $0.167$ mm. Note that the fluid in the main tanks is continually disturbed by the rotating cylinders, which makes it always turbulent for the range of Reynolds numbers considered here. For this reason the inlets of the test section are the sources of the natural perturbations. Even though these operating conditions are similar to those presented in Ref. for Poiseuille flow, we recall here that the mean advection velocity is nearly zero, so the turbulent flow is not advected from the inlets to the test section. The flow is laminar in the entire test section up to ${\mbox{\textrm{Re}}}\simeq 420$ (as in Fig. \[fig:2\]a). For higher Reynolds numbers some turbulent structures appear at the left entry ($x_<0$) generated by the turbulence in the main tank. Up to ${\mbox{\textrm{Re}}}\simeq 480$, the amplitude of these perturbations is not strong enough to trigger the transition and the turbulent structures present in the main tank do not propagate further into the test section. For ${\mbox{\textrm{Re}}}\gtrsim 480$ the Couette-Poiseuille flow is no longer stable and the test section is occasionally invaded by transient patches of turbulence. However, up to ${\mbox{\textrm{Re}}}\simeq 510$ these events rarely occur and the undisturbed laminar base flow can persist for most of the time. In figure \[fig:4\]a,b we present a sequence of images illustrating such a localized turbulent spot surrounded by laminar flow, which is slowly advected to the right with a very small advection speed of $U_{\rm advection}\simeq 0.095 U_{\rm belt}$. As we increase the Reynolds number even further (to ${\mbox{\textrm{Re}}}\simeq 670$), the spots expand obliquely to form a turbulent structure reminiscent of laminar-turbulent bands, one example of which is presented in Fig. \[fig:3\]. Finally, at high enough Reynolds number (${\mbox{\textrm{Re}}}\simeq 780$), the flow is uniformly turbulent (Fig. \[fig:2\]b). In order to demonstrate the transition to turbulence in more detail, we measure the instantaneous streamwise velocity as a function of $y_$ with the PIV configuration shown in Fig. \[fig:5\]. We acquire double-frame images with the sampling frequency of 2 Hz, which are cross-correlated to determine the instantaneous velocity fields. In Fig. \[fig:STturb\]a we present the spatio-temporal diagram of a single streamwise velocity profile $U_(t_,y_)$ for low Reynolds number (${\mbox{\textrm{Re}}}=480$) measured at $(x_=0,z_=0)$. The isocontours on this diagram are nearly horizontal, which shows that the flow is laminar and does not depend on time. This corresponds to the visualization of a laminar flow shown in Fig. \[fig:2\]a). We determine the base flow profile $U_{\rm base}(y_)$ by time-averaging the results within the entire sequence of measurements. The resulting profile (black solid line in the center of Fig. \[fig:STturb\]a, see also Fig. \[fig:InterpolationProfiles\]b) is a quadratic polynomial. As we increase the Reynolds number to ${\mbox{\textrm{Re}}}= 520$, we observe a transition to turbulence triggered by intrinsic noise of the installation. In Fig. \[fig:STturb\]b the flow becomes locally time-dependent/intermittent for $t_* \in (170, 1550)$, due to the passage of the localized turbulent spot through the PIV measurement section (compare also with the visualizations of the turbulent spot in Fig. \[fig:4\]). However, for $t_* \in (0, 170)$ the flow is stationary and laminar. We calculate the time-averaged profile in this range (parabolic profile on left side of Fig. \[fig:STturb\]b), which we consider as the base flow without perturbation $U_{\rm base}(y_)$. Then we subtract it from the measured instantaneous streamwise velocity component $U(t,y_)$ to calculate the streamwise velocity fluctuations $u'(t,y_)=U(t,y_)-U_{\rm base}(y_)$ (Fig. \[fig:STturb\]c). We can clearly observe the unsteady structure of these fluctuations, a signature of a turbulent spot in plane Couette-Poiseuille flow. The transition starts in the vicinity of the moving wall, in the high shear region ($t_ \in (250, 500)$), and then, as the turbulent structure grows, it gradually spreads across the whole gap. Finally the flow relaxes back to the laminar state ($t_>1550$). We also show two examples of turbulent streamwise velocity and fluctuation profiles ($t_=415$ and $t_=1100$ in Fig. \[fig:STturb\]b,c), which are calculated by time averaging the data within the range delimited by white dashed lines. The profile at $t_=415$ shows the wall-normal transfer of fluid with negative velocity from the low-shear region toward the high-shear region, whereas the profile at $t_=1100$ shows the opposite. Note in Fig. \[fig:STturb\]c that near the moving wall, the streamwise velocity fluctuations are negative, whereas away from the wall (low shear region) the streamwise velocity fluctuations are positive. These PIV measurements demonstrate that plane Couette-Poiseuille flow can be also regarded as asymmetric Poiseuille flow with one active, high shear region near the moving belt. This is in contrast to the classical symmetric Poiseuille profile with two active regions, one next to each wall. Transition to turbulence triggered by a permanent perturbation {#sec:4} ============================================================== ![Instantaneous flow visualizations for different Reynolds numbers. The transition to turbulence is triggered by a sphere placed in the test section near the moving wall. The $(x_,z_)$ origin is now located at the center of the sphere. The stationary/moving wall is closer to/further from the reader and the direction of the moving wall is toward the right. We also superpose on these images the time-averaged envelope contours that represent the total area of the spot (black contours) and its active turbulent core (white contours).[]{data-label="fig:FT_Vis8a"}](Fig12 "fig:")\ ![Time-averaged envelope representing the total area of the turbulent spot: a) spatial extent of the envelope along the streamwise direction for $z=0$; b) spanwise extent of the envelope, where the $x$-coordinate has been selected for each Reynolds number to maximize the spanwise extent.[]{data-label="fig:ENV_RES"}](Fig13 "fig:")\ ![Dependence on ${\mbox{\textrm{Re}}}$ of the envelopes of the total area and the active core of the turbulent spot: a) the area; b) the $x_$ position of the centroids; c),d) the extent in $z$ and $x$ directions. Crosses correspond to the active area (containing wavy streaks), while circles represent the region which includes in addition the weak undisturbed streaks and oblique waves at the laminar-turbulent interface.[]{data-label="fig:ENVplots"}](Fig14 "fig:")\ ![Decomposition of the flow visualization of the turbulent spot into the structures that move upstream and downstream: a) full image of flow visualization; b) pattern which moves upstream (right); c) pattern which moves downstream (left). All three pictures correspond to the same instant of time.[]{data-label="fig:LeftRight"}](Fig15 "fig:")\ In this section we present flow visualizations for the transition forced by an external and permanent perturbation. For this we insert a ferromagnetic sphere of diameter $6.2$ mm, which is held at a fixed position within the test section by a strong magnet. The sphere touches the moving wall and so is within the high shear region. In addition, the friction with the moving belt causes rotation of the obstacle. However, this imposed frequency is higher than the typical frequencies observed in the flow and thus can be neglected. This obstacle locally modifies the flow [@bottin_discontinuous_1998], creating a steady, localized disturbance. For each Reynolds number, we take a sequence of 90 images with sampling frequency $f=1$ Hz. We recall that we acquire the images with very high spatial resolution (3800 $\times$ 2800 pixel matrix). We shift the origin of the coordinate system with respect to Fig. \[fig:2\]-\[fig:4\] by placing it at the center of the sphere. We call the left/right side of the sphere the downstream/upstream direction, taking the direction of the back flow (Poiseuille component) as the reference. In Fig. \[fig:FT\_Vis8a\] we present flow visualizations representing the flow structure as the Reynolds number is increased. For ${\mbox{\textrm{Re}}}= 165$ we observe a few stationary streamwise vortices which expand towards the left (Fig. \[fig:FT\_Vis8a\]a). The vortical structure observed on the left side of the sphere is probably due to a pair of streamwise counter-rotating vortices generated in the wake of the sphere, which, in uniform background flow, appears at ${\mbox{\textrm{Re}}}_{\rm sphere}\simeq 210$ [@johnson_flow_1999; @gumowski_transition_2008], where ${\mbox{\textrm{Re}}}_{\rm sphere}=(U_{\rm freestream}\,d_{\rm sphere})/\nu$. In our case $d_{\rm sphere}=6.2$ mm $\simeq h$, which implies that ${\mbox{\textrm{Re}}}_{\rm sphere} \simeq {\mbox{\textrm{Re}}}$. For ${\mbox{\textrm{Re}}}= 255$ the turbulence starts to invade the right side of the sphere (note the appearance of small vortices for $x_>0$ in Fig. \[fig:FT\_Vis8a\]c). As the Reynolds number is further increased, this streamwise extent to the right becomes increasingly important. We have also observed that for ${\mbox{\textrm{Re}}}\lesssim 480$ the spot stays in a fixed location pinned to the sphere, but for higher Reynolds numbers the size of the spot fluctuates and it moves toward the right. This can be compared with the front speeds in pipes for puffs, where the upstream front (with respect to the direction of the Poiseuille component) travels more slowly downstream than the average velocity of the base flow [@barkley_rise_2015]. The analogue of this situation in our case is the motion to the right (upstream). The spots have a preferred inner structure (a spanwise-periodic pattern of streamwise streaks) with wavelength $\lambda_z$ about $2.5h$ (representing the wave vector ($k_{x}=0, k_{z}=2.52$) in Fourier space). However the spot structure also includes oblique waves (i.e. straight streaks which are oriented slightly obliquely with respect to the streamwise direction) at the laminar-turbulent interface and undulated (or wavy) rolls in the center of the spot, which broaden the spatial Fourier spectrum. In order to describe the dependence of the area of the turbulent spot on Reynolds number, we use the two-dimensional Hilbert transform to compute the envelope of the modulated function of gray levels representing the spot. First, we normalize the pixel intensity of each image by dividing it by the background reference corresponding to the laminar flow without the sphere. Then we compute its two dimensional FFT spectrum and we filter it, retaining the range $\|k_{x}\| < 1.19$ and $k_{z} \in (0.76, 4.28)$. Next, we use the two-dimensional inverse FFT transform to compute the filtered spot and we get its envelope/amplitude $\rho(x,z)$. Finally, we compute for each Reynolds number the time-averaged spatial envelope for all images in the sequence. This is justified because the global dynamics of the turbulent spot is nearly stationary, as the forcing is constant in time and the turbulent region is pinned to the sphere until ${\mbox{\textrm{Re}}}\simeq 480$. We also estimate the size of the more dynamically active, turbulent region at the core of the turbulent spot. To do this we use the observation that this region is directly related to streamwise waviness of the streaks, which resembles travelling waves. The streamwise dependence of wavy streaks is manifested by the appearance of modes with $k_{x} \neq 0$ in the spatial spectrum. However, such modes generate higher harmonics. This effect is further increased by the fact that we analyse the pixel intensity of flow visualizations, which adds spurious nonlinear content. As a result we are not able to identify a single mode which corresponds to the streak waviness. Instead, we consider the spectral range $\|k_{x}\| \in (1.13,1.85)$ and $k_{z} \in (1.60,3.44)$, which is related to the harmonics of this structure. In this way, we can insure that the envelope computed from this spectral region corresponds to the short wavelength streamwise undulation rather than to the long oblique straight waves. In order to describe a spatial distribution of both regions (namely the active core related to the waviness of the streaks and the total area, which in addition includes the surrounding region with oblique waves at the laminar-turbulent interface) for different Reynolds numbers, we superpose iso-contours of both envelopes on the flow visualization pictures (Fig. \[fig:FT\_Vis8a\]). Note in Fig. \[fig:FT\_Vis8a\]a that there is no active region for ${\mbox{\textrm{Re}}}= 165$, since the structures there are wake vortices generated by the sphere and unrelated to turbulence. In order to better illustrate how the size of the turbulent spot changes with increasing Reynolds number, we show spatial profiles of the total spot envelope along the $x_$ (Fig. \[fig:ENV\_RES\]a) and $z_ $ (Fig. \[fig:ENV\_RES\]b) directions. The former is plotted for $z_=0$, and the latter for the value of $x_$ which maximizes the size along the $z_$ direction. For low Reynolds numbers (${\mbox{\textrm{Re}}}=165$ in Fig. \[fig:FT\_Vis8a\]a and \[fig:ENV\_RES\]a) all of the activity takes place on the left side of the sphere. The upstream front is steep, whereas in the downstream direction the envelope slowly decays to zero with a large tail extending toward $x_<0$. As we increase the Reynolds number the turbulent region extends further and further upstream. In Fig. \[fig:ENVplots\] we present several quantities to further characterize this dependence. Fig. \[fig:ENVplots\]a shows that the area of both the total and the active regions increase monotonically with Reynolds number. In Fig. \[fig:ENVplots\]b we show the dependence on Reynolds number of the $x_$ centroid position for both total and active regions. First, we observe that both centroids follow the same evolution. For low Reynolds numbers they are located on the left side of the sphere, at ${\mbox{\textrm{Re}}}\simeq 380$ they cross zero and for higher ${\mbox{\textrm{Re}}}$ they continue to shift upstream. This indicates that the high-shear (Couette) region near the moving wall becomes increasingly important as the Reynolds number is increased. Finally, at ${\mbox{\textrm{Re}}}=510$ almost all activity takes place within the high shear (Couette) part. This agrees with the numerical observations in plane Poiseuille flow with zero net flux that the turbulent structures move with/against the direction of the Poiseuille component for low/high Reynolds numbers (see Ref. , note that in that paper the direction of the Poiseuille component is in the positive $x_$ direction, opposite to our case). However, recall that instead of measuring the propagation speed of the turbulent structure, we are measuring the direction in which the turbulent spot extends. One should think of this as continuous advection of the turbulence, which decays as it moves downstream and is simultaneously continuously regenerated by a permanent perturbation. As mentioned in the discussion of Fig. \[fig:FT\_Vis8a\]a, for sufficiently low ${\mbox{\textrm{Re}}}$, there is no active region; the perturbations seen are vortices in the wake of the perturbing sphere, which are located downstream/left from the sphere. Since the right side ($x_>0$) is less affected by the sphere, we plot in Fig. \[fig:ENVplots\]a the part of the active region located only on the right side of the sphere (green crosses). We note that the area of this portion of the active area remains nearly equal to zero up to ${\mbox{\textrm{Re}}}=330$ and then starts to grow. The Reynolds number dependence of the streamwise and spanwise size of turbulent spots is presented in Fig. \[fig:ENVplots\]c,d. Both of them grow monotonically with Reynolds number up to ${\mbox{\textrm{Re}}}=470$. At ${\mbox{\textrm{Re}}}=510$ the spanwise extent seems to saturate as a result of the finite size of our test section. The streamwise extent is less affected, as the streamwise dimension of our installation is bigger ($L_x/h = 2000/5.75 = 347.8$) than the spanwise one ($L_z/h = 520/5.75 = 90.4$). The spanwise extent grows with Reynolds number by adding new streaks in the $z$ direction, which can be observed in Fig. \[fig:ENV\_RES\]b. Similar behaviour has been observed numerically in plane Couette flow [@duguet_stochastic_2011]. Finally, in order to separate the turbulent structures that move downstream and upstream, we record a video of a turbulent spot generated by the sphere for ${\mbox{\textrm{Re}}}=470$. To do this, we use the video camera with acquisition frequency $f=25$ Hz, which enables us to calculate the two dimensional spatio-temporal ($x,t$) FFT transform for each $z$ location. Motivated by the decomposition of travelling waves in thermal convection [@croquette_nonlinear_1989; @kolodner_complex_1990], we introduce the following procedure: we calculate the inverse FFT of the two dimensional spectrum ($k_x,\omega$) for each of the quadrants I ($k_x>0,\omega>0$) and II ($k_x<0,\omega>0$) separately. These two quadrants represent the travelling waves that go to the right/upstream (with phase $k_xx-\omega t$) and left/downstream (with phase $k_xx+\omega t$) respectively. In Fig. \[fig:LeftRight\] we present the resulting fields for a given instant of time. Video frames are normalized by dividing their intensity by that of the image of the laminar flow (Fig. \[fig:LeftRight\]a), whereas in Fig. \[fig:LeftRight\]b,c we plot the fluctuations of the normalized pixel intensity. Fig. \[fig:LeftRight\]a shows a turbulent spot with a characteristic V-shape pointing to the left. The dominant pattern of the right-going structures (Fig. \[fig:LeftRight\]b) are oblique waves at the tips of turbulent spot (similar to those found in plane Poiseuille flow [@carlson_flow-visualization_1982; @henningson_wave_1987]). In contrast, the downstream pattern (Fig. \[fig:LeftRight\]c) contains the wavy streaks, which defines our active region. Thus it is associated with the turbulent core of the spot (see also Supplemental Material at \[URL will be inserted by publisher\] for the full video showing this propagation). This difference between two patterns indicates the role of the streamwise pressure gradient (absent in plane Couette flow), which breaks the left/right symmetry. One can observe the tape joining the two ends of the plastic belt ($x_=35$ in Fig. \[fig:LeftRight\]a), which at this instant moves to the left, is visible only in the pattern moving downstream/left (Fig. \[fig:LeftRight\]c). This confirms that our decomposition of the flow visualization separates the upstream and downstream patterns. Conclusions {#sec:5} =========== We have presented a new installation to investigate plane Couette-Poiseuille flow. We have achieved this by combining a moving belt with a streamwise pressure gradient forcing the back-flow in the opposite direction. The mean velocity of the resulting base flow is nearly zero, which enables us to generate turbulent structures which are nearly stationary in the laboratory frame. This is the first time that stationary structures have been generated experimentally in a shear flow with a non-zero streamwise pressure gradient. We describe the observation of the following sequence as Reynolds number is increased: laminar state, localized spots which grow with Reynolds number (both in the streamwise and spanwise directions) and finally, oblique expansion of the spot, which forms a turbulent band. We note that we have characterized the linear stability of our particular velocity profile, given by $U(y_)=\frac{3}{4} ({y_}^2-1)+\frac{1}{2}(y_* +1)$ by solving the Orr-Sommerfeld/Squire equations for two dimensional, wall-bounded, parallel shear flow, using a Matlab code [@computer_Hoepffner] and have not found any linear instability up to ${\mbox{\textrm{Re}}}=10^8$. Our velocity profile is equivalent (under the combined operations of reflection in $x$ and $y$ and a Galilean transformation) to the profile $U(y_)=\frac{3}{4} (1-{y_}^2)+\frac{1}{2}(y_* +1)$, which was shown to be linearly stable [@balakumar_finite-amplitude_1997], similarly to plane Couette or pipe flow. We present the first demonstration that the transition to turbulence in plane Couette-Poiseuille flow is subcritical in nature and occurs through localized turbulent structures (spots), similarly to other shear flows such as boundary layer, pipe, pure plane Poiseuille and Couette flows. We have measured with Particle Image Velocimetry (PIV) the flow structure inside the gap, showing that the domain is divided into high-shear (Couette) and low-shear (Poiseuille) regions of activity (see Fig. \[fig:STturb\]a). The plane Couette-Poiseuille flow has thus only one high-shear layer near the moving wall, which differentiates it from the classical symmetric plane Poiseuille flow with two high-shear regions. We have also measured the perturbation of the flow due to the passage of a spot, which is located initially in the high-shear region near the wall. As time proceeds the spot fills the whole gap. Finally, the flow in the high-shear region becomes less turbulent but the low-shear (Poiseuille) region remains active. In addition, we can observe (see Fig. \[fig:4\]a,b and Fig. \[fig:STturb\]c) that the spot moves to the right, which is the upstream direction with respect to the backflow induced by the pressure gradient (Poiseuille component). We have also investigated the ${\mbox{\textrm{Re}}}$ dependence of the size of the turbulent spot triggered by a constant, localized perturbation. Two regions of the turbulent spot can be distinguished: the active turbulent core characterized by waviness of the streaks (reminiscent of travelling waves) and the total area, which also includes the weak undisturbed streaks and oblique waves at the laminar-turbulent interface. We have shown that the area of both regions and their streamwise/spanwise extents grow monotonically with Reynolds number. Analyzing the evolution of the centroid positions, we have shown that for low Reynolds numbers the turbulent spot extends downstream. As the Reynolds number is increased the spot shifts further upstream, which means that the high-shear (Couette) region becomes increasingly dominant. Similar behaviour has been observed numerically in Poiseuille flow with zero mean flux, where the turbulent structures move with/against the direction of the Poiseuille component for low/high ${\mbox{\textrm{Re}}}$ [@tuckerman_turbulent-laminar_2014]. We isolate the right and left going waves, showing that the former are dominated by oblique waves located mainly at the tips of the turbulent spot, while the latter are related to the active core of the turbulent spot. This left-right symmetry breaking is due to the streamwise pressure gradient. This new experiment to study subcritical transition to turbulence in wall-bounded flows is capable of producing high-quality detailed information on the dynamics of turbulent spots. The only existing investigation of Couette-Poiseuille flow with zero mean velocity was reported in Ref. , however they operated only in the fully turbulent regime, with sparse spatial resolution inside the gap and without visualizations. In contrast, in our installation we have measured the streamwise fluctuations (the basic flow modifications) produced during the passage of the spot by very precise PIV measurements with high spatial resolution inside the narrow gap of the facility. In addition, the wide geometry of this experimental set up gives us a large field for clear and very high contrast flow visualizations. This has allowed us to obtain the first quantitative and systematic results on the spatial evolution of turbulent spots, marking an important advance with respect to previous experiments. We thank Matthiew Chantry and Tomasz Bobinski for fruitful discussions, as well as Arnaud Prigent for help with image processing. We also acknowledge Konrad Gumowski and Tahar Amorri for technical assistance. This work was supported by a grant, TRANSFLOW, provided by the Agence Nationale de la Recherche (ANR).	{ "pile_set_name": "ArXiv" }
Two-parameter asymptotics in the Cauchy problem for a parabolic equation S.V. Zakharov Institute of Mathematics and Mechanics,\ Ural Branch of the Russian Academy of Sciences,\ 16, S.Kovalevskaja street, 620990, Ekaterinburg, Russia Abstract. The Cauchy problem for a quasi-linear parabolic equation with a small parameter at a higher derivative is considered. The initial step-like function contains another small parameter. Formal asymptotic solutions of the problem in small parameters are constructed. Mathematics Subject Classification: 35K15, 35K59. Introduction ============ In the present work, we consider the Cauchy problem for a quasi-linear parabolic equation: $$\begin{aligned} \label{eq} \frac{\partial u}{\partial t} + \frac{\partial \varphi(u)}{\partial x} = \varepsilon \frac{\partial^2 u}{\partial x^2},& \quad t\geqslant 0, & \quad \varepsilon>0, \\ \label{ic}\phantom{\frac{1}{1}} u(x,0,\varepsilon,\rho) = \nu ( {x}{\rho}^{-1}),& \quad x\in\mathbb{R}, & \quad \rho>0. \end{aligned}$$ We assume that $\varepsilon>0,$ the function $\varphi$ is infinitely differentiable and its second derivative is strictly positive. The initial function $\nu$ is bounded and smooth. This model is used for studying the evolution of a wide class of physical systems with a small dissipation and probabilistic processes [@bu; @wh; @ib]. The interest to problem under consideration is explained by physical applications and the fact that its solutions allow one to obtain viscous generalized solutions of the limit equation. This problem had been studied by N.S. Bakhvalov, I.M. Gelfand, A.M. Il’in, E. Hopf, O.A. Ladyzhenskaya, O.A. Oleinik, and many other mathematicians. It is strictly proved [@lsu], that there exists a unique bounded infinitely differentiable with respect to $x$ and $t$ solution $u(x,t,\varepsilon)$. The aim of the present paper is to construct asymptotics solutions $u(x,t,\varepsilon,\rho)$ of problem (\[eq\])–(\[ic\]) as $\varepsilon\to 0$ and $\rho\to 0$. The structure of asymptotic series essentially depends on the relation between parameters $\varepsilon$ and $\rho$ as shown below. First, note that the change of variables $$x = \rho \sigma, \qquad t = \rho \theta$$ in equation (\[eq\]) and the initial condition (\[ic\]) leads to the following problem: $$\frac{\partial u}{\partial \theta} + \frac{\partial \varphi(u)}{\partial \sigma} = \delta \frac{\partial^2 u}{\partial \sigma^2}, \qquad \delta = \frac{\varepsilon}{\rho}, \qquad u(\sigma,0) = \nu (\sigma).$$ An asymptotic approximation of the solution of such a problem up to an arbitrary power of parameter $\delta$ is obtained directly from [@ib Ch. VI]. For example, the expansion of the solution in a neighborhood of the singular point $(x,t)=(0,\rho)$ has the form $$\sum\limits_{k=1}^{\infty} \delta^{k/4} \sum\limits_{j=0}^{k-1} w_{k,j}(\xi,\tau) \ln^j \delta,$$ where coefficients $w_{k,j}(\xi,\tau)$ depend on the inner variables, which are determined using change $$x = \varepsilon^{3/4} \rho^{1/4} \xi, \qquad t = \rho + \varepsilon^{1/2} \rho^{1/2} \tau.$$ The leading term of this expansion is found with the help of the Cole–Hopf transform: $$w_{1,0}(\xi,\tau)= - \frac{2}{\varphi''(0) \Lambda(\xi,\tau)} \frac{\partial \Lambda(\xi,\tau)}{\partial \xi},$$ $$\Lambda(\xi,\tau) = \int\limits_{-\infty}^{\infty} \exp ( -2 z^4 + z^2 \tau + z \xi) \,dz.$$ Now, let the relation between parameters ${\varepsilon}$ and ${\rho}$ be such that $$\displaystyle\frac{\rho}{\varepsilon}\to 0.$$ Assume that the function $\nu$ in (\[ic\]) satisfies the following asymptotic relations: $$\label{na} \nu(\sigma) = \sum\limits_{n=0}^{\infty} \frac{\nu^{\pm}_n}{\sigma^{n}}, \qquad \sigma\to \pm\infty.$$ In papers [@2ps; @zz], it is shown that for the solution of problem $(\ref{eq})$–$(\ref{ic})$ as $\varepsilon\to 0$, $\rho\to 0$, $\mu=\rho/\varepsilon\to 0$ in the strip $$\{ (x,t) : x\in\mathbb{R},\ 0 \leqslant t\leqslant T\}$$ there holds the asymptotic formula $${u}(x,t,\varepsilon,\rho)= h_0\left( \frac{x}{\rho}, \frac{\varepsilon t}{\rho^2}\right)- R_{0,0,0}\left( \frac{x}{2\sqrt{\varepsilon t}} \right)+ \Gamma\left( \frac{x}{\varepsilon}, \frac{t}{\varepsilon} \right) +O(\mu^{1/2}\ln\mu),$$ where $$h_0(\sigma,\omega) = \frac{1}{2\sqrt{\pi\omega}} \int\limits_{-\infty}^{+\infty} \nu(s) \exp\left[ -\frac{(\sigma-s)^2}{4\omega}\right]ds,$$ $$R_{0,0,0}(z) = \frac{\nu^{-}_{0}}{\sqrt{\pi}} \int\limits_{z}^{+\infty} \exp(-y^2)\, dy + \frac{\nu^{+}_{0}}{\sqrt{\pi}} \int\limits_{-\infty}^{z} \exp(-y^2)\, dy,$$ $\Gamma$ is the solution of the equation $$\frac{\partial \Gamma}{\partial \theta} + \frac{\partial \varphi(\Gamma)}{\partial \eta} - \frac{\partial^2 \Gamma}{\partial \eta^2} = 0$$ in the inner variables $\eta = x/\varepsilon$, $\theta = t/\varepsilon$ with the initial condition $$\Gamma(\eta,0) = \begin{cases} \nu^-_0, & \eta<0, \\ \nu^+_0, & \eta\geqslant 0. \end{cases}$$ In the present paper, formal asymptotic solutions of problem $(\ref{eq})$–$(\ref{ic})$ are constructed in the form of infinite series. The outer expansion =================== The behavior of the solution of problem $(\ref{eq})$–$(\ref{ic})$ is mainly determined by the solution of the limit problem $$\label{lcp} \frac{\partial u}{\partial t} + \frac{\partial \varphi(u)}{\partial x} =0, \qquad u(x,0) = \begin{cases} \nu^-_0, & x<0, \\ \nu^+_0, & x\geqslant 0. \end{cases}$$ For $\nu^-_0>\nu^+_0$ using the method of characteristics, we find its generalized solution $$u_{0,0}(x,t) = \begin{cases} \nu^-_0, & x<ct, \\ \nu^+_0, & x>ct, \end{cases} \qquad {c}= \frac{\varphi(\nu_0^+)-\varphi(\nu_0^-)}{\nu_0^+ -\nu_0^-}.$$ This solution is discontinuous on the line of the shock wave $x=ct$. First, let us find the outer expansion in the domain $$\Omega^+_0 = \{ (x,t)\,: \, x>ct+\varepsilon^{1-\delta_0}, \ 0<\delta_0<1\}.$$ Taking into account (\[na\]), we will construct the outer asymptotic expansion in the form of the series $$\label{outp} U^+(x,t,\varepsilon,\rho) = \nu^+_0 + \sum\limits_{m = 1}^{\infty}\sum\limits_{n=0}^{m-1} \rho^{m-n}\varepsilon^{n} u^+_{m,n}(x,t).$$ In the domain $$\Omega^-_0 = \{ (x,t)\,: \, x<ct-\varepsilon^{1-\delta_0}\}$$ we will construct an analogous series $$\label{outm} U^-(x,t,\varepsilon,\rho) = \nu^-_0 + \sum\limits_{m = 1}^{\infty}\sum\limits_{n=0}^{m-1} \rho^{m-n}\varepsilon^{n} u^-_{m,n}(x,t).$$ Formally substituting series (\[outp\]) and (\[outm\]) into equation (\[eq\]) and collecting coefficients at $\rho^{m-n}\varepsilon^{n}$, we arrive at the recurrence system of initial value problems $$\label{os} \frac{\partial u^{\pm}_{m,n}}{\partial t} + \varphi'(\nu^{\pm}_0) \frac{\partial u^{\pm}_{m,n}}{\partial x} =F^{\pm}_{m,n}, \quad u^{\pm}_{m,n}(x,0)= \delta_{n,0}\nu^{\pm}_{m} x^{-m},$$ where $\delta_{0,0}=1$, $\delta_{n,0}=0$ for $n\neq 0$, $$\label{fmnp} F^{\pm}_{m,n} = \frac{\partial^2 u^{\pm}_{m-1,n-1} }{\partial x^2} - \sum\limits_{q=2}^{m-n} \frac{\varphi^{(q)}(\nu^{\pm}_0)}{q!} \sum\limits_{\substack{i_1+\dotsc+i_q = m\\ j_1 +\dotsc+ j_q = n}} \frac{\partial \left(u_{i_1,j_1}\cdot \dotsc\cdot u_{i_q,j_q} \right) } {\partial x}.$$ Using the method of characteristics, we find the coefficients of the outer expansion: $$\label{upmn} u^{\pm}_{m,n}(x,t) = \frac{\delta_{n,0}\nu^{\pm}_{m} }{[x-\varphi'(\nu^{\pm}_0)t]^{m}} + \int\limits_{0}^{t} F^{\pm}_{m,n}(x-\varphi'(\nu^{\pm}_0)(t-t'),t') dt'.$$ Thus, for $t=0$ formal series (\[outp\]) and (\[outm\]) become asymptotic series for the initial function: $$U^{\pm}(x,0,\varepsilon,\rho)=\sum\limits_{m=0}^{\infty} \nu^\pm_m \left(\frac{\rho}{x} \right)^m.$$ Using relations (\[fmnp\]) and (\[upmn\]), by induction we arrive at the following statement. Theorem 1. [For $m\geqslant 1$ and $0\leqslant n\leqslant m-1$, there holds formula $$\label{umn} u^{\pm}_{m,n}(x,t) = \sum\limits_{s=n}^{m-1} \frac{ \alpha^{\pm}_{m,n,s}\, t^s}{[x-\varphi'(\nu^{\pm}_0)t]^{m+s}},$$ [where]{} $\alpha^{\pm}_{m,n,s}$ are constants. ]{} The inner expansion =================== We make the change of variables $$\label{cv} x = \rho\sigma, \qquad t = \frac{\rho^2}{\varepsilon} \omega.$$ Then equation (\[eq\]) becomes $$\label{ea} \frac{\partial h}{\partial \omega} - \frac{\partial^2 h}{\partial \sigma^2} = - \mu \frac{\partial \varphi(h)}{\partial \sigma},$$ where $h(\sigma,\omega)\equiv u(\rho\sigma,\rho^2 \omega/\varepsilon)$, $$\mu = \frac{\rho}{\varepsilon} \to 0.$$ We seek the inner expansion in the form of the series $$\label{ao} H(\sigma,\omega,\mu)=\sum\limits_{n=0}^{\infty} \mu^{n} h_n(\sigma,\omega),$$ for whose coefficients from equation (\[ea\]) and condition (\[ic\]) we obtain the recurrence chain of initial value problems $$\begin{aligned} \frac{\partial h_0}{\partial \omega}-\frac{\partial^2 h_0}{\partial \sigma^2} &= 0, & h_0(\sigma,0) &=\nu(\sigma), \label{ph0} \\ \frac{\partial h_1}{\partial \omega}-\frac{\partial^2 h_1}{\partial \sigma^2} &= - \frac{\partial \varphi(h_0)}{\partial \sigma}, & h_1(\sigma,0) &= 0, \label{ph1} \\ \frac{\partial h_{n}}{\partial \omega}-\frac{\partial^2 h_{n}}{\partial \sigma^2} &= - \frac{\partial E_{n}}{\partial \sigma}, & h_{n}(\sigma,0) &= 0, \label{phn} \end{aligned}$$ where $$E_{n} = \sum\limits_{q=1}^{n-1} \frac{\varphi^{(q)}(h_0)}{q!} \sum\limits_{n_1+\dotsc+n_q=n-1} \prod\limits_{p=1}^{q} h_{n_p}, \qquad n\geqslant 2.$$ To find where series (\[ao\]) makes sense and to construct asymptotics in other domains, it is necessary to know the behavior of functions $h_n(\sigma,\omega)$ at infinity. As shown in paper [@hd], the function $$\label{fh0} h_0(\sigma,\omega) = \frac{1}{2\sqrt{\pi\omega}} \int\limits_{-\infty}^{\infty} \nu(s) \exp\left[ -\frac{(\sigma-s)^2}{4\omega}\right] ds,$$ i.e., the solution of problem (\[ph0\]), has the following asymptotics as $\|\sigma\|+\omega\to \infty$: $$\label{ah0} h_0(\sigma,\omega) = R_{0,0,0}(z) + \sum\limits_{m=1}^{\infty} \omega^{-m/2} \left[ R_{0,m,0}(z) + \ln\omega R_{0,m,1}(z)\right],$$ where $$\label{r000} R_{0,0,0}(z) = \nu^{-}_{0} \mathrm{erfc}(z) + \nu^{+}_{0} \mathrm{erfc}(-z),$$ $$\mathrm{erfc} (z) = \frac{1}{\sqrt{\pi}} \int\limits_{z}^{+\infty} \exp(-y^2)\, dy, \qquad z = \frac{\sigma}{2\sqrt{\omega}},$$ $R_{0,m,0}$ and $R_{0,m,1}$ are smooth functions. The asymptotics of solutions to problems (\[ph1\])–(\[phn\]), which can be expressed in the form of convolution $$\label{hnr} h_n(\sigma,\omega) = - \int\limits_{0}^{\omega} \int\limits_{-\infty}^{\infty} \frac{1}{2\sqrt{\pi(\omega-v)}} \exp\left[ -\frac{(\sigma-s)^2}{4(\omega-v)}\right] \frac{\partial E_n}{\partial s} ds dv,$$ as $\|\sigma\|+\omega\to \infty$ are found by the same method of paper [@hd]. Proceeding by induction, one can show that the following statement is valid. Theorem 2. [ For solutions of problems $(\ref{ph0})$–$(\ref{phn}),$ which are determined recursively by formulas $(\ref{fh0})$ and $(\ref{hnr}),$ for all $n\geqslant 0$ there holds the asymptotic expansion $$\label{ahn} h_n(\sigma,\omega) = \omega^{n/2} \sum\limits_{m=0}^{\infty} \omega^{-m/2} \sum\limits_{l=0}^{m} (\ln\omega)^l R_{n,m,l}(z).$$ ]{} Acknowledgments. This work was supported by the Russian Foundation for Basic Research, project no. 14-01-00322. [99]{} J. Burgers, A Mathematical Model Illustrating the Theory of Turbulence, Advances in Applied Mechanics, Academic Press, New York, 1948. G.B. Whitham, Linear and Non-Linear Waves, Wiley-Interscience, New York, 1974. A.M. Il’in, Matching of Asymptotic Expansions of Solutions of Boundary Value Problems, Am. Math. Soc., Providence, R.I., 1991. O.A. Ladyzhenskaya, V.A. Solonnikov, and N.N. Ural’tseva, Linear and Quasilinear Equations of Parabolic Type, Am. Math. Soc., Providence, R.I., 1968. Two-parameter asymptotics in the Cauchy problem for a quasi-linear parabolic equation, Asymptotic Analysis, 63 No.1-2, (2009), 49-54. Cauchy problem for a quasilinear parabolic equation with a large initial gradient and low viscosity, Comp. Math. Math. Phys., 50 Issue 4, (2010), 665-672. S.V. Zakharov, Heat distribution in an infinite rod, Mathematical Notes, 80 (3), (2006), 366-371.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Spectroscopy of 333 NGC 6819 stars and [Gaia]{} astrometry are used to map Li evolution from the giant branch tip to 0.5 mag below the Li dip. Isochrone comparison with $[Fe/H] = -0.04$, based upon neural network spectroscopic analysis, produces an age of 2.25 (2.4) Gyr for $E(B-V)$ = 0.16 (0.14) and $(m-M)$ = 12.40 (12.29). Despite originating outside the Li dip, only 10% of single subgiants/giants have measurable Li. Above the Li dip, the limiting A(Li) for single stars is 3.2 $\pm$ 0.1 but the lower range is comparable to that found within the dip. The F-dwarf Li dip profile agrees with the Hyades/Praesepe, evolved forward. The Li level among stars populating the plateau fainter than the Li dip is A(Li) = 2.83 $\pm$ 0.16; the dispersion is larger than expected from spectroscopic error alone. Comparison of Li and $V_{ROT}$ distributions among turnoff stars in NGC 7789, NGC 2506, NGC 3680, and NGC 6819 indicates that rotational spindown from the main sequence is critical in defining the boundaries of the Li dip. For higher mass dwarfs, spindown is likewise correlated with Li depletion, creating a second dip, but at higher mass and on a longer time scale. The Li distribution among evolved stars of NGC 6819 is more representative of the older M67, where subgiant and giant stars emerge from within the Li dip, than the younger NGC 7789, where a broad range in $V_{ROT}$ among the turnoff stars likely produces a range in mass among the giants.' author: - 'Constantine P. Deliyannis' - 'Barbara J. Anthony-Twarog' - 'Donald B. Lee-Brown' - 'Bruce A. Twarog' title: 'Li Evolution and the Open Cluster NGC 6819: A Correlation Between Li Depletion and Spindown in Dwarfs More Massive Than the F-Dwarf Li-Dip' --- Introduction ============ The challenges to understanding stellar evolution are numerous. For the majority of stars change occurs at a prohibitively slow pace, necessitating comparisons among supposedly similar stars of different ages, under circumstances where the degree of similarity may be questionable and the relative ages indeterminate. Direct observation of stars is limited to the surface properties, requiring inference based upon theoretical models of the stellar interior to confirm or contradict the plausibility of the models. While asteroseismology has provided an increasingly reliable probe of the structure and evolutionary state of stars within the [Kepler]{} field [see, e.g. @ST11; @WU14; @WL14; @SI15; @HA17], the more traditional approach to assess what goes on beneath the surface has relied on detecting the changes wrought by the mixing of partially processed materials from the interior to the stellar atmosphere. Among the elements adopted for this purpose, Li has proven invaluable due its propensity for destruction above a well-defined temperature and the expectation that it will vary over time for any star where a convective or mixed envelope of sufficient depth can access the Li-depleted stellar interior. This surface signature can be altered by extending the mixing zone deeper, by extending the Li-depleted zone higher, or by the right combination of both. Because they have the potential to minimize the range of variables that need to be considered, as well as supplying more precise values for those that do, star clusters remain an ideal environment for testing all aspects of stellar evolution, not just Li variation. However, when it comes to Li evolution, comparison between theory and observation has not been kind to the models. It is strikingly clear that at least some of the physical mechanisms left out of the Standard Stellar Evolution Theory (SSET), such as rotation, diffusion, mass loss or gain, magnetic fields, etc., significantly affect the surface Li abundance. Understanding the nature of the shortcomings of the SSET can help inform us about which additional mechanisms are important. Perhaps the best success of the SSET comes from the agreement between the predicted degree of subgiant Li dilution in metal-poor stars [@DDK] and observations of field [@RY95] and cluster [@LI09] subgiants. For main sequence stars, observation supports only qualitatively the SSET prediction [@DDK; @PI97] that lower-mass stars have deeper surface convention zones (SCZs) and have thus depleted more Li. Quantitatively, the best case scenario for the SSET is the [possible]{} agreement between the models and slowly-rotating cluster dwarfs with ages less than about 150 Myr (e.g. see discussion of the Pleiades in @AT18b [@CU17; @SOP15]). However, whereas the SSET predicts that stars that are now G dwarfs depleted their Li only during the early pre-main-sequence, real open cluster G dwarfs continue to deplete their Li during the main sequence (MS) [@JE97; @SR05; @CU17]. On average, the older the star the worse the discrepancy, with the Sun being one of the most egregious offenders, having depleted a factor of 50 more Li than the factor of $\sim 3$ predicted by the SSET [@KI97; @PI97; @AS09; @TH17]. Another example is that of lower mass stars in the pre-main sequence and early MS phase, which exhibit large dispersions in Li at the same mass and age, with rapid rotators exhibiting higher Li abundances [@SOP14; @SOP15; @BO18; @AT18b]. A particularly distressing example is F dwarfs, in which deep envelope SCZs should not exist. In sharp contradistinction to the SSET, F dwarfs develop severe Li depletions during the MS near $T_{\mathrm{eff}} =$ 6600 K, a phenomenon commonly known as the Li Dip [@BT86]. Making sense of these discrepancies occupies an important role in mapping out post-MS evolution, and much progress has been made. Rotation-induced radius inflation is the leading contender to explain the large Li dispersions in young G and K dwarfs [@SOP15; @AT18b; @JDJ18]. For the non-SSET Li depletion in F and G dwarfs, a variety of evidence points to rotationally-induced mixing due to instabilities triggered by angular momentum loss [@PI90] as the dominant mechanism [@CU17]. This evidence includes the Li/Be depletion ratio [@DBS98; @BAK04] the Be/B depletion ratio [@BO05; @BO16], higher Li in Short-Period-Tidally-Locked-Binaries (SPTLB) [@RY95; @CU17], the timing of the Li depletion [@SD04], and the steepening of the Li-rotation relation with age in F dwarfs [@ST03], among others; diffusion and slow mixing due to gravity waves might also play a role. Understanding these discrepancies may also play a crucial role in cosmology. The Spite Li plateau [@SP82a; @SP82b] among the older, more metal-poor dwarfs populating the Galactic halo highlights a discrepancy of about a factor of three between the Li abundances of these stars and the inferred Big Bang Li value from Planck [@CO14], [if]{} we assume that these stars have not depleted their Li. However, until we have a better handle on diffusion, mixing, and Li-destroying processes among lower mass stars of all metallicities, such claims seem premature [e.g. @NO12; @GR13; @GR14; @GR16]. For example, although direct evidence remains elusive, rotationally-induced mixing is a very reasonable way to deplete Li in these stars by a factor of three. With the goal of using atmospheric Li to probe stellar structure and evolution among low mass stars, the authors have undertaken an extensive spectroscopic program to survey members of a key set of open clusters from the tip of the giant branch to as far down the main sequence as the technology allows. Results have been published for the clusters NGC 3680 (age = 1.75 Gyr) [@AT09], NGC 6253 (3.0 Gyr) [@AT10; @CU12], and, most recently, the metal-deficient open cluster NGC 2506 (1.85 Gyr) [@AT16; @AT18a]. The current investigation reports on the analysis of over 330 stars in the older open cluster, NGC 6819 (2.3 Gyr). The cluster is defined by a unique combination of characteristics. Its location within the [Kepler]{} field has made it the focus of asteroseismic studies reaching down the giant branch [@ST11], with a rapidly expanding literature related to the cluster and its members [@AT13; @JE13; @PL13; @YA13; @WU14; @WL14; @MI14; @LB15; @BR16; @HA17]. The age of NGC 6819 situates it in a key evolutionary phase where the red giants come from stars on the hotter, supposedly undepleted, side of the Li dip, but the turnoff stars are still in a mass range where partial degeneracy at hydrogen exhaustion slows the evolutionary rate enough to populate both the subgiant branch and the first-ascent giant branch below the red giant clump, a trait it shares with the slightly younger but metal-deficient NGC 2506 [@AT18a]. Preliminary spectroscopic analysis of the sample discussed in this investigation led to the discovery of a unique Li-rich giant fainter than the level of the clump [@AT13], below the point where standard stellar evolution models predict the initiation of mixing assumed to create Li-rich atmospheres [@CH10]. Since Li-rich stars at any location along the giant branch remain rare, at present their existence requires either a relatively specialized and restricted mixing or mass loss process or the merger of a planet of significant mass with its companion star [@CA16; @AG16]. Spectroscopic [@CA15] and asteroseismic [@HA17] evidence suggests that this Li-rich star has a substantially lower mass than other cluster members in close propinquity on the HR diagram, which might suggest a severe He-core-flash at the RGB tip as the origin of both the extra Li and the mass loss. As we will discuss below, uncertainty about its cluster membership has been eliminated by the astrometric information supplied by the [Gaia]{} DR2 release [@GA18a]. In addition to the Li-rich giant, astrometric [@PL13 hereinafter PL], photometric [@AT14 hereinafter Paper I], and spectroscopic [@LB15 hereinafter Paper II] investigations have found the cluster to have slightly subsolar metallicity, confirmed below from a new analysis using a neural network approach and by the high-dispersion spectroscopic work of @SL19, in contrast with claims of \[Fe/H\] comparable to the Hyades from earlier analysis of three red giants [@BR01], and to be affected by variable reddening. The latter discovery is relevant because traditional high dispersion spectroscopic analysis requires reliable input parameters for the models used in interpreting the spectra. Stellar temperatures, if derived from photometric indices, and surface gravities, if derived using precise estimates of the cluster distance via comparison of the observed color-magnitude-diagram (CMD) to theoretical isochrones of appropriate age and metallicity, are both dependent upon the assumed reddening. Fortunately, a neural network approach has the capability of circumventing these issues. The outline of the paper is as follows: Section 2 summarizes the spectroscopic data, discussed in detail in Paper II, and revisits the cluster membership taking the [Gaia]{} DR2 astrometry for NGC 6819 into account. Section 3 lays out the reddening corrections, age and distance estimates through comparison of the CMD to theoretical isochrones. Section 4 reanalyzes the cluster metallicity using a neural network approach to the spectroscopy, and details the parameters leading to the spectroscopic Li abundances. Section 5 explores the patterns among the NGC 6819 Li abundances for the dwarfs and giants, while Section 6 discusses the trends among Li and the rotational distributions of various clusters. Section 7 summarizes our conclusions. Spectroscopic Observations and Data Reduction ============================================= Observations ------------ NGC 6819 was an open cluster targeted for comprehensive analysis, including identification of cluster members, by the WIYN Open Cluster Study [@MA00]. Our initial spectroscopic sample of probable cluster members was constructed using the radial-velocity survey of NGC 6819 by @HO09 [hereinafter H09]. All stars brighter than $V \sim 16.75$ with radial-velocity membership probabilities greater than 50% were identified as spectroscopic candidates, while stars classed as double-lined spectroscopic binaries were eliminated. Single-lined systems were retained since the existence of the companion would have minimal impact on spectral line measurement. Stars were not eliminated based upon their position in the CMD to avoid biasing the sample against stars undergoing potentially anomalous evolution. Spectroscopic data were obtained using the WIYN 3.5-meter telescope[^1] and the Hydra multi-object spectrograph over 13 nights from September and October 2010, June 2011 and February 2013. Six configurations were designed to position fibers on a total of 333 stars. Detailed discussion of the processing and reduction of these spectra is presented in Paper II. Cluster Membership - Radial Velocities -------------------------------------- Comparison of our radial velocities ($V_{RAD}$) for 304 likely single-star members with those of H09 showed excellent agreement, with a difference of -0.27 km s$^{-1}$, in the sense $(H09 - Paper II)$, and a dispersion consistent with the predicted scatter from the individual measurements (Paper II). @MI14 updated their high-precision radial-velocity studies in NGC 6819, so we have revised the comparison of our radial velocities (Paper II) to the expanded data set with essentially the same result. From over 300 single stars common to the two surveys, our radial velocities are larger by $0.2 \pm 1.1$ km s$^{-1}$ (sd), confirming the minor offset discussed in Paper II, but an insignificant difference with respect to either the variance among the residual values or the estimated error for a single star’s radial velocity in our study, 1.1 km s$^{-1}$. As noted in Paper II, the comparison using single-lined binaries shows a dramatically larger offset and scatter, as expected. The discussion by @MI14 incorporates membership probabilities using both the proper motions of PL, who provided membership information for over 15,000 stars in this rich cluster field with the highest precision within 10 of the cluster center, and the distribution of the WOCS radial velocities. Since our sample was compiled prior to PL, it relied heavily upon the radial-velocity work of H09, with the result that of the 333 stars in our sample, only 1 has a radial-velocity membership probability below 50%. Not surprisingly, the astrometric work of PL tagged 59 of the remaining 332 stars as proper-motion nonmembers, eliminating these interlopers from the spectroscopic abundance analysis of Paper II. With the advent of [Gaia]{} [@GA16], a new level of precision has been added to the astrometric database, requiring a reevaluation of the earlier, ground-based astrometric classifications. Cluster Membership - [Gaia]{} ------------------------------- To identify potential cluster members within NGC 6819, we will follow the simple prescription adopted in @AT18b. While [Gaia]{} DR2 astrometry has been used by @CA18 to select highly probable astrometric members, only a handful of these stars have [Gaia]{} DR2 radial velocities, and a subset of our stars have poor to inadequate [Gaia]{} DR2 astrometric measures. Stars will be classified initially as proper-motion members or nonmembers based upon their position within the proper-motion vector-point diagram, taking into account the dispersion among the cluster members and the individual uncertainty in the measured proper motions. From the identified members, a second check is made using the derived cluster parallax, eliminating those stars which deviate from the cluster mean value by more than three times the quoted uncertainty in the parallax. As seen below, this simple approach is more than adequate for our current needs. Cross-matching our spectroscopic sample with the [Gaia]{} DR2 catalog, all stars were found but four retained only coordinate positions and no astrometric information (5006, 10010, 14002, 16005; all numbers refer to the ID on the WOCS system, HO09). For these stars, we have defaulted to the membership classification from the radial-velocity and proper-motion probabilities compiled by @MI14; all are probable members. As a first cut on the [Gaia]{} DR2 sample, the quoted uncertainties in the positions for each of the 329 stars were combined to identify stars where the astrometry was likely to be unreliable since the positional errors invariably translated into large uncertainties in the proper motion and parallax. Nine stars (4008, 7004, 11014, 13007, 15002, 22005, 22020, 47007, 49023) were found to have combined positional errors above 0.1 mas. From PL, 7004, 15002 and 22005 are nonmembers and will be eliminated. Of the 6 remaining members (PL), three (4008, 13007, 22020) are SB1 binaries [@MI14]. To define the cluster reference motion and parallax, the [Gaia]{} DR2 cross-match was restricted to 190 stars with both high-precision DR2 proper motions and PL probabilities above 90%. Mean cluster values of the proper motion in both right ascension and declination were determined and the radial vector distance of each star from the cluster mean calculated. Eight stars with total proper motion placing them more than 0.30 mas yr$^{-1}$ away from the cluster proper-motion vector point were removed and the centroid rederived. The adopted cluster proper-motion center is -2.9159 $\pm$ 0.0072 (sem) mas yr$^{-1}$ and -3.8584 $\pm$ 0.0074 (sem) mas yr$^{-1}$ in right ascension and declination, respectively. If we then restrict our parallax sample to only 123 stars within 0.15 mas yr$^{-1}$ of the cluster proper motion, the mean cluster parallax becomes 0.3552 $\pm$ 0.0025 (sem) mas. As expected, these are all in excellent agreement with the values derived by @CA18 from 1589 probable (above 50%) members from astrometry alone. These will be adopted for the cluster in the discussion which follows. We note that there is significant evidence that the [Gaia]{} DR2 parallaxes suffer from a zero-point error, leading to an underestimate of the parallax and overestimate of the distance, a point we will return to in detail in Section 3. Returning again to the 320 stars with reliable [Gaia]{} DR2 coordinates, 26 stars with proper-motion vectors placing them more than 0.6 mas yr$^{-1}$ away from the cluster motion were classed as nonmembers. Of the 26, 19 have $\Delta$$\pi$/$\sigma$$_\pi$ $>$ 3, where $\Delta$$\pi$ is defined as the absolute value of ($\pi_{clus}$ - $\pi_{star}$), confirming in part the field-star classification. Of the 26, only 2 had PL proper-motion membership above 31% and and 20 had 0% probability. Among the proper-motion distribution, 17 stars have proper-motion vectors which place them radially between 0.3 and 0.6 mas yr$^{-1}$ away from the cluster motion. Only 2 of these stars have $\Delta$$\pi$/$\sigma$$_\pi$ $>$ 3. Of the 17, PL assigns proper-motion membership at 98% or above for 12. The remaining 5 are at 14% or less, including both stars which are probable parallax nonmembers; only the the parallax nonmembers will be excluded from our final sample. Of the remaining 277 stars with proper-motion vectors within a radial vector distance of 0.3 mas yr$^{-1}$ from the cluster’s motion, only 7 have $\Delta$$\pi$/$\sigma$$_\pi$ $>$ 3; these will be excluded from the membership list. This last set of 270 members includes 25 stars with ground-based proper-motion probabilities below 50% and 206 with probabilities at or above 90% (PL); 20 stars are SB1 binaries. The full set of probable members sits at 295, of which 26 are SB1. Table 1 contains a complete listing of the membership classification for each star based upon the [Gaia]{} DR2 data, detailing if it is consistent with membership via proper motion, parallax, or both. Only stars meeting both criteria will be treated as probable cluster members in the discussions which follow. WOCS 7017 - Li-Rich Giant ------------------------- @CA15 and @HA17 have discussed the evidence for and against the membership of the Li-rich giant, 7017, in NGC 6819 with admirable thoroughness. As suggested by @AT13, the large errors associated with PL’s proper motion for this fascinating star implied that the older, more positive estimation of membership probability provided by @S72 should keep 7017 on the table as a possible cluster member. @CA15 strengthened that case by providing an explanation for the atypically large error in PL’s proper motion, as well as bolstering the membership credentials of 7017 through an independent spectroscopic estimation of the star’s gravity and abundance pattern. An intriguing feature of their discussion of this Li-rich giant was the derivation of a spectroscopically-based, anomalously low mass, possibly related to its surface Li abundance, but confirming the structural uniqueness of this red clump giant demonstrated by asteroseismology [@ST11; @HA17]. The analysis above places 7017 in the category of definite member from both proper motion and parallax. Stellar Properties: Variable Reddening, Isochrone Ages, and Distance Moduli =========================================================================== One important challenge for any investigation of the stellar properties in NGC 6819 is the variable reddening across the face of the cluster, as demonstrated by PL. The sense and amplitude of the variable reddening were validated in Paper I; Fig. 10 of that paper showed the tightening of the Strömgren CMD that results from applying broadly-derived spatial reddening estimates to colors and magnitudes, while Fig. 11 used broadband photometry [@RV98] and Yale-Yonsei [$Y^2$; @DE04] isochrones for the determination of age and distance. For Figs. 10 and 11 of Paper I, individual deviations from the average foreground reddening were determined for three spatial zones, with appropriate adjustments to the photometric indices, e.g. individual $(B-V)$ colors were corrected by an amount equal to $\delta$$E_{(B-V)}$, defined as $E(B-V)_{star} - 0.16$, with a corresponding adjustment of $3.1 \delta$$E_{(B-V)}$ to the $V$ magnitudes, placing all stars under a uniform reddening of $E(B-V)$ = 0.16, the cluster mean as derived from extended Strömgren photometry of stars at the turnoff (Paper I). For future reference, $(B-V)'$ will denote $(B-V)-$$\delta$$E_{(B-V)}$ and $V'$ will represent $V - 3.1$$\delta$$E_{(B-V)}$. The $V$ and $(B-V)$ data for the spectroscopic sample are the same as compiled and discussed in Paper II. For the additional members used to define the CMD, especially below the magnitude limit of the spectroscopy, $V$ magnitudes are taken from Paper I, while the initial $(B-V)$ indices are those of @RV98, the same system used as the standard for the merger of the multiple color sources in Paper II. Isochronal Constraints ---------------------- Fig. 1 of Paper II emphasized the CMD locations for the spectroscopic sample, with different symbols indicating membership and binarity from H09. Fig. 1 of the current paper provides a dramatic update, making use of the radial-velocity membership determinations from @MI14 and the astrometric analysis of Section 2. Large symbols designate the stars of Table 1 with blue open symbols, black crosses, red solid squares, and black solid squares denoting probable single-star members, nonmembers, member binaries, and nonmember binaries, respectively. For each star in the spectroscopic sample, individual reddening estimates were derived from the reddening map of PL, with values listed in Table 1 of Paper II. To extend the sample beyond the depth limits of our spectroscopy while minimizing the impact of the variable reddening, stars within 6 of the cluster center were processed through the same [Gaia]{} DR2 astrometric procedure as the spectroscopic sample. Because of increasing astrometric errors at fainter magnitudes, there is an artificial decline in the number of stars retained as members, with a cutoff near $V$ $\sim$ 17.5. Those stars bright enough to be included in @MI14 were checked and all stars classed as nonmembers or uncertain members were eliminated. All stars were individually reddening corrected and plotted as small open triangles in Fig. 1; no distinction is made between single stars or binaries. We emphasize that the membership selection is optimized to delineate the cluster CMD sequence, so eliminating nonmembers is a clear priority over completeness. For the comparison of Fig. 1, we have adopted the isochrones from @VR06 [hereinafter VR], constructed for an abundance of $[Fe/H] = -0.04$, rather than the $Y^2$ models of @DE04 from Papers I and II. The adopted $[Fe/H]$ is the quadratic sum of results from Papers I and II, but has been confirmed through the use of a neural network abundance analysis applied to the high-resolution spectra which form the basis of the current investigation, as discussed in Section 4. Because of its location within the [Kepler]{} field, the NGC 6819 CMD, age, and distance modulus have garnered significantly more discussion than most clusters. While a comprehensive overview of the cluster properties is neither necessary nor desirable for our purposes, a few points regarding the CMD match should be made. While use has been made of the [Gaia]{} DR2 parallaxes for individual stars in weighing the probability of cluster membership, the success of this technique is almost wholly dependent upon the relative precision of the astrometric measures and tells us little, if anything, about potential systematics in the parallax scale. If $(m-M)$ = 12.40 and $A_V$ = 0.5, then $(m-M)_o$ = 11.90, $d$ = 2400 pc, and $\pi$ = 0.417 mas, noticeably larger than obtained above (0.355 mas) using a simple average of highly probable members. A more elaborate approach as illustrated by @CA18 derives the identical result within the errors, $\pi$ = 0.356 mas. However, as discussed by @ST18 [@ZI18] and @RI18, there is growing evidence for a zero-point offset to the [Gaia]{} DR2 parallax scale at the level of 0.05 to 0.08 mas, with the offset size potentially dependent upon position in the sky. @CA18, through comparison with cluster distances derived by the BOCCE project [@BR06], find a typical systematic offset to the [Gaia]{} DR2 parallaxes equal to -0.05 mas. Applying this to the data for NGC 6819 produces $\pi$ = 0.405 mas, in excellent agreement with the main sequence fit given the uncertainty in the parallax zero-point. Keeping in mind that the detailed intermediate-band analysis of the cluster (Paper I) defines $E(B-V)$ = 0.16 as the cluster mean reddening, we have compiled two alternative matches defined by lower reddening to illustrate the trend. The three isochrones shown in Fig. 1 have been adjusted for an optimum fit to the turnoff and unevolved main sequence under the assumption that the reddening is $E(B-V)$ = 0.16 (blue curve), 0.14 (green curve), or 0.12 (red curve). We have incremented the VR $(B-V)$ colors by +0.01, in conformity with our past usage of isochrones zeroed to a solar color of $(B-V)$ = 0.65 at an age of 4.6 Gyrs [e.g. @TW09; @RA12b]. The apparent moduli have been set to assure identical fits to the observed main sequence at $(B-V)' = 0.85$. As expected, the highest reddening leads to the youngest age (2.25 Gyr vs. 2.55 Gyr) and a larger apparent distance modulus (12.40 vs. 12.18). While the subgiants appear too faint relative to the models, independent of the adopted reddening, the isochrones of higher reddening nicely bracket the giant branch from the base to above the level of the clump. We note that the fit of the VR isochrone to the photometry is essentially identical to that presented in Paper I, in which $Y^2$ isochrones were used; with $E(B-V)$ = 0.16 and $[Fe/H] = -0.06$, an age of 2.3 to 2.5 Gyr was derived for an apparent modulus of 12.4. Use of a higher adopted metallicity for the cluster would require an even younger isochrone and a larger distance modulus. This prediction can be tested using the analysis of multiple isochrone sets and NGC 6819 in Fig. 10 of @JE13, who assumed $E(B-V)$ = 0.12, $(m-M)$ = 12.3 and $[Fe/H]$ = +0.09; if an isochrone of the exact metallicity was not available, the one closest to +0.09 was selected. For VR, the closest match was $[Fe/H]$ = +0.13. With their lower reddening (0.12) relative to 0.16, our apparent modulus from Fig. 1 is 12.18; partial compensation comes from a metallicity higher by 0.17 dex, which should boost the modulus by $\sim$0.17 mag [@TW09], leading to a final value of $(m-M)$ = 12.35, the same within the uncertainties as adopted by @JE13. In agreement with our Fig. 1, the color of their turnoff best matches their age of 2.25 Gyr at the lower reddening due to the higher metallicity, while the subgiants appear fainter than their models at this age. Note also that their unevolved main sequence lies increasingly above the cluster photometry as one moves down the main sequence, while the isochrone fit in Fig. 1 remains consistently within or at the lower edge of the distribution at fainter magnitudes. This difference reflects the changing slope of the main sequence with changing metallicity. By contrast, the $Y^2$ match [@JE13] indicates an age midway between 2.25 and 2.5 Gyr, with a good fit from the lower main sequence through to the subgiant branch. The result should be the same for higher reddening and lower metallicity, with the expectation that the unevolved main sequence models should lie increasingly fainter than the photometry at fainter magnitudes, as confirmed in Fig. 11 of Paper I. For completeness, the BaSTI isochrones [@PI04] with $[Fe/H] = +0.06$ supply a good match from the unevolved main sequence through the subgiant branch for an age of 2.25 Gyr, while the DSEP isochrones [@DO08] imply an age between 2.75 and 3.0 Gyr from the turnoff, with the subgiant models too faint compared to the photometry. For isochrone comparisons based upon $VI$ rather than $BV$, the reader is referred to Fig. 8 of @BR16. As is obvious, age and distance estimates through isochrone fitting depend strongly upon the adopted reddening and metallicity, as well as the choice of isochrones. For fixed metallicity and reddening, the latter becomes the dominant source of scatter among investigations. As noted earlier, we have consistently attempted to minimize the impact of different approaches to stellar models, to different transformations of the models from the theoretical to the observational plane, and to differences in the assumed solar composition by requiring that a one-solar-mass star with $[Fe/H] = 0.0$ have a specific $B-V$ and $M_{V}$ at an age of 4.6 Gyr. Such simple zero-point offsets become less reliable as $[Fe/H]$ deviates from solar but, as exemplified by the comparison of the results from the $Y^2$ and VR isochrones for NGC 6819, they can lead to greater consistency in both age and distance. Eclipsing Binary Constraints ---------------------------- An alternative which minimizes the role of metallicity and the adopted isochrones is the use of eclipsing binaries, as detailed in the exquisite analysis of three systems in NGC 6819 by @BR16, an expansion and revision of the earlier work by @JE13 and @SA13. With radius and $T_{\mathrm{eff}}$ known from the binary analysis and photometric temperatures, one can derive $M_{V}$ with minimal impact due to reddening and/or metallicity uncertainty. The weighted average in $(m-M)$ is 12.38 $\pm$ 0.04, where the quoted uncertainty is the error in the mean, in excellent agreement with the original estimate from Paper I and the VR fit in Fig. 1 for $E(B-V)$ = 0.16. Equally important, with the masses and radii known, @BR16 derive an age for NGC 6819 through an extensive set of isochrone and model comparisons. To account for the possibility that the metallicity could be either approximately solar or higher by $\sim$0.1 dex, the analysis is done under two different assumptions for the composition. A key difference in their approach is the method by which they determine the choice of isochrones. To avoid concerns about the assumed Z$_{\sun}$ for each isochrone source, the isochrones selected from each set are picked to have Z = 0.012 and 0.015, irrespective of the adopted Z$_{\sun}$ for the models. The isochrones are normalized, independent of distance and reddening, by forcing a match between the observed position of 24009 C in the CMD and a star of the same mass at a given age as defined by the isochrones. In effect, 24009 C takes the equivalent role of the sun in fixing the position of the isochrones to the CMD. The age of the cluster can then be set by mapping how well the other members of the binary systems and the evolved stars at the top of the turnoff match the predicted position of the isochrones. Using this CMD-based approach, @BR16 derive an age of 2.21 $\pm$ 0.10 $\pm$ 0.20 Gyr for the cluster, in excellent agreement with the fit in Fig. 1. The agreement is relevant because the adopted isochrone match will be used to define the relationship between the stars’ positions within the CMD and the predicted masses used to delineate the trends of Li with age and mass, as discussed below. However, the challenges posed by the differences in the construction of theoretical isochrones and their transformation to the observational plane can be seen in the range of values obtained by using the forced match of the isochrones to the CMD to derive $E(B-V)$ and $(m-M)$. For Z = 0.015, @BR16 find $E(B-V)$ between 0.19 and 0.22 and $(m-M)$ between 12.46 and 12.57; for Z = 0.012, both $E(B-V)$ and $(m-M)$ are systematically larger by 0.03 and 0.1, respectively. Returning to the distribution of stars on the CMD, the inclusion of the astrometric constraints imposed by [Gaia]{} DR2 has significantly reduced the scatter in the CMD, both among the spectroscopic sample and the fainter main sequence. Of the spectroscopically observed stars between the base of the giant branch and the clump, 7 stars which scatter away from the mean relation are eliminated though, as noted earlier, the anomalous giant, 7017, is now a definite member. Only three remaining member stars scatter blueward of the giant branch and two of these are binaries. The third star, 8005, is a definite astrometric, single-star member, but its radial velocity places it at 59% membership probability. Comparison of the radial velocities from @MI14 and Paper II shows virtually identical values, consistent with a lack of variability and increasing the likelihood that the velocity deviation of 8005 from the cluster mean is real. At the turnoff region, keeping in mind that double-lined binaries are excluded, the select sample of triangles nicely illustrates the location of the binary sequence as expected for systems with two identical stars. This sequence crosses the evolving turnoff at $V$ $\sim$ 15.5. It is potentially significant to note that all but three of the numerous single-lined binaries at the turnoff lie at or above this location. Spectroscopic Abundances ======================== Metallicity Estimation: ANNA ---------------------------- As mentioned above, metallicity plays a major role in the estimation of the key cluster parameters of age and distance modulus, thereby impacting the specific values derived for individual stellar masses, evolutionary phases, and, more directly, equivalent width corrections for the Fe line located near Li in the determination of the Li abundances among the cooler stars. As an alternative to our EW-based spectroscopic \[Fe/H\] estimates and our photometric $T_{\mathrm{eff}}$ values, we have attempted to derive \[Fe/H\] and $T_{\mathrm{eff}}$ for each cluster member in our sample using ANNA [@LB17; @LB18a; @LB18b], a new, flexible, Python-based code for automated stellar parameterization. ANNA utilizes a feed-forward, convolutional neural network [@AR02] trained on synthetic spectra, a machine-learning technique, to infer stellar parameters of interest from input spectra. Multiple tests show that ANNA is capable of producing accurate metallicity estimates with precision competitive with our EW-based analysis. Additionally, ANNA is capable of accurately inferring $T_{\mathrm{eff}}$ from our spectra alone, providing an alternate temperature determination for each star. A deep discussion of ANNA’s design and capabilities can be found in @LB18a, but a short summary of its operation can be found in @AT18a. ANNA is freely available for download; the version of ANNA used in this investigation can be found at Zenodo [@LB17], while the current version of the current version of the code can be found at GitHub.[^2] As a starting point for deriving the cluster metallicity and for comparing the effectiveness of ANNA relative to the traditional EW technique of Paper II, all stars classed as nonmembers and/or binaries in Table 1 were removed from the sample. This cut reduced the sample to 268 stars, higher than the 251 stars of Paper II restricted using ground-based astrometry. For future reference, the average \[Fe/H\] for all these stars using Paper II abundances is \[Fe/H\] = -0.038 $\pm$ 0.104 (sd). If, as in Paper II, the four reddest/coolest stars with $(B-V)_o \geq 1.35$ are removed, the remaining 264 stars have \[Fe/H\] = -0.033 $\pm$ 0.091 (sd). By comparison, ANNA generated a mean \[Fe/H\] = -0.049 $\pm$ 0.099 (sd) from 267 stars; star 13002 failed to converge to a coherent solution and was dropped from the analysis. Removal of the 4 coolest stars has a negligible impact upon the average or the dispersion. In fact, the only exclusion from the sample which has any impact on the the average is the removal of the two stars with the most deviant abundances, 12002 and 35008 at \[Fe/H\] = 0.67 and 0.38, respectively. For the remaining 265 stars of all colors, \[Fe/H\] = -0.053 $\pm$ 0.085 (sd). It should be noted that this is an improvement over the pattern identified in the metal-poor cluster, NGC 2506, where, with a limited wavelength range ($\sim$400 Å) and $R \sim$ 13000, ANNA’s reliability declined noticeably for hotter stars. The reduction in spectral features at higher $T_{\mathrm{eff}}$ minimized the sensitivity to changes in \[Fe/H\] and $T_{\mathrm{eff}}$; for metal-rich stars at the very cool $T_{\mathrm{eff}}$ end, the reverse issue, too rich a spectroscopic palette, can reduce the applicability of the code. Consisent with ANNA analysis of stars in NGC 2506 [@AT18a], the stellar parameters generated by ANNA, \[Fe/H\], $T_{\mathrm{eff}}$, log $g$, and $v_t$, exhibit increasing scatter compared to values from alternative. This trend is interpreted as an indicator that the parameters which carry the most weight in defining the final optimal match to the observed spectra follow a similar order, i.e. the dominant parameter in constraining the neural network is the metallicity while the microturbulent velocity is the least impactful. As illustrated in Fig. 2, the ANNA abundances show no trend with $T_{\mathrm{eff}}$. The larger scatter among the hotter stars is defined in part by a lower signal-to-noise ratio (S/N) among a few of the stars at the turnoff, coupled with the weaker line spectrum within increasing $T_{\mathrm{eff}}$. It is important to note, however, that, as shown in Fig. 3, the derived mean \[Fe/H\] remains independent of the S/N, as was also found for the traditional metallicity analysis of Paper II. The two anomalous points in both figures are the aforementioned 12002 and 35008. The ANNA-based \[Fe/H\] values for all members are listed within Table 1. Temperature: ANNA ----------------- The estimation of $T_{\mathrm{eff}}$ in Paper II came from reddening-corrected $(B-V)$ indices, with separate calibration relations for turnoff stars [@DE02] and red giants [@RA05], blended to supply a smooth transition across the subgiant branch, a procedure adopted in our earlier investigations [@AT09; @CU12]. With $T_{\mathrm{eff}}$ from ANNA, we can first check if the derived values exhibit a plausible correlation with increasing color index, noting that with the exception of added scatter due to the additional correction caused by variable reddening, the precision of the $(B-V)$ indices is high enough that the typical scatter in $T_{\mathrm{eff}}$ from photometric errors alone should be just under $\pm$80 K; for the giants, the comparable number is $\pm$50 K. Scatter caused by uncertainties in the reddening could easily double these estimates. Fig. 4 shows the trend of ANNA $T_{\mathrm{eff}}$ as a function of reddening-corrected $(B-V)$. The trend of decreasing $T_{\mathrm{eff}}$ with increasing $(B-V)_o$ is obvious, but the relation loses sensitivity for giants redder than $(B-V)_o$ = 1.2 or $\sim$4200 K. This is consistent with the $T_{\mathrm{eff}}$ pattern found for NGC 2506 [@AT18a] and is tied to the increasing complexity of the line structure in the spectra for the coolest giants over too restricted a range in bandpass. At the hotter end, there is a trend of increased $T_{\mathrm{eff}}$ with decreasing $(B-V)_o$, with an asymmetric scatter toward lower $T_{\mathrm{eff}}$ at a given $(B-V)_o$. A quick glance at Fig.1 suggests a possible explanation for the scatter. At the turnoff of the CMD, stars in the $(B-V)$ = 0.6 to 0.7 ($(B-V)_o$ = 0.44 to 0.54) range are a mixed population of moderately unevolved stars, stars within the red hook prior to the hydrogen-exhaustion-phase, and even brighter subgiants. Unlike the usual photometrically-defined $T_{\mathrm{eff}}$, where stars at the turnoff with the same color but slightly different evolutionary phases are treated identically, ANNA has the option of modifying the surface gravity and $T_{\mathrm{eff}}$ to optimize the match of the true spectrum to a synthetic analog. Unfortunately, the stars which define the low $T_{\mathrm{eff}}$ extension do not fall within a specific category of evolutionary phase at a given $(B-V)_o$. It should be noted, however, that the typical log $g$ as derived by ANNA for the excessively cool stars is lower on average by approximately 0.2 dex than that for the stars that fall upon the mean relation, despite a similar distribution in $V$. If we exclude the 10 stars with the largest discrepancies in $T_{\mathrm{eff}}$ between the ANNA values and the photometric estimates, the mean offset, in the sense (PHOT - ANNA) is +117 $\pm$ 144 K for 257 giants and dwarfs. As an external check, perhaps the best comparison among multiple sources, at least for the cooler giants, comes from the data of @HA17, who derived $T_{\mathrm{eff}}$ from $(V-K)$, adopting a fixed value of $E(B-V)$ = 0.15 for all stars, and the color-$T_{\mathrm{eff}}$ relation of @CV14. Comparison of their $T_{\mathrm{eff}}$ to that from the $(b-y)$ of @CA14 using the color-$T_{\mathrm{eff}}$ relation of @RA05, to the $T_{\mathrm{eff}}$ from @CA14, to the photometric $T_{\mathrm{eff}}$ of Paper II, and to the $T_{\mathrm{eff}}$ from APOGEE [@PI14] produces offsets, in the sense (HA17 - LIT), of -48 $\pm$ 51 K from 52 giants, +14 $\pm$ 35 from 52 giants, +35 $\pm$ 41 from 50 giants, and +55 $\pm$ 48 from 30 giants, respectively. It should be noted that the residuals between the @HA17 and Paper II show a clear trend with $E(B-V)$ in that stars with larger $E(B-V)$ corrections show a smaller residual $T_{\mathrm{eff}}$ than stars with smaller $E(B-V)$; adjustment for this effect reduces the scatter among the residuals to $\pm$33 K. For the current discussion, however, the agreement among the multiple modes of deriving $T_{\mathrm{eff}}$ for the giants gives us some encouragement that the original $T_{\mathrm{eff}}$ scale of Paper II for the giants is a more reliable representation of the true system for evaluating the Li abundances. Since the Li abundance is being derived using the same EW-based approached as the \[Fe/H\] determination of Paper II and the ANNA abundances for Fe are in excellent agreement with those of Paper II, we will retain the color-based $T_{\mathrm{eff}}$ of Paper II for both the giants and dwarfs throughout the remainder of this Li analysis. Spectroscopic Analysis: Lithium Abundances ------------------------------------------ We employ two methods to estimate the abundance of Li, a curve-of-growth-based computation using a direct measurement of the equivalent width of the line at 6708 Å and spectrum synthesis. As the latter technique is particularly helpful for cooler stars with more blended spectra, spectrum synthesis was employed for all stars with $(B-V) \geq 0.70$, making use of the 2010 version of the MOOG software suite [@SN73] and a line list used and discussed in @CU12. In spectrum synthesis, a model spectrum is constructed for each star, employing specifically chosen $T_{\mathrm{eff}}$, log $g$ and microturbulent velocity parameters as input; values for these parameters for each star are identical to those used for spectroscopic analysis described in Paper II and are found in Table 1 of that paper. Examples of spectra illustrating the region near the Li line may be found in Fig. 1 of Paper I. For the warmer stars, equivalent widths were measured using the SPLOT utility within the IRAF spectroscopic data reduction packages [^3]. We use each star’s temperature and the cluster iron abundance to numerically deblend the nearby Fe I line at 6707.4 Å from the Li line at 6707.8 Å , then use the “corrected” equivalent width, the star’s $T_{\mathrm{eff}}$ and a grid of curve-of-growth abundances developed by @ST03 from MOOG model atmospheres and employed by @SD04. The spectrograph pixel-wavelength scale, measured Gaussian full-width of the line, and the S/N per pixel are used to compute an equivalent width error for each star, utilizing a prescription originally proposed by @CA88 and reformulated by @DP93. For a significant detection of Li, we require that the Li equivalent width, following subtraction of the Fe I contribution, be at least three times the estimated error in the equivalent width. Table 1 includes a summary of our Li abundances for all stars and the final errors in the abundance for stars with measurable Li. Photometric $T_{\mathrm{eff}}$, surface gravities, microturbulent velocities, and rotational velocities for all stars are contained in Paper II and will not be repeated. [rccccrrrrr]{} 1002 & Y & Y & N & Y & 57 & 3967 & -0.28 & 0.30 & 0.08\ 1004 & Y & Y & N & Y & 66 & 4242 & -0.06 & -1.20 & 0.00\ 1007 & Y & Y & N & Y & 53 & 4231 & -0.13 & -1.50 & 0.00\ 1014 & Y & Y & N & Y & 57 & 4270 & -0.02 & -1.50 & 0.00\ 1016 & Y & Y & N & Y & 19 & 4119 & 0.06 & &\ 2001 & Y & N & N & N & & & & &\ 2003 & Y & Y & N & Y & 60 & 4676 & -0.05 & 0.60 & 0.00\ 2004 & Y & Y & N & Y & 100 & 4372 & -0.04 & -1.00 & 0.00\ 2006 & Y & Y & N & Y & 72 & 4594 & -0.04 & 0.35 & 0.00\ 2007 & Y & Y & N & Y & 76 & 4558 & -0.01 & 0.35 & 0.00\ & & & & & & & & &\ 2012 & Y & Y & SB1 & Y & & & & -0.50 & 0.00\ 2016 & Y & N & N & N & & & & &\ 3001 & Y & Y & N & Y & 129 & 4496 & -0.11 & 0.35 & 0.00\ 3003 & Y & Y & N & Y & 44 & 4506 & -0.24 & 0.00 & 0.00\ 3004 & Y & Y & N & Y & 113 & 4489 & 0.02 & 0.35 & 0.00\ 3005 & Y & Y & N & Y & 63 & 4459 & 0.00 & 0.35 & 0.00\ 3007 & Y & Y & N & Y & 74 & 4596 & 0.01 & 0.35 & 0.00\ 3009 & Y & Y & N & Y & 139 & 4481 & -0.06 & 0.35 & 0.00\ 3021 & Y & Y & N & Y & 98 & 4598 & -0.03 & 0.30 & 0.00\ 4001 & Y & Y & N & Y & 85 & 4497 & -0.01 & 0.30 & 0.00\ Li Patterns =========== Derived values of A(Li) are shown in Figs. 5 through 10 for stars other than those designated as nonmembers. We now discuss the apparent trends among the stars as a function of their evolutionary phase. Li: The Turnoff Region ---------------------- Fig. 5 shows the trend of Li with reddening-corrected $V$-magnitude for all stars bluer than $(B-V)_o$ = 0.54 in Fig. 1. Fig. 6 illustrates the same sample plotted as a function of $(B-V)_0$. The adopted mean reddening for these two figures is $E(B-V)$ = 0.16 and $A_V$ = 0.50. The first feature of importance is the limiting A(Li) for stars on the brighter or higher mass side of the Li dip. All single-star members have A(Li) = 3.35 or less. Of the six stars with A(Li) above 3.2, only two are single and the remaining four are classed as single-lined binaries (red and green points). The five stars that are situated both brighter and bluer than the cluster isochrone turnoff, supposedly blue stragglers of potentially binary origin, all exhibit upper limits to A(Li) of 2.5 or less. One star, 10010, with measurable A(Li) at 3.19 is located in the CMD at the tip of the isochrone blue hook for single stars about to enter the subgiant branch. A key property which makes NGC 6819 invaluable for probing the nature of the Li dip is its age. While older clusters such as NGC 6253, M67, and NGC 188 [@CU12; @PA12; @RA03] have been studied in the mass range populating the Li dip, they are too old for mapping the high-mass edge of the distribution. The stars at the top of the turnoff feeding the subgiant and giant branches in these clusters come from the Li dip itself, making the exact boundary impossible to define and ensuring that the giants exhibit no significant Li abundance, even before convection and mixing kick in. The value of the stars more massive than the Li dip stems from the often-made assumption and prediction from SSET that if any stars within a cluster retain the signature of the primordial cluster abundance, it should be these stars. However, near 7000 K, upwards diffusion may be enriching the surfaces of slower rotators [@RM93], as evidenced by the super-Li-rich dwarf J37 in the Hyades-aged cluster, NGC 6633 [@DE02]. Among the clusters of intermediate age studied to date, NGC 752, NGC 3680, IC 4651 [@AT09], NGC 2506 [@AT18a], NGC 7789 [@AT19 in prep.], and now NGC 6819, there is evidence in each cluster for a limiting A(Li) value typically between A(Li) = 3.2 and 3.35, consistent with the primordial solar system value of A(Li) = 3.30 [@AG89]. However, it is also true that every cluster exhibits a range of A(Li) which often extends to 2.8 or lower. Since the normal stars within the hydrogen exhaustion phase at the top of the turnoff and beyond have evolved off the main sequence en route to their current locations, it is perhaps unsurprising that some of the brighter turnoff stars have A(Li) well below their cluster limit. What is surprising, however, is the changing fraction of turnoff stars which fall below the given cluster limit, depending upon the age. For NGC 6819, among stars in the red hook at essentially identical magnitudes and supposedly similar evolutionary phase, A(Li) can range from the detection limit of $\sim$ 3.2 to an upper limit of less than 1.6. We will return to this issue in Section 6. Moving down the turnoff toward fainter $V$, the next striking feature is the sharp transition from detectable A(Li) near 3.2 to stars with detections or upper limits below 2.3. This edge occurs over a magnitude range of less than 0.1 mag near $V_0$ $\sim$ 14.75, a range comparable to the combined photometric and reddening uncertainties alone. The Li dip remains deep to $V_0$ $\sim$ 15.25, where a more gradual rise in detectable A(Li) begins, plateauing to a fixed value near $V_0$ $\sim$ 15.45. For stars fainter than this edge, the degree and range of evolution off the ZAMS should be significantly less than for the stars on the bright side of the Li dip. Despite this, if we bin the stars fainter than the dip by $V_0$ using bins 0.1 mag wide between $V_0$ = 15.50 and 15.90, excluding binaries and upper-limits, the mean A(Li) for the four bins is 2.83, with a dispersion among the bin averages of only 0.05 dex. By contrast, the dispersion within each bin ranges from 0.15 to 0.22 dex, with a dispersion among all stars between 15.5 and 15.9 of 0.17 dex; the predicted dispersion from the spectroscopic errors is $\pm$0.05 dex. We conclude that the abundance scatter among the stars on the less evolved portion of the main sequence is real. If the Li dip profile in $V$ is symmetric [@CU17], the center is located near $V_0$ = 15.1 $\pm$ 0.1 mag or $M_V$ = 3.2 for $E(B-V)$ = 0.16; the analogous numbers for $E(B-V)$ = 0.14 and 0.12 are $M_V$ = 3.3 and 3.4, respectively. Determination of the exact profile of the dip is challenging since the majority of Li measures at its center are defined only by upper limits and the rise to a Li plateau among the lower mass stars appears more gradual than the sharp boundary among the higher mass stars. Using the isochrones of Fig. 1, the mass at the center of the symmetric dip is 1.348 $\pm$ 0.025 $M_{\sun}$; for the lower reddening values the masses are 1.317 and 1.288 $M_{\sun}$, respectively. Under the assumption that the profile of the Li dip in NGC 6819 has the same shape as the more asymmetric curve found in the Hyades and Praesepe by @AT09 evolved to the age of NGC 6819, the A(Li) minimum shifts to a mass of 1.36 $\pm$ 0.02 $M_{\sun}$. A mapping of the reddening-corrected $V$ magnitude to initial mass for $E(B-V)$ = 0.16 from the VR isochrones is illustrated in Fig. 7; we emphasize that the mapping of mass to magnitude is virtually indistinguishable if the $Y^2$ isochrones are used. The $Y^2$ isochrones generate masses typically larger by 0.008 $M_{\sun}$. The distribution is cut off at the red edge of the hydrogen-exhaustion phase since the spread in mass among stars on the subgiant branch through the tip of the giant branch is less than 0.03 $M_{\sun}$ and would crowd all the evolved stars into a single narrow vertical band at the right of the figure. The Li dip, central-mass-$[Fe/H]$ relation derived by @AT09 from several open clusters and the $Y^2$ isochrones predicts that for $[Fe/H] = -0.04 \pm 0.03$, the mass for the Li dip center should be $1.36 \pm 0.04$ $M_{\sun}$, entirely consistent with results displayed in Fig. 7 after adjustment for the small mass offset between $Y^2$ and VR. For $[Fe/H] = +0.09$, the predicted central mass is 1.42 $M_{\sun}$, essentially the same as found in the Hyades. @CU12 present a different relationship between $[Fe/H]$ and Li dip mass, relevant for older clusters for which stars on the hotter side of the dip are no longer present on the main sequence. This alternative relation pegs the masses of stars in the Li plateau on the cool side of the Li dip, with a similar dependence on metallicity; for the lower abundance cited above, stars on the cool side of the Li dip in NGC 6819 should have initial masses at 1.20 $M_{\sun}$, in excellent agreement with the initial mass associated with the peak A(Li) value on the low-mass side of the main sequence Li dip. Two critical independent checks on the validity of the mass scale, independent of the Li dip, are currently available. First, one can make use of the mass estimates for three stars in the eclipsing binaries analyzed by @BR16 which overlap in apparent magnitude with the data in Figs. 1 and 5. For eclipsing binary stars 23009A, 24009A, and 40007A, @BR16 derive $V$ = 15.13, 15.74, and 16.11, respectively. Neglecting the small effect due to the possible variation in reddening, we can translate these stars to the appropriate location in the CMD and derive their individual masses from the same isochrones used to construct the mass trend in Fig. 7. The isochronal masses are 1.23, 1.31, and 1.47 M$_{\sun}$, while the binary mass determinations are 1.218 $\pm$ 0.008, 1.251 $\pm$ 0.057, and 1.464 $\pm$ 0.011 M$_{\sun}$, respectively. Second, building upon the asteroseismological data for red giants in NGC 6819, @HA17 derive masses for first-ascent red giants and red clump stars, obtaining 1.61 $\pm$ 0.02 and 1.64 $\pm$ 0.02 M$_{\sun}$, respectively. While the VR isochrones do not include clump stars, we can assign the mass at the base of the vertical red giant branch as typical of the red giants and the stars at the tip of the RGB to have masses similar to the red clump stars. For $E(B-V)$ = 0.16 and an age of 2.25 Gyr, the red giants and clump stars have isochronal masses of 1.609 and 1.633 M$_{\sun}$. By contrast, if we lower the reddening to $E(B-V)$ = 0.14 and 0.12, with corresponding shifts in age and distance, the paired masses become 1.574-1.598 M$_{\sun}$ and 1.542-1.566 M$_{\sun}$, respectively. A more specific question beyond the typical mass of the stars populating the Li dip is the actual profile of the feature. Does the range of stars within the Li dip evolve over time, i.e. do the boundaries of the dip expand over time, encroaching on stars of higher and lower mass or temperature than found at the boundaries of the dip at an earlier age? To test this possibility, we make use of the Li dip profile defined by the Hyades and Praesepe clusters as discussed in @AT09 and revised by @CU17. To minimize the impact of metallicity, we transfer the $(B-V)$-based relation illustrated in Fig. 7 of that paper to a $T_{\mathrm{eff}}$-based relation, building upon the long-standing observation that the physical mechanism controlling the Li dip is solely temperature-dependent [@BA95; @CH01; @AT09; @CU12; @RA12a], explaining why higher metallicity stars in the Li dip have higher masses. For internal consistency, we have converted the A(Li) vs $(B-V)_0$ to A(Li) vs $T_{\mathrm{eff}}$ using the color-temperature relation from a VR isochrone of age 0.6 Gyrs and \[Fe/H\] = +0.13, the isochrone set with a metallicity closest to that derived for the Hyades and Praesepe [@CU17], \[Fe/H\] = +0.15. Using an isochrone of any age between 0.3 Gyr and 0.9 Gyr leaves the conclusions unchanged. Fig. 8 shows the Hyades/Praesepe data with no Li upper limits or binaries included, superposed on the A(Li) trend with $T_{\mathrm{eff}}$ for NGC 6819 at the age of the Hyades , i.e. the masses of the stars occupying the vertical turnoff of NGC 6819 have been used to derive their $T_{\mathrm{eff}}$ at the time the cluster had the same approximate age as the Hyades. The $T_{\mathrm{eff}}$ scale for the A(Li) profile has been adjusted by adding 150 K to the temperatures to align the Li-dip with that of the Hyades; use of the $Y^2$ isochrones would lead to a smaller shift of 120 K. The need for the shift can have multiple origins tied to the theoretical evolutionary rates for stars of varying mass as predicted by the isochrones, the $(B-V)_0$ - $T_{\mathrm{eff}}$ conversion relation, and the adopted reddening and distance modulus. The temperature adjustment could imply that the masses of the stars populating the vertical turnoff in NGC 6819 are too large compared to the unevolved main sequence stars by about 0.05 $M_{\sun}$ and/or the adopted distance modulus which defines $M_V$ for stars in the Li dip is too large. As noted above, a reduction in $E(B-V)$ from 0.16 to $\sim0.13$ would produce the appropriate change in the derived stellar masses. However, the related changes in both the distance modulus and the masses are contradicted by the excellent agreement with the eclipsing binary analysis [@BR16] and asteroseismology [@HA17]. What is striking about the comparison is the virtually identical $T_{\mathrm{eff}}$ range in the two profiles. The most Li-rich stars at the cool and the hot edges of the Li dip sit on or just outside the Hyades/Praesepe profile. At the hot edge, as already noted, the A(Li) limit based upon a pair of stars in NGC 6819 is as high as that found in the younger clusters (A(Li) $\sim$ 3.35), but the majority of stars scatter toward lower A(Li), unlike the minimal spread at a given $T_{\mathrm{eff}}$ in the combined Hyades/Praesepe sample. At the cool edge, only one star in NGC 6819 sits within the Li dip, i.e. systematically above the band defined by the younger clusters. In the mean, the stars between 6400 K and 6600 K fall on or below the trend defined by the younger clusters, potentially indicating that the Li-depletion mechanism in this temperature regime continued to reduce the surface abundance beyond the value predicted at 0.65 Gyr. The viability of this claim ultimately depends upon the assumed difference in the initial cluster A(Li) for the comparison clusters with significantly different metallicities. One could shift the Li data for the Hyades/Praesepe sample down by 0.2 dex to remove the separation between the two samples but this would place the coolest stars beyond the dip in the older cluster systematically above the younger stars. The consistency of the boundaries of the Li dip when comparing NGC 6819 and the Hyades/Praesepe data, particularly at the hot edge, is important because taken individually, the statistical samples defining these rapid transitions in A(Li), especially for Hyades/Praesepe, are modest, at best. @CU17 referred to the hot edge as “the wall" but cautioned against reading too much into a trend defined by a handful of stars. However, as discussed in Section 6, the cumulative sample afforded by the merger of data from Hyades/Praesepe [@CU17], NGC 752, NGC 3680, IC 4651 [@AT09] and now NGC 6819, leaves little doubt that the transition from Li-rich to Li-poor among the hotter stars occurs over a very small range in mass. Li: Subgiants and Giants ------------------------ Figs. 9 and 10 show the run of Li abundances across the subgiant and giant branches as a function of $V_0$ and $(B-V)_0$, respectively. Because of its unique status, star 7017 is included in all cases by a magenta starred symbol. ![A(Li) as a function of $V_0$ for stars brighter than $V_0 = 14.75$. Symbol colors and types are same as in Fig. 2. 7017 is shown as a magenta star.](fig9.pdf){width="\columnwidth"} ![A(Li) as a function of $(B-V)_0$ for stars brighter than $V_0 = 14.75$. Symbols have the same meaning as in Fig. 9](fig10.pdf){width="\columnwidth"} Choosing $(B-V)_0$ = 0.54, the coolest point on the red hook at the turnoff, as the color boundary of the subgiant branch leading to the vertical giant branch, the trend from the subgiant branch and beyond is apparent, but limited. The limitation is imposed by the fact that only upper limits to A(Li) are available for the majority of the stars. Limits determined through spectrum synthesis can be set to lower values at cooler temperatures, so the trend defined in Fig. 10 does not supply insight into how deep the Li depletion goes and at what rate for most stars leaving the main sequence. However, even at the start of the subgiant evolution, of the six stars between $(B-V)_0$ = 0.55 and 0.6, only two have detectable Li with A(Li) near 3.0 for the single star and below 2.5 for the binary. The remaining four, including one binary, already have upper limits below A(Li) = 1.4, 40 times lower than the detections still possible at the turnoff. Beyond this phase, all upper limits lie below 1.4, with the exception of one star near 1.6. Of the 65 stars populating the post-turnoff phase ($(B-V)_0$ $>$ 0.54), 8, including two binaries, have detectable Li; all stars with detectable Li on the vertical giant branch, with the exception of 7017, have A(Li) at 1.4 or less. Li Evolution: Comparison of Theory and Observation ================================================== Throughout the entire range of stellar masses where Li is observed, evidence points to rotation as the dominant non-standard mechanism affecting the surface A(Li). In most cases it is through the relationship between angular momentum loss and mixing, which is shown here for the first time to be relevant for stars more massive than those in the F-dwarf Li dip, although in young G/K dwarfs it can also be through structural and related effects due to rapid rotation. We relate our findings from NGC 6819 to most of this mass range Li Dip: Origins --------------- To decipher the internal structure of the lower mass stars currently populating the vertical turnoff of NGC 6819 and beyond, no feature is more important than the Li dip, particularly its boundaries. A key to the nature and origin of the Li dip centers on a critical question: what stellar parameter(s) determines the edge on the hot (high mass) side of the profile? The order-of-magnitude decline in Li across the hot edge of the dip within the Hyades/Praesepe sample [@AT09; @CU17] occurs over a color range of $\sim$0.02 mag in $(B-V)_0$, equivalent to a change in $T_{\mathrm{eff}}$ of $\sim$80 K. For the intermediate-age (1.45 Gyr to 1.75 Gyr) composite sample of NGC 752, IC 4651, and NGC 3680, the transition in the vertical turnoff occurs over a range in $V$ of less than $\sim$0.1 mag [@AT09], the same as NGC 6819. Equally important, the physical mechanisms controlling dip evolution on the cool edge may be distinctly different from those among the hotter dwarfs. As illustrated in Fig. 8, the stars on the cool side of the Li dip in NGC 6819 may have undergone greater depletion over this time interval but, within the uncertainties, the width of the Li-profile remains relatively static between 0.65 Gyr and 2.25 Gyr. The uncertainties include the limitation that the majority of A(Li) measures at the centers of the Li dips in multiple clusters are only upper limits and the fact that the translation of the A(Li) profile from the evolved turnoff region of the older clusters to the comparable age of the Hyades/Praesepe sample requires an offset in the $T_{\mathrm{eff}}$ scale to bring the A(Li) profiles into alignment. Since its delineation by @BT86, following on earlier work by @WA65, explanations for the origin of the Li dip have focused on multiple options: mass loss exposing deeper, Li-diluted layers for stars shedding their outer atmospheres [@SC90; @SW92], diffusion which allows Li atoms to settle below the visible surface without the need to reach layers hot enough for its destruction [@MI86; @RM93; @CH95], and some form of mixing or circulation mechanism, driven by gravity waves [@GA91; @MA96; @TC03; @TC05], rotation [@CM88; @DE93; @CT99; @PI10], or some combination thereof that takes the atmospheric Li to interior layers where the temperature crosses 2.5 x 10$^6$ K, Li is destroyed, and the Li-depleted material can then be dredged back up to the surface, usually via convection, meridional circulation or other mixing mechanism. The Yale rotational models [@PI89; @PI90], for example, show that stars deplete surface Li as a result of mixing induced by the loss of angular momentum and the resulting instabilities that are triggered in the stellar interior. Of these options, the mass loss rates required to produce the level of depletion found in the sun or within the Li dip are excessive, if not implausible, while failing to explain the apparent correlation among lithium, beryllium, and/or boron depletion for hotter stars, where beryllium and boron survive to progressively greater depth [@SW92; @DE97; @RA07; @BO05; @BO16]. Some evidence for diffusion comes from chemically peculiar A dwarfs [@RM09]; some of these show Li depletion [@BU97; @BU98; @BU00]. However, for cooler stars within the F-dwarf Li dip, below the surface convection zones, diffusion timescales into deeper layers are longer, resulting in an increasing Li abundance as a function of depth. This results in Li enhancement as the stars leave the main sequence and evolve across the subgiant branch when convection mixes deeper layers back to the observable atmosphere. This pattern has not been seen for Li in M67 [@PI88; @BA95; @PA12], with the possible exception of one star [@SD00], or NGC 6253 [@CU12], clusters old enough that the subgiant stars come from the mass range which defines the Li dip among unevolved main sequence stars. The large changes in A(Li) seen in the subgiants of these clusters seem to be caused exclusively or primarily by rotational mixing. Similarly, no more massive, post-turnoff stars in NGC 6819 show evidence of diffusion. Li aside, small and subtle diffusion effects have been claimed in M67 for a range of other metals through comparison of abundances for stars below, at, and above the turnoff region [@BE18; @GA18; @SO19] with elemental differentials ranging from $\sim0.05$ to 0.3 dex. Li Evolution: the Role of Stellar Rotation and the Cooler Dwarfs ---------------------------------------------------------------- @KR67 showed that the distribution of rotational velocities for dwarfs hotter than mid-F drops by an order of magnitude for dwarfs cooler than mid-F. @BO87 noted that the minimum in A(Li) in the Hyades coincides in $T_{\mathrm{eff}}$ with this break in the Kraft curve. The link between the drop in $V_{ROT}$ and the center of the Li dip has been confirmed in NGC 2516, M34, and NGC 6633 [@TE02; @JE02], among other clusters, and redefined more precisely for the Hyades by @BO16 and the combined Hyades/Praesepe sample by @CU17. As illustrated in Fig. 15 of @CU17, the minimum in the depth of the Li dip occurs at $T_{\mathrm{eff}}$ just below 6700 K, where the $V_{ROT}$ typically approaches 60 km s$^{-1}$. $V_{ROT}$ then declines in linear fashion with declining $T_{\mathrm{eff}}$ to below 10 km s$^{-1}$ near $T_{\mathrm{eff}}$ = 6200 K, where A(Li) returns to a value between 3.1 and 3.2 at the cool edge of the Li dip. A brief outline of current understanding of the evolution of the rotation-$T_{\mathrm{eff}}$ relation is relevant. Rotation periods can be determined from measurements of chromospheric activity (see, e.g. @NO84 and @SO91), and increasingly through the photometric variation of stars caused by nonuniform surface flux, specifically star spots, as seen through satellite surveys like [CoRoT]{} [@BA06], [Kepler]{} [@BO10], and, most recently, [Gaia]{} DR2 [@LA18a]. Periods do not suffer from the lack of information about inclination angle ([sin i]{}) inherent in measures of line broadening [@AF12; @ME11b; @MC13a; @MC13b; @MC14; @RE15]. [The broad pattern that has emerged from decades of observational analysis is a period-mass-age relation.]{} In clusters below 100 Myr in age, the relation is bimodal, with short-period and long-period rotation rates in the same cluster, though this bifurcation is less distinct for stars with masses below 0.4 $M_{\sun}$ and/or younger than 10 Myr [@BO14]. As clusters age, the long-period sequence evolves to even longer rotation periods, while the short-period sequence becomes subsumed within the long-period pattern by the age of the Hyades [@TE02; @BA03; @ME09; @ME11a]. Exposing the physical processes, particularly angular momentum loss, magnetic field structure, and convection, underlying the time evolution of rotation as a function of stellar mass has been the ongoing goal of many studies [e.g. @BA10; @CR11; @RE12; @BR14; @ST16; @GR18] since the determination of an empirical power-law relation by @SK72 for G dwarfs. A large degree of scatter in A(Li) has been observed in G/K dwarfs of the Pleiades (roughly 100 Myr-old) [@BU87; @SO93; @BO18], and also in the slightly older cluster, M35 [@AT18b]. A(Li) is correlated with $V_{ROT}$, in that stars that have retained A(Li) closer to their cluster primordial value have a significantly higher probability of falling within the bifurcated short-period category or in transition to the long-period track, i.e. they have yet to spin down to the long-period rotation rate. Evidence suggests rapid rotation leads to radius inflation which leads to less depletion of Li [@JDJ18]. It is unclear how this scatter at a given $T_{\mathrm{eff}}$ evolves. For example, the Hyades/Praesepe G dwarf sample of [@CU17] shows little scatter except near 6000K and 5200K, while the K dwarf sample (where the effect is most pronounced in the Pleiades) shows only upper limits. Moving to slightly higher mass, Fig. 8 suggests the intriguing possibility that the early-G dwarfs of NGC 6819 have depleted more Li than those in the younger Hyades: the only two stars cooler than 6150 K lie below the mean trend of the Hyades, one severely so. With the inclusion of NGC 6819 [@ME15], the period-mass-age relation was extended beyond the Hyades and NGC 6811 ($\sim$1 Gyr) to an age of 2.3 Gyr, confirming the continued spindown of cooler, lower mass stars with a precision permitting potential age estimation for individual field stars. Moving up the main sequence to higher mass, stellar models that incorporate enhanced angular momentum transport below the convective zone at a level that increases with increasing mass for stars between 0.95 $M_{\sun}$ and 1.15 $M_{\sun}$ can generally reproduce the observed trend of surface rotation and Li abundance with age for stars of different mass [@SOP16]. According to the models, for stars just redward of the cool side of the Li dip, the critical factor dominating Li evolution is the convective zone - radiative zone interface, with significant differential rotation early on producing a rapid decline in Li, but giving way to almost solid body rotation as the star ages, producing a flattening of the Li trend for stars older than $\sim$2 Gyr. Of particular relevance for the current analysis is the convergence of the period-mass relations for clusters of all ages as the mass of the main sequence star increases, leading to a unimodal, short-period trend for stars with $(B-V)_0$ below $\sim$0.47. In both NGC 6811 and NGC 6819, the rotation period declines precipitously between $(B-V)_0$ = 0.55 and 0.45, the Kraft break [@KR67], but the limiting periods for the two clusters are 1.3 days for NGC 6811 and just under 4.8 days for NGC 6819 [@ME15]. If the stars in NGC 6819 had the same rotational distribution as those in NGC 6811 at the same age, they have spun down by almost a factor of 4 over 1.3 Gyr. This decline in rotation speed among the cooler stars in the Li dip may be the continuation of a pattern of gradual spindown exhibited by stars in the Pleiades cluster (100 Myr) and continuing through M35 [@GE10], M34, NGC 2516, to at least the age of the Hyades [@TE02; @CU17], i.e. the range in rotational speed among stars with $(B-V)_0$ between 0.47 and 0.55 declines more gradually over time, in contrast with the more rapid change found among stars of lower mass. Fig. 8 illustrates that stars in NGC 6819 in this range of $T_{\mathrm{eff}}$ ($6600 - 6300 K$) have clearly depleted more Li, on average, than those in the Hyades. This additional Li depletion in NGC 6819 may thus be correlated with the additional angular momentum loss these stars have suffered relative to the Hyades. The sharp break in the period (rotation speed) – color (spectral type) relations near $(B-V)_0$ $\sim$ 0.47 is generally attributed to the absence of effective magnetic braking among the hotter stars due to the transition from stars with convective to radiative atmospheres. Attempts at more realistic 3-D modelling of the convective layer for a star at 1.47 $M_{\sun}$, almost exactly the mass of a star just beyond the high mass edge of the Li-dip boundary, show that a convective overshoot layer is created at the base of the traditional mixing-length model, extending the convective zone deeper and hotter than predicted by more traditional approaches, but still well above the Li-destroying zone [@KI16]. Yet rapidly-rotating stars on the hot side of the Li dip have suffered Li depletion. If angular momentum redistribution caused by internal mixing and magnetic braking coupled to a shallow convective atmosphere are the predominant sources of the spindown on the main sequence, stars blueward of the Kraft break [@KR67] may simply take longer ($\sim$200 Myr) to develop a more gradual degree of braking from a higher initial rotation rate. Another insight about the relationship of angular momentum loss and Li depletion, or lack thereof, comes from short-period, tidally-locked binaries (SPTLBs)\]. According to tidal circularization theory [@ZA89], close binaries with periods less than about 8 days would have tidally locked during the early pre-main sequence phase, before the stellar interior was hot enough to destroy any Li. Therefore, SPTLBs could exhibit higher Li abundances than normal single stars at a given phase of evolution [@SO90; @DE90]. Indeed, SPTLBs within the disk and among moderately metal-poor stars [@DE97; @RY95], X-ray binaries [@MA05], and V505 Per [@BA13], as well as members of the Hyades [@TH93] and M67 [@DE94], exhibit higher A(Li) than comparable single stars do. A number of caveats should be noted. In SPLTBs other complications could come into play, e.g. meridional circulation in very close binaries, so it is not expected that all SPTLBs would be better preservers of Li than non-binaries. Consistent with theoretical predictions, certain classes of SPTLBs do not show high Li: a) SPTLBs in the Pleiades (100 Myr) are Li-normal, as expected since the Pleiades are too young for rotationally-induced mixing to have become effective in depleting the surface Li; b) binaries with P $>$ 8 days are Li-normal; and c) short-period binaries with mid-F or earlier spectral types are Li-normal, as expected, since such hotter stars would not have been able to tidally lock during the early pre-main-sequence phase [@ZA89; @RY95]. We encourage study of SPTLBs in NGC 6819. Rotation of Hotter Dwarfs in NGC 6819 and Other Clusters -------------------------------------------------------- Before expounding on what we learn from NGC 6819, we point out varied additional evidence favoring rotationally-induced mixing as the primary cause of the Li dip over other proposed mechanisms. This evidence includes the Li/Be ratio, the Be/B ratio, and the timing of the formation of the Li dip. Since Li, Be, and B survive to different depths, combined knowledge from two or more of these elements can offer extremely powerful constraints. For example, if the efficiency of mixing depends on depth in different ways, the resulting surface Li/Be ratio will be affected in different ways (greater shallower mixing will affect Li more than it does Be). @DBS98 found that the depletion of Li and Be in F dwarfs is closely correlated, and the well-defined slope of A(Li) versus A(Be) strongly favors the predictions of the Yale rotational models [@DE97] and rotational models from @Char94, while strongly arguing against models with diffusion [@MI86; @RM93] and and mass loss [@SC90]. These data also argue against the gravity-wave-induced mixing models of @GA91, a non-rotational slow mixing mechanism in which Li is depleted more severely relative to Be than in the rotational models. These conclusions have been supported by a number of additional Li/Be studies in field stars and open clusters (e.g., @BDKS [@BK02; @BAK04]. Boron survives to a greater depth than does Be and can thus provide additional invaluable constraints on the character of the Li-dip-forming mechanism, especially if it is mixing. As with the Li/Be depletion correlation, B depletions detected in severely Li-depleted and Be-depleted field [@BDSL; @BO05] and Hyades [@BO16] F dwarfs leads to a B/Be depletion correlation that also strongly favors rotational mixing over diffusion, mass loss, and other types of mixing. Finally, [timing]{} is important: rotational mixing begins closer to the age of the Pleiades whereas diffusion and mass loss become more prominent closer to the age of the Hyades. Detection of the beginning of the formation of the Li dip in M35 [@SD04], a cluster just slightly older (about 150 Myr) than the Pleiades thus favors rotational mixing; arguably, a few F dwarfs in the Pleiades may already be depleting their Li. Fig. 11 presents the rotational velocity distributions among the turnoff stars for four clusters processed and analyzed in the same way, NGC 6819 at 2.25 Gyr, NGC 3680 at 1.75 Gyr [@AT09], NGC 2506 at 1.85 Gyr [@AT16; @AT18a], and NGC 7789 at 1.5 Gyr [@BR13; @RI13; @AT19]. The spectroscopic samples for NGC 3680, NGC 2506, and NGC 7789 have all been matched with [Gaia]{} DR2 astrometry to eliminate probable nonmembers, following the same procedure adopted for NGC 6819. Because NGC 2506 is more metal-poor, the mass of the stars occupying the Li dip should be lower than in NGC 3680. Qualitatively, one therefore needs to observe NGC 2506 at a greater age than NGC 3680 to place the stars in the vertical turnoff above the Li dip in the same relative evolutionary phase as a more metal-rich cluster like NGC 3680. In short, despite being older than NGC 3680, from a Li-dip perspective at the turnoff, NGC 2506 is qualitatively younger. To place the cluster comparison on a common scale while minimizing issues of reddening, metallicity, and distance, stars have been defined in $V$ based upon their magnitude difference relative to $V$ at the high mass (luminous) edge of the Li dip, where $\Delta$$V$ = ($V_{star}$ - $V_{dip}$). Thus, stars with negative values (open black circles) lie at higher mass, outside the Li dip, while positive and increasing $\Delta$$V$ extends across and beyond the cool, low-mass edge of the Li dip (red crosses). For NGC 2506, the spectroscopy did not reach the level of the dip. Using the cluster parameters derived in @AT16 and @AT18a, coupled with the discussion of the central mass of the Li dip and its relative boundaries in $V$ [@AT09], stars at the center of the evolved NGC 2506 Li dip should have a mass $\sim$1.28 $M_{\sun}$, with the luminous edge positioned $\sim$0.3 mag brighter than the center, placing the predicted cliff in Li abundance for NGC 2506 at $V$ = 15.75, just below the faint limit of the spectroscopic sample. We will adopt this magnitude as $V_{dip}$ for NGC 2506. For NGC 7789, NGC 6819, and NGC 2506 only single-star members are included in the analysis; due to the smaller statistical sample for NGC 3680, all members, binaries or not, were retained. For NGC 7789, one star lies beyond the high-velocity edge of the plot at $V_{ROT}$ = 126 km s$^{-1}$ and $\Delta$$V$ = -0.9. Stars with even higher rotation speeds in NGC 7789 have not been included because the distortion caused by the extreme rotation made reliable estimation of their speeds and their Li abundances impossible. All stars observed in NGC 2506 have been included. Four stars in NGC 3680, KGP 988, 1410A, 1347 and 1506 [@KO95], with measurable rotation speeds between $V_{ROT}$ = 40 and 60 km s$^{-1}$ and $\Delta$$V$ between -0.35 and -1.5 are not plotted. All four stars originally were classed as possible binaries due to the unusual width of the spectral lines [@AT09]; if evaluated as single stars with rapid rotation, 988 and 1347 have A(Li) = 3.24 $\pm$ 0.04 and 3.05 $\pm$ 0.06, respectively. If processed as SB2s, the paired stars in 988 and 1410a generate A(Li) = 2.8 $\pm$ 0.1 and 2.75 $\pm$ 0.15, respectively. Equally important, the lower bound near $V_{ROT}$ $\sim$ 20 km s$^{-1}$ in NGC 7789 is not an artifact of the analysis; red giants in the same cluster with rotational velocities below this limit are readily measured, thus [all]{} stars in our sample rotate faster than 20 km s$^{-1}$. While the turnoff of the younger NGC 7789 extends to higher mass (more negative $\Delta$$V$) than in NGC 2506, the spread in $V_{ROT}$ and the mean $V_{ROT}$ are greater in the younger cluster for stars above the Li dip. The shifts toward lower $V_{ROT}$ in the distribution above and within the Li dip for NGC 3680 and especially for NGC 6819 relative to NGC 7789 are dramatic. By the age of NGC 6819 there is almost no statistically significant difference in the distribution of $V_{ROT}$ between the higher mass stars outside the Li dip and those within, implying that the more rapid rotators among the higher mass stars, as illustrated by the black circles in NGC 2506 and NGC 3680, have spun down to the lower rotation levels of the cooler dwarfs by the age of NGC 6819. For the single-star red giants beyond $(B-V)_0$ = 0.54, the mean $V_{ROT}$ of the NGC 6819 members is cut in half again to values no greater than 8.4 $\pm$ 3.3 km s$^{-1}$. The relevance of the $V_{ROT}$ pattern for the evolution of Li becomes apparent in Fig. 12, where the distributions of Li among [only]{} the stars brighter than the Li dip (black circles of Fig. 11) are illustrated. To avoid potential issues with the zero-points of the A(Li) scale from one cluster to the next, the A(Li) distribution is based upon $\Delta$A(Li), the difference in A(Li) between a given star and the highest value in a cluster sample. Histograms are based upon the percentage of the stars relative to the total cluster sample above the Li dip. The solid lines represent the fraction of stars with detectable Li in a given $\Delta$A(Li) bin, while the dashed curve is the percentage counting stars with either measured A(Li) or upper limits within the A(Li) range. Thus, the dashed curve should always sit on or above the solid histogram. For NGC 7789, with a significantly larger sample size than the other three clusters, scatter in A(Li) led to an initial $\Delta$A(Li) bin between 0 and 0.2 with only five stars, creating an artificial offset in the distribution relative to the other three clusters. For comparison purposes, we have reset the upper bound in A(Li) in NGC 7789 to be 0.2 dex lower than the maximum observed and counted the five resulting negative values of $\Delta$A(Li) into the first bin between 0.00 and 0.2. Once again, the change from NGC 7789 to NGC 6819 is obvious. The fraction of stars with measurable Li is peaked at low $\Delta$A(Li) values for NGC 7789 and NGC 2506, flattens slightly for NGC 3680, and is almost bimodal for NGC 6819, with a secondary peak near $\Delta$A(Li) = 1.2. Even more apparent, the fraction of stars with upper limits well below the cluster maximum grows significantly. [Clearly, in going from NGC 7789 to NGC 6819, the spin-down of these A dwarfs correlates with increasing depletion of lithium.]{} A Potential Li - $V_{ROT}$ Link ------------------------------- The stars within the red hook and the phase entering the subgiant branch are well evolved from their main-sequence state, so their current rotation characteristics should not reflect those of stars on the unevolved main sequence. The masses of the typical stars entering the giant branch are 1.9 $M_{\sun}$, 1.65 $M_{\sun}$, 1.75 $M_{\sun}$, and 1.6 $M_{\sun}$ in NGC 7789, NGC 2506, NGC 3680, and NGC 6819, respectively. Of particular relevance for the current discussion is the series of papers [@RO02a; @RO02b; @RO07; @ZO12] detailing the rotational characteristics of early type stars from B through late A. @ZO12 found that, unlike the hotter/higher mass stars which showed a clear bimodal distribution of rotational speeds, the late A stars, ranging from 2.5 $M_{\sun}$ to 1.6 $M_{\sun}$, had a unimodal distribution with a well-defined peak tied to the mass. Stars in the 1.6 to 2.0 $M_{\sun}$ range had a broad distribution in $V_{ROT}$[sin i]{} which peaked near 125 km s$^{-1}$; correcting for the inclination effect shifted this peak to 145 km s$^{-1}$. The predicted range in the peak over a change of 0.3 $M_{\sun}$ is only 12 km s$^{-1}$, with stars of lower mass spinning slower. While a fraction of the stars in NGC 7789 have $V_{ROT}$ above 100 km s$^{-1}$, making reliable rotational speed and Li estimation impossible, the averages of the observed $V_{ROT}$ distributions in NGC 7789, NGC 2506, and NGC 3680 are less than 1/2 to 1/3 of the predicted value for the late A stars. A potential factor which could impact the discrepancy between the @ZO12 sample and the clusters is that the rapidly rotating stars in this mass range may have evolved to a temperature which carries them out of the sample range studied by @ZO12. However, models which include rapid rotation indicate that these stars live longer at higher luminosity, occupying the same region of the CMD as lower mass, slower rotating stars [@BR15b]. For stars in the mass range of interest, @ZO12 conclude that the stars evolve as differential rotators during their entire main-sequence lifetime, with the mean equatorial velocity accelerating during the first third of the star’s main-sequence life, then remaining high or slowing down mildly beyond that point. If correct, this analysis implies that the obvious spindown seen through the comparison of NGC 7789 with the late A-star distribution takes place in the limited time between the main-sequence evolution and the subgiant branch, i.e. when the star is near the cool end of the red hook and beyond. If the turnoff stars of NGC 7789 and the 1.45 Gyr-old cluster NGC 752 [@TW15] are evolved equivalents of the late A stars in @ZO12, then these stars spend the majority of their main-sequence lives with true rotational speeds between 100 km s$^{-1}$ and 250 km s$^{-1}$. Starting with the insightful analyses of @BR15a [@BR15b; @BR15c], the impact of a large range of rotational velocities on the interpretation of the ages of star clusters and field stars below 2 Gyr in age has become an extensive and vibrant area of research, in large part driven by the desire to explain the extended main sequences in CMDs of young and intermediate-age clusters of the Magellanic Clouds [e.g., @MA07; @MA08; @GO09; @GI11]. The implication that a spread in initial rotation rates among stars of a given mass within a cluster will lead to a range of mass among giants of the same age fits well with the so-called double clump identified within intermediate-age open clusters [@GI99; @GI00; @GI09; @GO14]. The standard interpretations of this feature are tied to the fact that the giants in these clusters lie near the boundary where He-ignition occurs under non-degenerate or degenerate conditions. If the internal structure of the stars of a fixed mass evolving up the giant branch and beyond can be altered by a range in rotation and/or mass loss prior to He-flash or by a range in mass among the stars leaving the main sequence, potentially due again to a range in rotation speeds on the main sequence, the red giant branch will contain He-core-burning stars populating two clumps. Beyond an age of $\sim$2 Gyr, as in NGC 6819, the rapid rotators have minimal impact and the CMD clump returns to the expected appearance for a single-mass population. In addition to rotation [@YA13b; @NI15; @NI16; @BA16; @LI17; @MA18], the impacts of binaries [@YA18], variable stars [@SA18], stellar age spread [@GO11; @GO15; @GO17], and convective core overshoot [@YA17] have been investigated, with varying degrees of success. Of particular relevance for the current discussion, however, is the study by @WU16. CMD evolution is studied in the context of an initially rapidly rotating population of B and A stars that slow down over time due predominantly to their evolutionary expansion from the main sequence to the red giant branch. For the populous Magellanic Cloud clusters, this eliminates the contradiction between a large color spread among the stars at the turnoff feeding into a subgiant branch with a narrow range in luminosity [@LI14]. With an age of 1.5 Gyr, the highest mass star above the Li dip in NGC 7789 has already spun down by about a factor of two. From Fig. 12, the fraction of turnoff stars brighter than the Li dip in NGC 6819 with detectable Li is comparable to those with only upper limits, in distinct contrast with the other 3 clusters. Closer examination of the NGC 6819 pattern for these stars in Fig. 5 reveals that, [in comparison with the samples in NGC 7789, NGC 2506 and NGC 3680, the NGC 6819 turnoff stars are developing a second Li dip at higher mass than the F-dwarf Li dip, with the greatest concentration of stars with detectable Li located just above the high mass boundary of the Li dip.]{} Since the brighter stars above the Li dip in NGC 6819 populate the red hook, it is tempting to assume that the growth of the convective zone as the stars expand into the red hook and beyond is the primary source for this Li depletion. This argument fails, however, when applied to NGC 7789, NGC 2506, and NGC 3680. As noted earlier, the striking feature about the blue edge of the Li dip is the rapid change in the Li-depletion process as one changes the mass ($T_{\mathrm{eff}}$) of the main-sequence stars by a very small amount. If $T_{\mathrm{eff}}$ were the sole determinant of the effectiveness of the process driving Li-depletion, as soon as stars above the Li dip evolved redward off the main sequence toward a vertical turnoff, they should trigger the same process which depletes Li among stars within the Li dip. As an example, using the VR isochrones and Fig. 10, stars defining the high mass edge of the Li dip have a mass of 1.42 $M_{\sun}$. Tracking back to the unevolved main sequence as defined by the CMD location of the boundary at the age of the Hyades, these stars had a temperature of approximately 6900 K. Evolution off the main sequence moves this boundary to 6870 K at 0.9 Gyr, 6830 K at 1.2 Gyr, 6760 K at 1.5 Gyr (the age of NGC 752 and NGC 7789), 6680 K at 1.75 Gyr (NGC 3680), and 6480 K by 2.25 Gyr (NGC 6819). Between the age of the Hyades when the Li dip is fully developed and the age of NGC 6819, the stars defining the hot boundary evolve in $T_{\mathrm{eff}}$ across a range which encompasses the entire Li dip. By the age of NGC 7789, stars at the hot boundary have spent at least 0.3 Gyr within the Li dip and all stars more massive than this boundary have been there longer. Despite this, the large majority of stars above the turnoff in NGC 7789, NGC 2506, and NGC 3680 still retain A(Li) near their supposedly original values. Even at the age of NGC 6819, a significant fraction of the stars at the higher-mass boundary still have detectable and high values of A(Li). As discussed in detail for NGC 2506 [@AT18a], a plausible answer to this delayed reaction comes from the mechanisms commonly used to explain the Li dip itself. The depth of the convection zone among stars hotter than the Li dip ranges from nonexistent to inadequate for driving Li depletion. As demonstrated by the models of @CH10, inclusion of significant rotation and rotationally-induced mixing can produce an immediate and continuous decline in A(Li) as stars evolve across the subgiant branch, especially in light of the deepening convective zone at cooler $T_{\mathrm{eff}}$. The pattern revealed in the turnoff regions of the clusters under discussion implies that Li depletion only becomes significant for stars entering the main sequence red hook if the time spent in this phase, the growth of the convective zone, and the degree of rotational spindown combine to induce a serious depletion in the atmospheric Li level. A high degree of depletion inevitably occurs for the majority of stars evolving along the giant branch, but the initiation point for significant depletion shifts to earlier phases of post-main-sequence evolution and lower mass as a cluster ages. Evolved Stars: Li in Giants and Subgiants ----------------------------------------- The striking change in the A(Li) distributions for the turnoff stars in going from NGC 7789 to NGC 6819 is reflected in the giants, though the pattern is somewhat modified by the relative change in the evolutionary phases occupied by post-main-sequence stars with increasing age. As discussed earlier, the transition from NGC 7789 to NGC 6819 is important in part because it covers the mass and age range where stars leaving the main sequence change from nondegenerate to partially degenerate cores, and He-ignition at the tip of the giant branch switches from a quiescent phase to He-flash under degenerate conditions. In fact, a mixture of red giants undergoing both forms of He-ignition within the same cluster has long been a potential explanation for the peculiar red giant clumps in both NGC 7789 and NGC 752 [@GI00], as discussed earlier. The state of the star in the hydrogen-exhaustion phase determines the rate of evolution across the subgiant branch and up the first-ascent red giant branch. The state of the star at He-ignition should seriously impact the degree of mixing, raising the prospect of serious Li-depletion among stars undergoing He-flash, i.e., most if not all red clump stars should exhibit greatly reduced Li abundances. Testing this idea is complicated. For example, nearly all stars on the first ascent RGB in NGC 6819 already have unmeasurably low A(Li), making it impossible to know whether there are further reductions in A(Li) after the He flash on the way to the red clump. These low A(Li) values in NGC 6819 are largely due to Li depletion during the MS, so the distribution of A(Li) in all phases from the MS through the helium flash must be taken into account. The measured Li distribution among the subgiants and giants will depend upon the relative population of subgiants, first-ascent red giants and red clump stars, as well as the masses of the stars feeding the giant branch. To understand the evolutionary impact, a simple comparison of NGC 7789 and NGC 2506 is in order. As seen in Fig. 12, 87% and 88% of the stars at the turnoff above the Li dip in NGC 7789 and NGC 2506, respectively, have measurable Li within 1 dex of the cluster maximum. Since A(Li) is easier to measure among stars cooler than this color range, any decline in the fraction of subgiant and giant stars with comparable A(Li) must be a reflection of a real reduction in the elemental abundance. However, as shown in the CMD distribution for clusters with distinctly different ages, the subgiant branch and the giant branch below the clump in the younger NGC 7789 are sparsely populated [@GI98; @TW12; @BR13] while the comparable regions in the older NGC 2506 are easily delineated [@AT16], a primary factor in its selection for Li analysis to test the degree of mixing occurring across the subgiant branch [@AT18a]. Thus, the giants in NGC 7789 are almost totally dominated by stars at or above the level of the red giant clump while in NGC 2506, the majority of stars are first-ascent red giants below the level of the clump. As expected for more evolved stars, both clusters show a reduction in the percentage of red giants with measurable Li, dropping to one-third for NGC 7789 and one-half for NGC 2506, with the greater drop for NGC 7789 tied to the more advanced evolutionary state of the giants in that sample. By contrast, among the single stars within the giant branch of NGC 6819 (Fig. 10), only 10% have measurable A(Li), despite a first-ascent giant branch as well populated as that within the younger NGC 2506. The key is that, in addition to any Li depletion which might occur across the subgiant branch or at the first dredge-up, the stars leaving the main sequence are already reduced to a supply of only 35% with A(Li) measurable within 1 dex of the cluster maximum. The majority of giants within NGC 3680 do have measurable A(Li), but the significance is minimal given a sample of only nine stars, four of which are binaries [@AT09]. Summary and Conclusions ======================= As emphasized in Paper II, for any attempt to delineate the evolution of Li with stellar age and mass, NGC 6819 displays a rare blend of critical characteristics. It is young enough for the turnoff stars to cover the full range of mass defining the Li dip, with the higher mass stars above the Li dip still retaining limiting values that should be representative of the initial cluster abundance, but old enough that the stars evolving through the subgiant branch and ascending the giant branch for the first time have partially degenerate cores, leading ultimately to He-flash at the tip of the giant branch. Despite an age significantly larger than the typical open cluster evaporation timescale, the cluster is rich in members bright enough to allow high dispersion spectroscopic analysis below the level of the Li dip. The precision radial-velocity, proper-motion, and photometric surveys of the cluster field have greatly enhanced identification of probable cluster members while allowing for detection of and correction for variable reddening across the face of the cluster, a necessity for reliable temperature determination as a basis for spectroscopic and photometric analysis. Finally, the inclusion of the cluster within the [Kepler]{} field allows asteroseismic tests of the internal structure and evolutionary phase for the evolved cluster members, confirming in the case of star 7017 the anomalous nature of this star relative to the other cluster members. Building upon the precision photometry and spectroscopy of Papers I and II, as well as the current abundance re-evaluation from ANNA, it is found that for a mean cluster reddening in the range of $E(B-V)$ = 0.14 to 0.16, having corrected for variable reddening across the cluster face and adopting a slightly subsolar mean metallicity of $[Fe/H] = -0.04$, the best estimate for the cluster age is 2.40 to 2.25 Gyr, with an apparent distance modulus between 12.29 and 12.40, in excellent agreement with the zero-point-adjusted parallax measures from [Gaia]{} DR2. Ages closer to 2.6 Gyr can be obtained by adopting artificially lower reddening or by selecting different sets of isochrones, but the latter alteration is a reflection of the range among theoretical models rather than a problem with the observational data for the cluster. It should be noted that adoption of the older spectroscopic abundance of $[Fe/H] = +0.09$ for the cluster [@BR01] would reduce the age below the currently derived value of 2.25 to 2.4 Gyr. Turning to Li, the limiting value of A(Li) among the single stars with the highest mass at the turnoff is A(Li) = 3.2 $\pm$ 0.1, consistent within the errors with that found for the primordial solar system. The majority of the stars with A(Li) above 3.2 are within binary systems, potentially indicating that they have retained their primordial Li value while the remaining single stars, having evolved well off the ZAMS and now approaching the hydrogen-exhaustion-phase (HEP) and beyond, may have undergone Li depletion to varying degrees. The depletion range in A(Li) among the single turnoff stars above the high-mass boundary of the Li dip is as large as that found within the dip. For stars on the low-mass side of the Li dip, the mean A(Li) = 2.83, with no trend with magnitude over the range from $V_0$ = 15.5 to 15.9. Equally important, the dispersion in A(Li) among these stars is 0.16 dex, three times larger than expected from spectroscopic errors alone. So, while there are stars fainter than the Li dip with A(Li) approaching the limiting value for stars on the hot side of the Li dip, the significant lack of differential evolution predicted among the cooler stars implies that these stars merely represent the high-Li end of an intrinsic scatter centered near A(Li) = 2.85. In short, the cooler stars have depleted their initial Li from A(Li) $\sim$ 3.3 to the current mean of 2.83, with a scatter tied to some intrinsic property of the sample. The Li dip within the vertical turnoff of the cluster, when translated back to the age of the Hyades/Praesepe and adjusted for a modest shift in the $T_{\mathrm{eff}}$ scale, shows a profile with $T_{\mathrm{eff}}$ that bears a striking resemblance to that defined by the younger clusters. On the hot or high-mass side of the profile, the transition from measurable and high Li to depleted or unmeasurable Li takes place over a ZAMS $T_{\mathrm{eff}}$ range of less than 80 K. On the hot side above the Li dip, the limiting value of A(Li) for NGC 6819 lies below the limiting estimate for the much younger Hyades for the reasons noted previously. Taking this into account, there is little evidence for a widening in the mass or $T_{\mathrm{eff}}$ range of the Li dip between 0.6 Gyr and 2.3 Gyr. The stars on the cool slope of the Li dip may show some depletion in NGC 6819 relative to the Hyades, but the cool boundary of the Li dip is an excellent match to that within the Hyades. Within the center of the Li dip, solid conclusions are more difficult to achieve because the majority of stars within this mass range have only upper limits to A(Li). Of the 65 stars with $(B-V)_0$ greater than 0.54 on the subgiant branch and beyond, only 8, including 2 binaries, have detectable Li and of these, only 3, including 7017, overlap with the range of detectable Li among the stars at the turnoff above the Li dip. The implied pattern of a significant depletion of Li among stars just leaving the main sequence but well before the the phase of the first dredge-up is consistent with the need for an additional mixing mechanism among stars with masses in the range leading to partially degenerate cores after the HEP, as illustrated by IC 4651, NGC 752, NGC 3680 [@PA04; @AT09] and NGC 2506 [@AT18a]. The greater cluster age and lower turnoff mass also correlates with the absence of an apparent double clump associated with clusters like NGC 752 and NGC 7789 and is more representative of the pattern in the older cluster M67, where stars leaving the main sequence emerge from within the Li dip [@PA12]. As noted, while the boundary of the Li dip on the high-mass edge is sharply defined, stars above this edge display a significant range in Li, from an approximately primordial solar value to upper limits competitive with those found within the Li dip. Equally important, the distribution of A(Li) among these stars is heavily skewed toward values an order of magnitude lower than the cluster limiting/primordial value, in striking contrast with younger clusters like NGC 7789 and NGC 2506 where the majority of stars have measurable A(Li) within 1 dex of the cluster limit. In short, as the stars higher in mass than the Li dip evolve through the HEP, they are reproducing the pattern of depletion found within the Li dip itself, but at a much later phase of evolution. It should be reemphasized, however, that the mechanism defining the Li dip cannot be purely $T_{\mathrm{eff}}$-dependent for the stars of higher mass; the stars more luminous than the Li dip boundary enter the $T_{\mathrm{eff}}$ range defining the Li dip on the unevolved main sequence long before significant depletion becomes detectable, in stark contrast with the traditional Li dip which exhibits depletions in clusters by 100 Myr and is fully formed by the age of the Hyades/Praesepe [@CU17]. Adjusting for the fact that the mass ranges among the turnoff stars within NGC 7789 and NGC 2506 extend to higher values than in NGC 6819, when comparing stars of comparable relative mass above the Li dip, the mean and dispersion of $V_{ROT}$ among the stars in NGC 7789 and NGC 2506 are larger by a factor of 2-3 than among the stars in NGC 6819. Moreover, the mean and dispersion of $V_{ROT}$ among the stars above and below the high-mass boundary of the Li dip at the turnoff of NGC 6819 are essentially identical. When coupled with the minimization of Li-depletion among SPTLBs, it is concluded that the second critical factor controlling mixing and Li-depletion appears to be the rate of stellar spindown, rather than the absolute size of $V_{ROT}$. Within the greater context of Li-depletion along the entire main sequence, the observations indicate that the Li dip potentially arises as a consequence of the convolution of two mass-temperature-dependent functions. Moving along the main sequence from low to high mass, the average depth of the surface convection zone grows shallower, leading to slower rates of Li-depletion until stars like the sun and higher mass should exhibit little if any Li-depletion in their atmospheres. The second, competing factor is stellar spindown which potentially induces mixing at the base of the convection zone and/or within the atmosphere itself. For stars in the 0.6 to 1.0 $M_{\sun}$ range, the initial bimodal distribution of $V_{ROT}$ converges by $\sim$0.4 Gyr to a unimodal profile dominated by slow rotators. Observations of stars with masses just below solar show a clear trend of decreasing magnetic field strength with age up to the age of the Hyades [@FO18], while analysis of single-star solar analogs shows a clear correlation between declining $V_{ROT}$ and declining A(Li) [@BE17]. However, between $(B-V)_0$ = 0.55 and 0.4, the trend of $V_{ROT}$ with increasing $T_{\mathrm{eff}}$ undergoes a sharp rise, and a rapid decline in $V_{ROT}$ by 0.4 Gyr is no longer applicable. For stars on the cool side of the Li dip and just beyond, the initial spindown does occur, but it takes longer and starts later, generating the red edge of the Li dip in lower mass stars between 0.2 and 0.6 Gyr. Thus, even lower mass, solar analogs deplete their surface Li by a factor of 10 over the first Gyr. The spindown-driven mixing and the convection zone continue to deplete Li at a slow but detectable rate over the lifetime of the star [@TU15], leading to the Li-plateau among cooler dwarfs, without reaching deep enough to destroy Be. Clearly, this empirical explanation for the observed pattern remains qualitative at best, in part because current theoretical stellar models which attempt to reproduce the observed spindown among solar-type stars as a function of age using a variety of angular momentum transport mechanisms can do so, but not without failing to simultaneously reproduce the Li abundance and the internal rotational structure of the sun [@AM16]. Somewhat surprisingly, the best models for reproducing the evolution of the rotation rate with age often deplete too much Li over time. Among the hotter stars in the Li dip, the significantly higher initial $V_{ROT}$ distribution ultimately compensates for an even shallower convection zone, driving the mixing zone deep enough to cause Li-depletion and Be-depletion when brought into contact with the thin convective layer of the atmosphere. As the rotation rate slowly declines, the level of rotational mixing subsides and the convection zone no longer maintains contact with the depleted layers below the atmosphere, leaving A(Li) at a fixed level until evolution beyond the main sequence and subsequent spindown during the HEP drive the mixing layer down to where additional Li-depleted gas can be accessed. Extensive use was made of the WEBDA[^4] database, maintained at the University of Brno by E. Paunzen, C. Stutz and J. Janik. NSF support for this project was provided to BJAT, DLB and BAT through NSF grant AST-1211621, and to CPD through NSF grant AST-1211699. The authors gratefully acknowledge the thoughtful comments of the referee which led to greater clarity in the extensive discussion of the text. Affer, L., Micela, G., Favata, F., & Flaccomio, E. 2012, , 424, 11 Aguilera-Gómez, C., Chanamé, J., Pinsonneault, M., & Carlberg, J. 2016, , 829, 127 Amard, L., Palacios, A., Charbonnel, C., Gallet, F., & Bouvier, J. 2016, , 587, 105 Anders, E., & Grevesse, N. 1989, Geochim. Cosmochim. Acta, 53, 197 Anthony-Twarog, B. J., Deliyannis, C. P., Lee-Brown, D. B., & Twarog, B. A. 2019, in preparation Anthony-Twarog, B. J., Deliyannis, C. P., Rich, E., & Twarog, B. A. 2013, , 767, L19 Anthony-Twarog, B. J., Deliyannis, C. P., & Twarog, B. A. 2014, , 148, 51 (Paper I) Anthony-Twarog, B. J., Deliyannis, C. P., & Twarog, B. A. 2016, , 152, 192 Anthony-Twarog, B. J., Deliyannis, C. P., Twarog, B. A., Croxall, K. V., & Cummings, J. 2009, , 138, 1171 Anthony-Twarog, B. J., Deliyannis, C. P., Twarog, B. A., Cummings, J. D., & Maderak, R. M. 2010, , 139, 2034 Anthony-Twarog, B. J., Lee-Brown, D. B., Deliyannis, C. P., & Twarog, B. A. 2018a, , 155, 138 Anthony-Twarog, B. J., Deliyannis, C. P., Harmer, D., et al. 2018b, , 156, 37 Arbib, M. A. 2002, The Handbook of Brain Theory and Neural Networks, 2nd edn. (Cambridge, MA: MIT Press) Asplund, M., Grevesse, N., Sauval, A. J., & Scott, P. 2009, , 47, 481 Baglin, A., Auvergne, M., Barge, P., et al. 2006, [Proc. The CoRoT Mission Pre-Launch Status: Stellar Seismology and Planet Finding]{} (ESA Special Publication), Vol. 1306, ed. M. Fridlund, A. Baglin, J. Lochard, & L. Conroy (Noordwijk: ESA), 33 Balachandran, S. 1995, , 446, 203 Barnes, S. A. 2003, , 586, 484 Barnes, S. A., & Kim, Y. C. 2010, , 721, 625 Baugh, P., King, J. R., Deliyannis, C. P., & Boesgaard, A. M. 2013, , 125, 753 Bastian, N., Niederhofer, F., Kozhurina-Platais, V., et al. 2016, , 460, L20 Beck, P. G., do Nascimento, J.-D., Jr., Duarte, T., et al. 2017, , 602, 63 Bertelli Motta, C., Pasquali, A., Richer, J., et al. 2018, , 478, 425 Boesgaard, A. M. 1987, , 99, 1067 Boesgaard, A. M., Armengaud, E., King, J.R., Deliyannis, C. P. & Stephens, A. 2004, , 613, 1202 Boesgaard, A. M., Deliyannis, C. P., King, J. K. & Stephens, A. 2001, , 553,754 Boesgaard, A. M., Deliyannis, C. P., & Steinhauer, A. 2005, , 621, 991 Boesgaard, A. M., Deliyannis, C. P., Stephens, A. & Lambert, D. .L. 1998, , 492, 727 Boesgaard, A. M. & King, J. R. 2002, , 565, 587 Boesgaard, A. M., Lum, M. G., Deliyannis, C. P., et al. 2016, , 830, 49 Boesgaard, A. M., & Tripicco, M. J. 1986, , 302, 49 Borucki W. J., Koch D., Basri G., et al. 2010, [Sci]{}, 327, 977 Bouvier, J., Barrado, D., Moraux, E., et al. 2018, , 613, 33 Bouvier, J., Matt, S. P., & Mohanty, S., et al. 2014, Protostars and Planets VI, eds. H. Beuther, R. S. Klessen, C. P. Dullemond, & T. Henning (Tucson: University of Arizona Press), 433 Bradley, P. A., Guzik, J. A., Miles, L. F., et al. 2015, , 149, 68 Bragaglia, A., Carretta, E., Gratton, R. G., et al. 2001, , 121, 327 Bragaglia, A. & Tosi, M. 2006, , 131, 1544 Brandt, T. D., & Huang, C. X. 2015a, , 807, 24 Brandt, T. D., & Huang, C. X. 2015b, , 807, 25 Brandt, T. D., & Huang, C. X. 2015c, , 807, 58 Brewer, L. N., Sandquist, E. L., Mathieu, K., et al. 2016, , 151, 66 Brown, T. 2014, , 789, 101 Brunker, S., Anthony-Twarog, B. J., Deliyannis, C. P., & Twarog, B. A. 2013, BAAS, 221, 25028 Burkhart, C., & Coupry, M. F. 1997, , 318, 870 Burkhart, C., & Coupry, M. F. 1998, , 338, 1073 Burkhart, C., & Coupry, M. F. 2000, , 354, 216 Butler, R. P., Marcy, G. W., Cohen, R. D., & Duncan, D. K. 1987, , 319, L19 Cantat-Gaudin, T., Jordi, C., Vallenari, A., et al. 2018, , 618, 93 Carlberg, J. K., Smith, V. V., Cunha, K., et al. 2015, , 802, 7 Casagrande, L., Silva Aguirre, V., Stello, D., et al. 2014, , 787, 110 Casagrande, L., & VandenBerg, D. L. 2014, , 444, 392 Casey, A. R., Ruchti, G., Masseron, T., et al. 2016, , 461, 3336 Cayrel de Strobel, G. 1988,in [IAU Symposium 132, The Impact of Very High S/N Spectroscopy on Stellar Physics]{}, ed. G. Cayrel de Strobel & M. Spite (Dordrecht: Kluwer), 345 Chaboyer, B., Demarque, P., & Pinsonneault, M. H. 1995, , 441, 865 Charbonneau, P., & Michaud, G. 1988, , 334, 746 Charbonnel, C., & Lagarde, N. 2010, , 522, 10 Charbonnel, C., & Talon, S. 1999, , 351, 635 Charbonnel, C., Vauclair, S., Maeder, A., Meynet, G. & Schaller, G. 1994, , 283, 155 Chen, Y. Q., Nissen, P. E., Benoni, T., & Zhao, G. 2001, , 371, 943 Coc, A., Uzan, J.-P., & Vangioni, E. 2014, , 10, 50 Cranmer, S. R., & Saar, S. H. 2011, , 741, 54 Cummings, J. D., Deliyannis, C. P., Anthony-Twarog, B. J., Twarog, B. A., & Maderak, R. M. 2012, , 144, 137 Cummings, J. D., Deliyannis, C. P., Maderak, R. M., & Steinhauer, A. 2017, , 153, 128 Deliyannis, C. P. 1990, PhD. Thesis, Yale University Deliyannis, C. P., Boesgaard, A. M., Stephens, A., King, J. R., Vogt, S. S. & Keane, M. J. 1998, , 498, L147 Deliyannis, C. P., Demarque, P. & Kawaler, S. D. 1990, , 73, 21 Deliyannis, C. P., King, J. R., Boesgaard, A. M., & Ryan, S. G. 1994, , 434, L71 Deliyannis, C. P., & Pinsonneault, M. H. 1993, [IAU Colloquium 137, Inside the Stars]{}, ed. W. W. Weiss & A. Baglin (ASP Conf. Ser. 40; San Francisco: ASP), 174 Deliyannis, C. P., & Pinsonneault, M. H. 1997, , 488, 836 Deliyannis, C. P., Pinsonneault, M. H., & Duncan, D. K. 1993, ApJ, 414, 740 Deliyannis, C. P., Steinhauer, A., & Jeffries, R. D. 2002, , 577, L39 Demarque, P., Woo, J.-H., Kim, Y.-C., & Yi, S. K. 2004, , 155, 667 ($\rm{Y}^2$) Dotter, A., Chaboyer, B., Jevremović, D., et al. 2008, , 178, 89 Dupret, M.-A., Grigahcéne, A., Garrido, R., Gabriel, M., & Scuflaire, R. 2004, , 414, L17 Folsom, C. P., Bouvier, J., Petit, P., et al. 2018, , 474, 4956 Gaia Collaboration, Brown, A. G. A., Vallenari, A., et al. 2016, , 595, A2 Gaia Collaboration, Brown. A. G. A., Vallenari, A., et al. 2018a, , 616, 1 (DR2) Gaia Collaboration, Babusiaux, C., van Leeuwen, F., et al. 2018b, , 616, 10 Garraffo, C., Drake, J. J., & Cohen, O. 2015, , 813, 40 Garraffo, C., Drake, J. J., Dotter, A., et al. 2018, , 862, 90 Garcia Lopez, R. J., & Spruit, H. C. 1991, , 377, 268 Gao, X., Lind, K., Amarsi, A. M., et al. 2018, , 481, 2666 Geller, A. M., Mathieu, R. D., Braden, E. K., et al. 2010, , 139, 1383 Gim, M., VandenBerg, D. A., Stetson, P. B., Hesser, J. E., & Zurek, D. R. 1998, , 110, 1318 Girardi, L. 1999, , 308, 818 Girardi, L., Eggenberger, P., & Miglio, A. 2011, , 412, L103 Girardi, L., Mermilliod, J.-C., & Carraro, G. 2000, , 354, 898 Girardi, L., Rubele, S., & Kerber, L. 2009, , 394, L74 Goudfrooij, P., Girardi, L., & Correnti, M. 2017, , 846, 22 Goudfrooij, P., Girardi, L., Kozhurina-Platais, V., et al. 2014, , 797, 35 Goudfrooij, P., Girardi, L., Rosenfeld, P., et al. 2015, , 450, 1693 Goudfrooij, P., Puzia, T. H., Kozhurina-Platais, V., & Chandar, R. 2009, , 137, 4988 Goudfrooij, P., Puzia, T. H., Kozhurina-Platais, V., & Chandar, R. 2011, , 737, 3 Gruyters, P., Korn, A. J., Richard, O., et al. 2013, , 555, 31 Gruyters, P., Lind, K., Richard, O., et al. 2016, , 589, 61 Gruyters, P., Nordlander, T., & Korn, A. 2014, , 567, 72 Guzik, J. A., Kaye, A. B., Bradley, P. A., Cox, A. N., & Neuforge, C. 2000, , 542, 57 Handberg, R., Brogaard, K., Miglio, A., et al. 2017, , 472, 979 Handler G. 1999, , 309, L19 Hole, K. T., Geller, A. M., Mathieu, R. D., et al. 2009, , 138, 159 (H09) Jackson, R. J., Deliyannis, C. P., Jeffries, R. D. 2018, , 476, 3245 Jeffries, R. D. 1997, , 292, 177 Jeffries, R. D., Totten, E. J., Harmer, S., & Deliyannis, C. P. 2002, , 336, 1109 Jeffries, M. W., Sandquist, E. L., Mathieu, R. D., et al. 2013 , 146, 58 King, J. R., Deliyannis, C. P., Hiltgen, D. D. et al. 1997, , 113, 1871 Kitiashvili, N., Kosovichev, A. G., Mansour, N. N., & Wray, A. A. 2016, , 821, L17 Kozhurina-Platais, V., Girard, M., Platais, I., & Van Altena, W. F. 1995, , 109, 672 Kraft, R. P. 1967, , 150, 551 Lanzafame, A. C., Distefano, E., Messina, S., et al. 2018, , 616, 16 Lee-Brown D. B. 2017 Artificial Neural Network Abundances (ANNA) v0.1.0, Zenodo, doi:10.5281/zenodo.1043585 Lee-Brown, D. 2018a, Ph.D. Thesis, University of Kansas Lee-Brown, D., Anthony-Twarog, B. J., Deliyannis, C. P., Rich, E., & Twarog, B. A. 2015, , 149, 121 (Paper II) Lee-Brown, D., Anthony-Twarog, B. J., & Twarog, B. A. 2018b, BAAS, 231, 213.05 Li, C., de Grijs, R., & Deng, L. 2014, , 516, 367 Li, C., de Grijs, R., Deng, L., & Milone, A. P. 2017, , 834, 156 Lind, K., Primas, F., Charbonnel, C., Grundahl, F. & Asplund, M. 2009, , 503, 545 Maccarone, T. J., Jonker, P. G., & Sills, A. I. 2005, , 435, 671 Mackey, A. D., & Broby Nielsen, P. 2007, , 379, 151 Mackey, A. D., Broby Nielsen, P., Ferguson, A. M. N., & Richardson, J. C. 2008, , 681, L17 Martocchia, S., Niederhofer, F., Dalessandro, E., et al. 2018, , 477, 4696 Mathieu, R. D. 2000, in ASP Conf. Series 198, [Stellar Clusters and Associations: Convection, Rotation, and Dynamos]{}, ed. R. Pallavicini, G. Micela, & S. Sciortino (San Francisco, CA: ASP), 517 McQuillan, A., Aigrain, S., & Mazeh, T. 2013, , 432, 1203 McQuillan, A., Mazeh, T., & Aigrain, S. 2013, , 775, L11 McQuillan, A., Mazeh, T., & Aigrain, S. 2014, , 211, 24 Meibom, S., Barnes, S. A., Latham, D. W., et al. 2011a, , 733, L9 Meibom, S., Barnes, S. A., Platais, I., et al. 2015, , 517, 589 Meibom, S., Mathieu, R. D., & Stassun, K. G. 2009, , 695, 679 Meibom, S., Mathieu, R. D., Stassun, K. G., Liebesny, P., & Saar, S. H. 2011b, , 733, 115 Michaud, G. 1986, , 302, 650 Milliman, K. E., Mathieu, R. D., Geller, A. M., et al. 2014, , 148, 38 Montalban, J., & Schatzman, E. 1996, , 305, 513 Newton, E. R., Irwin, J., Charbonneau, D., et al. 2016, , 821,93 Niederhofer, F., Bastian, N., Kozhurina-Platais, V., et al. 2016, , 586, 148 Niederhofer, F., Georgy, C., Bastian, N., & Ekström, S. 2015, , 453, 207 Nordlander, T., Korn, A. J., Richard, O., & Lind, K. 2012, , 753, 48 Noyes, R. W., Hartmann, L. W., Baliunas, S. L., Duncan, D. K., & Vaughan, A. H. 1984, , 279, 763 Pace, G., Castro, M., Meléndez, J., Theado, S., & Do Nascimento, J. D. Jr. 2012, , 541, 150 Pasquini, L., Randich, S., Zoccali, M., Hill, V., Charbonnel, C., & Nordström, B. 2004, , 424, 951 Pietrinferni, A., Cassisi, S., Salaris, M., & Castelli, F. 2004, , 612, 168 Pilachowski C. A., Saha A., & Hobbs L. M. 1988, , 100, 474 Pinsonneault, M. 1997, , 35, 557 Pinsonneault, M. 2010, in IAU Symp. 268, [Light Elements in the Universe]{}, ed. C. Charbonnel, M. Tosi, F. Primas, & C. Chiappini (Cambridge: Cambridge University Press), 375 Pinsonneault, M. H., Elsworth, .Y, Epstein, C., et al. 2014, , 215, 19 Pinsonneault, M. H., Kawaler, S. D., & Demarque, P. 1990, , 74, 501 Pinsonneault, M. H., Kawaler, S. D., Sofia, S., & Demarque, P. 1989 , 338, 424 Platais, I., Gosnell, N. M., Meibom, S., et al. 2013, , 146, 43 (PL) Ramírez, I., Fish, J. R., Lambert, D. L., & Allende Prieto, C. 2012, , 756, 46 Ramírez, I., Michel, R., Sefako, R., et al. 2012b, , 752, 5 Ramírez, I., & Meléndez, J. 2005, , 626, 465 Randich, S., Primas, F., Pasquini, L., Sestito, P., & Pallavicini, R. 2007, , 469, 163 Randich, S., Sestito, P., & Pallavicini, R. 2003, , 399, 133 Reiners, A., & Mohanty, S. 2012, , 746, 43 Reinhold, T., & Gizon, L. 2015, , 583, 65 Renson, P., & Manfroid, J. 2009, , 498, 961 Rich, E., Anthony-Twarog, B. J., Deliyannis, C. P., & Twarog, B. A. 2013, BAAS, 221, 250.27 Richer. J., & Michaud, G. 1993, , 416, 312 Riess, A. G., Casertano, S., Yuan, W., et al. 2018, , 861, 126 Rosvick, J., & VandenBerg, D. A. 1998, , 115, 1516 Royer, F., Gerbaldi, M., Faraggiana, R., & Gómez, A. E. 2002a, , 381, 105 Royer, F., Grenier, S., Baylac, M.-O., Gómez, A. E., & Zorec, J. 2002b, , 393, 897 Royer, F., Zorec, J., & Gómez, A. E. 2007, , 463, 671 Ryan, S. G., & Deliyannis, C. P. 1995, , 453, 819 Salinas, R., Pajkos, M. A., Vivas, A. K., Strader, J., & Contreras Ramos, R. 2018, , 155, 183 Sanders, W. L. 1972, , 19, 155 Sandquist, E. L., Mathieu, R. D., Brogaard, K., et al. 2013, , 762, 58 Schramm, D. N., Steigman, G., & Dearborn, D. S. P. 1990, , 259, 55 Sills, A., & Deliyannis, C. P. 2000, , 544, 944 Sestito, P., & Randich, S. (2005), , 442,615 Silva Aguirre, V., Davies, G. R., Basu, S., et al. 2015, , 452, 2127 Skumanich, A. 1972, , 171, 565 Slumstrup, D., Grundahl, F., Silva Aguirre, V., & Brogaard, K. 2019, , 622, 111 Sneden, C. 1973, , 184, 839 Soderblom, D. R., Duncan, D. K., & Johnson, D. R. H. 1991, , 375, 722 Soderblom, D. R., Jones, B. F., Balachandran, S., et al. 1993, , 106, 1059 Soderblom, D. R., Oey, M. S., Johnson, D. R. H., & Stone, R. P. S. 1990, , 99, 595 Somers, G., & Pinsonneault, M. H. 2014, , 790, 72 Somers, G., & Pinsonneault, M. H. 2015, , 449, 4131 Somers, G., & Pinsonneault, M. H. 2016, , 829, 32 Souto, D., Allende Prieto, C., Cunha, K., et al. 2019, , 874, 97 Spite, M., & Spite, F. 1982a, , 297, 483 Spite, F., & Spite, M. 1982b, , 115, 357 Stassun, K. G., & Torres, G. 2018, , 862, 61 Steinhauer, A. 2003, PhD thesis, Indiana University Steinhauer, A., & Deliyannis, C. P. 2004, , 614, L65 Stello, D., Meibom, S., Gilliland, R. L., et al. 2011, , 739, 13 Stelzer, B., Damasso, M., Scholz, A., & Matt, S. P. 2016, , 463, 1844 Swenson, F. J., & Faulkner, J. 1992, , 395, 654 Tkachenko, A., Aerts, C., Yakushechkin, A., et al. 2013, , 556, A52 Talon, S., & Charbonnel, C. 2003, , 405, 1025 Talon, S., & Charbonnel, C. 2005, , 440, 981 Terndrup, D. M., Pinsonneault, M., Jeffries, R. D., et al. 2002, , 576, 950 Thévenin,F., Oreshina, A. V., Baturin, V. A., et al. 2017, , 598, 64 Thorburn, J. A., Hobbs, L. M., Deliyannis, C. P., & Pinsonneault, M. H. 1993, , 415, 150 Tody, D. 1986, , 627,733 Tucci Maia, M., Meléndez, J., Castro, M., et al. 2015, , 576, L10 Twarog, B. A., Anthony-Twarog, B. J., Deliyannis, C. P., & Thomas, D. T. 2015, , 150, 134 Twarog, B. A., Anthony-Twarog, B. J., & Edgington-Giordano, F. 2009, , 121 1312 Twarog, B. A., Anthony-Twarog, B. J., & Schafer, B. S. 2012, BAAS 219, 438.09 VandenBerg, D. A., Bergbusch, P. A., & Dowler, P. D. 2006, , 162, 375 Wallerstein, G., Herbig, G. H., & Conti, P. S. 1965, , 141, 610 Wu, T., Li, Y., & Hekker, S. 2014a, , 781, 44 Wu, T., Li, Y., & Hekker, S. 2014b, , 786, 10 Wu, X., Li, C., de Grijs, R., & Deng, L. 2016, , 826, L14 Yang, S.-C., Sarajedini, A., Deliyannis, C. P., et al. 2013a, , 762, 3 Yang, W. 2018, , 860, 132 Yang, W., Bi, S., Meng, X., & Liu, Z. 2013b, , 776, 112 Yang, W., & Tian, Z. 2017, , 836, 102 Zahn, J.-P., & Bouchet, L. 1989, , 223, 112 Zinn, J. C., Pinsonneault, M. H., Huber, D., & Stello, D. 2018, arXiv: 1808.02650 Zorec, J., & Royer, F. 2012, , 537, 120 [^1]: The WIYN Observatory was a joint facility of the University of Wisconsin-Madison, Indiana University, Yale University, and the National Optical Astronomy Observatory. [^2]: https://github.com/dleebrown/ANNA [^3]: 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. [^4]: http:// webda.physics.muni.cz	{ "pile_set_name": "ArXiv" }
--- abstract: \| In this talk I describe recent joint works with R.Schoen and with G.Gibbons and R.Schoen which prove the non-existence of certain asymptotically flat, stationary solutions of the Einstein equations with more than one body. The basic restriction is for example satisfied when spacetime has an isometry reversing the sign of the timelike Killing vector and fixing a hypersurface in the space of Killing trajectories which is disjoint from the bodies. I also give a detailed treatment of the Newtonian situation. Keywords: stationary $n$-body problem address: 'Vienna University, Gravitational Physics, Faculty of Physics, Boltzmanngasse 5, A-1090 Vienna, Austria' author: - 'Robert Beig${}^{\dagger}$' date: ' [File:]{}' title: 'The stationary n-body problem in general relativity' --- [^1] Introduction ============ ’A configuration of several gravitating bodies can not be in stationary equilibrium’: this plausible expectation is in the context of GR still awaiting a rigorous proof. We will in this talk start by explaining the Newtonian situation. This can be summarized by the statement that there are no static solutions with more than one body, and where there exists a plane separating these bodies. This statement is well-known, but I have not seen a printed version. That some separation condition is needed in the statement of that result is clear from the fact that without it there are in fact solutions for the static 2 body problem, with source an elastic solid, both in the Newtonian [@bs] and in the Einstein theory [@as], for example with one small body put near a critical point of the potential of a large body, this critical point lying in a suitably hollow region near the large body.\ In the next section we consider a nonlinear generalization of the Newtonian theory due to Giulini. Here we are able to obtain a weaker result in the spirit of the general relativistic result in [@bgs]. Namely, we show there can not exist solutions where the potential has zero normal derivative on a plane lying in the vacuum region. This result, which is in the spirit of the GR result in [@bgs], in particular rules out several-body solutions which are reflection symmetric across a plane in the vacuum region. In the final section we describe one of the results in [@bgs]. This states that there can be no stationary spacetime having a reflection symmetry which inverts the orientation of the orbits of the timelike Killing vector and maps a timelike hypersurface tangential to these orbits into itself. This corresponds to a situation where the spin-spin interaction between matter on the two sides of this timelike hypersurface is attractive.\ In the case of two axially symmetric black coaxial black holes the problem treated here has also been studied [@w94], and an analogous result has been proved in [@lt]. In the absence of the discrete symmetry we are imposing, i.e. where there is the a priori possibility for spin-spin forces to balance the gravitational attraction, a full proof of nonexistence has been found very recently in [@nh], using methods from the theory of completely integrable systems. The Newtonian situation ======================= Let $\Omega_{(\alpha)}$ ($\alpha=1,2,..N$) be bounded open connected sets with smooth boundary $\partial \Omega_{(\alpha)}$ in $\mathbb{R}^3$ with Euclidean metric $\delta_{ij}$. The sets $\Omega_{(\alpha)}$ play the role of support of bodies. Each body has associated with it a positive function $\rho_{(\alpha)}$, the mass density, and the Cauchy stress tensor, a symmetric tensor field $\sigma_{\!(\alpha)ij}$. In a theory of continuum mechanics such as ideal elasticity, the quantities $\rho_{(\alpha)}, \sigma_{\!\!(\alpha)ij}$ are given, not as functions on $\mathbb{R}^3$, but as functions of some further fields such as maps from $\mathbb{R}^3$ into some ’material manifold’ $\mathcal{B}_\alpha$ and of the first derivatives of these maps. The basic equations then take the following form ($\Delta = \delta^{ij}\partial_i \partial_j$) $$\label{U} \Delta U = 4 \pi G \sum_\alpha \rho_{(\alpha)}\, \chi_{\Omega_{(\alpha)}}\,,\hspace{0.7cm}U \rightarrow 0\,\, \mathrm{at}\,\,\infty$$ with $\rho_{(\alpha)} > 0$ and $$\label{sigma} \partial^j \sigma_{\!(\alpha)ij} = \rho_{(\alpha)}\partial_i U\hspace{0.3cm}\mathrm{in}\,\,\Omega_{(\alpha)}\,,\hspace{0.7cm} \sigma_{\!(\alpha)i}^j n_j\|_{\partial \Omega_{(\alpha)}} = 0$$ The detailed structure of the matter equations (\[sigma\]) - which form a quasilinear second-order system of equations, usually elliptic, for the underlying map with Neumann-type boundary condition - need not occupy us here. The important thing is that these equations imply $$\label{total} M_\alpha(\xi,U) \Doteq \int_{\Omega_{(\alpha)}}\!\!\!\!\! \xi^i \partial_i U \,\,dx = 0\,,$$ where $\xi^i$ is any of the 6 Euclidean Killing vector, i.e. satisfies $\partial_{(i}\,\xi_{j)}=0$. The meaning of (\[total\]) is of course that the total force and the total torque acting on each body be zero. One easily deduces (\[total\]) from (\[sigma\]) by using integration by parts, the symmetry of the Cauchy stress together with the Killing equation and the boundary condition of vanishing normal stress in (\[sigma\]).\ The next question is if perhaps Eq.(\[U\]) alone implies these conditions. The answer is ’yes’ for $N=1$. Namely the ’Newtonian stress tensor’ defined by $$\label{theta} \Theta_{ij} = \frac{1}{4 \pi G}\,[(\partial_i U)(\partial_j U) - \frac{1}{2}\delta_{ij}(\partial U)^2]$$ satisfies, by virtue of (\[U\]), the relation $$\label{distr} \partial^j \Theta_{ij} = \sum_\alpha \rho_{(\alpha)} \chi_{\Omega_{(\alpha)}}\, \partial_i U\,.$$ Thus the gravitational equation (\[U\]) together with the matter equations (\[sigma\]) - including the boundary conditions - imply that $$\label{distr1} \partial^j(\Theta_{ij} - \sum_\alpha \sigma_{\!(\alpha)ij} \chi_{\Omega_{(\alpha)}})= 0$$ holds in the sense of distributions.\ Integrating (\[distr\]) against $\xi^i$ over $\mathbb{R}^3$ we get zero on the l.h. side: integrate by parts and use the Killing equation. The boundary term gives zero, since $\partial U = O(\|x\|^{-2})$. Thus $$\label{sum} \sum_\alpha M_\alpha(\xi,U) = 0$$ In fact, defining $U_{(\alpha)}$ to be the unique solution of $$\label{Ualpha} \Delta U_{(\alpha)} = 4 \pi G \rho_{(\alpha)}\, \chi_{\Omega_{(\alpha)}}\,,\hspace{0.7cm}U_{(\alpha)} \rightarrow 0\,\, \mathrm{at}\,\,\infty$$ we find, by applying the same argument first to $U_{(\alpha)}$ and then to $U_{(\alpha)} + U_{(\beta)}$, that $$\label{actreact} M_\alpha(\xi,U_{(\beta)}) + M_\beta(\xi,U_{(\alpha)}) = 0$$ This says that the total force and torque exerted by the gravitational field of body 1 on body 2 is minus that exerted by body 2 on body 1 - which of course is nothing but the law ’actio = reactio’. Eq.(\[actreact\]) also entails, for $\alpha = \beta$, the statement that the self-force and self-torque due to the gravitational field of a body vanishes.\ We now observe that (\[actreact\]) still falls short of guaranteeing (\[total\]) when $N>1$, because, for $N=2$ say, we need, in order for (\[total\]) to hold, in addition to $M_1(\xi,U_{(1)})$, $M_2(\xi,U_{(2)})$ and $M_1(\xi,U_{(2)}) + M_2(\xi,U_{(1)})$ all being zero, that also $M_1(\xi,U_{(2)})$ and $M_2(\xi,U_{(1)})$ vanish separately.\ Now, finally, we show that $M_1(\xi,U_{(2)})$ is indeed non-zero when $\Omega_{(1)}$ and $\Omega_{(2)}$ are separated by a plane. More precisely, when $\Omega_{(1)}$ lies to the left and $\Omega_{(2)}$ to the right of some separating plane $S$, we will show that $$\label{sign} \partial_n U_{(2)}\|_\Sigma > 0\,$$ on every plane $\Sigma$ parallel to and to the left of $S$ and where $n$ is the left normal. This, by (\[total\]), will clearly imply $M_1(\xi,U_{(2)})>0$, when $\xi$ is taken to be a translational Killing vector pointing along the left-normal of $\Sigma$. Inequality (\[sign\]) is nothing but the statement that the gravitational force due to body 2 on any body separated from body 2 by a plane is attractive. To prove (\[sign\]), we use an elegant argument due to R.Schoen. Consider, in region 2, i.e. to the right of $\Sigma$, the field $\bar{U}(x) = U_{(2)}(x) - U_{(2)}(\rho \circ x)$, where $\rho$ is the reflection across $\Sigma$. The field $\bar{U}$ satisfies $$\label{region2} \Delta \bar{U} = 4 \pi G \rho_{(\alpha)}\chi_{\Omega_{(\alpha)}}\hspace{0.4cm}\mathrm{in\,\,region\,\, 2}$$ Furthermore $\bar{U}$ goes to zero at infinity and vanishes on $\Sigma$. Thus, by the maximum principle, it follows that $\bar{U} < 0$. Then, by the Hopf lemma, $\partial_n \bar{U}\|_\Sigma = 2 \,\partial_n U_{(2)}\|_\Sigma > 0$. We have thus proved the\ There exists no solution to the equations (\[U\],\[sigma\]) with $N>1$ and where there is a plane $S$ separating the bodies. Let us note that all the results obtained here could have been proved using the explicit Green’s function of the flat Laplacian, namely that, for compactly supported $\rho$, the unique solution $U$ of $\Delta U = 4 \pi G \rho$ with $U$ decaying at infinity has the form $$\label{explicit} U(x) = - \,G \int_{\mathbb{R}^3} \frac{\rho(x')}{\|x - x'\|}\,\, dx'$$ We have avoided this and tried to make the argument as conceptual as possible. It is clear for example that the above result remains to be true, when the Laplacian $\Delta$ is replaced by ’meson-type’ operator $\Delta - \mu^2$, with $\mu = \mathrm{const}$. We have however made crucial use of the linearity of the Poisson equation, which allows to consider the concept of ’force on a body due to the gravitational field of some other body’. This is not any longer available in the following nonlinear modification of the Newtonian theory. The Giulini theory ================== This is a nonlinear modification of the Newtonian theory where $1/c^2$ of the gravitational self-energy has an active gravitational mass associated with it (the constant $c$ being the speed of light). The equations now are $$\label{g1} \Delta U = \frac{4 \pi G}{c^2}\,\sum_\alpha \epsilon_{(\alpha)}\chi_{\Omega_{(\alpha)}} - \frac{1}{2c^2}\,(\partial U)^2\,,\hspace{0.7cm}U \rightarrow 0\,\, \mathrm{at}\,\,\infty$$ $$\label{g2} e^{-\frac{U}{c^2}} \partial^j \left(e^{\frac{U}{c^2}} \sigma_{(\alpha)ij}\right) = \frac{1}{c^2}\,\, \epsilon_{(\alpha)}\,\partial_i U\hspace{0.3cm}\mathrm{in}\,\,\Omega_{(\alpha)}\,,\hspace{0.7cm} \sigma_{\!(\alpha)i}^j n_j\|_{\partial \Omega_{(\alpha)}} = 0$$ with $\epsilon_{(\alpha)} > 0$ playing the role of ’rest energy plus internal energy density’. The equation (\[g1\]) can be written in the linear form $\Delta \Psi = \frac{2 \pi G}{c^4}\, \Psi\,\sum_\alpha \epsilon_{(\alpha)}\chi_{\Omega_{(\alpha)}}$ in terms of $\Psi = e^\frac{U}{2 c^2}$, but that will play no role for us. The potential $U$ has the expansion $$\label{U1} U = - \frac{G m}{r} + O^\infty(\frac{1}{r^2})$$ The constant $m$ in (\[U1\]) is positive. To prove this, observe that $$\label{sphere} m = \frac{1}{4 \pi G}\lim_{R \rightarrow \infty} \int_{S_R}\!\! e^\frac{U}{2 c^2}\,\partial_i U\,dS^i\,$$ where $S_R$ is a large sphere. Now use the Gauss theorem together with (\[g1\]) on the r.h. side of (\[sphere\]).\ Again, in this theory, there is a law ’actio = reactio’ in the form $$\label{g3} \partial^j\, [e^\frac{U}{c^2}(\Theta_{ij} - \sum_\alpha \sigma_{\!(\alpha)ij} \chi_{\Omega_{(\alpha)}})]= 0$$ Now let $S$ be a plane disjoint from all bodies. We can then integrate (\[g3\]) against $\xi^i$ over one of the half-spaces bounded by $S$, where $\xi$ is the translational Killing vector coinciding with the unit normal $n$ of $S$. Using integration by parts together with (\[U1\]) and (\[theta\]) there results $$\label{g4} \frac{1}{2} \int_S e^\frac{U}{c^2}\,[(\partial_n U)^2 - (\partial^A U)(\partial_A U)]\, dS = 0\,$$ where $\partial_A$ is a partial derivative tangential to $S$. Suppose that $\partial_n U\|_S = 0$. It then follows that $S$ is an equipotential for $U$. But this is impossible due to (\[U1\]) and the fact that $m \neq 0$. Thus we have obtained the\ If a solution of (\[g1\],\[g2\]) contains a plane $S$ disjoint from the bodies on which $\partial_n U$ is zero, this has to be the trivial (’no-body’-) solution with $U=0$. The hypothesis of the above theorem is fulfilled when there is a reflection across some plane $S$ disjoint from the bodies which maps the set of bodies into itself. It is this theorem - which is weaker than that available in the ’pure’ Newtonian case - for which there exists an analogue in GR, to which we now turn. The GR situation ================ The result I am now explaining works for the stationary Einstein equations with sources compactly supported in space. The nature of these sources is completely irrelevant except that they should meet the requirements of the positive-energy theorem, i.e. obey the dominant-energy condition[^2]. Otherwise the field equations will only be used in the vacuum region. For simplicity we will restrict to $3+1$ dimensions, although an analogous results for $n+1$ with $n>3$ is also true. We consider stationary - not just static - case, to allow for spin-spin interactions. Thus we assume $(M, ds ^2)$ has a timelike Killing vector $\xi^\mu$ with complete orbits. Then $ds^2$ can be written as (in distinction to the previous section, we set $c$ equal to 1) $$\label{metric} ds^2 = - e^{2 U} (dt + \psi_i dx^i)^2 + e^{- 2 U}\, h_{ij} dx^i dx^j$$ with $h$ being a Riemannian metric on $N$, the quotient space under the action of $\xi = \partial_t$. We refer to $U$ as the gravitational potential and define $1/2$ the curvature of the Sagnac connection $\psi$ by $$\label{sagnac} \omega = \frac{1}{2}\, d \psi,\hspace{1.2cm}\omega_{ij} = \partial_{[i} \psi_{j]}$$ the vacuum field equations turn out to be ($\mathcal{G}_{ij} = \mathcal{R}_{ij} - \frac{1}{2} h_{ij} \mathcal{R}$) \[field\] $$\begin{aligned} \mathcal{G}_{ij} - 8 \pi G\,(\Theta_{ij} + \Omega_{ij})&= 0\\ \Delta_h U + e^{4 U}\,\omega_{kl}\omega^{kl}&=0\\ D^j\left(e^{4 U} \,\omega_{ij}\right)& = 0\end{aligned}$$ In (\[field\]a) we are using the definitions $$\label{def1} 8 \pi G \,\Theta_{ij} = 2\, [(D_i U) (D_j U) - \frac{1}{2}h_{ij}(DU)^2]$$ and $$\label{def2} 8 \pi G\,\Omega_{ij} = 2 \,e^{4 U}\,[\, - \omega_{ik}\omega_{j}{}^k + \frac{1}{4} \,h_{ij}\, \omega_{kl}\omega^{kl}]$$ (Using $d \omega = 0$, Eq.(\[field\]c) follows from Eq.’s (\[field\]a,b).) We consider asymptotically flat solutions of (\[field\]). They can be shown [@bsi] to have the form \[as\] $$\begin{aligned} h_{ij}&=\delta_{ij} + O^\infty\left(\frac{1}{r^2}\right)\\ U &= - \frac{G m}{r} + O^\infty\left(\frac{1}{r^2}\right)\\ \omega_{ij} &= - G\, \frac{L_{ij} + 3 \,a_{[i}L_{j]k}a^k}{r^3} + O^\infty\left(\frac{1}{r^4}\right)\end{aligned}$$ in suitable coordinates. Here $a^i = \frac{x^i}{r}$. Furthermore the constants $m$ and $L_{ij}=L_{[ij]}$ are respectively the ADM mass and the (mass-centered) spin tensor of the configuration. When $m=0$, the ADM energy of the initial data set induced on $t=\mathrm{const}$ is also zero. Then, when the source satisfies the dominant energy condition, the positive energy theorem implies that spacetime is flat[^3]. Let $(N,h)$ have a hypersurface $S$ disjoint from the matter region, which is non-compact, closed and totally geodesic w.r. to the unrescaled metric $e^{- 2 U} h_{ij}$. Suppose in addition that the pull-back-to-$S$ of $\omega$ is zero. Then spacetime is flat. When spacetime has more than one asymptotically flat end, all the statements above hold separately w.r. to any such end. We first have to comment on the surface $S$ in the hypothesis of the theorem. The typical situation where our hypothesis on $S$ will be satisfied is when there is a reflection isometry $\Psi$ of spacetime which maps $\xi$ into $ - \xi$ (and thus projects down to an isometry of $(N, h_{ij})$ and $(N, e^{- 2 U} h_{ij})$) and leaves $S$ invariant [^4].\ In [@bs] it is shown in detail that $S$ has to have a finite number of ends in the asymptotically flat region and $S$ approaches a plane there in a precise sense. Thus, for each end of $S$ asymptotic coordinates $(x^1, x^A)$ in (\[as\]) can be chosen so that each end of $S$ is given by $x^1 = 0$. It then follows from the hypothesis on $\omega$ that $L_{AB}$ in (\[as\]c) vanishes. There are then good physical reasons detailed in [@bgs] to believe that this corresponds to a situation where the spin-spin interaction between the bodies on the two sides of $S$ is attractive.\ For the proof of the theorem we contract (\[field\]a) with $n^i n^j$, with $n^i$ being the unit normal to $S$ in $(N,h)$. The result is $$\label{nn} \mathcal{G}_{nn} = 2 \left[\frac{1}{2}\,(D_n U)^2 - \frac{1}{2}\,(D_A U)(D^A U)\right] - \,e^{4 U} \,\omega_{A n} \omega^A{}_n\,\,.$$ Here $D_A$ is the intrinsic derivative on $S$. In (\[nn\]) we have used that $\omega_{AB}$ is zero. Now, from the Gauss equation, $$\label{gauss} \mathscr{R} = - 2\, \mathcal{G}_{nn} + (\mathrm{tr}\, k)^2 - \mathrm{tr}(k^2)\,,$$ where $\mathscr{R}$ is the Ricci scalar of $S$ and $k$ its extrinsic curvature. Since $S$ is totally geodesic w.r. to $e^{- 2 U} h$, we have that $$\label{bar} k_{AB} = \,\, q_{AB} \,D_n U$$ where $q$ is the intrinsic metric on $S$ induced from $h$. Inserting (\[bar\],\[nn\]) into Eq.(\[gauss\]), the terms involving $D_n U$ cancel so that finally $$\label{pos} \mathscr{R} = 2\, (D_A U)(D^A U) + 2 \,e^{4 U}\,\omega_{A n}\omega^A {}_n\;.$$ In particular $\mathscr{R}$ is non-negative. One now integrates (\[pos\]) over $S$. Using the version of the Gauss-Bonnet formula for surfaces with boundary and the asymptotically planar nature of $S$, one arrives at a contradiction except when $\mathscr{R}$ vanishes and $D_A U$, whence $U$, is zero on $S$. Thus by (\[as\]) $m$ is zero, whence spacetime is flat, and we are done.\ We remark that for the $n+1$-dimensional version of this theorem one uses, instead of the Gauss-Bonnet formula, the positive energy theorem in dimension $n-1$[@s]. [XXX]{} Andersson, L. and Schmidt, B.G. (2009) [[Static self-gravitating many-body systems in Einstein gravity]{}]{}, Class.Quant.Grav. [[26]{}]{} 165007 (15pp) \[arXiv:0905.1243 \[gr-qc\]\] Bartnik, R.A. and Chrusciel, P.T., (2005) [[Boundary value problems for Dirac-type equations]{}]{} J. Reine Angew. Math.[[579]{}]{}, 13-73 \[arXiv:math/0307278\] Beig, R. and Simon, W. (1980) [[The stationary gravitational field near spatial infinity]{}]{}, Gen.Rel.Grav. [[12]{}]{}, 1003-1013 Beig, R. and Schmidt, B.G. (2008) [[Celestial mechanics of elastic bodies]{}]{}, Math. Z. [258]{} 381–394, \[arXiv:gr-qc/0612189\] Beig, R. and Schoen, R.M. (2009) [[On Static n-Body Configurations in Relativity]{}]{}, Class.Quant.Grav.[[26]{}]{} 075014 (7pp), \[arXiv:0811.1727 \[gr-qc\]\] Beig, R., Gibbons, G.W., and Schoen, R.M., (2009) [[Gravitating Opposites Attract]{}]{}, Class.Quant.Grav.[[26]{}]{} 225013 (12pp), \[arXiv:0907.1193 \[gr-qc\]\] Gibbons, G.W., Hawking, S.W., Horowitz, G.T., and Perry, M.J. (1983) [[Positive mass theorems for black holes]{}]{}, Commun.Math.Phys. [[88]{}]{}, 295-308 Giulini, D. (1997) [[Consistently Implementing the Fields Self-Energy in Newtonian Gravity]{}]{}, Phys.Lett. [[A232]{}]{} 165-170, \[arXiv:gr-qc/9605011\] Li, T.G. and Tian, G. (1991) [[Nonexistence of axially symmetric stationary solution of Einstein vacuum equation with disconnected symmetric event horizon]{}]{}, manuscripta math. [[73]{}]{} 83-89 Neugebauer, G. and Hennig, J. (2009) [[Non-existence of stationary two-black-hole configurations]{}]{}, Gen.Rel.Grav. [[41]{}]{} 2113-2130, \[arXiv:0905.4179 \[gr-qc\]\] Schoen, R.M., in preparation Weinstein, G. (1994) [[On the Force Between Rotating Coaxial Black Holes]{}]{}, Trans.Amer.Math.Soc. [[343]{}]{} 899-906 [^1]: ${}^{\dagger}$ Supported in part by Fonds zur Förderung der Wissenschaftlichen Forschung project no. P20414-N16. [^2]: Actually, all we need is that the ADM energy in the asymptotic rest system of the Killing vector be nonzero. [^3]: One can also allow for the presence of horizons, see [@ghp],[@bc]. [^4]: In fact the presence of the above isometry also implies $D_n U$ to vanish on $S$ - a property we do not require in the theorem.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We present a high angular resolution ($\sim 0.2\arcsec$), high sensitivity ($\sigma \sim 0.2$ mJy) survey of the 870 $\mu$m continuum emission from the circumstellar material around 49 pre-main sequence stars in the $\rho$ Ophiuchus molecular cloud. Because most millimeter instruments have resided in the northern hemisphere, this represents the largest high-resolution, millimeter-wave survey of the circumstellar disk content of this cloud. Our survey of 49 systems comprises 63 stars; we detect disks associated with 29 single sources, 11 binaries, 3 triple systems and 4 transition disks. We present flux and radius distributions for these systems; in particular, this is the first presentation of a reasonably complete probability distribution of disk radii at millimeter-wavelengths. We also compare the flux distribution of these protoplanetary disks with that of the disk population of the Taurus-Auriga molecular cloud. We find that disks in binaries are both significantly smaller and have much less flux than their counterparts around isolated stars. We compute truncation calculations on our binary sources and find that these disks are too small to have been affected by tidal truncation and posit some explanations for this. Lastly, our survey found 3 candidate gapped disks, one of which is a newly identified transition disk with no signature of a dip in infrared excess in extant observations.' author: - 'Erin G. Cox' - 'Robert J. Harris' - 'Leslie W. Looney' - 'Hsin-Fang Chiang' - Claire Chandler - Kaitlin Kratter - 'Zhi-Yun Li' - Laura Perez - John Tobin title: 'Protoplanetary Disks in $\rho$ Ophiuchus as Seen From ALMA' --- Introduction {#sec:intro} ============ In recent years, there has been an explosion in the detections of extra-solar planets. As of early 2017, there are nearly 3000 exoplanets confirmed and another 2500 candidate exoplanets (e.g., exoplanets.org). These planets show a great diversity of properties including masses, sizes, and architectures. In fact, many of these systems have planets that are unlike our solar system, such as super Earths, hot Jupiters, or hot Neptunes [@ch13]. The diversity of the planet population is likely some combination of differences in the initial conditions during the evolution of the circumstellar disk in which the planetary system forms and the necessary random interactions or scattering events during the planetary growth process [@bi11; @bi15]. To better understand the origin of the planet diversity, we therefore need to explore the inherent diversity in the circumstellar disks around young stellar objects (hereafter, YSOs). By directly observing the environments in which young planetesimals are expected to form, we can characterize the initial conditions of these other worlds. To explore these early conditions, we must observe the protostar at the evolutionary phases that likely have the largest impact on planet evolution. While the exact phase is still unknown, the protostar must have evolved to the point where a large mass reservoir, i.e., a protoplanetary disk, surrounds the star. A protostar’s evolutionary path can be divided into 4 parts- Class 0 – III [e.g., @la87; @an93; @du14]. During the initial collapse, i.e. the Class 0 phase, the protostar is engulfed in a large envelope full of nascent dust and gas. By the Class I phase, most of the envelope material has been funneled onto the central protostar through a circumstellar disk. During the Class II phase, the protostar no longer has its nascent envelope surrounding it, and the majority of the circumstellar material is in a large disk. Lastly, during the final phase of the protostar, Class III, the protostar has essentially accreted all of its final mass, leaving a very tenuous (if any) circumstellar disk left [@an05; @an07 e.g.]. It is well known that planets form in the disk surrounding forming protostars, and it is commonly thought that most of planet formation happens during the Class II phase of evolution. This is due to the fact that, by this time, the majority of the remaining gas and dust are surrounding the central protostar in a disk, allowing a large reservoir for planetesimals to form and evolve. While there is overwhelming indirect evidence for planet formation in disks, direct imaging of forming protoplanets has been scarce, with few examples in the literature (LkCa 15, @ki12; FW Tau; ROXs 12; ROXs 42B, @kr14). However, recent Atacama Large Millimeter/submillimeter Array (ALMA) observations of protoplanetary disks are beginning to reveal likely indicators of ongoing planet formation, such as the gaps in the millimeter disks of HL Tau [@al15], a Class I/II protostar, and of TW Hydra [@an16], a Class II protostar. $\rho$ Ophiuchus is an ideal laboratory for studying star and planet formation for several reasons. First, it is relatively close ($d \sim 137$ pc; @or17); second, it is relatively young (between 0.5 - 2 Myr; @wi08); finally, it has a large number of confirmed/candidate members ($\gtrsim 300$; @wi08). Despite these advantages, there are few millimeter-wave studies of its disk population that are representative of the disk content of Oph. One reason for this is that the stellar population is not well-characterized or studied: with $A_v$ ranging from 1 to 100 across the cloud, an accurate/representative stellar census has not been possible to date, despite many optical/IR surveys of different parts of the cloud (see, e.g, @ba03 and references therein), making connection to host star properties difficult. Another reason is that Oph lies far in the southern hemisphere, making it somewhat challenging to observe with northern instruments. Of the few large-scale surveys toward Oph, most have been done with single-dish telescopes, and thus are potentially confused by cloud contamination, companion stars, etc. The first studies of the Oph cloud core [@an90; @le91] showed an abundance of millimeter/centimeter-bright, deeply embedded objects residing in the dense core. Subsequent systematic studies of both the cloud core and surrounding regions [@an94; @an07] demonstrated that millimeter flux tends to decline with class, signifying circumstellar mass depletion during evolution (either through accretion or outflow or dispersion, by, e.g., photo-evaporation), and also that the millimeter spectral index tends to decline as well, most likely indicating grain growth in the circumstellar dust [e.g., @ri10]. Subsequent work at high resolution with both the Submillimeter Array (SMA) and ALMA have yielded more details by probing the disk structures in Oph at sub-arcsecond resolution. These studies have, however, focused principally on either the detailed structure of the transitional disk population of Oph [@an09; @an10; @pe14] or other special (i.e., bright) objects [@pe12; @sa14]. Despite these studies of special sub-populations of Oph disks, there has, to date, been no systematic study at high-resolution ($<$ 0.2) of the disks of the $\rho$ Ophiuchi cloud complex. In this article, we present the results from our ALMA 870 $\mu$m survey of $\sim$50 evolved disks in $\rho$ Ophiuchus. Sample Selection {#sec:sam} ================ One of the main goals of this program is to observe the compact disk dust emission toward a large sample of sources that does not have the inherent biases of previously known millimeter flux detections. To achieve such a large and representative sample of sources, we used the Spitzer c2d catalog of YSO candidate sources in $\rho$ Ophiuchus [@ev03], which requires S/N $\geq$ 3 in all the 4 IRAC bands and the 24 $\mu$m MIPS band. This criteria yields 297 protostellar sources. To increase the likelihood of detectable circumstellar mass (i.e. long wavelength dust emission), we narrowed the sample to sources with 70 $\mu$m MIPS band detection S/N $>$ 2. This requirement removed mostly the older source population (e.g. Class III objects based on SED fitting between 2-24 $\mu$m) and other sources that have low-mass disks due to other factors (i.e. environment, system mass, etc.), including 18 Flat and 10 Class I sources, which left 64 sources. Finally, as this project is focusing on the more evolved sources without significant envelope emission, we also removed the sources that were known embedded sources from Young et al. (2006). This resulted in a total of 50 sources in our sample. These sources were then compared to Herschel PAC continuum maps at 70 and 100 $\mu$m to verify that the sources all had far-infrared emission. While doing this, it was realized that one of the sources was a clear galaxy (J163524.3-243359) and another was offset by exactly 1 arcminute (J162646.4-241160), which was likely a typo in the c2d catalog and is now corrected. The final source list of 49 sources with their YSO class from the c2d catalog are given in Table 1. [llcc\|llccc]{} &\ (r)[1-4]{} (r)[5-9]{} Field name & Alt name & & Ref & Field name & Alt name & Separation (arcsec) & PA ($^\circ$) & Ref\ ROph3 & IRAS 16201-2410 & & e & ROph2 & V 935 Sco& 0.02 & …& d\ ROph4 & & & e & ROph5 & WSB 19 & 1.49 & 262.9 & a\ ROph8 & DoAr 25 & & a,c& ROph6 & DoAr 21 & $\sim 0.005$ & …& g\ ROph9 & El 24 & & a,b,c & ROph7 & DoAr24 E& 2.03 & 150 & a\ ROph10 & GY 33 & & a & ROph12 & WSB 40 & 0.017 & …& d\ ROph14 & GY 211 & & c & ROph21 & SR 9 & 0.638 & 353.3 & a\ ROph15 & GY 224 & & a,b & ROph26 & ROXs 42C & 0.277 & 151 & a\ ROph16 & GY 235 & & a & ROph27 & WSB 71 & 3.56 & 35.0 & a\ ROph17 & GY 284 & & a & ROph32 & WSB 74 & $\lesssim 0.043$ & …& e\ ROph18 & YLW 47 & & a,c & ROph33 & DoAr 51 & 0.784 & 79.3 & a\ ROph19 &DoAr 33 & &a,c & ROph34 & L1689-IRS7 & 7.56 & 334.9 & h\ ROph20 & GY 314 & &a,c & ROph36 & & 0.025 & …& d\ ROph22 & SR 20 W & &a,c & ROph45 & IRS 54 & 7.17 & 323.1 & b\ ROph24 & WSB 63 & &a,c &\ (r)[5-9]{} ROph25 & WSB 67 & &a & Field name & Alt name & Separation (arcsec) & PA($^\circ$) & Ref\ (r)[5-9]{} ROph29 & DoAr 44 & &a,c,f & ROph11 & WSB 38 Aa-Ab & 0.098 & 24.2 & a\ ROph35 & Haro 1-17& &a & & WSB 38 Aab-B & 0.577 & 105.4 & a\ ROph40 & & & a & ROph13 & SR 24 Aab & 0.197 & 84 & a\ ROph41 & WL6 & &a,b & & SR 24 Aab-B & 5.065 & 349 & a\ ROph42 & GY 312 & &b & ROph23 & SR 13 Aa-Ab & 0.013 & ...$^{\dagger}$ & a\ ROph43 & & &b & & SR 13 Aab-B & 0.399 & 96 & a\ ROph44 & GY 344 & &b & ROph31 & L1689-IRS5 A-Bab & 3.0 & 241 & a\ ROph46 & WSB 60 & &a & & L1689-IRS5 Ba-Bb & 0.14 & 84.4 & a\ ROph48 & IRS 63 & &a &\ ROph50 & Haro 1-11 & &a,c &\ \ (r)[1-4]{} Field name & Alt name & Field Name & Alt Name\ (r)[1-4]{} ROph1 & & ROph39 &\ ROph28& & ROph47 &\ ROph30& & ROph49 &\ ROph38& WSB 82 & &\ $^{\dagger}$ binary orbits with a period of $\sim$ years, so the position angle depends sensitively on observation epoch. Reference key: \[a\] @ra05, \[b\] @du04 , \[c\] @ch15, \[d\] @rr16, \[e\] @ko16, \[f\] @wi16, \[g\] @lo08, \[h\] This work \[tab:multtab\] Because we select for sources that have infrared excesses in each of the IRAC and MIPs bands, we preferentially observe sources with a substantial disk reservoir. Since mass estimates at longer wavelengths are less affected by optical depth than those at shorter wavelengths, we attempt to quantify this bias by computing model disk fluxes at 70 $\mu$m and comparing them to the observed MIPS 70 $\mu$m fluxes in our sample. To do this, we assume the standard analytic prescription for a viscously-evolving, geometrically-thin disk [@ly74; @ha98], a radial power-law in temperature, and a power-law in frequency for the total (i.e., gas + dust) opacity [@hi83], i.e. $$\begin{aligned} \Sigma(r) &\propto& \left(\frac{r}{r_c} \right)^{-\gamma} \exp{-(\frac{r}{r_c})^{2-\gamma}}\\ T(r) &=& T_{\mathrm{1 au}} \left(\frac{r}{\mathrm{1 AU}}\right)^{-q} \\ \kappa_{\nu} &=& 0.03 \left( \frac{\lambda}{870 \mu\mathrm{m}}\right)^{-\beta}\mathrm{cm}^2\mathrm{g}^{-1}\\\end{aligned}$$ with $T_{\mathrm{1 AU}} = 280$ K, $q = 0.5$, $\beta = 1$, $r_c = 100$ au, and $\gamma = 1$. These values and expressions are roughly appropriate for these disks as observed in the (sub)-millimeter (e.g., @hu08 [@an09; @an10]), although their applicability to the mid/far-infrared is uncertain. The median uncertainty for the c2d survey of Ophiuchus at 70 $\mu$m is approximately 25 mJy, so our 70 $\mu$m selection criteria selects sources with fluxes in excess of $\sim 50$ mJy at 70 $\mu$m. Using these relations, we estimate that our sources all have $\gtrsim 0.2 - 1$ Jupiter mass worth of circumstellar material (gas + dust), depending on the exact values for the quoted values above, as well as the relatively uncertain gas-to-dust ratio used for the computation of the opacity. ---------------------------------- ------------ ------------- --------------------------- ----------------- ------------- -------------------------- (r)[1-3]{} (r)[4-7]{} Field name Alt name IR band/ref Field name Alt name band/ref True disk or binary/ ref ROph1 N \[a\] ROph2 V 935 Sco N \[a\] CB \[2\] ROph5 WSB 19 N \[a\] ROph 3$^{\dagger\dagger}$ IRAS 16201-2410 N,S \[c,f\] T \[3\] ROph7 DoAr 24 E N \[a\] ROph4 N \[a\] T \[3\] ROph9 El 24 N,M \[a,b\] ROph6 DoAr 21 N,M \[a,b\] CB \[4\] ROph10 GY 33 N \[a\] ROph8 DoAr 25 N,M \[a,b\] T \[2,3\] ROph14 GY 211 N \[a\] ROph11 WSB 38 N \[a\] T \[3\] ROph15 GY 224 N \[a\] ROph12 WSB 40 M \[b\] T \[2\] ROph16 GY 235 N \[a\] ROph13 SR 24 M,S \[b,d\] T \[2\] ROph25 WSB 67 N \[a\] ROph17 GY 284 M \[b\] T \[7\] ROph27 WSB 71 N \[a\] ROph18 YLW 47 M \[b\] T \[5\] ROph28 N \[a\] ROph19 DoAr 33 N \[a\] T \[2,3,5\] ROph30 N \[a\] ROph20 GY 314 M \[b\] T \[5\] ROph31 L1689-IRS5 N \[a\] ROph21 SR 9 N,M \[a,b\] T \[3,5\] ROph33 DoAr 51 N \[a\] ROph22 SR 20 W M \[b\] T \[5\] ROph34 L1689-IRS7 N \[a\] ROph23 SR 13 M \[b\] CB \[1\] ROph35 Haro 1-17 N \[a\] ROph24 WSB 63 N \[a\] T \[2,3,5\] ROph39 N \[a\] ROph26 ROXs 42C N \[a\] T \[5\] ROph40 ISO-Oph 51 N,M \[a,b\] ROph29$^\dagger$ DoAr 44 S \[e\] T \[2,5,6\] ROph41 WL 6 N \[a\] ROph32 WSB 74 N \[a\] CB \[3\] ROph42 GY 312 N \[a\] ROph36 N \[a\] CB \[2\] ROph43 N \[a\] ROph38 WSB 82 S \[f\] T \[7\] ROph44 GY 344 N \[a\] ROph46 WSB 60 M \[b\] T \[7\] ROph45 IRS 54 N \[a\] ROph50 Haro 1-17 N \[a\] T \[7\] ROph47 N \[a\] ROph48 IRS 63 N \[a\] ROph49 N \[a\] ---------------------------------- ------------ ------------- --------------------------- ----------------- ------------- -------------------------- : Transition vs. non-transition disks Abbreviation key: N - near/mid-infrared colors; M - mid/far-infrared colors; S - (sub)mm-wave imaging of cavities; T - no indication of interloping circumbinary disk, CB - indication that disk is circumbinary, not transitional. $\dagger$: this source was classified as a pre-transitional disk by @es10 but did not meet the color criterion to be a transitional disk according to @ci10. We treat it as a non-transition disk here, for consistency. $\dagger\dagger$: this source was classified as a transitional disk on the basis of Spitzer IRS spectra by [@fu09] but, as with ROph 29, the colors did not meet the criteria of @ci10. Reference key: \[a\] @ci10, \[b\] @re15, \[c\] @fu09, \[d\] @an05b, \[e\] @an09, \[f\] this work, \[1\] @ra05 \[2\] @rr16 , \[3\] @ko16, \[4\] @lo08, \[5\] @ch15, \[6\] @wi16, \[7\] assumed transition based on lack of data. \[tab:multtab2\] Out of the 49 targets selected, 12 – ROph 2, 4, 6, 8, 11, 19, 21, 24, 26, 32, 36, and 50 – were identified by [@ci10] to be candidate transitional disks on the basis of Spitzer near-/mid-infrared colors, eleven sources – ROph 6, 8, 12, 13, 17, 18, 20, 21, 22, 23, and 46 – were identified by [@re15] on the basis of Spitzer/Herschel mid-/far-infrared colors. Of these nineteen total transitional disk candidates, three – ROph 2, 12, and 36 – were discovered by [@rr16] to harbor tight stellar binaries ($a \lesssim 3$ au) whose infrared signature mimicked that of transitional disks, and one – ROph 32 – was discovered by [@ko16] to be a spectroscopic binary with $a \sim 0.6$ au. ROph 6 was found to be a very tight ($\sim 5$ milliarcsecond) binary by @lo08; its mid-infrared color is due to the presence of a hot ring of material at a large distance from the star, and is most likely not indicative a true transitional disk. This leaves fourteen candidates that are ‘bona fide’ transition disks with no evidence of being binary interlopers, see Table \[tab:multtab2\]. One caveat to keep in mind for this survey is the impact of unresolved (or unknown) multiplicity in the targets. $\rho$ Ophiuchus has been the target of several optical and infrared surveys of varying completeness in the past three decades, both targeting Class I/Flat [@ba04; @ha04; @du04; @ha06; @du07] and Class II [@gh93; @ra05] sources. It is known that stellar companions can have a dramatic effect on circumstellar material via tidal interactions that preferentially strip away outer disk material in circumstellar disks and inner disk material in circumbinary disks [@ar94; @pi05; @pi08]. Observationally, truncation manifests as a decreased likelihood of an infrared excess in multiple systems as opposed to isolated stars [reflecting absence of an inner disk, @ci09; @kr11] or decreased millimeter-wave continuum emission . Such signatures, if unrecognized, can bias the results of infrared and millimeter surveys of protoplanetary disks. To mitigate the effect of this in our sample, we have surveyed the available literature on multiplicity in Ophiuchus to identify which of our targets are multiple systems. Unfortunately, the principal surveys we used provide different sensitivities to various separations on the sky and give fairly heterogeneous coverage. @ra05 conducted a magnitude limited (K $\leq$ 10.5) speckle imaging survey of 158 principally Class II objects and is sensitive to companions with separations between roughly 0.1$\arcsec$ and 6.4, down to a contrast ratio of 0.1. @du04 [@du07] conducted a direct imaging survey of principally class I/Flat objects in the mid-infrared with coverage ranging from 0.8$\arcsec$ to 10.0. Because most surveys are flux-limited, several of the lower-luminosity sources in our sample have not been observed in these surveys. Of the 49 sources in our sample, 10 have not been observed in any available survey (ROph1, 3, 4, 28, 30, 38, 39, 42, 47, 49). Table 1 summaries the multiplicity status of each system in our sample. Observations and Data Reduction {#sec:obs} =============================== The ALMA Band 7 observations were taken under proposal 2013.1.00157.S using a continuum only setup to maximize the dust continuum sensitivity in two configurations for the snapshot survey. The lower resolution observations were obtained on 2015 April, 4 in ALMA configuration C34-1/(2) for $\sim$30 minutes of total time, which was about 12 seconds of integration time on each source. The C34-1/(2) configuration baselines ranged from 14 to 356.3 meters with typical recoverable scale of 8.4. The higher resolution observations were obtained on 2015 July, 24 in ALMA configuration C34-7/(6) for $\sim$47 minutes of total time, which was about 24 seconds of integration time on each source. The C34-7/(6) configuration baselines ranges from 42 to 1574 meters with typical recoverable scale of 2.6. In both observations, the 4 continuum bands were centered at 336.5, 338.4, 348.5, and 350.5 GHz. The quasars J1517-2422 and J1625-2527 were used for bandpass and phase calibration, and Titan was used for flux calibration. In this paper, we assume an absolute flux calibration uncertainty of $\sim$10%, but only statistical uncertainties are considered. The observations were reduced using the Common Astronomy Software Applications (CASA) package [@mc07] using the 4.7.0 CASA and ALMA pipeline package. Briefly, the pipeline first applies a priori calibrations, such as baseline corrections and phase corrections from water vapor radiometer measurements. Then, it conducts a standard interferometric reduction: bandpass calibration, flux calibration, and antenna gain calibrations. These calibrations are performed separately for each of the two observations and each science target field was subsequently split out of its parent dataset. For the C34-1/(2) configuration observations, we performed a phase only selfcal over the integration time for the sources with $\geq$10$\sigma$ detections to improve the S/N in the maps. The C34-7/(6) configuration map S/N was not improved from selfcal, so the selfcal gains were not used. After the final images were made for each configuration, we checked to ensure that the fluxes measured for each observation were consistent; finding that they were, we combined the two datasets and used the combined datasets to produce the images analyzed in this work. To produce the final images, we imaged the data using the `CLEAN` task in CASA. The data were imaged with natural weighting to produce a typical resolution of 0.21$\arcsec$ by 0.18. Many of our sources are relatively compact (see Fig. 1 and 2) and standard `CLEAN` was sufficient to deconvolve the sources successfully. However, for many of the more extended disks, standard `CLEAN` left substantial deconvolution errors in the residual maps, so we used the multi-scale version of the algorithm to produce images of some of the disks in Figures \[fig:singles\] and \[fig:triples\] as well as all of the transition disks in our sample (Figure \[fig:trans\]). The use of the multi-scale `CLEAN` algorithm yielded residual maps that were dominated by Gaussian noise. Results {#sec:res} ======= This survey provided very well-resolved images of the diverse population of protoplanetary disks in $\rho$ Oph YSOs. Figures \[fig:singles\]-\[fig:trans\] show the different YSOs divided into single sources, binaries, triple systems and transition disks. The sources that do not have multiplicity information are considered single, unless they show evidence of being a transition disk. Since transition disks are separated into their category, we do not include them in any other categories (i.e., singles or triples). Each figure uses the same stretch for flux values, such that the brightest sources show the deepest red color. At a glance, it is obvious that our sample is not only composed of different types of systems, but also each type shows great diversity in size, brightness, and flux distribution. In the 49 stellar systems that we targeted, there were 63 stars, and disks associated with 13 stars were not detected: 4 around single stars, 4 around components of binaries, and 5 around components of triple systems. Table \[tab:source\] summarizes this information for all sources, including classifications of the YSOs from c2d. The disk sizes and position angles, as well as the peak and integrated fluxes, were estimated by fitting a gaussian in the image plane using the CASA task `imfit`. Disk masses were estimated from the integrated fluxes by assuming optically thin emission and an isothermal disk with a dust temperature of $T_d = 20$ K, i.e. $$\begin{aligned} M_d &=& \frac{F_{\nu} d^2}{\kappa_{\nu} B_{\nu}(T_d)}\\\end{aligned}$$ where $F_{\nu}$ is the integrated flux at 870 $\mu$m, $d$ is the estimated distance to Ophiuchus (137 pc), $\kappa_{\nu} = 0.03$ cm$^2$/g is the total opacity at 870 $\mu$m assuming a [@hi83] dust opacity and a 100:1 gas-to-dust ratio [@bo78], and $B_{\nu}$ is the Planck function. An important caveat here is that this mass is calculation is only an estimate at best. Recent studies have suggested that the gas mass might be considerably lower than the often prescribed 100:1 ratio [@wi14]. ![image](f1.pdf) ![image](f2.pdf) ![image](f3.pdf) ![image](f4.pdf) To test for significant differences in flux and radius among the different sub-populations of our sample, we used the implementation of the Kaplan-Meier (KM) product estimator in the `lifelines` Python package [@dp17] [^1] to estimate the Cumulative Distribution Functions (CDFs) for all of the sub-populations (see Figures \[fig:single\_flux\_cdfs\] - \[fig:tauclass\]). The KM estimator is akin to an empirical cumulative probability distribution, but it has the advantage of being able to account for the non-detections in our sample by incorporating $\sigma$ upper limits when appropriate. For all the distributions we compute for fluxes, upper limits are incorporated. However, for the radii KM estimators, we only incorporate detections, as the radius of a non-detected object is ill-defined. Confidence intervals for each bin in the KM estimator are computed using Kalbfleisch and Prentice’s modification of the result of [@gw26] (see p. 18 of @kp02 for details). After the KM estimators are computed for each sub-population, we use the non-parametric log-rank test to determine whether or not it is likely that the two cumulative distributions in flux or radius are different for the pairwise combinations of sub-populations. Figure \[fig:single\_flux\_cdfs\] shows the CDFs comparing the flux of the single sources in our survey with the fluxes of the other populations (binary, triples, multiples, and transition disks). Perhaps most striking of these comparisons, is that of the binary population. The binary sources in this survey show systematically lower flux values than the isolated population. In Table 3 we report the $p$-values of the different comparisons, as well as the median flux and radius of the different populations. The $p$-value represents the probability that, given our data, the two populations compared are drawn from a single distribution. Thus, the higher our $p$-value, the more likely that this is the case; conversely, the lower the $p$-value, the less likely it is for the two populations to be from the same distribution. We define two populations to have a significant difference if the log-rank on their respective KM estimators yields $p\lesssim0.05$. When comparing the binaries with the singles, we find that both the flux and radius $p$-values show a suggestive trend with both $p$-values $<0.1$ ($p_{flux}=0.06946$ and $p_{radius}=0.01766$). Our binary sample includes three circumbinary disks (disks that encompass both components of the binary system), that we find to be quite bright compared to the rest of the binary sample. Since @ha12 also found this to be true in Taurus, we looked at the same comparison without these sources. We find that without the circumbinary disks, we get $p_{flux}=0.00876$ and $p_{radius}=0.00075$, which is lower than our cutoff. Figures \[fig:single\_flux\_cdfs\] and \[fig:single\_radii\_cdfs\] show the CDF comparing the two populations in the top left panel for flux and radius, respectively, and Figure \[fig:singles\_vs\_bin\_nocb\] shows the same plot excluding circumbinary disks. Each binary component was counted as one source, and in the case of a non-detection, the 3$\sigma$ value of the map was used as an upper limit for fluxes. All known, non-spectroscopic, (i.e., Oph 6, 12, 32, and 36) binaries in our sample are resolved, therefore blending of component fluxes is not an issue in our sample. It can be visually seen in these plots (see top left panel of both Figure \[fig:single\_flux\_cdfs\] and Figure \[fig:single\_radii\_cdfs\], and Figure \[fig:singles\_vs\_bin\_nocb\]), that there is hardly any overlap between the isolated YSOs and the binaries. The binary components are systematically dimmer as well as smaller, than their isolated counterparts. In Figure \[fig:primary\_vs\_secondaries\], we show the comparison between the components of the binary YSOs. The brighter component has a median flux value of 27.74 mJy, while the dimmer component is at 6.45 mJy. This a factor of 5 different, although we note that the large uncertainty in each individual bin of the KM estimator makes any observed difference between the populations not significant. The difference in the median radius for either component is $<$ 2 au, meaning that there is not a discernible difference in the sizes of the two. Across star forming regions, the inner disk fraction for single stars and for wide binaries (i.e., binaries with projected separation $>$ 40 au) is comparable, with $\sim$ 50% of these systems harboring enough material to make them Class II objects. On the other hand, tighter systems ($<$ 40 au) are preferentially less likely to have evidence of an infrared excess, with only $\sim$ 20% of those systems harboring enough material to make them Class II objects [@ci09]. Folding these data into our analysis of the millimeter emission for singles vs. binaries would most likely make the difference between the two much starker. The triple systems in our sample are slightly more complicated than the binaries. Two of the three systems (ROph11 and ROph31) are treated in the same way as the binaries, where we use the 3$\sigma$ value for the non-detections. The third system, ROph23, is also treated in this way, however this system has a circumbinary disk. Since this cannot be divided into two different systems, we count this as one source and use the 3$\sigma$ value of the map as the upper limit twice. When comparing these with the singles, we find $p_{flux}=0.73140$ and $p_{radius}=0.03613$. A caveat to keep in mind when looking at the triple systems in this sample, is that we did not detect all three sources in any of the systems. These systems consist of a tight pair that will resemble binary systems, with a single star further away. In ROph 11 and in ROph 31, the distance of the third component from the tight pair is much larger than the separation of the tight pair itself. Therefore, the disk associated with the distant object more closely resembles a disk from a single source. This is likely the case in ROph 23 as well, though the orbit superimposes the distant companion onto the circumbinary disk. The transition disks we used in comparing with the isolated sources were ones that show a depletion of millimeter emission in their inner cavity in this dataset, not necessarily those listed in Table 2. Visually, our transition disk population (see Figure \[fig:trans\]) seems to be the most unique in both flux and size. It was somewhat surprising that the fluxes of these disks did not show $p<0.05$ when compared to the singles ($p_{flux}=0.10204$ and $p_{radius}=0.04363$). We only have 4 transition disks in our sample, so the small numbers may contribute to the higher $p$-values. The median flux for this population is a factor of 5 brighter than any other population and the median radius is 3.5 times as large as the isolated population, suggesting that the transition disks come from a different distribution. We did use the two different populations of Table 2 to see if there was anything statistically different between sources that either have sub-/millimeter cavities or infrared colors indicating they are transition disks, and those that do not. We find $p_{flux}=0.12846$ and $p_{radius}=0.09715$. ------------------------------------- -- ![image](f5a.pdf) ![image](f5b.pdf) ![image](f5c.pdf) ![image](f5d.pdf) ------------------------------------- -- ------------------------------------- -- ![image](f6a.pdf) ![image](f6b.pdf) ![image](f6c.pdf) ![image](f6d.pdf) ------------------------------------- -- ------------------------------------- -- ![image](f7a.pdf) ![image](f7b.pdf) ------------------------------------- -- ------------------------------------- -- ![image](f8a.pdf) ![image](f8b.pdf) ------------------------------------- -- ------------------------------------- -- ![image](f9a.pdf) ![image](f9b.pdf) ------------------------------------- -- --------------------------------------- -- ![image](f10a.pdf) ![image](f10b.pdf) --------------------------------------- -- Figure \[fig:cdfs\_class1\_vs\_2\] shows the CDF plots for the different classifications of YSOs. Our sample consisted of mostly Class II YSOs, followed by Flat and then Class I objects. Since the Flat sources are thought to be on average less evolved than the the Class II sources, we combined these with the Class I sources to more easily compare the two. The less evolved population shows a higher ($\sim$ 13 mJy) median flux, while also having a slightly lower ($\sim$ 1 au) median radius. This is as expected, since as the YSO evolves into a Class II object, its peak energy output moves to shorter wavelengths [@la87] and, as the nascent material from the envelope falls in, the disk surrounding the protostar will grow [@du14]. We report the detection of a $\sim 1600$ au millimeter-wave companion to ROph34, L1689-IRS 7. The system L1689 IRS 7 has only sparsely been surveyed for companions. It was included in the Ratzka et al. (2005) survey area, but the source was determined to be single. The separation regime that the Ratzka survey was sensitive to ranged from 0.1 to 6.4 arcseconds, and the companion that we report is located outside of 7 arcseconds. The companion can be seen in 2MASS. The $JHK_s$ magnitudes of the northern component are uniformly $\sim$ 2 mag lower than the corresponding magnitudes for the southern component $K_s \sim 8.5$ for the primary and $10.5$ for the secondary. Since the colors are the same, it is likely that the companion is also a class II low mass star that is a bona fide member of the Oph complex. From the K band contrast, we estimate a stellar mass ratio of 0.1-0.3 based on Seiss (2001) models for a 1 Myr old object (roughly consistent with the fact that the source is a Class II object). Comparison Flux $p$-value Radius $p$-value ------------------------------------------- ------------------- -------------------- Singles & Binaries 0.06946 0.01766 Singles & Binaries (no circumbinary disk) 0.00876 0.00075 Singles & Triples 0.73140 0.03613 Singles & Multiples 0.11562 0.00368 Singles & Transition Disks 0.10204 0.04363 Binary Components 0.21271 0.15456 Binary Components (no circumbinary disk) 0.53125 0.15456 Class II & Class I/Flat 0.24451 0.79303 Population Median Flux (mJy) Median Radius (au) Singles 46.3 17.9 Binaries 27.74 7.1 Binaries (no circumbinary disk) 19.6 6.85 Bright Binary component 27.74 6.45 Bright Binary (no circumbinary disk) 21.29 6.17 Dim Binary component 6.17 7.54 Dim Binary (no circumbinary disk) 6.45 7.54 Triples 15.41 8.08 Multiples 19.6 7.8 Transition Disks 262 62.34 Class I/Flat Sources 30.55 12.6 Class II Sources 18.73 13.426 : Statistics of different populations. Comparison of the various $p$-values obtained from each CDF. Note that ’multiples’ represents a combination of both binary and triple systems. We used a distance of 137 pc to compute the radius [clclccccll]{} 1 & 2MASS J16213192-2301403 & II &16:21:31.923 -23:01:40.761& 0.193 $\pm{0.015}$ $\times$ 0.07 $\pm{0.032}$ & 164 $\pm{6.6}$ & 6.2 $\pm{0.21}$ & 9.63 $\pm{0.49}$ & 0.59\ 2$^{a,}$ & V935 Sco & II &16:22:18.523 -23:21:48.549& 0.215 $\pm{0.011}$ $\times$ 0.125 $\pm{0.009}$ & 80.5 $\pm{4.7}$ & 41.98 $\pm{0.89}$ & 72.9 $\pm{2.3}$ & 4.43\ 3$^{c}$ & IRAS 16201-2410 & II &16:23:09.219 -24:17:05.364& 0.456 $\pm{0.043}$ $\times$ 0.283 $\pm{0.029}$ & 81.4 $\pm{8.2}$ & 26 $\pm{1.8}$ & 114.3 $\pm{9.7}$ & 6.96\ 4 & 2MASS J16233609-2402209 & II &16:23:36.113 -24:02:21.227& 0.160 $\pm{0.014}$ $\times$ 0.072 $\pm{0.016}$ & 6.7 $\pm{6.5}$ & 5.27 $\pm{0.15}$ & 7.12 $\pm{0.31}$ & 0.43\ 5$^{a}$a & WSB 19A & II &16:25:02.119 -24:59:32.798& 0.198 $\pm{0.011}$ $\times$ 0.188 $\pm{0.013}$ & 145 $\pm{60}$ & 14.11 $\pm{0.27}$ & 27.74 $\pm{0.77}$ & 1.69\ 5$^{a}$b & WSB 19B & II &16:25:02.011 -24:59:33.004& 0.127 $\pm{0.017}$ $\times$ 0.107 $\pm{0.019}$ & 106 $\pm{39}$ & 14.45 $\pm{0.33}$ & 19.60 $\pm{0.7}$ & 1.20\ 6 & DoAr 21 & II &16:26:03.300 -24:23:36.000& – $\times$ –& ... & ... & $<$ 0.89 & $<$ 0.05\ 7$^{a}$a & GSS 31a & II &16:26:23.362 -24:20:59.997& 0.100 $\pm{0.004}$ $\times$ 0.060 $\pm{0.007}$ & 169 $\pm{5}$ & 39.34 $\pm{0.21}$ & 46.75 $\pm{0.4}$ & 2.85\ 7$^{a}$b & GSS 31b & II &16:26:23.432 -24:21:01.749& 0.071 $\pm{0.004}$ $\times$ 0.051 $\pm{0.008}$ & 147 $\pm{15}$ & 34.67 $\pm{0.21}$ & 38.36 $\pm{0.38}$ & 2.33\ 8 & DoAr 25 & II &16:26:23.680 -24:43:14.303& 1.071 $\pm{0.038}$ $\times$ 0.487 $\pm{0.015}$ & 110 $\pm{1.4}$ & 38 $\pm{1.2}$ & 515 $\pm{18}$ & 31.33\ 9$^{d}$ &Elias 24 & II &16:26:24.078 -24:16:13.855& 0.558 $\pm{0.049}$ $\times$ 0.474 $\pm{0.042}$ & 45 $\pm{81}$ & 63 $\pm{4.5}$ & 489 $\pm{39}$ & 29.77\ 10 & GY 33 & II &16:26:27.540 -24:41:53.882& 0.337 $\pm{0.013}$ $\times$ 0.097 $\pm{0.014}$ & 160.5 $\pm{1.7}$ & 10.22 $\pm{0.27}$ & 23.45 $\pm{0.84}$ & 1.43\ 11$^{t}$ & WSB 38 Aa & II &16:26:46.471 -24:12:00.39& ... & ... & $<$0.8 & $<$0.8 & $<$0.04\ 11$^{t}$ & WSB 38 Ab & II &16:26:46.474 -24:12:00.30& ... & ... &$<$0.8 & $<$0.8 & $<$0.04\ 11$^{t}$ & WSB 38B & II &16:26:46.427 -24:12:00.443& 0.086 $\pm{0.013}$ $\times$ 0.033 $\pm{0.020}$ & 108 $\pm{14}$ & 13.82 $\pm{0.2}$ & 15.41 $\pm{0.38}$ & 0.94\ 12$^{a,?}$ & WSB 40 & II &16:26:48.651 -23:56:34.589& 0.082 $\pm{0.012}$ $\times$ 0.056 $\pm{0.018}$ & 167 $\pm{32}$ & 7.78 $\pm{0.13}$ & 8.83 $\pm{0.24}$ & 0.53\ 13$^{b}$ & SR 24aa & II &16:26:58.438 -24:45:32.24& ... & ... & $<$ 1.2 & $<$ 1.2& $<$ 0.07\ 13$^{b}$$ $ & SR 24ab & II &16:26:58.453 -24:45:32.21& ... & ... & $<$ 1.2 & $<$1.2 & $<$ 0.07\ 13$^{b,c}$ & SR 24b & II &16:26:58.504 -24:45:37.220& 0.984 $\pm{0.118}$ $\times$ 0.563 $\pm{0.070}$ & 22.5 $\pm{8.3}$ & 42.9 $\pm{4.7}$ & 624 $\pm{73}$ & 37.96\ 14 & GY 211 & II &16:27:09.096 -24:34:08.708& 0.265 $\pm{0.005}$ $\times$ 0.127 $\pm{0.004}$ & 33.1 $\pm{1.2}$ & 45.79 $\pm{0.43}$ & 91.2 $\pm{1.2}$ & 5.55\ 15 & GY 224 & F &16:27:11.168 -24:40:47.100& 0.428 $\pm{0.009}$ $\times$ 0.148 $\pm{0.005}$ & 92.23 $\pm{0.79}$ & 42.13 $\pm{0.62}$ & 126.2 $\pm{2.4}$ & 7.67\ 16 & GY 235 & F &16:27:13.813 -24:43:32.053& 0.208 $\pm{0.010}$ $\times$ 0.172 $\pm{0.010}$ & 177 $\pm{13}$ & 26.18 $\pm{0.41}$ & 51 $\pm{1.1}$ & 3.11\ 17 & GY 284 & F &16:27:30.841 -24:24:56.528& – $\times$ – & ...& 2.78 $\pm{0.11}$ & 2.79 $\pm{0.19}$ & 0.17\ 18 & YLW 47 & II &16:27:38.314 -24:36:58.997& 0.155 $\pm{0.005}$ $\times$ 0.144 $\pm{0.005}$ & 105 $\pm{20}$ & 51.68 $\pm{0.39}$ & 82.73 $\pm{0.94}$ & 5.03\ 19 & DoAr 33 & II &16:27:39.004 -23:58:19.149& 0.225 $\pm{0.006}$ $\times$ 0.176 $\pm{0.005}$ & 78.2 $\pm{5.6}$ & 37.68 $\pm{0.49}$ & 76.4 $\pm{1.4}$ & 4.64\ 20 & GY 314 & II &16:27:39.422 -24:39:15.940& 0.258 $\pm{0.005}$ $\times$ 0.145 $\pm{0.006}$ & 138.9 $\pm{2.1}$ & 72.12 $\pm{0.86}$ & 151.9 $\pm{2.5}$ & 9.23\ 21$^{a}$a & SR 9A & II &16:27:40.275 -24:22:04.568& 0.073 $\pm{0.016}$ $\times$ 0.065 $\pm{0.021}$ & 69 $\pm{83}$ & 13.15 $\pm{0.26}$ & 14.83 $\pm{0.48}$ & 0.90\ 21$^{a}$b & SR 9B & II &16:27:40.272 -24:22:03.888& – $\times$ – & ... & 3.02 $\pm{0.036}$ &3.02 $\pm{0.036}$ & 0.18\ 22 & SR 20 W & II &16:28:23.337 -24:22:41.070& 0.420 $\pm{0.022}$ $\times$ 0.145 $\pm{0.011}$ & 65.7 $\pm{1.7}$ & 22.16 $\pm{0.76}$ & 64.1 $\pm{2.9}$ & 3.90\ 23$^{b,}$Aab & EM\ SR 13Aab & II &16:28:45.266 -24:28:19.358& 0.412 $\pm{0.040}$ $\times$ 0.329 $\pm{0.034}$ & 90 $\pm{27}$ & 31.7 $\pm{2.2}$ & 148 $\pm{12}$ & 9.00\ 23$^{b}$ & EM\* SR 13B & II &16:28:45.28 -24:28:19.318& ... & ... & ... &$<$ 0.85 &$<$ 0.04\ 24 & WSB 63 & II & 16:28:54.071 -24:47:44.694 & 0.266 $\pm{0.005}$ $\times$ 0.107 $\pm{0.006}$ & 0.07 $\pm{1.28}$ & 15.53 $\pm{0.17}$ & 30.55 $\pm{0.48}$ & 1.86\ 25 & WSB 67 & II &16:30:23.398 -24:54:16.511& 0.175 $\pm{0.011}$ $\times$ 0.112 $\pm{0.013}$ & 12.8 $\pm{8.4}$ & 10.83 $\pm{0.21}$ & 16.82 $\pm{0.49}$ & 1.03\ 26$^{a}$a & ROXS 42Ca & II &16:31:15.738 -24:34:02.487& – $\times$ –& 115.6 $\pm{3.7}$ & 3.96 $\pm{0.16}$ & 4.14 $\pm{0.18}$ & 0.25\ 26$^{a}$b & ROXS 42Cb & II &16:31:15.748 -24:34:02.72& – $\times$ –& ... & $<$ 0.89 & $<$ 0.89 & $<$ 0.4\ 27$^{a}$a & DoAr 43a & II &16:31:30.873 -24:24:40.288& 0.267 $\pm{0.008}$ $\times$ 0.110 $\pm{0.006}$ & 38.3 $\pm{1.6}$ & 18.8 $\pm{0.28}$ & 36.29 $\pm{0.78}$ & 2.21\ 27$^{a}$b & DoAr 43b & -&16:31:31.025 -24:24:37.484& 0.121 $\pm{0.021}$ $\times$ 0.111 $\pm{0.025}$ & 97 $\pm{85}$ & 4.83 $\pm{0.15}$ & 6.58 $\pm{0.32}$ & 0.40\ 28 & 2MASS J16313124-2426281 & II &16:31:31.245 -24:26:28.438& 1.301 $\pm{0.029}$ $\times$ 0.157 $\pm{0.005}$ & 49.05 $\pm{0.21}$ & 14.75 $\pm{0.25}$ & 124.8 $\pm{2.4}$ & 7.60\ 29$^{c}$ & DoAr44 & II &16:31:33.455 -24:27:37.515& 0.911 $\pm{0.147}$ $\times$ 0.821 $\pm{0.134}$ & 63 $\pm{59}$ & 12.4 $\pm{1.8}$ & 262 $\pm{41}$ & 15.99\ 30 & 2MASS J16314457-2402129 & II &16:31:44.577 -24:02:13.475& 0.110 $\pm{0.007}$ $\times$ 0.069 $\pm{0.014}$ & 133.8 $\pm{8.9}$ & 15.30 $\pm{0.16}$ & 18.81 $\pm{0.32}$& 1.14\ 31$^{b}$a & LDN 1689 IRS 5A & F &16:31:52.111 -24:56:16.030& 0.117 $\pm{0.013}$ $\times$ 0.113 $\pm{0.013}$ & 79 $\pm{84}$ & 69.9 $\pm{1.3}$ & 94.4 $\pm{2.7}$ & 5.75\ 31$^{b}$ba & LDN 1689 IRS 5Ba & -&16:31:51.929 -24:56:17.44& ... & ... & $<$ 1 & $<$ 1 & $<$ 0.06\ 31$^{b}$bb & LDN 1689 IRS 5Bb & -&16:31:51.915 -24:56:17.376& 0.129 $\pm{0.025}$ $\times$ 0.047 $\pm{0.037}$ & 117 $\pm{17}$ & 4.39 $\pm{0.16}$ & 5.47 $\pm{0.33}$ & 0.34\ 32 & WSB 74 &- &16:31:54.700 -25:03:24.000&– $\times$ – & ... & ... & $<$ 0.71 & $<$ 0.04\ 33$^{a}a$ & DoAr 51A &- &16:32:11.848 -24:40:21.90&– $\times$ – & ... & ... & $<$ 0.75 & $<$ 0.05\ 33$^{a}b$ & DoAr 51B &- &16:32:11.904 -24:40:21.76&– $\times$ – & ... & ... & $<$ 0.75 & $<$ 0.05\ 34$^{a}$a& L1689-IRS 7A & II &16:32:21.047 -24:30:36.309& 0.080 $\pm{0.012}$ $\times$ 0.066 $\pm{0.021}$ & 144 $\pm{66}$ & 29.17 $\pm{0.5}$ & 33.43 $\pm{0.96}$ & 2.03\ 34$^{a}$b & L1689-IRS 7B & II &16:32:20.811 -24:30:29.487& 0.110 $\pm{0.015}$ $\times$ 0.066 $\pm{0.038}$ & 156 $\pm{28}$ & 5.16 $\pm{0.13}$ & 6.32 $\pm{0.25}$ & 0.39\ 35 & Haro 1-17 & II &16:32:21.928 -24:42:15.208& 0.135 $\pm{0.016}$ $\times$ 0.064 $\pm{0.023}$ & 79 $\pm{12}$ & 6.97 $\pm{0.2}$ & 8.87 $\pm{0.41}$ & 0.53\ 36$^{a,}$ & 2MASS J16335560-2442049AB & II &16:33:55.610 -24:42:05.370& 0.670 $\pm{0.084}$ $\times$ 0.461 $\pm{0.059}$ & 77 $\pm{15}$ & 24.9 $\pm{2.7}$ & 233 $\pm{28}$ & 14.17\ 38$^{c,d}$ & WSB 82 & II &16:39:45.440 -24:02:04.250& 1.301 $\pm{0.057}$ $\times$ 0.632 $\pm{0.028}$ & 171.6 $\pm{2.2}$ & 17.44 $\pm{0.72}$ & 437 $\pm{19}$ & 26.52\ 39 & 2MASS J16214513-2342316 & I &16:21:45.122 -23:42:32.182& 0.628 $\pm{0.032}$ $\times$ 0.118 $\pm{0.014}$ & 174.33 $\pm{0.97}$ & 21.73 $\pm{0.81}$ & 94.3 $\pm{4.3}$ & 5.73\ 40 & 2MASS J16263682-2415518a & F &16:26:36.827 -24:15:52.298& 0.390 $\pm{0.046}$ $\times$ 0.323 $\pm{0.041}$ & 6.1 $\pm{28.3}$ & 10.17 $\pm{0.83}$ & 46.3 $\pm{4.5}$ & 2.82\ 41 & WL6 & I &16:27:21.791 -24:29:53.826& 0.106 $\pm{0.019}$ $\times$ 0.072 $\pm{0.027}$ & 16 $\pm{28}$ & 15.33 $\pm{0.37}$ & 18.73 $\pm{0.73}$ & 1.14\ 42 & GY 312 & I &16:27:38.936 -24:40:21.058& 0.390 $\pm{0.009}$ $\times$ 0.121 $\pm{0.007}$ & 168.27 $\pm{0.94}$ & 34.45 $\pm{0.56}$ & 92.9 $\pm{2}$ & 5.66\ 43 & 2MASS J16274161-2446447 & I &16:27:41.601 -24:46:45.082& 0.267 $\pm{0.011}$ $\times$ 0.139 $\pm{0.009}$ & 99.8 $\pm{3.1}$ & 14.75 $\pm{0.32}$ & 31.3 $\pm{0.95}$ & 1.90\ 44 & GY 344 & - &16:27:45.800 -24:44:54.000& – $\times$ – & ... & ... & $<$ 0.74 & $<$ 0.04\ 45$^{a}a$ & YLW 52a & F &16:27:51.796 -24:31:46.048& 0.216 $\pm{0.043}$ $\times$ 0.106 $\pm{0.066}$ & 129 $\pm{17}$ & 3.13 $\pm{0.28}$ & 5.46 $\pm{0.72}$ & 0.34\ 45$^{a}b$ & YLW 52b & F &16:27:51.479 -24:31:40.33& ... & ... & $<$ 0.7 & $<$ 0.7 & $<$ 0.03\ 46 & WSB 60 & II &16:28:16.503 -24:36:58.463& 0.554 $\pm{0.026}$ $\times$ 0.512 $\pm{0.025}$ & 135 $\pm{27}$ & 26.31 $\pm{0.98}$ & 232.2 $\pm{9.6}$ & 14.17\ 47 & 2MASS J16313383-2404466 & F &16:31:33.831 -24:04:47.036&$<$ 0.16 $\times$ 0.057 & ...& 4.4 $\pm{0.24}$ & 5.52 $\pm{0.48}$ & 0.34\ 48 & IRS 63 & F &16:31:35.659 -24:01:29.893& 0.521 $\pm{0.024}$ $\times$ 0.359 $\pm{0.018}$ & 150 $\pm{5.2}$ & 123.5$\pm{4.8}$ & 776 $\pm{35}$ & 47.19\ 49 &2MASS J16442430-2401250 & - &16:44:24.300 -24:01:25.000& – $\times$ – & ... & ... & $<$ 0.80 & $<$ 0.05\ 50 & Haro 1-11 & II &16:27:38.325 -23:57:32.936& 0.151 $\pm{0.020}$ $\times$ 0.113 $\pm{0.021}$ & 84 $\pm{30}$ & 7.61 $\pm{0.28}$ & 11.21 $\pm{0.62}$ & 0.68\ $^{a}$[Field is a binary source]{} $^{b}$[Field is a triple source]{} $^{c}$[Transition Disk]{} $^{d}$[Evidence of gap in disk]{} ${}$[Circumbinary Disk]{} ${,?}$[Potential circumbinary Disk]{} \[tab:source\] Discussion {#sec:dis} ========== Comparison with the Taurus-Auriga Molecular Cloud ------------------------------------------------- In this work, we have used ALMA to map the distribution of 870 $\mu$m emission from 49 selected pre-main sequence stellar systems in the $\rho$ Ophiuchus molecular cloud and used these maps to construct the distribution of disk fluxes and radii from various subpopulations. A natural question to ask is how the systems in one molecular cloud compare to those of another. To do this, we have compiled a target list of sources in the Taurus-Auriga molecular cloud to which we compare our sample. Taurus represents an obvious choice for such a comparison. First, it has a well-characterized stellar population and disk population due to its proximity (145 pc; @lo07 [@to07; @to09]) as well as a relatively uniformly low extinction across the whole cloud (@lo10). Second, $\rho$ Oph has a relatively low stellar density across much of its volume and very few UV/X-ray luminous O/B-type stars, much like Taurus and opposed to clusters such as Orion. Such environmental impacts are known to have severe and deleterious effects on protoplanetary disk masses and radii (e.g., @ma14). Finally, the two clusters are close to the same age: $\rho$ Oph is between 0.5 - 2 Myr [@wi08] old, while Taurus is in the vicinity of 1-2 Myr old (@lu10). In order to quantify how common our sample is, we have constructed a sample of Taurus sources to which we compare our Oph sample. To do this, we used the results of the Spitzer* survey performed by @re10. They surveyed approximately 44 square degrees of Taurus in each of the 7 different IRAC/MIPS bands. To ensure that our comparison stars were in Taurus, we restricted our selection to the subsample of their survey that had already previously been identified as Taurus members, rather than those sources that were inferred to be Taurus members based on colors from their survey. As in our survey, we only included sources with detections in all of the IRAC bands as well as the 24 and 70 $\mu$m bands in the same fashion as was done for our present survey. The sensitivities of the @re10 survey are similar to those of the c2d survey, so this is probably a fair comparison. After selecting candidate sources in Taurus, we restricted the sample to those sources that had (sub)-millimeter flux information in the literature. Where multi-band photometry was available, we used the derived spectral index to infer the 870 $\mu$m flux density; where it was not available, we assumed that the intrinsic spectral index was 3. The qualitative results for this work do not depend on the precise value of $\alpha$ we assume. We use the same KM estimators to compare the corresponding subpopulations of Taurus-Auriga objects with Oph objects. The $p$-values for these comparisons are found in Table 4. In Figures \[fig:taupop\] and \[fig:tauclass\] we show the CDF comparisons of the Oph and Taurus populations. We find that the single sources have different median fluxes (31 mJy vs. 57 mJy), with their corresponding low $p$-value (0.00282), most likely due to the Taurus population having a high flux tail in its distribution. One possibility for this dichotomy is the difference in the environments between the two clouds. $\rho$ Oph tends to have more clustered YSOs while Taurus’s YSOs are more dispersed. We find that Ophiuchus typically has dimmer binary and triple systems, as well as Class II protostars, with its Class I population being much dimmer ($\sim$31 mJy vs. $\sim$116 mJy) than that of Taurus. Due to the low number of Class I YSOs in our survey, this is likely due to small number statistics. --------------------------------------- -- ![image](f11a.pdf) ![image](f11b.pdf) ![image](f11c.pdf) ![image](f11d.pdf) --------------------------------------- -- --------------------------------------- -- ![image](f12a.pdf) ![image](f12b.pdf) --------------------------------------- -- Comparison Flux $p$-value ------------------- ------------------- Singles 0.00434 Binaries 0.70718 Triples 0.18367 Multiples 0.23964 Class I 0.00122 Class II 0.50860 Taurus Population Median Flux (mJy) Singles 57.4 Binaries 10.7 Triples 29.2 Multiples 12.9 Class I Sources 115.8 Class II Sources 21.4 : Statistics of various comparisons. Comparison of the various $p$-values obtained from each CDF. Note that ’multiples’ is a combination of both binary and triple systems. \[tab:taur\] Disks in Binary Systems and Tidal Truncation -------------------------------------------- Protoplanetary disks in binary systems are subject to far more interactions than disks in single systems, due to the manner in which disks around stars and stellar companions interact. The disks surrounding these protostars can only grow to a certain radius before that material is stripped away by its companion. This is likely due to the interactions with their companions, yielding a loss of disk material [@je96; @ha12]. The lower disk fluxes can be interpreted as being due to lower disk masses. Theory indicates that disk truncation in binaries is particularly sensitive to the binary’s semimajor axis $a$ and eccentricity $e$. Essentially, the closer the periastron distance $d = a(1-e)$, the more severe the truncation. We use the analytic model described in @pi05 to estimate ------------------- -- ![image](f13.pdf) ------------------- -- the equilibrium truncated radius of our binary sources. This model yields a prediction for a circumstellar disk’s truncation radius given its host binary’s orbital elements $a$ and $e$, as well as the mass ratio $q$, which we assume to be unity (this has little effect, as the truncation radius depends only very weakly on $q$ for reasonable values of $q$). Because we have no orbital information on our binary systems outside of a projected separation on the sky, we implement a statistical method to estimate the true orbital elements $a$ and $e$ based on the projected separation of the two stars; for details, see @ha12. We then convert this to a prediction for the tidal radii. The detailed predictions are somewhat sensitive to the choice of the eccentricity probability density function; we choose a uniform distribution picking $e$ between 0 and 1 for this. As seen in Figure \[fig:trunc\], all of our sources, barring the two with upper limit detections, are well below the equilibrium line. This means that for the binary systems we observed in Oph, truncation is not responsible for the disk size observed. This is in contrast to what @ha12 (see their Figure 11) found for the Taurus binary systems. The Taurus systems have a much more scattered distribution with points both above and below the equilibrium. Figure \[fig:trunc\] shows that the measured dust disk radius and that predicted from our statistical modeling disagree. However, there are two caveats to this analysis. First, the gas and dust extents are not necessarily the same. Dust-size dependent aerodynamic effects such as radial drift can lead to differences in the structure of the gas (which comprises the bulk of the disk mass) and that of the large particles responsible for the millimeter continuum emission (e.g., @we77 [@pe12]). Due to these effects, the dust emission extent is expected to be more compact than the gas-line emission, with theoretical estimates of the ratio of 0.88 mm continuum extent to CO emission line extent ranging between 1.5 to about 4 (e.g., @fa17). Observational evidence also suggests this to be the correct range (e.g., @an12 [@va17]). Accordingly, the measured radii could be corrected by a typical correction factor of $\sim 2-3$ and be brought into good agreement with truncation models. Alternatively, because our predictions for the tidal radii are dependent on the (unknown) eccentricity distribution for pre-main sequence binaries, it is plausible that an eccentricity distribution weighted more towards moderate to high eccentricity would alleviate the discrepancy we note. For main-sequence stars with periods $P \gtrsim$ 100 days, observations are consistent with a uniform distribution between 0 and 1 [@du13]. It is plausible that, in the past, the progenitors of these systems (and the analog of the disk-bearing systems we focus on here) had higher eccentricities that were subsequently damped due to star/disk interaction (e.g., @al01), making the higher eccentricity distribution the more appropriate one to use here. Transition and Gapped Disks --------------------------- An interesting outcome of this ALMA survey is how diverse the YSO population we observed is. As discussed in §\[sec:sam\], our aim was to probe more evolved protostars, to characterize their disks. Of the 49 stellar systems we observed, 5 include transition disks (see Figure \[fig:trans\]). Our ALMA observations were only $\sim$36 seconds and provide unprecedented detail in all 5 of these sources. Three of these are known transition disks that have been heavily observed and studied in both the infrared and the sub-/millimeter regimes (ROph 13, ROph 29, ROph 36). One source, ROph 3, does not have existing sub-/millimeter observations, but was determined to be a transition disk from IR data. Finally, ROph 38 has no existing millimeter data, and, unlike the other transition disks observed in this survey, there is no indication of a central cavity in the broadband Spitzer near to mid-infrared photometry taken during the C2D survey [@ev03]. We, however, detect a large millimeter cavity as well as a gap and a ring-like structure of low-level emission surrounding it. This indicates that, while the central cavity maybe devoid of mm-sized particles, it is not devoid of small particles. The detection of disks that show evidence for narrow gaps in their emission is a particularly exciting result from our survey. Such gaps in the millimeter emission from the disk have been directly imaged previously in the young Class I/II object HL Tau [@al15] and the nearby older Class II TW Hya [@no16; @an16], as well as in the higher mass Herbig Ae stars HD 163296 [@is16] and HD 169142 [@fe17]. Modeling of ALMA continuum data at 0.87 and 1.3 mm of the young Class II star AA Tau also suggests multiple gaps in this star’s disk [@lo17]. The leading candidates for how the gaps open are either a forming protoplanet/gas-giant core gravitationally torques material around it, effectively repelling some disk material away from it [@lp86], or through enhanced grain growth due to pressure bumps caused by planets [@bi10]. Other suggestions from theorists for forming rings and gaps the millimeter emission include dust sintering [@oz16] and disk surface density variation driven by inhomogeneous magnetic field distribution (e.g., @fl15) or magnetic disk-winds [@su17]. The exact details of the gap opening, including the gap structure’s dependence on planetary embryo mass and surrounding disk structure, have not been fully analytically described [@cr06]. In fact, it is uncertain whether a single planet per gap is required for gap formation or if a single planet can carve multiple gaps [@do17]. It is, however, generally agreed that higher embryo masses carve more substantial gaps. Numerical calculations indicate that a range of planetary masses $\gtrsim 0.2 M_{\mathrm{Jupiter}}$ can carve observable gaps in disks’ millimeter emission. In our sample, we find two sources, ROph 9 (Elias 24) and ROph 38 (WSB 82) that exhibit clear evidence in the images of substantial disk gaps (see Figures \[fig:singles\] and \[fig:trans\]), while another source, ROph 8 (DoAr 25; see Fig \[fig:singles\]) shows some evidence of a potential gap in the disk in its image. We present these sources again in Figure \[fig:gaps\] with an altered color-scale to emphasize the gaps and low-lying emission in each disk. To quantify the structure of the gaps, we follow the procedure used by [@al15] to study the gap structure in the millimeter emission of HL Tau and deprojected each image using the fit disk center, inclination, and position angle, and produce azimuthally averaged surface brightness profiles. These profiles are shown in Figure \[fig:profiles\]. ![image](f14.pdf){width="\linewidth"} ![image](f15a.pdf){width="0.31\linewidth"} ![image](f15b.pdf){width="0.31\linewidth"} ![image](f15c.pdf){width="0.31\linewidth"} In the case of ROph 9 and 38, there are obvious deficits of emission observed at approximately 65 and 170 au, respectively. These gaps appear to be either unresolved or only marginally resolved by the synthesized beam of the array. There is a hint of a plateau in the profile for ROph 8, which when combined with the imaging results, suggest a potential deficit in emission at approximately 95 au. Because the gaps are only marginally resolved, they must be less than about 10 au in annular extent. To ensure that we are not ‘missing’ sources that may have gaps that are not obvious in the images, we constructed these deprojected profiles for each source in our sample. Each source without obvious evidence for a millimeter cavity (see Fig. 4) shows a monotonically decreasing flux density with radius until the flux density starts to approach the noise level in the images. We will present a more in depth analysis of both the transitional disks and the gapped disks in a future work. One potentially interesting question we can begin to ask is the fraction of disks ($f$) that show ongoing, present-day evidence for planet formation. If we consider either disk gaps or large millimeter cavities (in the absence of other explanations, such as known binarity) as evidence of ongoing planet formation, we find that 6 out of 49 disks in our sample show evidence of forming planets that are massive enough to open up large gaps or cavities at the current epoch. This yields an estimate of $f = 0.122$ with a 95% confidence interval of $0.031 < f < 0.21$. Note that $f$ represents the fraction of systems that are estimated to have large ($\gtrsim 0.2 M_{\mathrm{Jupiter}}$) mass reservoirs that also have signposts of planet formation (i.e., gaps or central cavities). Asymmetric Dust Disks --------------------- Asymmetries in the millimeter continuum emission from circumstellar disks have recently become of interest due to their likely origin in dust traps that may enable rapid grain growth past the barriers that, e.g., radial drift may impose [@pi11; @ra17; @mi17]. ![image](f16.pdf){width="0.5\linewidth"} These asymmetries are sometimes observed in transitional disks as a potential sign of a planet forming an azimuthally asymmetric pressure gradient within the surrounding disk (e.g., IRS 48, @vdm13; SAO 206462 and SR 21, @pe14). In each of our sample of transitional disks, we also observe somewhat substantial asymmetries in the outer disk, with a typical contrast of about 20% from maximum to minimum in the profile of brightness vs azimuth as a given disk radius. In addition to the transitional disks we identify on the basis of a substantial millimeter cavity in Figure \[fig:trans\], we also identify the primary disk in one of our binaries, ROph 40, as having a large asymmetry in its continuum emission. It is shown in Figure \[fig:roph40\]. This source shows no infrared signature of being a transitional disk in either Spitzer or Herschel data. Furthermore, any cavity in the millimeter emission is not obvious, unlike in the analogous case of ROph 38. However, the asymmetry in the dust emission is reminiscent of what is observed in the transitional disks both we and others have mapped in the millimeter. Unfortunately, we do not have coverage of the requisite gas-line emission to be able to tell whether this is a true dust trap or a mere asymmetry in the overall mass distribution of the source. Summary {#sec:sum} ======= We have presented an ALMA imaging survey of the 870 $\mu$m dust continuum emission from the circumstellar material of 49 systems in the $\rho$ Ophiuchus molecular cloud complex. These systems, having been selected on the basis of excess in each of the Spitzer IRAC and MIPS bands, represent the stellar systems most likely to have sufficient circumstellar material to enable planet formation over the next few Myr. This survey, observing each source only for 36 seconds per source, shows the versatility and promise of the ALMA instrument for studies in star and planet formation. Many of these sources represent low-mass targets that have not been observed at millimeter-wavelengths before. We summarize our results and analysis below: - We divided the sources into several different populations (i.e., single stars, binaries, triple systems and transition disks) and computed Kaplan-Meier product limit estimators to estimate the cumulative probability distribution for both disk flux and disk radius for each population. We find significant differences in both flux and radius amongst the singles and binaries in Oph: disk fluxes and radii in binaries are significantly smaller than in single stars. Similar results about the fluxes have been noted previously (e.g @je94 [@je96; @ha12]), but disk radii at millimeter wavelengths have never explicitly been found to be smaller in disks in binaries compared to disks around isolated stars. Large differences in circumstellar mass (for which, assuming a single temperature and $\kappa$, flux can be a proxy for) and radius over a small range of ages illustrate the diversity of conditions in the disk, wherein planets are forming. - The lack of flux in the binary population is typically considered to be due to either disk truncation after formation, or caused by something that sets the disk radii during formation. Using a statistical model to convert from projected separation to semi-major axis and eccentricity, we computed the distribution of expected truncation radii using the analytic prescription of [@pi05] for each disk. We found that the (dust) disks in our sample are much too small to have been significantly affected by tidal truncation. This could have a natural explanation, as the gas disk extent is expected to a $\sim$ a few times larger than the dust disk extent. On the other hand, if this is not the case, it could suggest that the smaller disk radii in the binary systems are primordial, rather than a product of binary interaction after disk formation. This may be counter-intuitive, because binary systems tend to have larger angular momenta than single systems. One possibility is that most of the angular momentum of a binary system is stored in the binary orbit, leaving less for the circumstellar disks. In any case, this is an intriguing result that disk and binary formation theories should seek to address. - We detected several transition disks, two of which are the first ever millimeter observations (ROph3 = 2MASS J16230923-2417047; ROph 38 = WSB 82), whereas one (ROph38 = WSB 82) is being classified as a transition disk for the first time based solely on the presence of a millimeter cavity unexpected from the available infrared data. In particular, WSB 82 is a transition disk with a noticeable gap in the low surface brightness outer disk that resembles the gaps seen in ALMA images of other Class II disks so far (e.g., HL Tau, @al15, HD 163296, @is16, and TW Hya, @an16). Interestingly, we find an intriguing trend that the transition disks are on average much brighter and larger than both Class I and Class II disks. Theoretical studies of disk evolution need to account for this trend. - We have discovered an unexpected millimeter companion to the Class II source WLY 2-69 at 7.56$\farcs$ ($\sim 1000$ AU); given the density of millimeter-wave background sources, it is most likely physically associated with the source. A search of the literature on multiplicity in Oph yielded no reports of an optical or infrared companion. An examination of archival images from Spitzer, however, shows this companion source in the infrared. Due to the sheer number of baselines available, as well as the very sensitive receivers on the antennas, ALMA is producing exciting results almost daily, particularly in the study of protoplanetary disks. Surveys such as this one of 49 targets in $\rho$ Ophiuchi as well as that of a set of 92 sources in the $\sigma$ Ori cluster [@an17] demonstrate conclusively that ALMA as a rapid survey instrument is coming into its own. We thank the anonymous referee for their helpful comments. EGC is supported by SOSPA3-017. ZYL is supported in part by NASA NNX 14AB38G and NSF AST-1313083 and 1716259. This paper makes use of the following ALMA data: ADS/JAO.ALMA\# 2013.1.00157.S. ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada), NSC and ASIAA (Taiwan), and KASI (Republic of Korea), in cooperation with the Republic of Chile. The Joint ALMA Observatory is operated by ESO, AUI/NRAO and NAOJ. The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. ALMA Partnership, Brogan, C. L., P[é]{}rez, L. M., et al. 2015, , 808, L3 Andre, P., Montmerle, T., Feigelson, E. D., & Steppe, H. 1990, , 240, 321 Andre, P., Ward-Thompson, D., & Barsony, M. 1993, , 406, 122 Andre, P., & Montmerle, T. 1994, , 420, 837 Andrews, S. M., & Williams, J. P. 2005, , 631, 1134 Andrews, S. M., & Williams, J. P. 2005, , 619, L175 Andrews, S. M., & Williams, J. P. 2007, , 671, 1800 Andrews, S. M., Wilner, D. J., Hughes, A. M., Qi, C., & Dullemond, C. P. 2009, , 700, 1502 Andrews, S. M., Wilner, D. J., Hughes, A. M., Qi, C., & Dullemond, C. P. 2010, , 723, 1241 Andrews, S. M., Wilner, D. J., Hughes, A. M., et al. 2012, , 744, 162 Andrews, S. M., Wilner, D. J., Zhu, Z., et al. 2016, , 820, L40 Ansdell, M., Williams, J. P., Manara, C. F., et al. 2017, , 153, 240 Artymowicz, P., & Lubow, S. H. 1994, , 421, 651 Artymowicz, P., & Lubow, S. H. 2001, The Formation of Binary Stars, 200, 439 Barsony, M., Koresko, C., & Matthews, K. 2003, , 591, 1064 Barsony, M., Ressler, M. E., & Marsh, K. A. 2004, Bulletin of the American Astronomical Society, 36, 156.01 Barsony, M., Haisch, K. E., Marsh, K. A., & McCarthy, C. 2012, , 751, 22 Birnstiel, T., Dullemond, C. P., & Brauer, F. 2010, , 513, A79 Bitsch, B., & Kley, W. 2011, , 536, A77 Bitsch, B., Lambrechts, M., & Johansen, A. 2015, , 582, A112 Bohlin, R. C., Savage, B. D., & Drake, J. F. 1978, , 224, 132 Carpenter, J. M., Mamajek, E. E., Hillenbrand, L. A., & Meyer, M. R. 2006, , 651, L49 Cheetham, A. C., Kraus, A. L., Ireland, M. J., et al. 2015, , 813, 83 Chiang, E., & Laughlin, G. 2013, , 431, 3444 Cieza, L. A., Padgett, D. L., Allen, L. E., et al. 2009, , 696, L84 Cieza, L. A., Schreiber, M. R., Romero, G. A., et al. 2010, , 712, 925-941 Crida, A., Morbidelli, A., & Masset, F. 2006, , 181, 587 Davidson-Pilon et al. 2017, lifelines.py, v0.11.1, Zenodo, doi:10.5281/zenodo.815943, as developed on GitHub Dong, R., Li, S., Chiang, E., & Li, H. 2017, , 843, 127 Duch[ê]{}ne, G., Bouvier, J., Bontemps, S., Andr[é]{}, P., & Motte, F. 2004, , 427, 651 Duch[ê]{}ne, G., Bontemps, S., Bouvier, J., et al. 2007, , 476, 229 Duch[ê]{}ne, G., & Kraus, A. 2013, , 51, 269 Dunham, M. M., Stutz, A. M., Allen, L. E., et al. 2014, Protostars and Planets VI, 195 Espaillat, C., D’Alessio, P., Hern[á]{}ndez, J., et al. 2010, , 717, 441 Evans, N. J., II, Allen, L. E., Blake, G. A., et al. 2003, , 115, 965 Facchini, S., Birnstiel, T., Bruderer, S., & van Dishoeck, E. F. 2017, arXiv:1705.06235 Fedele, D., Carney, M., Hogerheijde, M. R., et al. 2017, , 600, A72 Flock, M., Ruge, J. P., Dzyurkevich, N., et al. 2015, , 574, A68 Furlan, E., Watson, D. M., McClure, M. K., et al. 2009, , 703, 1964 Greenwood Jr., M. (1926). The Natural Duration of Cancer. Reports of Public Health and Related Subjects, Vol. 33, HMSO, London Ghez, A. M., Neugebauer, G., & Matthews, K. 1993, , 106, 2005 Haisch, K. E., Jr., Greene, T. P., Barsony, M., & Stahler, S. W. 2004, , 127, 1747 Haisch, K. E., Jr., Barsony, M., Ressler, M. E., & Greene, T. P. 2006, , 132, 2675 Hartmann, L., Calvet, N., Gullbring, E., & D’Alessio, P. 1998, , 495, 385 Harris, R. J., Andrews, S. M., Wilner, D. J., & Kraus, A. L. 2012, , 751, 115 Hildebrand, R. H. 1983, , 24, 267 Hughes, A. M., Wilner, D. J., Qi, C., & Hogerheijde, M. R. 2008, , 678, 1119-1126 Isella, A., Guidi, G., Testi, L., et al. 2016, Physical Review Letters, 117, 251101 Jensen, E. L. N., Mathieu, R. D., & Fuller, G. A. 1994, , 429, L29 Jensen, E. L. N., Mathieu, R. D., & Fuller, G. A. 1996, , 458, 312 Kalbfleisch, J. D., & Ross, L. P., 2002, The Statistical Analysis of Failure Time Data (2nd ed.; Hoboken, NJ: John Wiley & Sons, Inc.) Kohn, S. A., Shkolnik, E. L., Weinberger, A. J., Carlberg, J. K., & Llama, J. 2016, , 820, 2 Kraus, A. L., Ireland, M. J., Martinache, F., & Hillenbrand, L. A. 2011, , 731, 8 Kraus, A. L., & Ireland, M. J. 2012, , 745, 5 Kraus, A. L., Ireland, M. J., Cieza, L. A., et al. 2014, , 781, 20 Lada, C. J. 1987, Star Forming Regions, 115, 1 Leous, J. A., Feigelson, E. D., Andre, P., & Montmerle, T. 1991, , 379, 683 Lin, D. N. C., & Papaloizou, J. 1986, , 307, 395 Loinard, L., Torres, R. M., Mioduszewski, A. J., et al. 2007, , 671, 546 Loinard, L., Torres, R. M., Mioduszewski, A. J., & Rodr[í]{}guez, L. F. 2008, , 675, L29 Lombardi, M., Lada, C. J., & Alves, J. 2010, , 512, A67 Loomis, R. A., [Ö]{}berg, K. I., Andrews, S. M., & MacGregor, M. A. 2017, , 840, 23 Lynden-Bell, D., & Pringle, J. E. 1974, , 168, 603 Luhman, K. L., Allen, P. R., Espaillat, C., Hartmann, L., & Calvet, N. 2010, , 186, 111 Mann, R. K., Di Francesco, J., Johnstone, D., et al. 2014, , 784, 82 Miranda, R., Li, H., Li, S., & Jin, S. 2017, , 835, 118 McMullin, J. P., Waters, B., Schiebel, D., Young, W., & Golap, K. 2007, Astronomical Data Analysis Software and Systems XVI, 376, 127 Nomura, H., Tsukagoshi, T., Kawabe, R., et al. 2016, , 819, L7 Ortiz-Le[ó]{}n, G. N., Loinard, L., Kounkel, M. A., et al. 2017, , 834, 141 Okuzumi, S., Momose, M., Sirono, S.-i., Kobayashi, H., & Tanaka, H. 2016, , 821, 82 Orellana, M., Cieza, L. A., Schreiber, M. R., et al. 2012, , 539, A41 P[é]{}rez, L. M., Carpenter, J. M., Chandler, C. J., et al. 2012, , 760, L17 P[é]{}rez, L. M., Isella, A., Carpenter, J. M., & Chandler, C. J. 2014, , 783, L13 Pichardo, B., Sparke, L. S., & Aguilar, L. A. 2005, , 359, 521 Pichardo, B., Sparke, L. S., & Aguilar, L. A. 2008, , 391, 815 Pinilla, P., Birnstiel, T., Ricci, L., et al. 2012, , 538, A114 Ragusa, E., Dipierro, G., Lodato, G., Laibe, G., & Price, D. J. 2017, , 464, 1449 Ratzka, T., K[ö]{}hler, R., & Leinert, C. 2005, , 437, 611 Rebollido, I., Mer[í]{}n, B., Ribas, [Á]{}., et al. 2015, , 581, A30 Ricci, L., Testi, L., Natta, A., et al. 2010, , 512, A15 Rebull, L. M., Padgett, D. L., McCabe, C.-E., et al. 2010, , 186, 259 Ru[í]{}z-Rodr[í]{}guez, D., Ireland, M., Cieza, L., & Kraus, A. 2016, , 463, 3829 Salyk, C., Pontoppidan, K., Corder, S., et al. 2014, , 792, 68 Suriano, S. S., Li, Z.-Y., Krasnopolsky, R., & Shang, H. 2017, , 468, 3850 Testi, L., Birnstiel, T., Ricci, L., et al. 2014, Protostars and Planets VI, 339 Torres, R. M., Loinard, L., Mioduszewski, A. J., & Rodr[í]{}guez, L. F. 2007, , 671, 1813 Torres, R. M., Loinard, L., Mioduszewski, A. J., & Rodr[í]{}guez, L. F. 2009, , 698, 242 Weidenschilling, S. J. 1977, , 180, 57 Williams, J. P., & Best, W. M. J. 2014, , 788, 59 Wilking, B. A., Gagn[é]{}, M., & Allen, L. E. 2008, Handbook of Star Forming Regions, Volume II, 5, 351 Willson, M., Kraus, S., Kluska, J., et al. 2016, , 595, A9 van der Plas, G., Menard, F., Canovas, H., et al. 2017, arXiv:1707.01688 van der Marel, N., van Dishoeck, E. F., Bruderer, S., et al. 2013, Science, 340, 1199 [^1]: This package is available at <https://github.com/CamDavidsonPilon/lifelines/>	{ "pile_set_name": "ArXiv" }
--- abstract: 'A long-standing conjecture of Farrell and Zdravkovska and independently S. T. Yau states that every almost flat manifold is the boundary of a compact manifold. This paper gives a simple proof of this conjecture when the holonomy group is cyclic or quaternionic. The proof is based on the interaction between flat bundles and involutions.' address: - 'Department of Mathematics, Indiana University, Bloomington, IN 47405, USA' - 'Department of Mathematics, Capital Normal University, Beijing, 100048, China' author: - 'James F. Davis' - Fuquan Fang title: An almost flat manifold with a cyclic or quaternionic holonomy group bounds --- [^1] Introduction {#sec:intro} ============ A closed manifold $M$ is [almost flat]{} if there is a sequence of metrics $g_i$ on $M$ so that $\vert K_{g_i} \vert \operatorname{{diam}}(M,{g_i})^2\to 0$ when $i\to \infty$, where $K_{g_i}$ is the sectional curvature and $\operatorname{{diam}}(M,{g_i})$ is the diameter of $M$ with respect to the metric $g_i$. In his celebrated paper [@Gr1], Gromov generalized the classical Bieberbach theorem for flat manifolds and proved that every almost flat manifold is finitely covered by a [nilmanifold]{}, that is, the quotient of a simply connected nilpotent Lie group by a uniform lattice. (Conversely, work of Farrell and Hsiang [@FH] showed that every manifold finitely covered by a nilmanifold is homeomorphic to an almost flat manifold.) Ruh [@Ruh] strengthened Gromov’s theorem and proved that an almost flat manifold is diffeomorphic to an [infranilmanifold]{}, that is, a double coset space ${{\Gamma}}\backslash L \rtimes \operatorname{{Aut}}(L) / \operatorname{{Aut}}(L)$ where $L$ is a simply connected nilpotent Lie group and ${{\Gamma}}$ is a torsion-free subgroup of the affine group $L \rtimes \operatorname{{Aut}}(L)$ so that the kernel of ${{\Gamma}}\to \operatorname{{Aut}}(L)$ has finite index in ${{\Gamma}}$ and is discrete and cocompact in $L$. In fact, Ruh produced a flat connection with parallel torsion on the tangent bundle of an almost flat manifold. The map $\Gamma \to \operatorname{{Aut}}(L)$ is the holonomy of this connection. Conversely, it is not difficult to see that every infranilmanifold is almost flat. The class of almost flat manifolds is much larger than flat manifolds; there are infinitely many almost flat manifolds in every dimension greater than two. Almost flat manifolds play a fundamental role in Riemannian geometry. By the profound Cheeger-Fukaya-Gromov theorem [@CFG], almost flat manifolds are the fibers in a collapsing sequence of Riemannian manifolds with bounded curvature and diameter. All cuspidal ends of a complete Riemannian manifold with finite volume and negative pinched sectional curvature are almost flat manifolds, generalizing the fact that the cuspidal ends of a finite volume hyperbolic manifold are flat manifolds. Pedro Ontaneda recently proved an amazing converse (see Theorem A of [@Ont]): If an almost flat manifold $M$ is the boundary of a compact manifold, then there is a compact manifold $X$ with boundary $M$ so that $X - M$ admits a complete, finite volume Riemannian metric with negative pinched sectional curvature. By rescaling, one sees that for an almost flat manifold $M$, there is a sequence of Riemannian metrics $g_i$ on $M$ so that $\operatorname{{diam}}(M,{g_i})=1$ for all $i$ and $\|K_{g_i}\|\to 0$ as $i \to \infty$. By Chern-Weil theory, the Pontryagin numbers of an oriented closed manifold are integrals of the Pfaffin forms on the curvature form, and for an almost flat manifold these integrals must converge to zero as $i \to \infty$, since by the volume comparison theorem the sequence $\operatorname{{vol}}(M,{g_i})$ is bounded above. Therefore the Pontryagin numbers of an oriented almost flat manifold $M$ all vanish. It follows that the disjoint union of $M$ with itself is an oriented boundary. Furthermore, if $M$ has an almost complex structure then, by the same reasoning, all the Chern numbers of $M$ vanish. In particular, this implies that $M$ bounds. (Throughout this paper when we say $M$ bounds we mean that $M$ is diffeomorphic to the boundary of a compact manifold.) This clearly suggests a natural and very interesting conjecture, [almost flat manifolds are boundaries,]{} due to Farrell and Zdravkovska [@FZ], which is posed independently in the famous problem list of S.T.Yau [@Ya]. It is a well-known theorem of Thom that a closed manifold bounds if and only if all its Stiefel-Whitney numbers vanish. Wall showed that a closed oriented manifold bounds an orientable manifold if and only if all Stiefel-Whitney numbers and all Pontryagin numbers are zero. The above discussion implies that if an oriented almost flat manifold bounds, then it bounds orientably. A remarkable theorem of Hamrick and Royster [@HR] shows that every flat manifold bounds. But the corresponding statement for almost flat manifolds remains a conjecture. In some special cases it has been proven. The [holonomy group]{} of an infranilmanifold is the finite group $G$ given by the image of the fundamental group ${{\Gamma}}$ in $\operatorname{{Aut}}(L)$. Farrell and Zdravkovska [@FZ] proved that almost flat manifolds bound provided either that the holonomy group $G$ has order two or that the holonomy group $G$ acts effectively on the center of $L$. Upadhyay [@Upad] proved that an almost flat manifold bounds if all of the following conditions hold: $G$ is cyclic, $G$ acts trivially on the center of $L$, and $L$ is 2-step nilpotent. This paper contains a new and quite simple proof of the above results and proves the more general statement: \[main\_theorem\] Let $M$ be an almost flat manifold and let $\operatorname{{Syl}}_2G$ be the 2-sylow subgroup of its holonomy group. If $\operatorname{{Syl}}_2G$ is cyclic or generalized quaternionic, then $M$ is the boundary of a compact manifold. Since all rational Pontryagin numbers vanish, from the above theorem and cobordism theory it follows that every oriented almost flat manifold with such a holonomy group bounds an oriented manifold. However, it remains difficult to answer: [Problem]{} (a). Does every almost flat Spin manifold (up to changing Spin structures) bound a Spin manifold? (b). Is the bordism class $[M,h]\in \Omega _(BG)$ given by the holonomy map of an almost flat manifold trivial? For an almost flat Spin manifold $M$ with holonomy map $h: M\to BG$, perhaps the $\eta$-invariant can be used to detect information on the bordism class of $[M, h]$ (cf. [@GMP].) A strong conjecture of Farrell and Zdravkovska [@FZ] asked whether an almost flat (or flat) manifold bounds a compact manifold whose interior admits a complete finite volume metric with negative (or constant negative) sectional curvature. Long and Reid [@LR] give counterexamples to the flat version of the conjecture using $\eta$-invariants. As a corollary of the recent result of Ontaneda mentioned above, we see that this strong conjecture is true in the case of cyclic or quaternionic holonomy: An almost flat manifold whose 2-sylow subgroup of the holonomy group is cyclic or quaternionic bounds a compact manifold whose interior admits a complete, finite volume Riemannian metric with negative pinched sectional curvature. Almost flat manifolds with cyclic or quaternionic holonomy bound {#setup} ================================================================ Recall that a [nilmanifold]{} is a quotient $N \backslash L$ of a simply connected nilpotent Lie group $L$ by a discrete cocompact subgroup $N$. A nilmanifold is parallelizable; indeed one projects a basis of left invariant vector fields on $L$ to the nilmanifold. Since $N \subset L$ are both nilpotent, their centers $Z(N) \subset Z(L)$ are nontrivial. Translation by an element of order two in $Z(N) \backslash Z(L)$ gives a fixed-point free involution on $N \backslash L$. We now have two proofs that a nilmanifold bounds. The first proof is that a nilmanifold is parallelizable, so its Stiefel-Whitney numbers are zero, hence by Thom’s theorem it bounds. The second proof is that a nilmanifold admits a fixed-point free involution, and any closed manifold with a fixed-point free involution $\tau : M \to M$ is a boundary: $$\partial \left( \frac{M \times [0,1]}{(m,t) \sim (\tau (m), 1-t)}\right) = M.$$ Here The proof of our main result involves a combination of these two ideas. Lemma \[+1 eigenbundle\] is key. An [infranilmanifold]{} is a double coset space $M = \Gamma \backslash L \rtimes G/ G$ where $L$ is a simply connected nilpotent Lie group, $G$ is a finite subgroup of $\operatorname{{Aut}}(L)$ and $\Gamma$ is a discrete torsion-free cocompact subgroup of $L \rtimes G$ which maps epimorphically to $G$ under the projection $L \rtimes G \to G$. We require an infranilmanifold to have positive dimension. Let $N = \Gamma \cap L$. Then $N$ is a normal subgroup of $\Gamma$ and the sequence $$1\to N\to\Gamma \to G\to 1$$ is short exact. Furthermore $N$ is a discrete cocompact subgroup of $L$ and hence is a finitely generated, torsion-free, nilpotent group. The group $N$ is called the [nillattice]{} and $G$ is called the [holonomy group]{} of the infranilmanifold. We will replace the tautological regular covers $$L \rtimes G/G \to N \backslash L \rtimes G/G \to {{\Gamma}}\backslash L \rtimes G/G,$$ using the diffeomorphisms $L \cong L \rtimes G/G$ and $N \backslash L \cong N \backslash L \rtimes G/G$. We instead consider the covers $$L \xrightarrow{p} N \backslash L \xrightarrow{\pi} {{\Gamma}}\backslash L \rtimes G/G .$$ with $p(l) = Nl$ and $\pi(Nl) = {{\Gamma}}(l,e) G$. The affine group $\operatorname{Aff}(L) = L \rtimes \operatorname{{Aut}}L$ acts on $L$ via $(l',g)l = l'g(l)$. Thus ${{\Gamma}}< \operatorname{Aff}(L)$ acts freely on $L$ and, in fact, $\pi \circ p$ is a regular ${{\Gamma}}$-cover. Likewise $G$ acts freely on $N \backslash L$ via $g(Nl) = N \gamma l$ where $\gamma \in {{\Gamma}}$ maps to $g \in G$ and $\pi$ is a regular $G$-cover. We call this $G$-action on $N \backslash L$ the [affine action]{}. The $G$-action on $L$ given by $G < \operatorname{{Aut}}L$ leaves the center $Z(L)$ invariant. The $G$-action on $L$ can be reinterpreted as conjugation in the affine group by using the short exact sequence $1 \to L \to L \rtimes G \to G \to 1$. By comparing with the short exact sequence $1 \to N \to {{\Gamma}}\to G \to 1$, we see that the $G$-action also leaves $Z(N)$ invariant but need not leave $N$ invariant. We call the $G$-actions on $L$, $Z(L)$, $Z(N)$, and $Z(N) \backslash Z(L)$ [conjugation actions]{}. They are all actions via group automorphisms. A [central involution]{} $\tau$ of an infranilmanifold $M = {{\Gamma}}\backslash L \rtimes G/G$ is an element $\tau \in Z(N) \backslash Z(L)$ of order 2 which is invariant under the conjugation action of $G$. This determines maps $\tau : M \to M, ~ {{\Gamma}}(l,g)G \mapsto \Gamma(Tl,g) G$ and $\tau : N {{\backslash}}L \to N {{\backslash}}L, ~ N(l,g)G \mapsto N(Tl,g) G$ where $T \in Z(L)$ is a representative for $\tau$. Central involutions were key in all previous work on this problem [@HR], [@FZ], [@Upad] and will be for us too. \[2-group\] 1. Let $M$ be a closed manifold with an epimorphism of its fundamental group to a finite group $G$. Then $M$ bounds if and only if $H \backslash \widetilde M$ bounds where $\widetilde M$ is the $G$-cover of $M$ and $H$ is a 2-Sylow subgroup of $G$. 2. Any infranilmanifold $M = {{\Gamma}}\backslash L \rtimes G/G$ with $G$ a 2-group has a central involution. \(i) Note that $H \backslash \widetilde M \to M$ is a odd-degree cover, hence the domain and target have the same Stiefel-Whitney numbers; thus one bounds if and only if the other does. \(ii) Let $\Sigma$ be the subgroup of $Z(N)\backslash Z(L)$ generated by the elements of order 2. Since the 2-group $G$ acts as group automorphisms of the abelian 2-group $\Sigma$, there must be at least two orbits of cardinality one, hence there must be a non-trivial element fixed by $G$. To analyze the fixed point set of a central involution, we need a group theoretic remark. \[two\_actions\] Let $X$ be an $(H,K)$-biset where both $H$ and $K$ act freely on $X$. Let $q : X \to H \backslash X$ be the quotient map, let $F$ be the fixed set of $K$ acting on $H \backslash X$, and let $\widetilde F = q^{-1}F$. For $x \in \widetilde F$, there is a function $\varphi_x : K \to H$ so that $xk = \varphi_x(k) x$. Since $H$ acts freely, this function is uniquely defined, since $X$ is a biset, this function is a homomorphism, and since $K$ acts freely, it is a monomorphism. For a monomorphism $\varphi: K \to H$, let $$\widetilde F_\varphi = \{x \in X \mid \forall k \in K, xk = \varphi(k) x \}.$$ Then $\widetilde F = \coprod \widetilde F_\varphi$. It is easy to see that $h \widetilde F_\varphi = \widetilde F_{c_h \circ \varphi}$ where $c_h : H \to H$ is conjugation. Thus $q(\widetilde F_\varphi) = q(\widetilde F_{c_h \circ \varphi})$. The group $H$ acts on the set of monomorphisms $K \to H$ by conjugation, let $[\varphi]$ denote an orbit. Let $F_{[\varphi]} = q(\widetilde F_\varphi)$. Then $F = \coprod F_{[\varphi]}$. Note that there is a bijection between $(H,K)$-bisets and left $(H \times K)$-sets, where $hxk$ corresponds to $hk^{-1}x$. Later we will apply Remark \[two\_actions\] to analyze the fixed point set of a central involution $\tau : M \to M$ by setting $X = N \backslash L$, $H = G$ and $K = \langle \tau \rangle$. A vector bundle $E \to B$ is [flat]{} if it has finite structure group, that is, $E \cong {\widetilde B} \times_G V$ for some finite, regular $G$-cover ${\widetilde B} \to B$ and some ${{\mathbb R}}G$-module $V$. We call such a flat structure a [$(G,V)$-structure]{}. Such a bundle (over a CW-complex) is the pullback of the flat bundle $EG \times_G V \to BG$ along a map $B \to BG$. The regular $G$-cover ${{\widetilde{B}}} \to B$ can also be specified by a homotopy class of map $B \to BG$ or by a $G$-conjugacy class of homomorphism $\pi_1 B \to G$. Tangent bundles of infranilmanifolds are flat: \[infranilsub\_flat\] Consider an infranilmanifold $M = {{\Gamma}}{{\backslash}}L \rtimes G /G$. Note that $T_eL$ is an ${{\mathbb R}}G$-module. The tangent bundle of $M$ is flat: $$TM = (N \backslash L ) \times_G T_eL$$ Note that $L$, like all Lie groups, is parallelizable. Indeed there is an isomorphism of vector bundles over $L$ with $\Phi : L \times T_eL \xrightarrow{\cong} TL$ given by $\Phi(l,v) = d(L_l)_{e}v$, where $L_l : L \to L$ is left translation, defined by $L_l(l') = ll'$. Note $L \rtimes G$ acts on $L$ via $(l',g)l = l'g(l)$, on $TL$ via the differential of this action, and on $T_eL$ via $v \mapsto (dg)_ev$. It is straightforward to check that $\Phi$ is $L \rtimes G$-equivariant with respect to the diagonal $L \rtimes G$-action on the domain and the action given by the differential on the target. Verify the $L$-equivariance and $G$-equivariance separately and use the identity $g \circ L_l = L_{g(l)} \circ g : L \to L$. Note $\Gamma \subset L \rtimes G$ acts freely as deck transformations on $L$ and the subgroup $N$ acts trivially on $T_eL$, so the map $ \Phi$ descends to the desired isomorphism $N \backslash L \times_G T_eL \cong TM$. \[+1 eigenbundle\] Let $\tau: TM \to TM$ be a nontrivial bundle involution on the tangent bundle of a closed connected manifold $M$. Let $F \subset M$ be the fixed set of the involution restricted to $M$. If $\tau$ restricted to $TM\|_F$ is the identity, then $M$ bounds. Let $I = [0,1]$. Extend the involution on $M$ to an involution on $M \times I$ by setting $(m,t) \mapsto (\tau(m), 1-t)$. Choose a closed involution invariant tubular neighborhood $N$ of the fixed set, thus $F \times \{1/2\} \subset N \subset M \times I$ (see Figure \[cylinder\]). Note that $(N, \partial N)$ is equivariantly diffeomorphic to the (disk bundle, sphere bundle) pair $(D(\nu \oplus 1), S(\nu \oplus 1))$ where $\nu = \nu(F \hookrightarrow M)$ is the normal bundle and where the involution on the bundle pair is given by fiberwise multiplication by $-1$. Let $P(\nu \oplus 1)$ be the projective bundle $S(\nu \oplus 1)/(v \sim -v)$. Then $W = \left(M\times I - \text{int }N \right)/\tau$ gives a cobordism of manifolds from $M$ to $P(\nu \oplus 1)$. (0,4) arc (0:360: 1.5cm and .4cm ) ; (0,2) arc (0:180 : 1.5cm and .4cm ) ; (0,2) arc (360:180: 1.5cm and .4cm ) ; (-1.5cm,1.6cm) circle (8pt) ; (-1.5cm,1.6cm) circle (8pt) ; (-1.5cm,1.6cm) circle (2pt); (-1.5cm,2.4cm) circle (2pt); (-1.5cm,2.4cm) circle (8pt); (-1.5cm,2.4cm) circle (8pt); (0,0) arc (0:180 : 1.5cm and .4cm ) ; (0,0) arc (360:180: 1.5cm and .4cm ) ; (0,0) – (0,4); (-3,0) – (-3,4); We now upgrade to a cobordism of bundles. Note that $TM \times I$ is a bundle over $M \times I$. The bundle involution on $TM$ extends to a bundle involution $TM \times I$ by setting $(v,t) \mapsto (\tau(v), 1-t)$. This descends to a bundle $\xi$ over $W$ which restricts to $TM$ over $M$. We wish to identify $\xi\|_{P(\nu \oplus 1)}$. By homotopy invariance of pullbacks of vector bundles (see Proposition 1.3 of [@Segal]), there is an equivariant isomorphism of vector bundles $TM \times I \|_{D(\nu \oplus 1)} \cong \pi_D^TM\|_F$ where $\pi_D : D(\nu \oplus 1) \to F$ is the bundle projection. Since the involution on $TM\|_F$ is trivial, there is an induced isomorphism $\xi \|_{P(\nu \oplus 1)} \cong \pi^_P~ TM\|_F$ where $\pi_P : P(\nu \oplus 1) \to F$ is the bundle projection. Thus $\xi$ gives a cobordism of bundles from $TM$ to $\pi^_P~ TM\|_F$. Since Stiefel-Whitney numbers are cobordism invariants, for all partitions $J$ of $\dim M$, $$\begin{aligned} w_J(TM)[M] &= w_J( \pi^_P~ TM\|_F )[P(\nu \oplus 1)] \\ & = (\pi^_P w_J( TM\|_F))[P(\nu \oplus 1)]\\ & =0\end{aligned}$$ since $w_J (TM\|_F)=0$ because $\|J\| = \dim M > \dim F$. Thus all Stiefel-Whitney numbers of $M$ vanish, so by Thom’s theorem $M$ bounds. One wonders if there could be a direct proof of Lemma \[+1 eigenbundle\] which avoids the use of Thom’s Theorem. Note that any involution $\tau: M \to M$ induces the involution $d \tau : TM \to TM$, but this involution does not restrict to the identity on $TM\|_F$. An almost flat manifold bounds provided that the 2-sylow subgroup of the holonomy group is cyclic or generalized quaternionic. Let $M = \Gamma \backslash L \rtimes G/ G$ be an infranilmanifold with the 2-sylow subgroup of $G$ cyclic or generalized quaternionic. By Lemma \[2-group\](i), we can pass to odd degree cover and assume that $G$ is a 2-group. Then there is a unique element $g \in G$ of order 2. (In fact, according to a theorem of Burnside, a 2-group has a unique element of order 2 if and only if it is cyclic or generalized quaternionic.) Since the center of a $p$-group is nontrivial, $g$ is central. By Lemma \[2-group\](ii), there is a central involution $\tau \in Z(N)\backslash Z(L)$. Since it is $G$-invariant, it induces a nontrivial involution $\tau : M \to M$. By Remark \[two\_actions\], applied with $X = N \backslash L$, $H = G$, and $K = \langle \tau \rangle$, the fixed set of $\tau$ is $$F = \{[x] \in M \mid x \in N\backslash L, \tau x = g x\}.$$ The involution $\tau$ on $M$ extends to the flat tangent bundle $TM = (N \backslash L) \times_{G} T_eL$ via $\tau[x,v] = [\tau x, gv]$. Then $\tau$ restricted to $TM\|_F$ is the identity since for $[x] \in F$, $\tau[x,v] = [\tau x, gv] = [gx, gv] = [x,v]$. Thus by Lemma \[+1 eigenbundle\] $M$ is a boundary. The question of whether an infranilmanifold with holonomy group the Klein 4-group bounds remains open. [99]{} J. Cheeger, K. Fukaya, M. Gromov, [Nilpotent structures and invariant metrics on collapsed manifolds]{}, J. Amer. Math. Soc. [5]{} (1992), 327–372, MR1126118, Zbl 0758.53022. F. T. Farrell and W. C. Hsiang, [Topological characterization of flat and almost flat Riemannian manifolds $M^n$ $(n\not = 3,4)$]{}, Amer. J. Math. [105]{} (1983), 641–672, MR0704219, Zbl 0521.57018. F. T. Farrell and S. Zdravkovska, [Do almost flat manifolds bound?]{} Michigan Math. J. [30]{} (1983), 199–208, MR0718265, Zbl 0543.53037. P. B. Gilkey, R. J. Miatello, R. A. Podestá, [The eta invariant and equivariant bordism of flat manifolds with cyclic holonomy group of odd prime order]{}, Ann. Global Anal. Geom. [37]{} (2010), 275–306, MR2595680, Zbl 1225.58011. M. Gromov, [Almost flat manifolds]{}, J. Differential Geom. [13]{} (1978), 231–241, MR0540942, Zbl 0432.53020. G. C. Hamrick and D. C. Royster, [Flat Riemannian manifolds are boundaries,]{} Invent. Math. [66]{} (1982), 405–413, MR0662600, Zbl 0468.57025. D. D. Long and A. W. Reid, [On the geometric boundaries of hyperbolic $4$-manifolds]{}, Geom. Topol. [4]{} (2000), 171–178, MR1769269, Zbl 0961.57011. P. Ontaneda, [Riemannian Hyperbolization]{}, <http://arxiv.org/pdf/1406.1730.pdf>. E. A. Ruh, [Almost flat manifolds]{}, J. Differential Geom. [17]{} (1982), 1–14, MR0658470, Zbl 0468.53036. G. Segal, [Equivariant $K$-theory]{}, Inst. Hautes Etudes Sci. Publ. Math. [34]{} (1968), 129–151, MR0234452, Zbl 0199.26202. S. Upadhyay, [A bounding question for almost flat manifolds]{}, Trans. Amer. Math. Soc. [353 ]{}(2001), 963–972, MR1804410, Zbl 0976.57031. S. T. Yau, [Open problems in geometry]{}, Proc. Sympos. Pure Math., Part 1 [54]{} (1993), 1–28, MR1754714, Zbl 0984.53003. [^1]: The first author is supported by a NSF Grant and the second author is supported a NSFC Key Grant. The second author wishes to thank Weiping Zhang for bring this problem to his attention.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We develop and calibrate a characteristic waveform extraction tool whose major improvements and corrections of prior versions allow satisfaction of the accuracy standards required for advanced LIGO data analysis. The extraction tool uses a characteristic evolution code to propagate numerical data on an inner worldtube supplied by a $3+1$ Cauchy evolution to obtain the gravitational waveform at null infinity. With the new extraction tool, high accuracy and convergence of the numerical error can be demonstrated for an inspiral and merger of mass $M$ binary black holes even for an extraction worldtube radius as small as $R=20M$. The tool provides a means for unambiguous comparison between waveforms generated by evolution codes based upon different formulations of the Einstein equations and based upon different numerical approximations.' author: - 'M. C. Babiuc${}^{1,3}$, B. Szilágyi${}^{2,3}$, J. Winicour${}^{3,4}$ and Y. Zlochower${}^{5}$' title: A Characteristic Extraction Tool for Gravitational Waveforms --- Introduction ============ The strong emission of gravitational waves from the inspiral and merger of binary black holes has been a dominant motivation for the construction of gravitational wave observatories. The computation of the precise details of the waveform by means of numerical simulation is a key theoretical tool to enhance detection and allow useful scientific interpretation of the gravitational signal. See [@lindbo] for a review of the accuracy required of numerically generated waveforms to fully complement the sensitivity of the LIGO [@ligo] and Virgo [@virgo] observatories. However, the waveforms are not easy to extract accurately. The radiation falls off as $1/r$ so that although it asymptotically dominates near field gravitational effects it is nevertheless small and can be contaminated by numerical error. It is common practice for the Cauchy codes used in simulating the binary black hole problem to introduce a large but finite artificial outer boundary. A combination of linearized and far field approximations are then used to extract the waveform from data on a smaller worldtube, which ideally is causally isolated from the outer boundary. Such perturbative wave extraction at a finite distance, rather than at null infinity which more faithfully represents the idealization of a distant antenna, introduces systematic errors associated with the effects of gauge, nonlinearities, nonradiative near fields and back reflection. (See [@lehnmor; @strat] for analyses of waveform errors arising from perturbative extraction at a finite distance.) An alternative approach called [Cauchy-characteristic extraction]{} (CCE) [@ccm; @cce] provides a fully nonlinear interface between Cauchy and characteristic codes which utilizes the characteristic evolution to extend the simulation to null infinity, where the waveform is computed. An earlier implementation of CCE has recently been applied to extract waveforms from binary black hole simulations [@reis1; @reis2], from rotating stellar core collapse [@reis3] and to explore the memory effect [@reis4]. In this work, we present details and tests of a redesigned CCE module whose accuracy and efficiency has undergone major improvement. The module has been designed to provide a standardized waveform extraction tool for the numerical relativity community which will allow CCE to be readily applied to a generic Cauchy code. The first attempts to simulate collisions of black holes by Hahn and Lindquist [@hahn], and then by Smarr et al [@smarr], were hampered by both a lack of computing power and a proper understanding of the mathematical formulation of Einstein’s equations required for a stable numerical solution. Their work formed the impetus for the Binary Black Hole Grand Challenge, which was formed to take advantage of the increasingly powerful computers introduced in the 1980’s. The main results of the Grand Challenge were limited to the axisymmetric head-on collision of black holes and the gravitational collapse of rotating matter [@gc]. However, the standard Arnowitt-Deser-Misner [@adm] formulation of the Einstein equations adopted by the Grand Challenge had instabilities at the analytic level which limited more general binary black hole simulations to the premerger stage. Only with new formulations was a full inspiral and merger successful, first by Pretorius [@fP05] using the harmonic formulation, and soon after by Campanelli et al [@mCcLpMyZ06] and Baker et al [@jBjCdCmKjvM06b] using the Baumgarte-Shapiro-Shibabta-Nakamura formulation [@bssn1; @bssn2]. Numerous groups now have codes which can simulate this binary inspiral problem by evolving the Cauchy problem for Einstein’s equations. In CCE, the Cauchy evolution is used to supply boundary data on a timelike inner worldtube to carry out a characteristic evolution extending to future null infinity ${\mathscr{I}}^+$, where the waveform can be unambiguously computed using the geometric methods developed by Bondi et al [@bondi], Sachs [@sachsr] and Penrose [@Penrose]. This initial-boundary value problem based upon a timelike worldtube [@tam] has been implemented as a characteristic evolution code, the PITT null code [@isaac; @highp], which incorporates a Penrose compactification of the space-time. It computes the Bondi news function at ${\mathscr{I}}^+$, which is an invariantly defined complex radiation amplitude $N= N_{\oplus}+i N_{\otimes}$, whose real and imaginary parts correspond to the time derivatives $\partial_t h_{\oplus}$ and $\partial_t h_{\otimes}$ of the “plus” and “cross” polarization modes of the strain incident on a gravitational wave antenna. The error in the PITT code was tested to be second order convergent in analytic testbeds ranging from the perturbative regime [@babiuc05] to highly nonlinear single black hole spacetimes [@highp]. One of the successes of the Grand Challenge was the successful application of the code to generic single black hole dynamical spacetimes [@stablett; @wobb; @fiss; @mod]. For a review, see [@livccm]. The propagation of gravitational waves to ${\mathscr{I}}^+$ from an astrophysically realistic source using the PITT code has in the past been limited to the simulation of an imploding neutron star using a fluid dynamic code incorporated into the characteristic code [@papadop2003; @linke]. These simulations were restricted to the axisymmetric case because of computational demands arising at the center of the star. For such systems, CCE offers a way to combine the strengths of the Cauchy and characteristic approaches. Recently this combined approach has been applied to extract the waveform from the fully 3-dimensional collapse of a rotating star [@reis3]. A global characteristic simulation of the full inspiral and merger of a relativistic binary system is not possible because of the interior caustics formed by gravitational lensing. But the application of CCE to this system has been shown to be possible [@reis1; @reis2]. The error in CCE arises from three independent sources: [(I)]{} the Cauchy evolution; [(II)]{} the worldtube module; and [(III)]{} the characteristic evolution to ${\mathscr{I}}^+$ and the computation of the waveform. [(I)]{} Errors in the Cauchy evolution can arise from numerical approximations, improper boundary treatments, extraneous radiation content in the initial data, instabilities and bugs. Errors introduced at the outer grid boundary present a special problem for BSSN formulations for which there is no theoretical understanding of the proper boundary condition. The standard practice is to extract the waveform at a finite worldtube which is large enough to justify a far field approximation but which is causally isolated from the outer boundary during the simulation. For example, perturbative extraction at $r=100M$ would require that the outer boundary be at $r>500M$ for a $t\approx 400M$ simulation. We have designed the new CCE module so that it can be applied to a generic Cauchy code with extraction radius as small as $r=20M$. However, since any universally applicable extraction module must be designed to be independent of error introduced by the Cauchy code, the extracted waveform cannot be any more reliable than the Cauchy code. [(II)]{} The main improvement described and tested in this paper is a complete overhaul of the worldtube module, which converts the output of the Cauchy evolution to boundary data on an inner worldtube for the characteristic evolution. The prior version of this module, which was used in the first applications of CCE to obtain binary black hole waveforms [@reis1; @reis2; @reis3], contained inconsistencies and bugs which prevented clean convergence tests. We have corrected this worldtube module so that the present version exhibits clean convergence to which Richardson extrapolation can be applied to produce waveforms whose numerical error due to CCE is extremely small. In addition to improvement in consistency and accuracy, we have also redesigned the module to be more efficient and user friendly. These revisions are described in Appendix \[sec:rev\]. [(IIII)]{} In addition to thoroughly scrutinizing the PITT null code for bugs, we made several major modifications. In previous applications requiring very high resolution, such as the inspiral of a particle into a black hole [@partbh], there was excessive short wavelength noise which affected the quality of the simulation. In addition, in [@reis1; @reis2] it was reported that one of the equations governing calculation of the waveform at ${\mathscr{I}}^+$ had to be linearized in order to obtain reasonable behavior. These problems have been eliminated as a result of the modifications described in Appendix \[sec:rev\]. In Sec’s. \[sec:chform\] and \[sec:tests\], we review the formalism underlying characteristic evolution and the computational structure of the PITT code. We include enough details to make clear the difficulties underlying extraction of an accurate waveform at ${\mathscr{I}}^+$ and to explain the code modifications that have been made. We also demonstrate how the use of 4th order accurate angular derivatives improves the previous test results of CCE presented in [@strat]. In Sec. \[sec:interface\], we describe the design of the new worldtube module, how it treats the Cauchy-characteristic interface and how it can be readily applied to a Cauchy evolution. In Sec. \[sec:bbhwaveforms\], we test the new extraction tool on the Cauchy evolution of the inspiral and merger of two equal-mass, non-spinning black holes. We show that CCE can now be carried out for a worldtube radius as small as $20 M$ for a mass $M$ binary system, for which perturbative extraction would not be meaningful, and which was not possible with the prior implementation of CCE. Convergence tests now demonstrate clean second order global accuracy of the evolution variables. The waveforms are only first order accurate as a result of the asymptotic limits required at ${\mathscr{I}}^+$. However, the clean first order convergence of the waveform now allows application of Richardson extrapolation to obtain higher order accuracy. In this way, in Sec. \[sec:rich\], we construct a third order accurate waveform, which was not possible with earlier versions of CCE. The ability to apply Richardson extrapolation to CCE waveforms makes it possible to show that their numerical error satisfies the standards required for application to advanced LIGO data analysis. The first derivation [@flanhugh] of the accuracy required for numerically generated black hole waveforms to be useful as templates for LIGO was carried out in the frequency domain, in which the interferometer noise spectrum is calibrated. There are two separate criteria: one ensures that the error in the model waveform does not impact wave detection and the other ensures that the error does not impact the scientific content of the signal. These criteria both depend upon the noise spectrum of the detector in a way not easily applied to a numerical simulation. This has recently prompted a translation of these requirements into the time domain in which the waveforms are computed [@lindbo; @lindob; @lind], so that they can be readily enforced in practice. In Sec. \[sec:crit\] we show that the numerical error introduced by CCE satisfies these time domain criteria for an advanced LIGO detector. We also analyze the error introduced by the choice of initial data, which has a dependency upon the size of the extraction worldtube. Characteristic Formalism {#sec:chform} ======================== The characteristic formalism is based upon a family of outgoing null hypersurfaces emanating from an inner worldtube and extending to infinity where they foliate ${\mathscr{I}}^+$ into spherical slices. We let $u$ label these hypersurfaces, $x^A$ $(A=2,3)$ be angular coordinates which label the null rays and $r$ be a surface area coordinate. In the resulting $x^\alpha=(u,r,x^A)$ coordinates, the metric takes the Bondi-Sachs form [@bondi; @sachsr] $$\begin{aligned} ds^2 & = & -\left(e^{2\beta}\frac{V}{r} -r^2h_{AB}U^AU^B\right)du^2 -2e^{2\beta}dudr -2r^2 h_{AB}U^Bdudx^A \nonumber \\ & + & r^2h_{AB}dx^Adx^B, \label{eq:bmet}\end{aligned}$$ where $h^{AB}h_{BC}=\delta^A_C$ and $det(h_{AB})=det(q_{AB})$, with $q_{AB}$ a unit sphere metric. In analyzing the Einstein equations, we also use the intermediate variable $$Q_A = r^2 e^{-2\,\beta} h_{AB} U^B_{,r}.$$ Because the Bondi variable $V=O(r^2)$ at ${\mathscr{I}}^+$, the code is written in terms or the renormalized variable $W=(V-r)/r^2$. Here $W=0$ for the Minkowski metric in null spherical coordinates. The PITT null code employs a spherical grid based upon an auxiliary unit sphere metric $q_{AB}$, with associated complex dyad $q_A$ satisfying $ q_{AB} =\frac{1}{2}\left(q_A \bar q_B+\bar q_Aq_B\right)$. The Bondi-Sachs metric $h_{AB}$ induced on the spherical cross-sections can then be represented by its dyad component $J=h_{AB}q^Aq^B /2$, with the spherically symmetric case characterized by $J=0$. The fully nonlinear $h_{AB}$ is uniquely determined by $J$, which is the principle evolution variable. The determinant condition implies that the dyad component $K=h_{AB}q^A \bar q^B /2$ is determined by $1=K^2-J\bar J$. We also introduce spin-weighted fields $U=U^Aq_A$ and $Q=Q_Aq^A$, as well as the complex spin-weight operators $\eth$ and $\bar \eth$ [@newp] which represent the angular derivatives. Refer to [@eth] for details regarding numerical implementation. In this formalism, the Einstein equations decompose into hypersurface equations, evolution equations and conservation conditions on the inner worldtube. As described in more detail in [@tam; @cce], the hypersurface equations take the form $$\begin{aligned} \beta_{,r} &=& N_\beta[J], \label{eq:beta} \\ (r^2 Q)_{,r} &=& -r^2 (\bar \eth J + \eth K)_{,r} +2r^4\eth \left(r^{-2}\beta\right)_{,r} + N_Q[J,\beta], \label{eq:wq} \\ U_{,r} &=& r^{-2}e^{2\beta}Q +N_U[J,\beta,Q], \label{eq:wua} \\ V_{,r} &=& \frac{1}{2} e^{2\beta}{\cal R} - e^{\beta} \eth \bar \eth e^{\beta} + \frac{1}{4} r^{-2} \left(r^4 \left(\eth \bar U +\bar \eth U \right) \right)_{,r} + N_W[J,\beta,Q,U], \label{eq:ww}\end{aligned}$$ where $${\cal R} =2 K - \eth \bar \eth K + \frac{1}{2}(\bar \eth^2 J + \eth^2 \bar J) +\frac{1}{4K}(\bar \eth \bar J \eth J - \bar \eth J \eth \bar J) \label{eq:calR}$$ is the curvature scalar of the 2-metric $h_{AB}$. The evolution equation for $J$ takes the form $$\begin{aligned} && 2 \left(rJ\right)_{,ur} - \left(r^{-1}V\left(rJ\right)_{,r}\right)_{,r} = \nonumber \\ && -r^{-1} \left(r^2\eth U\right)_{,r} + 2 r^{-1} e^{\beta} \eth^2 e^{\beta}- \left(r^{-1} V \right)_{,r} J + N_J[J,J_{,u},\beta,Q,U,W], \label{eq:wev}\end{aligned}$$ where $N_\beta[J]$, $N_Q[J,\beta]$, $N_U[J,\beta,Q]$, $N_W[J,\beta,Q,U]$ and $N_J[J,J_{,u},\beta,Q,U,W]$ are nonlinear terms which vanish for spherical symmetry and can be constructed from the hypersurface values of the variables appearing in their argument. Expressions for these nonlinear terms as complex spin-weighted fields and a discussion of the conservation conditions are given in [@cce]. The hypersurface equations have a hierarchical structure in the order $[J,\beta,Q,U,W]$ such that the right hand sides, e..g. $N_\beta[J]$ only depend upon previous variables and their derivatives intrinsic to the hypersurface. The finite difference grid used in the code is based upon the compactified radial coordinate $$x=\frac{r}{R_E +r} \label{eq:compx}$$ so that $x=1$ at ${\mathscr{I}}^+$. Here $R_E$ is a parameter which in the CCE module is chosen as the radius of the extraction worldtube as determined by $R^2=\delta_{ij}x^i x^j$ in terms of the Cartesian coordinates $x^i$ used in the Cauchy evolution code. The auxiliary variables $$\nu =\bar \eth J \, , \quad {\cal B}=\eth \beta \, , \quad k=\eth K \label{eq:aux}$$ are also introduced to eliminate all second angular derivatives. In certain applications this has been found to give rise to increased accuracy by suppressing short wavelength error [@gomezfo]. The finite difference scheme for integrating the hypersurface and evolution equations has been described in [@highp; @gomezfo; @luisdis]. Except for the start-up procedure described in Sec. \[sec:interface\], we follow this scheme with two modifications. First, the finite difference approximation for the $\eth$-operators is increased from 2nd order to 4th order accuracy. This can be expected to give better angular resolution but does not affect the overall 2nd order accuracy implied by the radial and time integration schemes. Second, when rewritten in terms of the compactified $x$-coordinate, the hypersurface equations for $Q$ and $W$ take the form $$x(1-x)\partial_x F + 2F = RHS \label{eq:sing}$$ where the right hand side is regular at ${\mathscr{I}}^+$. In order to deal with the degeneracy of this equation at $x=1$, we rewrite (\[eq:sing\]) in the form $$\frac {\partial (r^2 F)} {\partial (r^2)}= \frac {RHS}{2} \label{eq:singfd}$$ and construct a centered finite difference approximation with respect to $r^2$. Expressed in terms of the grid $x_i = x_{i-i} +\Delta x$, this leads to $$F_i =\bigg ( \frac {x_{i-1}(1-x_i)}{x_i(1-x_{i-1})} \bigg )^2 F_{i-1} +\frac{\Delta x(x_i+x_{i-1}-2x_i x_{i-1})}{x_i^2(1-x_{i-1})^2} \frac{RHS}{2} \label{eq:asymqw}$$ which enforces the correct asymptotic limit $F\|_{x=1} = RHS/2$ when $RHS$ is constant near ${\mathscr{I}}^+$. In practice, the variation of $RHS$ implies that this limit is only enforced to first order accuracy when $RHS$ is evaluated by the mid-point rule. This is consistent with global second order accuracy of $Q$ and $W$ when the numerical error is measured by an $L_2$-norm over the hypersurface, but only first order accuracy can be expected for their values at ${\mathscr{I}}^+$. However, the asymptotic values of $Q$ and $W$ do not enter directly into the calculation of the waveform at ${\mathscr{I}}^+$. Waveforms at ${\mathscr{I}}^+$ {#sec:waveforms} ------------------------------ For technical simplicity, the theoretical derivation of the waveform at infinity is best presented in terms of an inverse surface-area coordinate $\ell=1/r$, where $\ell=0$ at ${\mathscr{I}}^+$. In the resulting $x^\mu=(u,\ell,x^A)$ conformal Bondi coordinates, the physical space-time metric $g_{\mu\nu}$ has the conformal compactification $\hat g_{\mu\nu}=\ell^{2} g_{\mu\nu}$, where $\hat g_{\mu\nu}$ is smooth at ${\mathscr{I}}^+$ and, referring to (\[eq:bmet\]), takes the form [@tam] $$\hat g_{\mu\nu}dx^\mu dx^\nu= -\left(e^{2\beta}V \ell^3 -h_{AB}U^AU^B\right)du^2 +2e^{2\beta}dud\ell -2 h_{AB}U^Bdudx^A + h_{AB}dx^Adx^B. \label{eq:lmet}$$ As described in [@strat], the Bondi news function $N(u,x^A)$ and the Newman-Penrose Weyl tensor component [@NP] $$\Psi_4^0(u,x^A)=\lim_{r\rightarrow \infty} r \psi_4$$ which describe the waveform are both determined by the asymptotic limit at ${\mathscr{I}}^+$ of the tensor field $$\hat \Sigma_{\mu\nu} = \frac{1}{\ell}(\hat \nabla_\mu\hat \nabla_\nu -\frac{1}{4}\hat g_{\mu\nu} \hat \nabla^\alpha\hat \nabla_\alpha)\ell. \label{eq:Sigma}$$ This limit is constructed from the leading coefficients in an expansion of the metric in powers of $\ell$. We thus write $$h_{AB}= H_{AB}+\ell c_{AB}+O(\ell^2).$$ Conditions on the asymptotic expansion of the remaining components of the metric follow from the Einstein equations: $$\beta=H+ O(\ell^2) ,$$ $$U^A= L^A+2\ell e^{2H} H^{AB}D_B H+O(\ell^2)$$ and $$\ell^2 V= D_A L^A +\ell (e^{2H}{\cal R}/2 +D_A D^A e^{2H})+O(\ell^2),$$ where $H$ and $L$ are the asymptotic limits of $\beta$ and $U$ and where ${\cal R}$ and $D_A$ are the 2-dimensional curvature scalar and covariant derivative associated with $H_{AB}$. The expansion coefficients $H$, $H_{AB}$, $c_{AB}$ and $L^A$ (all functions of $u$ and $x^A$) completely determine the radiation field. One can further specialize the Bondi coordinates to be [inertial]{} at ${\mathscr{I}}^+$, i.e. have Minkowski form, in which case $H=L^A=0$, $H_{AB}=q_{AB}$ (the unit sphere metric) so that the radiation field is completely determined by $c_{AB}$. However, the characteristic extraction of the waveform is carried out in computational coordinates determined by the Cauchy data on the extraction worldtube so that this inertial simplification cannot be assumed. In order to compute the Bondi news function in the $\hat g_{\mu\nu}$ frame, it is necessary to determine the conformal factor $\omega$ relating $H_{AB}$ to a unit sphere metric $Q_{AB}$, i.e. to an inertial conformal Bondi frame [@tam] satisfying $$Q_{AB}=\omega^2H_{AB}. \label{eq:unsph}$$ (See [@quad] for a discussion of how the news in an arbitrary conformal frame is related to its expression in this inertial Bondi frame.) We can determine $\omega$ by solving the elliptic equation governing the conformal transformation of the curvature scalar (\[eq:calR\]) to a unit sphere geometry, $${\cal R}=2(\omega^2+H^{AB}D_A D_B \log \omega). \label{eq:conf}$$ Equation (\[eq:conf\]) need only be solved at the initial time. Then the geometrical properties of ${\mathscr{I}}^+$ determines the time dependence of $\omega$ according to $$2\hat n^\alpha \partial_{\alpha} \log \omega =-e^{-2H}D_AL^A, \label{eq:omegadot}$$ where $\hat n^\alpha =\hat g^{\alpha\beta}\nabla_\beta \ell$ is the null vector tangent to the generators of ${\mathscr{I}}^+$. We use (\[eq:omegadot\]) to evolve $\omega$ along the generators of ${\mathscr{I}}^+$ given a solution of (\[eq:conf\]) as initial condition. The news function $N(u,x^A)$ is first computed by the code in terms of the computational coordinates $(u,x^A)$, as opposed to the inertial coordinates $(\tilde u,y^A)$ on ${\mathscr{I}}^+$ corresponding to an idealized distant observatory. The transformation to inertial coordinates proceeds by introducing the conformally rescaled metric $\tilde g_{\mu\nu} = \omega^2 \hat g_{\mu\nu}$ in which the cross-sections of ${\mathscr{I}}^+$ have unit sphere geometry, in accord with (\[eq:unsph\]). The rescaled null vector $\tilde n^\nu = \omega^{-1} \hat n^\nu$ is then the generator of the inertial time translation on ${\mathscr{I}}^+$, i.e. $\tilde n^\nu \partial_\nu = \partial_{\tilde u}$. The inertial coordinates thus satisfy the propagation equations $$\hat n^\nu \partial_\nu \tilde u = \omega \, , \quad \hat n^\nu \partial_\nu y^A =0, \label{eq:inertialc}$$ where $\hat n^\nu\partial_\nu =e^{-2H}(\partial_u + L^A \partial_{x^A})$ in terms of the computational coordinates. The inertial coordinates are obtained by integrating (\[eq:inertialc\]), thus establishing a second pair of stereographic grid patches corresponding to $y^A$. Then the news function is transformed into $N(\tilde u, y^A)$. (More precisely, we should write $\tilde N(\tilde u, y^A) = \hat N(u,x^A)$ to distinguish the functional form of the news in the different coordinates but we forgo this complication of notation.) In addition, in order for the real and imaginary parts of $N$ to correspond to the “plus” and “cross” polarization modes of a distant observatory, we need the proper choice of complex polarization vector ${\cal Q}^{\beta}$, which in the inertial coordinates is related to the unit sphere metric on ${\mathscr{I}}^+$ by $Q^{AB} =({\cal Q}^A \bar {\cal Q}^B +\bar {\cal Q}^A {\cal Q}^B)/2$. We fix the spin rotation freedom ${\cal Q}^{\beta} \rightarrow e^{-i\eta}{\cal Q}^{\beta}$ by requiring ${\tilde n}^{\nu}{\tilde \nabla}_{\nu} {\cal Q}^{\beta}=O(\Omega)$, so that the polarization frame is parallel propagated along the inertial time flow on ${\mathscr{I}}^+$. This fixes the polarization modes determined by the real and imaginary parts of the news to correspond to those of inertial observers at ${\mathscr{I}}^+$. In order to carry this out in the computational frame we introduce the dyad decomposition $H^{AB}=(F^A{\bar F}^B+{\bar F}^A F^B)/2$ where $$F^A = q^A \sqrt{ \frac{(K+1)}{2 } } -\bar q^A J \sqrt{ 1 \over 2(K+1)} .$$ We then set ${\cal Q}^{\beta}=e^{-i\delta}\omega^{-1}F^\beta +\lambda {\tilde n}^{\beta}$, where $F^\alpha:=(0,0,F^A)$. The requirement of an inertial polarization frame, ${\tilde n}^{\nu}{\tilde \nabla}_{\nu} {\cal Q}^{\beta} =O(\Omega)$, then determines the time dependence of the phase $\delta$ according to $$2i(\partial_u +L^A\partial_A)\delta = D_A L^A +H_{AC} \bar F^C ( (\partial_u +L^B \partial_B) F^A - F^B \partial_B L^A) . \label{eq:evphase}$$ The Bondi news now takes the explicit form $$N={1\over 4}e^{-2i \delta}\omega^{-2}e^{-2H}F^A F^B \{(\partial_u+{\pounds_L})c_{AB}-{1\over 2}c_{AB} D_C L^C +2\omega D_A[\omega^{-2}D_B(\omega e^{2H})]\}, \label{eq:news}$$ where $\pounds_L$ denotes the Lie derivative with respect to $L^A$. In the inertial Bondi coordinates, the expression for the news function reduces to the simple form $$N={1\over 4}{\cal Q}^A {\cal Q}^B \partial_u c_{AB}. \label{eq:inews}$$ However, the general form (\[eq:news\]) must be used in the computational coordinates, which is challenging for maintaining accuracy because of the appearance of second angular derivatives of $\omega$. Alternatively, the waveform can be obtained from the asymptotic value of the Weyl tensor. Asymptotic flatness implies that the Weyl tensor vanishes at ${\mathscr{I}}^+$, i.e. $\hat C_{\mu\nu\rho\sigma}=O(\ell)$ in the $\hat g_{\mu\nu}$ conformal Bondi frame (\[eq:lmet\]). This is the conformal space version of the peeling property of asymptotically flat spacetimes [@Penrose]. Let $(\hat n^\mu, \hat \ell^\mu, \hat m^\mu)$ be an orthonormal null tetrad such that $\hat n^\mu=\hat \nabla^\mu \ell$ and $\hat \ell^\mu \partial_\mu=\partial_\ell$ at ${\mathscr{I}}^+$. The radiation is described in this frame by the limit $$\hat \Psi:=-\frac{1}{2} \lim_{\ell \rightarrow 0}\frac{1}{\ell} \hat n^\mu \hat m^\nu \hat n^\rho \hat m^\sigma \hat C_{\mu\nu\rho\sigma}, \label{eq:psi}$$ which in Newman-Penrose notation [@NP] corresponds to $$\hat \Psi=-(1/2)\bar \psi_4^0. \label{eq:psi40}$$ The limit is independent of how the tetrad is extended off ${\mathscr{I}}^+$. A major calculational result in [@strat] is that $$\hat \Psi=\frac{1}{2}\hat n^\mu \hat m^\nu \hat m^\rho \bigg( \hat \nabla_\mu \hat \Sigma_{\nu\rho} -\hat \nabla_\nu \hat \Sigma_{\mu\rho}\bigg )\|_{{\mathscr{I}}^+} , \label{eq:psisigma}$$ where $\hat \Sigma_{\alpha\beta}$ is given by (\[eq:Sigma\]) and where (\[eq:psisigma\]) is independent of the freedom $$\hat m^\nu \rightarrow \hat m^\nu +\lambda \hat n^\nu. \label{eq:mnfreedom}$$ In the same inertial polarization frame used in describing the news, $$\Psi=\frac{1}{2} \omega^{-3}e^{-2i\delta} \hat n^\mu F^A F^B \bigg( \partial_\mu \hat \Sigma_{AB} -\partial_A \hat \Sigma_{\mu B} - \hat \Gamma^\alpha_{\mu B}\hat \Sigma_{A \alpha} +\hat \Gamma^\alpha_{A B}\hat \Sigma_{\mu\alpha} \bigg)\|_{{\mathscr{I}}^+} . \label{eq:psia}$$ An explicit expression for $\Psi$ in terms of the asymptotic metric coefficients involves lengthy algebra which was carried out using a Maple script to write it in terms of $\eth$ operators acting on the spin-weighted computational fields and to construct the final Fortran expression for $\Psi$. In inertial Bondi coordinates, (\[eq:psia\]) reduces to the single term $$\Psi = \frac {1}{4} Q^A Q^B\partial_u^2 c_{AB} = \partial_u^2 \partial_l J\|_{{\mathscr{I}}^+} . \label{eq:psiinert}$$ This is related to the expression for the news function in inertial Bondi coordinates by $$\Psi =\partial_u N. \label{eq:PsiNu}$$ However, as in the case of the news, the full expression for $\Psi$ obtained from (\[eq:psia\]) must be used in the code. This introduces additional challenges to numerical accuracy due to the large number of terms and the appearance of third angular derivatives. These difficulties can be appreciated by considering the linearized approximation, for which considerable simplification arises. To first order in a perturbation off a Minkowski background, the nonlinear expression (\[eq:psia\]) for $\Psi$ reduces to $$\Psi=\frac{1}{2}\partial_u^2 \partial_\ell J -\frac{1}{2}\partial_u J -\frac{1}{2}\eth L -\frac{1}{8} \eth^2( \eth \bar L +\bar \eth L) + \partial_u \eth^2 H. \label{eq:linPsi}$$ In the same approximation, the news function is given by $$N =\frac{1}{2} \partial_u \partial_\ell J +\frac{1}{2} \eth^2(\omega +2H). \label{eq:linN}$$ The linearized Einstein equations imply that (\[eq:PsiNu\]), i.e. $\Psi =\partial_u N$, still holds in the linearized approximation. (In the nonlinear case, the derivative along the generators of ${\mathscr{I}}^+$ is $\hat n^\nu \partial_\nu =e^{-2H}(\partial_u +L^A \partial_A)$ and (\[eq:PsiNu\]) must be modified accordingly.) The linearized expressions (\[eq:linPsi\]) and (\[eq:linN\]) provide a starting point to compare the advantages between computing the radiation via the Weyl component $\Psi$ or the news function $N$. The troublesome terms involve $L$, $H$ and $\omega$, which all vanish in inertial Bondi coordinates. One main difference is that $\Psi$ contains third order angular derivatives, e.g. the term $\eth^3 \bar L$, as opposed to second angular derivatives in the case of $N$. This means that smoothness of the numerical error is more crucial in the $\Psi$ approach. Balancing this, another main difference is that $N$ contains the term $\eth^2 \omega$, which is a potential source of numerical error since $\omega$ must be propagated across the stereographic patch boundaries via (\[eq:omegadot\]). Test comparisons of waveforms obtained via $N$ and $\Psi$ are given in the next section. Tests of modifications to the stereographic grid {#sec:tests} ================================================ The characteristic evolution carried out by the PITT code integrates the Bondi-Sachs equations by means of a finite difference approximation [@isaac; @highp]. Stereographic coordinates $x^A=(q,p)$ are used to label the angles on the outgoing null cones. In the original code, two square stereographic patches were used, one centered about the North pole and the other about the South pole. In the new stereographic scheme introduced in [@strat], the patches were modified to have circular boundaries located just past the equator, and angular dissipation was introduced to suppress the short wavelength noise introduced by interpatch interpolation. In addition, in the original code $\eth$-derivatives were approximated by second order accurate finite difference approximations. In the present version used in this paper, the $\eth$-derivatives have been increased to fourth order accuracy. Although the overall second order convergence rate of the PITT code remains unchanged, these changes are expected to lead to more accurate waveforms. There has been extensive testing of the accuracy of past versions of the code in [@highp; @cce; @mod; @strat]. Here we repeat some of the linear wave tests presented in [@strat] in order to demonstrate the improvement obtained by fourth order accurate angular derivatives. First, in order to verify that the new treatment of stereographic patches is capable of producing a fourth order accurate evolution, we carry out a test of wave propagation on the sphere based upon solutions to the 2D wave equation $$- \partial_t ^2 \Phi + \eth \bar \eth \Phi = 0,$$ where $\Phi =cos(\omega t) Y_{lm}$, $\omega =\sqrt{l(l+1)}$ and $Y_{l m}$ are spherical harmonics. For the case $l=m=2$ we measure the convergence rate of the error. The simulations are run with $n+1$ grid points along the axes of each patch, with the grid sizes ranging from $n=80$ to $n=240$. For a given grid size, we use the $L_\infty$ norm to measure the error $${\cal E}(\Phi) =\|\|\Phi_{numeric}-\Phi_{analytic}\|\|_\infty$$ for the circular patches in each hemisphere. We measure the convergence rate for ${\cal E}(\Phi)$ at a given time $t$, for two consecutive grid sizes $n_1$ and $n_2$, by $${\cal R} = \frac{\log_2 \big ({\cal E}(\Phi)_{n_2} / {\cal E}(\Phi)_{n_1} \big )} {\log_2 \big (n_1 / n_2\big )}. \label{eq:conv}$$ Convergence rates for the derivatives are measured analogously. Excellent 4th order convergence of ${\cal E}(\Phi)$ was obtained. It is more important and challenging for assessing waveform extraction error to measure the error in the derivatives $\eth \Phi$, $\eth^2\Phi$, and $\eth^3\Phi$, since second angular derivatives enter in the computation of the Bondi news and third angular derivatives enter into the computation of $\Psi$. The convergence rates, measured with the $L_\infty$ norm over the North patch, are shown in Table \[table:2Dwave\] based upon the grid sizes $(n_1,n_2) = (80,120), (120,160), (160,200), (200,240)$. $ {\cal E}/n$ $n_1=80$ $n_1=120$ $n_1=160$ $n_1=200$ --------------------------- ---------- ----------- ----------- ----------- $ {\cal E}(\eth \Phi) $ $4.04$ $4.11$ $4.35$ $4.85$ $ {\cal E}(\eth^2\Phi)$ $4.07$ $4.24$ $4.80$ $3.95$ $ {\cal E}(\eth^3\Phi)$ $3.98$ $3.95$ $3.92$ $3.86$ \[1ex\] : Convergence rates for errors in $\eth \Phi$, $\eth^2\Phi$ and $\eth^3\Phi$ \[table:2Dwave\] For the coarser grids, good 4th order convergence is apparent for all the derivatives. As the grids are refined, the error eventually approaches (double precision) roundoff error and convergence becomes a moot question. Next we compare the accuracy of waveform extraction by computing the news function $N$ and the Weyl tensor component $\Psi$ in the test problem considered in [@strat], which is based upon a periodic, linearized gravitational wave on a Minkowski background (see Sec. 4.3 of [@BS-lin]). The linearized wave is expressed in Bondi-Sachs coordinates so that it allows direct measurement of the numerical error. The wave has period $T= \pi$ and $(l=2,m=0)$ spherical harmonic dependence, with the maximum value of $J\approx10^{-6}$. The data provided by the linearized solution at the extraction worldtube was propagated to ${\mathscr{I}}^+$ by the characteristic code, where the waveform was computed and compared to its analytic value. The computational error in the waveform was measured with the $L _2$ norm over the North patch using the $n=100$ grid. Figure \[fig:News2nd4th\] compares plots of the error in the real part of the news function $N$ (computed on the North patch) for the 2nd and 4th order accurate angular derivatives. The plots show roughly one order of magnitude improvement in accuracy for the $n=100$ grid. The corresponding plots of the error in the waveform measured by the Weyl component $\Psi$ show again roughly one order of magnitude improvement in accuracy. Further improvement in accuracy might be obtained by also increasing the radial derivatives to fourth order approximations but this could entail nonlinear complications which could affect the numerical stability of the evolution algorithm [@luisdis]. ![Plots of the $L_2$ errors ${\cal E}(N)$ vs $t$ in the real part of the news function extracted in a linearized gravitational wave test. The plots compare the errors obtained using the $2^{nd}$ and $4^{th}$ order accurate angular derivatives on an $n=100$ grid. The fourth order method reduces the error by an order of magnitude. The time variation of the error matches the period of the wave. []{data-label="fig:News2nd4th"}](Figure01){width="12cm"} In accord with (\[eq:PsiNu\]), the computation of the Weyl component $\Psi$ yields an alternative numerical value for the news $$N_\Psi = N\|_{u=0} +\int_0^u \Psi du, \label{eq:npsi}$$ where $N=N_\Psi$ in the analytic problem. Figure \[fig:NewsPsi4th\] compares these two extraction methods in terms of the errors in $N$ and $N_\Psi$ for the linearized wave test when using 4th order accurate angular derivatives. The plots show that the two methods are competitive although the error in $N_\Psi$ is slightly smaller in this case. ![Comparison plots of the $L_2$ errors ${\cal E}(N)$ and ${\cal E}(N_\Psi)$ vs $t$ in the news function computed directly and via the Weyl tensor for the linearized gravitational wave test. The results were obtained using $4^{th}$ order accurate angular derivatives on an $n=100$ grid. The two methods are competitive although ${\cal E}(N_\Psi)$ is slightly smaller in this case. []{data-label="fig:NewsPsi4th"}](Figure02){width="12cm"} Computational interface {#sec:interface} ======================= We have designed an interface that takes Cartesian grid data from a Cauchy evolution and converts it into boundary data for a characteristic evolution on a spherical grid extending to ${\mathscr{I}}^+$. We treat each component $g_{\mu\nu}(t,x^i)$ of the Cauchy metric as a scalar function in the $x^i$ Cartesian coordinates which are used in the $3+1$ evolution. In order to make the interface as flexible as possible for use as a community tool for waveform extraction, we have based it upon a spectral decomposition of the Cauchy data in the region between two world tubes or radii $R=R_1$ and $R=R_2$, where $R=\sqrt{\delta_{ij}x^i x^j}$ is the Cartesian coordinate radius. Then at a given time $t$, we decompose $g_{\mu\nu}(t,x^i)$ in terms of Chebyshev polynomials of the second kind $U_k(R)$ and spherical harmonics $Y_{l m}(\theta,\phi)$, where $(\theta,\phi)$ are related to $x^i/R$ in the standard way. The Chebyshev polynomials are conventionally defined as functions $U_k(\tau)$ on the interval $-1 \le \tau \le 1$. Here we map them to the interval $R_1 \le R \le R_2$ by the transformation $$\tau (R)=\frac{2R-R_1 -R_2}{R_2 -R_1} .$$ Thus, for $R_1<R<R_2$, we expand $$g_{\mu\nu} (t,x^i) = \sum_{kl m} C_{\mu \nu [kl m]}(t) U_k (R) Y_{l m}(\theta,\phi).$$ For the application to waveform extraction given in this paper, we choose $l\le l{Max}$, where $l{Max} =6$, and $k\le k_{Max}$, where $k_{Max} =6$. These values should be considered tentative and further experimentation is warranted to optimize accuracy. In tests of binary black holes with mass $M$ we use a relatively small range $R_2 -R_1 =10M$ and a larger value of $k_{Max}$ would certainly be needed if the range were expanded. Also, while $l{Max} =6$ might be sufficient for extraction at $R_E=100M$, a larger value might give improved results at $R_E=20M$. The coefficients $C_{\mu \nu [klm]}$ allow us to reconstruct a spherical harmonic decomposition of each component of the Cauchy metric on the extraction worldtube $R=R_E$, i.e. $$g_{\mu\nu [l m]}(t,R_E) = \sum_{k} C_{\mu \nu [kl m]}(t) U_k (R_E) .$$ This decomposition is carried out at a sequence of Cauchy time steps $t_n=t_0 +n\Delta t$, where $\Delta t$ is chosen to be much smaller than the physical time scales in the problem but, for purpose of economy, larger than the time step used for the Cauchy evolution. At each time step, the spectral coefficients are determined by a least squares fit to the Cauchy metric. The extraction module also requires the derivatives $\partial_t g_{\mu\nu}$ and $\partial_R g_{\mu\nu}$ at the extraction worldtube. The $R$-derivative is obtained analytically, at each time level $t_n$, by differentiation of the Chebyshev polynomials. In one option, the finite difference option, the $t$-derivative is constructed by a fourth-order accurate finite difference approximation based upon the sequence of Cauchy times $t=t_n$. In a second option, the fast-Fourier-transform option, we modify the Cauchy data by filtering each mode $f_n=C_{k l m} (t_n)$ to remove high-frequency noise (both numerical noise and high-frequency gauge waves). The filter works as follows. Let $f_n$ be the original data $(n=0, \dots ,N-1)$, and $g_n = f_n - a (n \Delta t) -b (n\ \Delta t)^2.$ The coefficients $a$ and $b$ are fixed by requiring that $g_N = g_0$ (where $g_N$ is extrapolated from $g_{N-1}$ and $g_{N-2}$) and $g_0 - g_{N-1} = g_2 - g_1$, i.e. the one sided derivatives taken at $n=0$ agree. This guarantees continuity of $g$ and its first derivative when periodically extended. We then perform a fast Fourier transform on $g_i$, truncate the transform at high frequencies, and perform an inverse Fourier transform to obtain a filtered $G_n$ and optionally, the inverse transform of $i \omega\ g_i$ to obtain a smooth time derivative of $G_n$. We then construct the filtered mode $C_{k l m} (t_n) = G_n + a (n\Delta t) + b (n\Delta t)^2$, as well as its time derivative. The stereographic coordinates $x^A=(q,p)$ used to label the outgoing null rays in the Bondi metric are matched to the spherical coordinates $(\theta,\phi)$ induced by the Cartesian Cauchy coordinates on the extraction worldtube by a standard transformation, using the conventions in [@eth]. The value of the surface-area coordinate $r$ in the Bondi-Sachs metric is obtained on the extraction worldtube from the 2-determinant of the Cartesian metric on the surfaces $t=t_n,R=R_E$. As a result $r_E(t_n,q,p):=r\|_{R=R_E}\ne const$ on the extraction worldtube. In order to deal with this complication, the transformation from Cartesian coordinates $(t, x^i)$ to Bondi-Sachs coordinates $(u,r,x^A)$ is carried out via an intermediate Sachs coordinate system $(u,\lambda,x^a)$ [@sachsl] where $\lambda$ is an affine parameter along the outgoing null rays. The affine freedom allows us to set $\lambda=0$ on the extraction worldtube. After carrying out the Jacobian transformation from $(t,x^i)$ to $(u,\lambda,x^A)$, the Cartesian metric and its first derivatives at the extraction worldtube provide a first order Taylor expansion in $\lambda$ (about $\lambda =0$) of the null metric in Sachs coordinates. The corresponding Taylor expansion of the metric in Bondi-Sachs coordinates then follows from the computed values of $r_E$ and $\partial_\lambda r$ at $\lambda=0$, which are obtained from the 2-determinant of the Cartesian metric [@ccm]. This allows us to build a grid based upon the characteristic coordinates $(x,q,p)$, with compactified radial coordinate $x$ given by (\[eq:compx\]). The grid values $x_i =x_{i-1} +\Delta x$, $1\le i \le n_x$, are adjusted so that $x_1 < x_E:= x\|_{R=R_E}$ and $x_{n_x}=1$. The characteristic time levels $u_n=u_{n-1}+\Delta u$ are chosen to coincide with the Cauchy times $t_n$ on the extraction worldtube by choosing $(u-t)\|_{R=R_E} =0$. In the previous version of the extraction module, the first order Taylor expansion for the Bondi metric was used to fill the gridpoints neighboring the extraction worldtube and thus initiate the radial integration of the hypersurface equations (\[eq:beta\]) - (\[eq:ww\]). However, the hypersurface equations require only 6 (real) integration constants, which can be supplied by their values at $R=R_E$. Using the Taylor expansion to fill the neighboring gridpoints leads to a potential inconsistency between the Bondi metric supplied by the Cauchy evolution and the radial derivatives determined by the characteristic hypersurface and evolution equations. In particular, we have found that such inconsistencies arising from error in the Cauchy data degrade the convergence rate of the characteristic extraction module. Because convergence of the extraction module is an important test of its reliability, we proceed here in a different manner which decouples the Cauchy and characteristic extraction errors. In the previous version of the extraction module, the Taylor expansions were also applied to the auxiliary variables $\nu=\bar \eth J$, ${\cal B}=\eth\beta$ and $k=\eth K$ by applying the $\eth$-operator to the Taylor expansions of the main variables. This was a complicated process because the $\eth$ operator intrinsic to the $\lambda=0$ extraction worldtube is not the same as the $\eth$ operator intrinsic to the $r=const$ Bondi spheres (as they differ by radial derivatives). In the process, several bugs were introduced in the radial start-up scheme. The present version of the extraction module streamlines the start-up of the auxiliary variables by avoiding the use of Taylor expansions. In this new approach, the hypersurface equations are integrated purely in terms of the values $\beta_E$, $Q_E$, $U_E$ and $W_E$ of the hypersurface variables on the extraction worldtube which are supplied by the Cauchy data. A mask is set up to identify those radial grid points $i \le B$ (referred to as “$B$ points”) for which $x_i - x_E \le \Delta x$. These grid points are “passive” points which do not directly enter in the evolution. Values of the hypersurface variables are assigned at the first active points $i=B+1$ (referred to as “$B+1$ points”) in the following manner, assuming that the values $J_E$ and $J_{B+1}$ of the evolution variable $J$ are known, as well as the values $\nu_{B+1}$ and $k_{B+1}$ of the auxiliary variables. (We address the latter assumption below in describing the start-up of the evolution algorithm.) Proceeding in the hierarchical order of the hypersurface equations, we first use (\[eq:beta\]) to determine $\beta_{B+1}$ according to $$\beta_{B+1} = \beta_E + N_\beta [J](r_{B+1} - r_E).$$ Because $N_\beta [J]$ only involves $J$ and $\partial_r J$ it may be evaluated at the mid-point between $x_E$ and $x_{B+1}$ so that the resulting error in $ \beta_{B+1}$ is $O(\Delta x^3)$. This also determines the auxiliary variable ${\cal B}=\eth \beta$ at the $B+1$ points provided the $B+1$ points on the neighboring rays have the same grid value $x_i$. However, in the case of an irregularly shaped extraction worldtube, there can be exceptions where this neighboring ray is a $B$ point. As a result, in cases where the $B$ points lie close to the boundary of the masked region they can couple to the B+1 points on neighboring rays through the $\eth$ operator. For this reason, we also update $B$ points by the same scheme used for the $B+1$ points. (If a $B$ point is within a small tolerance of the world-tube, we instead just copy the world-tube value rather than risk an ill-conditioned algorithm.) In this way, the start-up value of the auxiliary variable ${\cal B}_{B+1}$ is determined in all cases. Next in the hierarchy of hypersurface equations, we determine $Q_{B+1}$ in similar fashion. However, $N_Q[J,\beta]$ involves ${\cal B}=\eth \beta$ which cannot be determined on the extraction worldtube from the values of $\beta_E$ (because of the angular variation of $r_E$ discussed above). Consequently, in order to start up the $Q$-integration we evaluate $N_Q$ at $x_{B+1}$, where ${\cal B}_{B+1}$ is known. This results in an $O(\Delta x^2)$ error in the value of $Q_{B+1}$. Similar considerations apply to the start-up of the $U$ and $W$ integrations. As a result, the start-up leads to an overall $O(\Delta x^2)$ error in values at $x_{B+1}$, which is consistent with the global $O(\Delta x^2)$ error resulting from the remaining integration from $x_{B+1}$ to ${\mathscr{I}}^+$. This radial march to ${\mathscr{I}}^+$ proceeds in the same way as described in [@highp; @gomezfo; @luisdis] to determine all variables on the hypersurface. Having completed the radial march on the hypersurface at time $u_{N-1}$, the start-up of the integration scheme on $u_n=u_{n-1}+\Delta u$ begins with the determination of $J_{B+1}(u_n)$ from the worldtube data $J_E$, $\beta_E$, $Q_E$, $U_E$ and $W_E$ on $u_N$ and the fields already determined on $u_{N-1}$. We determine $J_{B+1}(u_n)$ using a null parallelogram algorithm [@nullp]. The evolution equation (\[eq:wev\]) for $J$ can be rewritten as $$2\partial_u \partial_r \Phi - \partial_r ( A \partial_r \Phi) = RHS$$ where $\Phi=rJ$ and $A=V/r = 1+ rW$. This can be integrated over the null parallelogram in the $(u,r)$ subspace bounded by the $u_n$ and $u_{n-1}$ hypersurfaces and by two ingoing characteristics. For constant $A$, the ingoing characteristics satisfy $r - (Au/2) = const$. As depicted in Fig. \[fig:Bp1alg\], by choosing one ingoing characteristic to pass through $r_E$ on the $u_n$ hypersurface and the other to pass through $r_{B+1}$ on the midpoint between the $u_n$ and $u_{n-1}$ hypersurfaces, we obtain the integral approximation $$\Phi(u_{n},r_-) = \Phi(u_{n},r_E) + \Phi(u_{n-1},r_+) - \Phi(u_{n-1 },r_0) +\frac { RHS (r_+ -r_0) \Delta u}{2}. \label{eq:nullpar}$$ Here the corners of the null parallelogram are located at $r_E$, $r_\pm = r_{B+1} \pm (A\Delta u/4)$ and $r_0=r_E + (A\Delta u/2)$; and the center of the null parallelogram is located at $r_c = (1/2)(r_E +r_+)$. This determines the start-up value $\Phi(u_{n},r_{B+1})$ through the second order accurate interpolation $$\Phi(u_{n},r_-)= \frac {[\Phi(u_{n},r_{B+1})-\Phi(u_{n},r_E)]r_-}{r_{B+1} -r_E}.$$ Using the worldtube data and field values on $u_{n-1}$, all other quantities can be approximated consistent with second order accuracy except for a term in $RHS$ which is proportional to $\partial_u J$. This term is treated to the required accuracy by a two-step Crank-Nicholson iteration, as is done in the main evolution scheme described in [@highp]. This leads to a value of $\Phi(u_{n},r_{B+1})$, and thus $J(u_{n},r_{B+1})$, with $O(\Delta x \Delta u^2)$ error. As in the case of the hypersurface equation, we also use this algorithm to update $J$ at the $B$ points to assure that the auxiliary variables $\nu_{B+1}$ and $k_{B+1}$ can be determined by application of the $\eth$ operator. Now the radial march continues to the $B+2$ points by a similar process. ![The start-up diagram in the $(u,r)$ subspace for the $B+1$ points (shaded squares). On the left, the extraction worldtube with fixed Cartesian radius $R_E$ moves with respect to the null grid. The null parallelogram for the start-up algorithm is bounded by the two outgoing characteristics at retarded times $u_{n-1}$ and $u_n$ and the two ingoing characteristics indicated by dashed lines. The labels for the radial null coordinate $r$ are indicated at the four corners (shaded circles). []{data-label="fig:Bp1alg"}](Figure03){width="12cm"} Convergence measurements ------------------------ In tests of the waveform and other variables obtained from a binary black hole evolution there are no exact values available for measuring error so that convergence rates cannot be obtained by use of (\[eq:conv\]). Instead, we obtain Cauchy convergence rates by using measurements obtained with three different gridsizes. For grids in the ratio $\Delta_3= \chi \Delta_2 =\chi^2 \Delta_1$ the Cauchy convergence rate of a measured quantity $F$ is given $${\cal R} = \frac{\log_2 \big ((F_3 -F_2)/(F_2-F_1) \big )} {\log_2 \chi}. \label{eq:cconv}$$ For quantities that approach the continuum value $F_0$ as $F=F_0 + G \Delta^2$, (\[eq:cconv\]) gives a convergence rate of 2 when $G$ is a smooth function independent of gridsize. In the main part of the characteristic evolution algorithm, $G$ is determined by the second derivatives of the evolution variables. However, a stochastic grid-dependent source of second order error occurs in the start-up algorithm due to the location of the $B+1$ points. The separation $x_{B+1}-x_E$ of this point from the extraction worldtube can vary discontinuously under a small change in gridsize, i.e. $x_{B+1}-x_E = (1+\epsilon) \Delta x$, where $\epsilon$ is a random number, $0 <\epsilon <1$. This random separation enters into the second order accurate approximations made in the start-up algorithm. The approach to the continuum value has the form $F=F_0 + (G+\tilde G \epsilon) \Delta^2$. Consequently, the stochastic part of the second order error can obscure the convergence rate determined by (\[eq:cconv\]) if $\tilde G$ is comparable in size to $G$. The only way to ensure clean convergence rates would be to implement a third order accurate start-up algorithm, which would involve a considerable amount of work. Fortunately, this source of error does not appear to be significant in the tests we have carried out. All the main variables exhibit second order convergence when measured at a finite radius for the results of the binary black hole inspiral presented in Sec. \[sec:bbhwaveforms\]. However, some asymptotic quantities at ${\mathscr{I}}^+$ display convergence rates intermediate between first and second order, for reasons discussed further in Sec. \[sec:bbhwaveforms\]. Constraints on the time step ---------------------------- Domain of dependence considerations place a constraint between the characteristic time step $\Delta u$ and the size of the characteristic grid analogous to the CFL condition for the Cauchy evolution. For a rough estimate, consider the Minkowski space case with the conformally rescaled metric $$ds^2=-\frac{(1-x)^2}{R_E^2} du^2 -\frac {2}{R_E} du dx + q_{AB} dx^A dx^B$$ in the compactified stereographic coordinates $(u,x,q,p)$ used in the code, for which the unit sphere metric takes the form $$q_{AB} dx^A dx^B = \frac{4}{1+p^2 +q^2} (dp^2 + dq^2).$$ The past light cone is determined by $$\frac {du}{R_E} = \frac {-dx - \sqrt {dx^2+(1-x)^2 q_{AB} dx^A dx^B}} {(1-x)^2}.$$ The restriction on the characteristic time step arising from domain of dependence considerations is strongest at the inner boundary, where $x=1/2$ (since $r_E=R_E$ in the Minkowski case); and it is also strongest at the equator, where $p^2+q^2 =1$. At these points $$\frac {\|du\|}{4 R_E} = dx + \sqrt{ dx^2+(1/4)(dp^2+dq^2 ) }.$$ For typical characteristic grid parameters, $\Delta p = \Delta q = \Delta x /4$, the resulting restriction is $$\frac {\|\Delta u\|}{R_E} < 8 K \Delta x \label{eq:cfl}$$ where $K \approx 1$ depends upon the details of the finite difference stencil. This restriction is strongest for a small extraction radius. The characteristic code monitors the corresponding restriction on $\Delta u$ determined by the curved space version of the compactified Bond-Sachs metric. For a Cauchy simulation of binary black holes with mass $M$ with timestep $\Delta t =M/32$ (sufficient to describe the frequencies typical of a binary system), (\[eq:cfl\]) leads to $$\frac {M}{256 R_E} < K \Delta x ,$$ for the choice of characteristic timestep $\Delta u = \Delta t$. The corresponding number of radial gridpoints must roughly satisfy $n_x < 128 R_E/M$. This places no limit of practical concern on the resolution of the characteristic evolution even for the small extraction radius $R_E =20 M$. Thus, for purposes of CCE, there are no demanding CFL restrictions on the characteristic grid and timestep. Initial characteristic data {#sec:initchar} --------------------------- The initial data for characteristic evolution consist of the values of $J$ on the hypersurface $u=T_0$. One way of attempting to suppress incoming radiation in this data is to set the Newman-Penrose Weyl tensor component $\psi_0=0$ on the initial null hypersurface. For a perturbation of the Schwarzschild metric, this condition implies no incoming radiation in the linearized approximation. The condition that $\psi_0$ vanish involves two-radial derivatives of $J$, which in the compactified coordinate $\ell=1/r$ takes the simple linearized form $\partial_\ell^2 J$ =0. Translated into the computational coordinate $x=1/(1+R_E \ell)$, we choose the solution $$J= \frac{J\|_{x_E} (1-x)x_E}{ (1-x_E)x},$$ which provides continuity of $J$ with its value determined by Cauchy data at the extraction worldtube. Since this choice of $J$ also vanishes at infinity, the initial slice of ${\mathscr{I}}^+$ has unit sphere geometry and equation (\[eq:conf\]) for the conformal factor has the simple solution $\omega=1$. Binary black hole measurements and waveforms {#sec:bbhwaveforms} ============================================ Here we present test results of waveform extraction from the inspiral and merger of two equal-mass, non-spinning black holes. For the Cauchy evolution we used the LazEv code [@Campanelli:2005dd; @Zlochower:2005bj] along with the Cactus framework [@cactus_web] and Carpet [@Schnetter:2003rb] mesh refinement driver. LazEV is a 8th order accurate finite difference code based upon the Baumgarte-Shapiro-Shibata-Nakamura (BSSN) formulation [@bssn1; @bssn2] of Einstein’s equations, which deals with the internal singularities by the moving puncture approach [@Campanelli:2005dd; @Baker:2005vv]. Our simulation used 9 levels of refinement with finest resolution of $h= M/80.64$, and outer Cauchy boundary at $400M$. The initial data consisted of a close quasicircular black-hole binary with orbital frequency $M\Omega = 0.050$, leading to more than a complete orbit before merger (See [@Campanelli:2006uy]). We output the metric data on the extraction worldtube every $\Delta t = M/20$. In the Cauchy evolution, we extract $\psi_4$ on spheres of Cartesian radius $R/M=50, 60, \cdots, 100$ and decompose in spin-weighted $(l,m)$ spherical harmonic modes. We use a perturbative formula [@Lousto:2010qx] to extrapolate the perturbative waveform $R \psi_4$ to $R=\infty$, $$\lim_{R\to\infty}[R \psi^{l m}_4(R,t) ] = r \psi^{l m}_4(r,t) - \frac{(l-1)(l+2)}{2}\int_0^t dt \psi_4^{l m}(r,\tau) d\tau + {\cal O}(r^2), \label{eq:extrap}$$ where $r$ is the areal radius corresponding to the Cauchy extraction radius R. The extrapolation of the perturbative waveform to infinity removes cumulative phase error which otherwise would be introduced by redshift effects. We present results for the characteristic extracted waveform either in terms of $\Psi$, related to the Bondi news by $\Psi=\partial_u N$, or, when comparing to the perturbative waveform, in terms of the Newman-Penrose component $\Psi_4=-2\bar \Psi$. For illustrative purposes, we concentrate on the dominant $(l=2,m=2)$ and sub-dominant $(l=4,m=4)$ spherical harmonic modes. The highest resolution black hole waveform extraction test was run with the following characteristic grid specifications: angular gridpoints $n_q=n_p=200$, radial gridpoints $n_x=224$. For convergence tests, we also used grids $n_q =n_p=100$, $n_x=112$ and $n_q =n_p=50$, $n_x=56$, so that the grid sizes were in the ratio $\chi = 2$. We refer to these as the $n=200$, 100 and 50 grids, respectively, The characteristic time steps used for these grids were $\Delta t = M/20 \,\, (n=200)$, $\Delta t = M/10 \,\, (n=100)$ and $\Delta t = M/5 \,\, (n=50)$. The characteristic extraction was carried out using worldtube radii $R_E=20M$, $50M$ and $100 M$. The Pitt null code was run on stereographic patches with circular boundaries using the auxiliary variables (\[eq:aux\]) to eliminate any second derivatives in the angular directions and using 4th order accurate angular derivatives. Angular dissipation was added with the coefficients $\epsilon_x=10^{-3}$, $\epsilon_u=10^{-4}$, $\epsilon_Q=\epsilon_W=10^{-6}$, in the notation of [@strat]. The best accuracy was obtained using the fast Fourier transform (FFT) option to obtain time derivatives of the worldtube data, as described in Sec. \[sec:interface\]. For strong signals, e.g. the dominant $(l=2,m=2)$ spherical harmonic mode, the finite difference (FD) and FFT options are in good agreement. However, for weak signals, e.g. the $(l=4,m=4)$ mode, the FD option can generate noticeable high frequency error. See Fig. \[fig:Psi4FFTFD\] for a comparison of waveforms computed with the two options. One likely source of the high frequency error with the FD option is the stochastic error introduced at each time level $t_n$ on the extraction worldtube by the least squares fit of the Cauchy data to the spectral expansion. In the FD option, this error is amplified when taking the time derivatives necessary to compute the waveform and becomes more prominent for short characteristic timesteps. It also becomes more prominent as the extraction radius is increased, in which case the extracted worldtube data is smaller. Similar high frequency noise is apparent in the worldtube data so that this error cannot be removed by refining the characteristic grid. However, the filtering intrinsic to the FFT option is effective in reducing this error. The remaining results reported in this paper were obtained with the FFT option. ![Comparison of the FD and FFT options for the $(2,2)$ (up) and $(4,4)$ (down) spherical harmonic modes of the real part of the characteristic waveform $\Psi$ obtained with the $n=200$ grid and extraction radius $R_E=20M$. The two options give comparable results for the $(2,2)$ mode but for the $(4,4)$, which is an order of magnitude smaller, high frequency error in the world tube data is noticeable in the waveform extracted with the FD option. []{data-label="fig:Psi4FFTFD"}](Figure04){width="12cm"} Convergence rates were measured for the $(l=2,m=2)$ spherical harmonic mode, which is the dominant mode in the waveform. Table \[bbhconv1\] gives the rates for the evolution variables measured on a sphere at Bondi radius $r=80M$ obtained with a small extraction radius $R_E=20M$ at a time corresponding to the peak of the signal ($t\approx 200M$). The rates are given for the real and imaginary part of the variables. All quantities are very close to second order convergent, including $J_{,x}$, which is the term which determines the waveform after transforming to inertial Bondi coordinates according to (\[eq:psiinert\]). $Variable$ $ Rate_{Re} $ $ Rate_{Im} $ ------------ --------------- --------------- $\beta$ $2.01$ $2.01$ $J$ $2.23$ $2.01$ $J_{,x}$ $2.03$ $2.33$ $Q$ $2.02$ $2.04$ $U$ $1.99$ $1.96$ $W$ $1.97$ $2.00$ : Convergence rates of the $(l=2,m=2)$ spherical harmonic mode on the sphere $r=80M$ for the metric variables measured at retarded time $u\approx 200M$ near the peak of the signal. The rates are given for the real and imaginary part of the variables. The extraction radius was $R=20M$. The results show that the evolution variables all display clean second order convergence. \[bbhconv1\] Table \[bbhconv2\] gives the corresponding convergence rates for these evolution variables measured at ${\mathscr{I}}^+$, again at the time corresponding to the peak of the signal and with extraction radius $R_E=20M$. In this case $Q$ and $W$ show deviation from second order convergence, consistent with the asymptotic error analysis presented in Sec. \[sec:chform\] in relation to (\[eq:asymqw\]). We also see that the derivative $J_{,x}$ deviates from second order convergence, which indicates a need for more accurate finite difference approximations near ${\mathscr{I}}^+$. There are several places in the present code where one-sided difference approximations are used for derivatives at ${\mathscr{I}}^+$. These convergence rates at the peak of the signal are representative of the rates over the entire run. This is illustrated in Fig. \[fig:ImJQConv\] which plots the rescaled errors of $Re J$ and $Im Q$ versus time at ${\mathscr{I}}^+$. $Variable$ $ Rate_{Re}$ $ Rate_{Im} $ ------------ -------------- --------------- $\beta$ $2.01$ $2.01$ $J$ $1.80$ $2.18$ $J_{,x}$ $1.23$ $1.20$ $Q$ $1.33$ $1.19$ $U$ $1.99$ $1.96$ $W$ $1.55$ $1.50$ : Convergence rates of the $(l=2,m=2)$ mode for the metric variables measured near the peak of the signal at ${\mathscr{I}}^+$, with an extraction radius $R=20M$. As expected from the analysis in Sec. \[sec:chform\], some asymptotic quantities only display first order convergence. \[bbhconv2\] ![Convergence plots of the asymptotic limits at ${\mathscr{I}}^+$ of the $(2,2)$ spherical harmonic modes of $Re J$ and $Im Q$ obtained with resolutions $n=50$, $n=100$ and $n=200$ with an extraction radius $R_E=20M$. The plots for $Re J$ are rescaled for 2nd order convergence (upper plot), while the plots for $Im Q$ (lower plot) are rescaled for 1st order convergence. The rescaled differences show that convergence rates at the peak of the signal given in Table \[bbhconv2\] are representative of the rates over the entire run. []{data-label="fig:ImJQConv"}](Figure05){width="12cm"} Table \[bbhconv3\] gives the corresponding convergence rates for the waveform as measured by the Bondi news $N$ and the Weyl component $\Psi$, again at a time corresponding to the peak of the signal and with extraction radius $R_E=20M$. We also show the convergence rate of the inertial time derivative $\partial_u N$ calculated directly from finite differencing $N$. All show roughly first order convergence. The rate for $\partial_u N$ is slightly better than that for $\Psi$, although $\Psi =\partial_u N$ in the continuum limit. The convergence rates of these quantities are affected by two chief factors: (i) the large number of terms involved in their calculation and (ii) their dependence on radial derivatives of the evolution quantities at ${\mathscr{I}}^+$. In all cases, one-sided approximations are used in several places to compute these radial derivatives. This is already apparent in the convergence rate for $J_{,x}$ shown in Table \[bbhconv2\]. $Variable$ $ Rate_{Re} $ $Rate_{Im}$ ---------------- --------------- ------------- $N$ $1.59$ $1.56$ $\partial_u N$ $1.57$ $1.55$ $\Psi$ $1.16$ $1.14$ : Convergence rates of the $(2,2)$ spherical harmonic mode for the Bondi news $N$, $\partial_u N$ obtained by finite difference, and the Weyl component $\Psi$, all measured near the peak of the signal with an extraction radius $R_E=20M$. \[bbhconv3\] Surface plots of the Bondi news $N$ and Weyl component $\Psi$, measured near the peak of the signal with an extraction radius $R_E=20M$, are shown in Fig’s \[fig:NewsB\_scri\] and \[fig:Psi4\_scri\]. Both figures display smooth angular dependence, showing that the angular dissipation is effective at removing short wavelength angular noise. In particular, there are no “spikes” near the equatorial patch boundary arising from interpatch interpolation. The main error in the waveform originates from intrinsically time dependent error in the data on the extraction worldtube. The time dependence of the real part of the characteristic extracted waveform and its comparison to the perturbative waveform are shown in Fig’s \[fig:Psi4\_Cc2\] and \[fig:Psi4\_Cc4\]. Figure \[fig:Psi4\_Cc2\] shows excellent agreement between these waveforms for the dominant $(l=2,m=2)$ mode, when both are extracted at $R=50M$. The insets show that this agreement exists in the early stages, when the amplitude is small, and persists throughout the final ringdown. Note that the perturbative extrapolation formula (\[eq:extrap\]) is essential to obtain this excellent phase agreement between the perturbative and characteristic waveforms. Figure \[fig:Psi4\_Cc4\] compares the characteristic and perturbative waveforms for the $(l=4,m=4)$ mode. In this case, the perturbative waveform is again extracted at $R=50M$ but the characteristic waveform is extracted at $20M$ to reduce the high frequency noise discussed previously. This high frequency noise can also be reduced by choosing a larger timestep for the characteristic evolution, again indicating that it arises from time derivatives of the stochastic error introduced in the worldtube data by the least squares fit. The characteristic and perturbative waveforms again show excellent agreement At very early times $t/M \approx 15$, the characteristic waveform shows an anomalous feature which can be attributed to “junk” radiation content in the initial data in the vicinity of $R_E=20M$. In addition, as shown in the insert, there is another anomalous feature, which is not understood, in the time interval about $t/M=90$. This feature is also evident in the extracted Cauchy data at $R_E=20M$. Possible sources for this feature are gauge modes excited in the interior region or numerical effects from the adaptive mesh refinement used in the Cauchy evolution. A better explanation would require further runs. ![Surface plot in the North $(q,p)$ stereographic patch of the real part of the Bondi news $N$ measured at the peak of the wave with an extraction radius $R_E=20M$. The equatorial patch boundary corresponds to the circle $p^2+q^2=1$. The smooth angular dependence near the equator shows that angular dissipation is effective at removing short wavelength noise arising from interpatch interpolation. []{data-label="fig:NewsB_scri"}](Figure06){width="12cm"} ![Surface plot in the North $(q,p)$ stereographic patch of the real part of the Weyl component $\Psi$ measured at the peak of the wave with an extraction radius $R_E=20M$. The smooth angular dependence near the equator $p^2+q^2=1$ shows that angular dissipation is effective at removing short wavelength noise arising from interpatch interpolation. []{data-label="fig:Psi4_scri"}](Figure07){width="12cm"} ![Comparison of the $(2,2)$ dominant spherical harmonic mode for $\Psi_4$ (characteristic) and $r \psi_4$ (perturbative Cauchy), both extracted at $R=50M$. The insets show that the excellent agreement extends to the early stages and the final ringdown when the amplitude is small.[]{data-label="fig:Psi4_Cc2"}](Figure08){width="12cm"} ![Comparison of the $(4,4)$ sub-dominant mode for $\Psi_4$ (characteristic) extracted at $R_E=20M$ and $r \psi_4$ (perturbative Cauchy) extracted at $R=50M$. There is good agreement in the strong amplitude regime of the wave $t/M>120$. At early times $t/M \approx 15$, the characteristic waveform exhibits effects of “junk” radiation in the initial data near $R_E=20M$. In addition, the insert magnifies an anomalous feature, which is not fully understood, in the interval about $t/M=90$. []{data-label="fig:Psi4_Cc4"}](Figure09){width="12cm"} Richardson extrapolation and convergence of the waveform {#sec:rich} ======================================================== The clean first order convergence results for the news $N$ and Weyl component $\Psi$ allows us to apply Richardson extrapolation to obtain higher order accuracy waveforms. We apply the results from the three different resolutions $n=(50,100,200)$, with grid spacing $(4\Delta, 2\Delta, \Delta)$ respectively, to obtain a third order accurate waveform as follows. The truncation error in a quantity $F$ can be represented by a power series $$F(\Delta)=F_0+F' \Delta + \frac{1}{2}F'' \Delta^2 +O(\Delta^3) .$$ We write $F_1=f(\Delta)$, $F_2=F(2\Delta)$ $F_4=F(4\Delta)$. Then the extrapolated value $$F_E = \frac{8}{3} F_1-2F_2 + \frac{1}{3} F_4$$ is 3rd order accurate, i.e. $$F_E = F_0 + +O(\Delta^3) .$$ In practice this can be confirmed as follows. Let $ F_I = 2F_1-F_2$ and $F_{II}=2F_2 -F_4$ be the second order accurate waveforms obtained using data from two resolutions. Then $ F_{II} -F_E = 4(F_{I}-F_E) +O(\Delta^3)$, i.e. $$\frac{1}{4}( F_{II} -F_E) = F_{I}-F_E \label{eq:check}$$ if we neglect the $O(\Delta^3)$ error, i.e. if we approximate the exact value $F_0$ by the third order accurate approximation $F_E$. Using the corresponding notation $(N_E, N_I, N_{II})$ for the news and $(\Psi_E, \Psi_I, \Psi_{II})$ for the Weyl component, we can check the validity of applying Richardson extrapolation to the waveform. Figure \[fig:Ncheck\] and Figure \[fig:Ncheck44\] graph the rescaled errors of the real and imaginary parts of $N_I(t) -N_E(t)$ and $\frac{1}{4}(N_{II}(t)-N_E(t))$ and Figure \[fig:Psicheck\] graphs the corresponding rescaled errors in $\Psi(t)$. In both cases, (\[eq:check\]) is confirmed. ![Plots confirming the validity of Richardson extrapolation to obtain higher order accuracy for the real and imaginary parts of the dominant $(2,2)$ spherical harmonic mode of the news $N(t)$. The rescaled errors show that $N_{I}$ and $N_{II}$ are second order accurate in accord with (\[eq:check\]). []{data-label="fig:Ncheck"}](Figure10){width="12cm"} ![Plots confirming the validity of Richardson extrapolation to obtain higher order accuracy for the sub-dominant $(4,4)$ spherical harmonic mode of the news $N(t)$. The rescaled errors show again that $N_{I}$ and $N_{II}$ are second order accurate in accord with (\[eq:check\]). []{data-label="fig:Ncheck44"}](Figure11){width="12cm"} ![Plots confirming the validity of Richardson extrapolation to obtain higher order accuracy for the real and imaginary parts of the $(2,2)$ spherical harmonic mode of the waveform $\Psi(t)$. The rescaled errors show that $\Psi_{I}$ and $\Psi_{II}$ are second order accurate in accord with (\[eq:check\]). Note that the second order error in $\Psi$ contains more high frequency noise than $N$ shown in Fig. \[fig:Ncheck\]. []{data-label="fig:Psicheck"}](Figure12){width="12cm"} These results validate the use of Richardson extrapolation to obtain third order accurate waveforms $N_E$ and $\Psi_E$. We can use $N_E$ and $\Psi_E$ as fiducial exact values and estimate the truncation error in the numerical waveforms by comparing them with the second order accurate approximates $N_I$ and $\Psi_I$. Thus the truncation errors in the news $N$and Weyl component $\Psi$ are conservatively given by $$\delta N = N_{I}-N_E =O(\Delta^2) \label{eq:Nerr}$$ and $$\delta \Psi = \Psi_{I}-\Psi_E = O(\Delta^2). \label{eq:Psierr}$$ Figure \[fig:NewsR20R50R100\] plots the differences between the dominant $(l=2,m=2)$ mode of the Richardson extrapolated waveform $N_E(t)$ obtained with extraction radii $R_E=20M$, $R_E=50M$ and $R_E=100M$. In the plot, the $R_E=20M$ waveform begins at $t=0$ and the other waveforms have been shifted backwards in time so that all three are in phase at the peak of the wave. Two sources of extraneous “junk” radiation can be seen in the figure. One arises from a [mismatch]{} between the initial characteristic and Cauchy data. The initial characteristic data $\psi_0 =0$ (see Sec. \[sec:initchar\]) implies the absence of initial radiation content on the assumption that the geometry of the initial null hypersurface is close to Schwarzschild. This assumption becomes valid as the extraction radius becomes large and the exterior Cauchy data can be approximated by Schwarzschild data. Thus this mismatch is largest for extraction at $R_E=20M$. This results in a noticeable difference at very early times between extraction at $R_E=20M$ and the other two radii. After $t/M \approx 100$, the three waveforms are in good agreement with their relative differences less than 0.6% at the peak of the wave. The second source of “junk” radiation apparent in Figure \[fig:NewsR20R50R100\] arises from the choice of conformally flat initial Cauchy data. This arises for all three extraction radii and accounts for the double hump in the news function in the interval from $t/M=0$ to $t/M =50$. ![Plots of the $(2,2)$ mode of $Re N$ obtained for extraction radii $R_E=20M$, $50M$, and $100M$. The $R_E=50M$ and $R_E=100M$ waveforms have been shifted backward in time so that they are in phase at the peak of the wave. The noticeable difference in the interval from $t/M=0$ to $t/M=100$ between the $R_E=20M$ waveform and the other two results from a mismatch between the initial characteristic and Cauchy data, which decreases with large extraction radii. For the waveforms extracted at all three radii, the double hump in the interval from $t/M=0$ to $t/M=100$ results from non-trivial “junk” radiation in the initial Cauchy data. The three waveforms are in good agreement in the inspiral and merger stage. At the peak of the wave, the relative difference between the $R_E=20M$ and $R_E=100M$ waveforms is less than 0.6% []{data-label="fig:NewsR20R50R100"}](Figure13){width="12cm"} It is of interest to measure the difference $$\delta \psi_4 = (\frac{1}{2} r \psi_4 +\bar \Psi) \label{eq:psi4err}$$ between the extracted waveform using the perturbative extrapolation formula (\[eq:extrap\]) and the Richardson extrapolated characteristic waveform, measured in accord with the normalization conventions indicated in (\[eq:psi40\]). Figure \[fig:magdiff\] plots the real part of the $(2,2)$ spherical harmonic component of $ \delta \psi_4(t)$, compared with the corresponding component of $Re \Psi$. The peak amplitude of $ \delta \psi_4(t)$ is approximately 1% of the peak amplitude of $\Psi$, which provides an estimate of the difference between perturbative and characteristic extraction. Figure \[fig:phasediff\] plots the phase difference $\delta \Phi$ between the $(2,2)$ spherical harmonic components of $r \psi_4(t)/2$ and $\Psi(t)$, i.e $$\delta\Phi = \Phi[\Psi]-\Phi [ \psi_4],$$ where e.g. $\Psi= \|\Psi\|e^{i\Phi[\Psi]}$. The phase difference is less than 0.05 radians in the interval $40M < t < 250M$ beginning after the initial burst of junk radiation and extending into the late ringdown. The phase errors become large at late times when the numerical noise in the waveform is comparable to the true signal amplitude and at very early times due to the inability of the codes to accurately model the relatively high-frequency initial data burst. Note that the magnitude and phase differences between $\psi_4$ extraction and characteristic extraction indicated in Figures \[fig:magdiff\] and \[fig:phasediff\] are based upon the perturbative extrapolation formula (\[eq:extrap\]). It is also common practice to compute $\psi_4$ at large radii and then extrapolate the values to infinity, cf. the waveform comparisons in the report of the Samurai project [@samurai]. In carrying out the extrapolation, the waveforms are translated by $r^$, where $r^ = r + 2 M \log (r/2M -1)$ is the tortoise coordinate obtained from the areal radius of the extraction sphere $r$ and $M$ is the ADM mass of the system [@Boyle:2007ft]. Here we use a linear extrapolation based upon waveforms at $R=50M$ and $R=100M$ to obtain an extrapolation $r\psi_4(lin)$ on ${\mathscr{I}}^+$ that is accurate to order $O(1/R^2)$. The deviations from the characteristic waveform are displayed in Figure \[fig:lmagdiff\], where we plot $Re[\delta \psi_4(R=50M)]$, $Re[\delta \psi_4(R=100M)]$ and $Re[\delta \psi_4(lin)]$ and in Figure \[fig:lphasediff\], where we plot the phase differences $\delta\Phi(R=50M)$, $\delta \Phi(R=100M)$ and $\delta\Phi(lin)$. The plots show how the deviations decrease with extraction radius. Linear extrapolation considerably reduces the deviation but it is interesting that perturbative extrapolation via (\[eq:extrap\]), which is based upon the single $R=100M$ result, gives the smallest deviation. As we discuss next, such time domain comparisons can be of deceptive value for gravitational wave data analysis. ![Plots of the time dependence of the real part of the $(2,2)$ spherical harmonic components of $\delta \psi_4(t)$, as defined in (\[eq:psi4err\]), and the characteristic waveform $\Psi(t)$. Here $\delta \psi_4(t)$ measures the difference between the perturbative and characteristic values of $\Psi(t)$. The approximate 1% ratio between the peak amplitudes of $\delta \psi_4(t)$ and $\Psi$ gives an estimate of the difference between perturbative and characteristic extraction. []{data-label="fig:magdiff"}](Figure14){width="12cm"} ![Plot of the time dependence of the difference in phase $\delta \Phi$, measured in radians, between the $(l=2,m=2)$ components of the characteristic waveform $\Psi(t)$ and the perturbative waveform $r\psi_4(t)/2$. The phase differences are less than 0.05 radians in the interval $40M < t < 250M$ beginning after the initial burst of junk radiation and extending into the late ringdown. The phase errors become large at late times when the numerical noise is comparable to the true signal amplitude and at very early times due the inability of the codes to accurately model the relatively high-frequency initial data burst. []{data-label="fig:phasediff"}](Figure15){width="12cm"} ![Plots of the time dependence of the difference $Re[\delta \psi_4]$ between the $(l=2,m=2)$ components of the characteristic waveform $\Psi(t)$ and the Cauchy $\psi_4$ waveforms extracted at $R=50M$ and $R=100M$, and their linear extrapolation to $R=\infty$ (denoted by “lin”). For comparison, we also include the corresponding difference (denoted by “pert”) using the perturbative extrapolation (\[eq:extrap\]). The plots show the expected trend toward smaller errors as $R\to\infty$. Interestingly, perturbative extrapolation, which uses only the $R=100M$ extraction sphere, gives the smallest deviation.[]{data-label="fig:lmagdiff"}](Figure16){width="12cm"} ![Plots of the time dependence of the phase difference $\delta\Phi $ between the $(l=2,m=2)$ components of the characteristic waveform $\Psi(t)$ and the Cauchy $\psi_4$ waveforms extracted at $R=50M$ and $R=100M$, and the corresponding linear extrapolation to $R=\infty$ (denoted by “lin"). For comparison, we also include the corresponding $\delta \Phi$ (denoted by “pert”) obtained by perturbative extrapolation (\[eq:extrap\]). The plots show the expected trend toward smaller errors as $R\to\infty$. Interestingly, perturbative extrapolation, which uses only the $R=100M$ extraction sphere, gives the smallest deviation.[]{data-label="fig:lphasediff"}](Figure17){width="12cm"} Advanced LIGO accuracy standards {#sec:crit} ================================ It has been emphasized [@linabuse] that the direct use of time domain errors would be an abuse of the accuracy standards required of model waveforms to be suitable for gravitational wave data analysis. The raw error envelopes $\delta N(t)$, $\delta \Psi(t)$ and $\delta \psi_4(t)$ cannot be used to test whether the accuracy standards are satisfied. Proper accuracy standards must take into account the power spectral density of the detector noise $S_n(f)$, which is calibrated with respect to the frequency domain strain $\hat h(f)$. Consequently, the primary accuracy standards must be formulated in the frequency domain in order to take detector sensitivity into account. Fortunately, for the purpose of calibrating waveforms from numerical simulations, it has been possible to translate the frequency domain accuracy requirements into requirements on the time domain $L_2$ error norms which meet all the needed criteria [@lindbo; @lindob]. There are two distinct criteria for waveform accuracy. First, if the numerical waveform were not sufficiently accurate then an unacceptable fraction of real signals would pass undetected through the corresponding filter. Second, the accuracy impacts on whether a detected waveform measures the physical properties of the source, e.g. mass and spin, to a level commensurate with the accuracy of the observational data. The accuracy standards for model waveforms have been formulated to prevent these potential losses in the detection and scientific measurement of gravitational waves. For a numerical waveform with strain component $h(t)$, the time domain error is measured by $${\cal E}_0 = \frac {\|\| \delta h\|\|}{\|\| h \|\|}, \label{eq:strainerr}$$ where $\delta h$ is the error in the numerical approximation and $\|\|F\|\|^2=\int dt \|F(t)\|^2$, i.e. $\|\|F\|\|$ is the $L_2$ norm, which in principle should be integrated over the complete time domain of the model waveform obtained by splicing a perturbative chirp waveform to a numerical waveform for the inspiral and merger. The error can also be measured in terms of time derivatives of the strain. In our case, the first time derivative corresponds to the error in the news $${\cal E}_1(Re N) = \frac {\|\|\delta Re N \|\|}{\|\| Re N \|\|} \, , \quad {\cal E}_1(Im N) = \frac {\|\|\delta Im N \|\|}{\|\| Im N \|\|} \label{eq:newserr}$$ and the second time derivative corresponds to the Weyl component error $${\cal E}_2 (Re \Psi) = \frac {\|\|\delta Re \Psi \|\|}{\|\| Re \Psi \|\|} \, , \quad {\cal E}_2 (Im \Psi) = \frac {\|\|\delta Im \Psi \|\|}{\|\| Im \Psi \|\|} . \label{eq:weylerr}$$ Here we measure $\delta N$ and $\delta \Psi$ according to (\[eq:Nerr\]) and (\[eq:Psierr\]) and use the 3rd order Richardson extrapolations to compute $\|\|Re N\|\|$, $\|\|Im N\|\|$, $\|\|Re \Psi \|\|$and $\|\|Im \Psi \|\|$. It is also of interest to measure the “error” $${\cal E}_2(Re \delta \psi) = \frac {\|\|Re \delta \psi_4\|\|} {\|\|Re \Psi \|\|} \, , \quad {\cal E}_2(Im \delta\psi) = \frac {\|\|Im \delta \psi_4\|\|} {\|\|Im \Psi \|\|} \label{eq:delpsi}$$ corresponding to the difference between perturbative and characteristic extraction, where $\delta{\psi_4}$ is normalized according to (\[eq:psi4err\]). In [@lindbo], it was shown that sufficient conditions to satisfy data analysis criteria for detection and measurement can be formulated in terms of any of the error norms ${\cal E}_k=({\cal E}_0, {\cal E}_1,{\cal E}_2)$, i.e. in terms of the strain, the news or the Weyl component. The accuracy requirement derived in [@lindbo] for detection is $${\cal E}_k \le C_k \sqrt{2\epsilon_{max}}, \label{eq:Ndet}$$ and the requirement for measurement is $${\cal E}_k \le C_k \frac {\eta_c}{\rho}. \label{eq:Nmeas}$$ Here $\rho$ is the optimal signal-to-noise ratio of the detector, defined by $$\rho^2 = \int_0^\infty \frac{4\|\hat h(f)\|^2}{S_n(f)} df ;$$ $C_k$ are dimensionless factors introduced in [@lindbo] to rescale the traditional signal-to-noise ratio $\rho$ in making the transition from frequency domain standards to time domain standards; $\epsilon_{max}$ determines the fraction of detections lost due to template mismatch, cf. Eq. (14) of [@lindob]; and $\eta_c\le 1$ corrects for error introduced in detector calibration. These requirements for detection and measurement, for either $k=0, 1,2$, conservatively overstate the basic frequency domain requirements by replacing $S_n(f)$ by its minimum value in transforming to the time domain. The values of $C_k$ for the inspiral and merger of non-spinning, equal-mass black holes have been calculated in [@lindbo] for the advanced LIGO noise spectrum. As the total mass of the binary varies from $0 \rightarrow \infty$, $C_0$ varies between $.65 > C_0 >0$, $C_1$ varies between $.24 < C_1 < .8$ and $C_2$ varies between $0 < C_2 < 1$. Thus only the error ${\cal E}_1$ in the news can satisfy the criteria over the entire mass range. The error in the strain ${\cal E}_0$ provides the easiest way to satisfy the criteria in the low mass case $M<<M_{\odot}$ and the error in the Weyl component ${\cal E}_2$ provides the easiest way to satisfy the criteria in the high mass case $M>>M_{\odot}$. We first concentrate on the error in the news, for which the accuracy requirement for detection is $${\cal E}_1 \le C_1 \sqrt{2\epsilon_{max}}, \label{eq:Ndet2}$$ and the requirement for measurement is $${\cal E}_1 \le C_1 \frac {\eta_c}{\rho}. \label{eq:Nmeas2}$$ Table \[Errors\] gives values of several versions of the $ {\cal E}_1$ error for the inspiral and merger of non-spinning, equal mass black holes described in Sec. \[sec:bbhwaveforms\]. For practical purposes, the error norms were computed over the time period of the simulation rather than for a complete model waveform obtained by splicing to a post-Newtonian chirp waveform. Assuming that the nonlinear error in the chirp waveform is small compared to the error in the numerical waveform, the effect is to overestimate the error norms by underestimating the denominators in (\[eq:strainerr\]), (\[eq:newserr\]) and (\[eq:weylerr\]). However, it has been pointed out in [@hannam; @pfeiffer] that splicing to a chirp waveform can produce significant error unless the numerical waveform extends to a large number of orbits, which can be computationally prohibitive; otherwise, only for binary masses $\gtrsim 100M_{\odot}$ is the splicing error negligible. An advantage of the $ {\cal E}_1$ error norm based upon the news function is that the denominator in (\[eq:strainerr\]) is directly related to the radiated energy. As a result, the factor by which $ {\cal E}_1(N)$ is overestimated is ${\cal F}^{-1/2}$, where $${\cal F}:= \frac { \Delta E({\rm Numerical}) } {\Delta E({\rm Chirp})+\Delta E(\rm{Numerical})}$$ and $\Delta E$ denotes the energy radiated in the indicated time periods. The total energy radiated during the post-Newtonian inspiral and merger can be estimated from the difference between the final black hole mass $M_H$ and the mass $M_0$ of the binary for a large initial orbit. The energy $ \Delta E({\rm Numerical})$ radiated during the numerically modeled time period can be obtained from the Bondi mass-loss formula. For the binary inspiral being considered here, the final black hole mass is $M_H\approx 0.965187$; the initial mass of the system at infinite separation (given by the sum of the individual black hole masses) is $M_0\approx 1.01447$; and $ \Delta E({\rm Numerical})\approx 0.0346 $. This leads to the fraction of energy $ {\cal F}\approx .702$ radiated during the numerical period, or $${\cal F}^{-1/2}\approx 1.19 \label{eq:enfact}$$ for the factor by which the $ {\cal E}_1$ errors in Table \[Errors\] are overestimated. We re-emphasize that the values in the Table do not include the error introduced by splicing the post-Newtonian and numerical waveforms. Besides the values $ {\cal E}_1(N)$ of the numerical truncation error in the real and imaginary part of the news function extracted at $R_E=20M$, $50M$ and $100M$, Table \[Errors\] includes the corresponding truncation error ${\cal E}_1(N_\Psi)$ obtained from integrating $\Psi$ via (\[eq:npsi\]). The Table also includes the modeling errors $ {\cal E}_1 (N_{\Delta R(20,100)})$ and $ {\cal E}_1 (N_{\Delta R(50,100)})$ in the news which results from the differences $N_{R=20}- N_{R=100}$ and $N_{R=50}- N_{R=100}$ obtained from extracting the waveform at radii $R_E=50M$ and $R_E=20M$ compared with extraction at $R_E=100M$. In computing these error norms, we integrate over the interval corresponding to $t/M \ge 100$ in Fig. \[fig:NewsR20R50R100\] to eliminate effects of the initial junk radiation. $Variable$ $ Re $ $Im $ ------------------------------------- --------------------- --------------------- ${\cal E}_1(N)_{R=20}$ $8.76\times10^{-4}$ $8.74\times10^{-4}$ ${\cal E}_1(N)_{R=50}$ $2.62\times10^{-4}$ $2.60\times10^{-4}$ ${\cal E}_1(N)_{R=100}$ $1.21\times10^{-4}$ $1.22\times10^{-4}$ ${\cal E}_1 (N_\Psi)_{R=20}$ $1.08\times10^{-3}$ $1.12\times10^{-3}$ ${\cal E}_1 (N_\Psi)_{R=50}$ $3.33\times10^{-4}$ $2.93\times10^{-4}$ ${\cal E}_1 (N_\Psi)_{R=100}$ $2.30\times10^{-4}$ $1.68\times10^{-4}$ ${\cal E}_1 (N_{\Delta R(20,100)})$ $5.41\times10^{-3}$ $5.55\times10^{-3}$ ${\cal E}_1 (N_{\Delta R(50,100)})$ $4.28\times10^{-3}$ $4.51\times10^{-3}$ : Error norms of the $(2,2)$ spherical harmonic mode for the Bondi news $N$, its counterpart $N_\Psi$ (obtained by time integral of the Weyl component $\Psi$) and for the differences $N_{\Delta R}$ comparing extraction at radii $R_E=20M$ and $R_E=50M$ to extraction at $R_E=100M$. \[Errors\] Consider first the criterion for detection where we set $\epsilon_{max} =.005$, which for advanced LIGO ensures less than a 10% signal loss, a target which is often adopted in LIGO searches for compact binaries [@lindob]. For this target, (\[eq:Ndet\]) reduces to ${\cal E}_1 \le 0.1C_1$, or ${\cal E}_1 \le .024$ for the low mass bound $C_1\approx.24$. This criterion is easily satisfied by all the error norms in Table \[Errors\]. Thus the advanced LIGO detection criterion is satisfied by CCE waveforms obtained from either the news or Weyl component throughout the entire binary mass range. In addition, the detection criterion is unaffected by modeling errors introduced by choice of extraction radius. Note that the ${\cal E}_1(N)$ and ${\cal E}_1 (N_\Psi)$ errors decrease with larger extraction radius. This is expected since the truncation error introduced by the characteristic evolution code depends upon the size of the integration region between the extraction worldtube and ${\mathscr{I}}^+$. The criterion for measurement is more stringent. For a calibration factor given by the expected lower bound $\eta_{min} =0.4$ and for the lower bound $C_1\approx.24$ corresponding to the small mass limit, (\[eq:Nmeas2\]) reduces to $${\cal E}_1 \le C_1 \frac {\eta_c}{\rho}= \frac{ 9.6\times 10^{-2}}{\rho}. \label{eq:Nmeas3}$$ For the most optimistic advanced LIGO signal-to-noise ratio, which is expected to be $\rho \approx 100$ for the strongest and best tuned events, the requirement for measurement is then $ {\cal E}_1 \le 9.6 \times 10^{-4}$. Thus, comparing (\[eq:Nmeas3\]) to the values in Table \[Errors\], the advanced LIGO measurement criterion is satisfied throughout the entire binary mass range by the numerical truncation error $ {\cal E}_1(N)$ in the CCE waveform obtained from the news function. The ${\cal E}_1 (N_\Psi)$ error obtained from the Weyl component for extraction radii $R_E\ge 50M$ also satisfy this full range of measurement standards. The value of ${\cal E}_1 (N_\Psi)$ for $R_E=20M$ would satisfy the full range of measurement standards for $\rho< 100$ if reduced by the factor ${\cal F}^{-1/2}$ given in (\[eq:enfact\]). Note also that the truncation error is being conservatively measured by the $O(\Delta^2)$ error (\[eq:Psierr\]), rather than the third order accurate error in the Richardson extrapolated waveform. These results can be compared with the measurement criterion for advanced LIGO data analysis reported in the Samurai project [@samurai], which was also based upon a non-spinning, equal-mass binary black hole inspiral and merger. There it was found that the mismatch between perturbative waveforms obtained using various Cauchy codes limited the measurement application to signal-to-noise ratios $\rho\lesssim 25$. This is consistent with our experience, and that reported in [@reis2], that the additional truncation error introduced by applying CCE to a Cauchy simulation of a binary inspiral is much smaller than the numerical error resulting from the Cauchy code. The values of ${\cal E}_1 (N_{\Delta R})$ in Table \[Errors\] give an estimate of the modeling error introduced by different choices of extraction radius. The error ${\cal E}_1 (N_{\Delta R(50,100)})$, introduced by extraction at $R_E=50M$ as compared to $R_E=100M$, only satisfies the full range of measurement standards for signal-to-noise ratios $\rho<21$, or $\rho<25$ if (\[eq:enfact\]) is taken into account. This would cover the most likely advanced LIGO events. These modeling errors primarily result from initialization effects which we have discussed and which would be less significant in simulations with a higher number of orbits. The results suggest that the choice of extraction radius should be balanced between a sufficiently large radius to reduce initialization effects and a sufficiently small radius where the Cauchy grid is more highly refined and outer boundary effects are better isolated. For the Cauchy grid setup in the present case, there is a factor of 2 in refinement at $r= 50M$ compared to $r=100M$, which for 8th order finite differencing has considerable impact on the error. Future experiments with longer runs involving more orbits will supply valuable guidance for optimizing the extraction radius. Table \[ErrorPsi4i\] gives some pertinent ${\cal E}_2$ error norms for the $(2,2)$ spherical harmonic component. Besides the numerical truncation error ${\cal E}_2 (\Psi)$ obtained for characteristic extraction at $R_E=20M$, $R_E=50M$ and $R_E=100M$, the Table includes the error norm ${\cal E}_2(\delta\psi)$ measuring the difference between perturbative and characteristic extraction, as defined in (\[eq:delpsi\]), obtained at $R_E=50M$ and $R_E=100M$. The Table also includes the modeling error $ {\cal E}_2 (\Psi_{\Delta R(50,100)})$ resulting from the difference $\Psi_{R=50}- \Psi_{R=100}$ obtained using characteristic extraction at radii $R_E=50M$ and $R_E=100M$, as well as the corresponding error norm $ {\cal E}_2 (\psi_{4,\Delta R(50,100)})$ resulting from the difference obtained using perturbative extraction at $R_E=50M$ and $R_E=100M$, i.e. $$\begin{aligned} {\cal E}_2 (Re \psi_{4,\Delta R(50,100)}) = \frac {\|\|Re [(r\psi_4/2)\|_{r=100M}- (r\psi_4/2)\|_{r=50M}]\|\|} {\|\| Re \Psi \|\|} \\ {\cal E}_2 (Im \psi_{4,\Delta R(50,100)}) = \frac {\|\|Im [(r\psi_4/2)\|_{r=100M}- (r\psi_4/2)\|_{r=50M}]\|\|} {\|\| Im \Psi \|\|} .\end{aligned}$$ All norms are again computed over the time interval $t/M \ge 100$ indicated in Fig. \[fig:NewsR20R50R100\] to reduce effects of initial junk radiation. Although the ${\cal E}_2$ norms are not effective for low mass binaries, they give some useful information for comparing extraction at various radii and comparing characteristic and perturbative extraction. In the high mass limit for which $C_2=1$ and with the same lower limits for $\epsilon_{max}$ and $\eta_c$ as for the ${\cal E}_1$ norms, the detection criterion (\[eq:Ndet\]) reduces to $${\cal E}_2 \le \sqrt{2\epsilon_{max}}= 0.1$$ and the measurement criterion (\[eq:Nmeas\]) reduces to $${\cal E}_2 \le \frac {\eta_c}{\rho} =\frac {0.4}{\rho} . \label{eq:2meas}$$ All the error norms in Table \[ErrorPsi4i\] satisfy the detection requirement for this high mass limit. The truncation errors ${\cal E}_2 (\Psi)$ decrease with extraction radius as in the case of the ${\cal E}_1 (N_\Psi)$ errors. The values at all three extraction radii satisfy the measurement requirement for the most optimistic advanced LIGO signal-to-noise ratio $\rho=100$. These results are consistent with the ${\cal E}_1 (N_{\Psi})$ error in Table \[Errors\] obtained by integrating $\Psi$. The norms ${\cal E}_2(\delta \psi)$ measure the difference between characteristic and perturbative extraction. The results in the Table show that this difference is fairly independent of whether the waveforms are extracted at $R_E=50M$ or $R_E=100M$. In the high mass limit in which (\[eq:2meas\]) is valid, these errors impact the measurement criterion only for signal to noise ratios $\rho>59$ but they could be expected to be more significant for low mass binaries. Whether the ${\cal E}_2(\delta \psi)$ error can be attributed to characteristic extraction or to perturbative extraction cannot be decided from this single test and deserves further investigation. A definitive answer would of course require knowledge of the “exact” waveform. The modeling error ${\cal E}_2(\psi_{4,\Delta R(50,100)})$, which results from comparing perturbative extraction at $R_E=50M$ and $R_E=100M$, is considerably larger than the corresponding modeling error ${\cal E}_2(\Psi_{\Delta R(50,100)})$ for characteristic extraction. This confirms the expectation that perturbative extraction requires a large extraction radius. Both of these modeling errors are substantial, which further emphasizes the importance of an optimal choice of extraction radius. $Variable$ $ Re $ $Im $ -------------------------------------------- --------------------- --------------------- ${\cal E}_2 (\Psi)_{R=20}$ $1.14\times10^{-3}$ $1.17\times10^{-3}$ ${\cal E}_2 (\Psi)_{R=50}$ $4.04\times10^{-4}$ $3.53\times10^{-4}$ ${\cal E}_2 (\Psi)_{R=100}$ $2.81\times10^{-4}$ $2.09\times10^{-4}$ ${\cal E}_2 (\delta \psi)_{R=50}$ $5.09\times10^{-3}$ $5.08\times10^{-3}$ ${\cal E}_2(\delta \psi)_{R=100}$ $6.81\times10^{-3}$ $6.32\times10^{-3}$ ${\cal E}_2(\Psi_{\Delta R(50,100)}) $ $1.94\times10^{-2}$ $1.91\times10^{-2}$ ${\cal E}_2(\psi_{4,\Delta R(50,100)}) $ $3.13\times10^{-2}$ $3.14\times10^{-2}$ : ${\cal E}_2$ error norms for the $(2,2)$ spherical harmonic mode of the CCE waveform $\Psi$ obtained using extraction worldtube radii $R_E=20M$, $R_E=50M$ and $R_E=100M$ and the norms of the difference $\delta \psi$ between $\Psi$ and the perturbative $\psi_4$ waveforms extracted at $50M$ and $100M$. We also tabulate the modeling error ${\cal E}_2(\Psi_{\Delta R(50,100)})$ resulting from the difference in extracting $\Psi$ at $50M$ and $100M$, and the corresponding modeling error ${\cal E}_2(\psi_{4,\Delta R(50,100)})$ for extraction via $\psi_4$ \[ErrorPsi4i\] Conclusion ========== We have developed a new characteristic waveform extraction tool. Bugs and inconsistencies in the previous version have been eliminated. The extracted waveform from a binary black hole inspiral now shows clean convergence. We have demonstrated that this allows the use of Richardson extrapolation to obtain third order accurate waveforms whose numerical truncation error satisfies the advanced LIGO standards for detection and measurement. Characteristic waveform extraction from a binary black hole inspiral can now be obtained without any recourse to linearization and from extraction radii as small as $R=20M$. The Cauchy interface has been simplified in terms of a spectral decomposition. There are still elements where accuracy could be improved. Some of these, such as more accurate start-up algorithms for the radial integrations at the extraction worldtube and more accurate asymptotic limits at ${\mathscr{I}}^+$, might be handled by small modifications but others, such as extending the overall accuracy to 4th or higher order, would entail a more major overhaul of the underlying PITT code. This is perhaps long overdue, but a proper treatment would require a better understanding of the underlying mathematical problem. The well-posedness of the gravitational worldtube-nullcone initial-boundary value problem upon which the code is based has not yet been established. Only recently has well-posedness been demonstrated for the corresponding nonlinear scalar wave problem [@kwsc]. The PITT code was developed in the early days of numerical relativity when considerations of well-posedness did not arise in the formulation of Cauchy as well as characteristic codes. The development of a stable characteristic code involved “educated guesses”. Today, the numerical relativity community is more aware of the benefits that a well-posed problem can bring. Most important, a proof of well-posedness of the continuum problem by means of energy estimates can be converted to ensure stability of the corresponding finite difference problem by the analogous estimates obtained by summation by parts. A new characteristic code based upon this approach would be of great value. Of equal value would be the implementation of Cauchy-characteristic matching (CCM), in which the characteristic evolution is used to supply outer boundary data for the Cauchy evolution. So far, CCM has only been successfully applied to a harmonic Cauchy code in the linearized regime [@harm]. Although there is room for further improvement in the CCE tool presented here, there also is pressing interest from several numerical relativity groups to apply the tool to extract waveforms from binary black hole inspirals. The emerging importance of this problem to the future of gravitational wave astronomy has created an urgency to make characteristic waveform extraction widely available. Simulations of binary black hole inspirals are too computationally expensive to be carried out solely for the purpose of wave extraction tests. This would conflict with the demands to apply computational resources to results of importance to gravitational wave astronomy and binary black hole astrophysics. However, the extra computational expense of adding characteristic extraction to a Cauchy simulation is fractionally small. For our tests, where we extracted twice as often as required, the interpolation, decomposition, and saving of the metric data used only $\approx6.9\%$ of the total simulation time. The application of characteristic extraction to simulations of astrophysical importance will at the same time provide a practical approach to improving the extraction tool by comparing results obtained with different formulations, different numerical techniques and different grid specifications. In particular, our test results emphasize the need for a better understanding of the optimal choice of extraction radius, which would balance between the discretization error in the Cauchy code, the initialization error, the error originating at the outer Cauchy boundary and the relatively small discretization error from CCE. We have demonstrated here how the module can be applied to the LazEv code, which is a finite-difference BSSN code, to produce calibrated binary black hole waveforms. We welcome applications to codes based upon other formulations of the Einstein equations, e.g. the harmonic formulation, and based upon other numerical methods, e.g. spectral methods. In particular, characteristic extraction offers a way to unambiguously compare binary black hole waveforms obtained from the same initial data using codes based upon different formulations of the Einstein equations, different numerical techniques, different evolution gauges and different methods of treating the internal singularities (by punctures or by excision). Such comparisons would be of especial importance in the case of precessing binaries composed of high spin black holes, where the reliability of perturbative extraction has not been extensively tested. We have made the present characteristic waveform extraction tool publicly available as part of the Einstein Toolkit [@einstk]. We thank L. Lindblom and C. Reisswig for many helpful discussions. B. S. acknowledges support from the Sherman Fairchild Foundation and NSF grants PHY-061459 and PHY-0652995 to the California Institute of Technology. J. W. acknowledges support from NSF grants PHY-0553597and PHY-0854623 to the University of Pittsburgh. Y. Z. acknowledges support from NSF grants PHY-0722315, PHY-0653303, PHY-0714388, PHY-0722703, DMS-0820923, PHY-0929114, PHY-0969855, and NSFâ€“CDI-1028087 and NASA grants 07-ATFP07-0158 and HST-AR-11763 to RIT and computational resources provided by the Ranger cluster at TACC (Teragrid allocations TG-PHY080040N and TG-PHY060027N) and by NewHorizons at RIT. M. C. B. acknowledges support from NSF grant PHY-0969709 to the Marshall University and computational resources provided by the Teragrid allocation TG-PHY090008. M. C. B. and B. S. thank the University of Pittsburgh for hospitality during the major part of this project. An essential component of this work is the PITT null code, to which N. T. Bishop, R. G[' o]{}mez, R. A. Isaacson, L. Lehner, P. Papadopoulos and J. Welling have made major contributions. Code revision {#sec:rev} ============= [Revisions to the worldtube module:]{} - The numerical error in the previous version of the worldtube module did not converge properly upon grid refinement. We have traced this problem to an inconsistency in the startup algorithm for the integration of the characteristic equations away from the extraction worldtube. Data from the Cauchy code had been used in an overdetermined manner to supply the integration constants for the characteristic equations. As a result, the Cauchy evolution introduced inconsistencies with the characteristic equations which degraded convergence of the numerical error. We have revamped this start-up algorithm so that the worldtube module now has clean second order accuracy with respect to grid size. - We have found and corrected bugs which had been introduced in the implementation of features designed to improve code performance. In particular, in parallelizing the code using the Cactus framework [@cactus_web], a complex spin-weighted term in the evolution module was incorrectly declared to be a real variable. In addition, it had been suggested that improved accuracy could be obtained by reducing second derivatives in the angular directions to first order form by the use of auxiliary variables [@gomezfo]. In the process of doing so, values of certain variables in the subroutines for the data at the extraction worldtube were inadvertently interchanged between the North and South stereographic patches. The introduction of these bugs made the resulting code inconsistent with the Einstein equations. (From tracing through the code archive, we determined that the bugs were introduced in 2002 or later so that they do not affect the validity of results prior to 2002. C. Reisswig has informed us that he recomputed some of the results in [@reis1; @reis2] using the corrected code and found good qualitative agreement with the original results.) - The matching interface has been simplified by introducing a pseudospectral decomposition of the Cauchy metric in the neighborhood of the extraction worldtube. This provides more economical storage of the inner boundary data for the characteristic code so that the waveform at ${\mathscr{I}}^+$ can be obtained with small computational burden compared to the Cauchy evolution. - Interpolation error arises because the characteristic grid points do not lie exactly on the extraction worldtube determined by the Cauchy coordinates. The interpolation stencils change in a discontinuous way when the grid is refined. Consequently, although this interpolation error is second order in grid size, there is a small stochastic component relative to the choice of grid. This can obscure the results of convergence tests. We have reduced such sources of error so that convergence tests can be used to validate the interface modules. - We have streamlined the start-up procedure at the extraction worldtube by initializing the auxiliary variables (introduced to remove second angular derivatives) directly in terms of the main variables. - In previous applications of the extraction module, it was expedient to set the characteristic data on the initial hypersurface to zero outside some radius. This necessitated a transition region to obtain continuity with the initial Cauchy data, which requires non-zero initial characteristic data at the extraction worldtube. Here we initialize the data by requiring that the Newman-Penrose component of the Weyl tensor intrinsic to the initial null hypersurface vanish, i.e. $\psi_0 =0$. For a linear perturbation of the Schwarzschild metric, this condition eliminates incoming radiation crossing the initial null hypersurface. Since $\psi_0$ consists of a second radial derivative of the characteristic data, this condition allows both continuity at the extraction worldtube and the desired asymptotic falloff of the characteristic data at infinity. [Modifications of the PITT code:]{} - A source of error in characteristic evolution is the intergrid interpolations arising from the stereographic patches used to coordinatize the spherical cross-sections of the outgoing null hypersurfaces. The previous version of the code used two square stereographic patches centered about the North and South poles, each overlapping the equator. This has now been modified by shrinking the overlap region so that each patch has a circular boundary located slightly past the equator, as is the practice in the use of stereographic grids in meteorology [@browning]. This eliminates the region near the corners of the square patch where the numerical error was most troublesome. Angular numerical dissipation has also been introduced and shown to be effective in controlling the short wavelength noise arising from the intergrid interpolations across the stereographic patches. Tests show that the resulting waveforms have smooth numerical error as functions on the sphere [@strat]. Characteristic codes based upon a six patch covering of the sphere [@reisswig; @roberto] offer the potential for better accuracy but they have not yet been developed to handle waveform extraction. See [@strat] for a comparison of the six patch and the stereographic approaches on a test problem. - The accuracy of the angular derivatives has been increased to a 4th order finite difference approximation, as opposed to the 2nd order accuracy in the original code. The radial derivatives and time integration remain second order accurate. - Some of the differential equations governing propagation along the characteristics become degenerate at ${\mathscr{I}}^+$ and affect the accuracy of asymptotic quantities such as the Bondi news function. The correct asymptotic behavior has now been incorporated into the finite difference approximation in order to increase accuracy. In addition, the accuracy of certain one-sided finite difference approximations necessary at ${\mathscr{I}}^+$ has also been improved. - In addition, the code has been extended to supply the waveform at ${\mathscr{I}}^+$ in terms of the radiative component of the Weyl tensor as well as the Bondi news function. For tests in the linearized regime, extraction via the Weyl tensor was found to be slightly more accurate than via the news function when large gauge effects are introduced in the characteristic coordinates [@strat]. On the other hand, the higher derivatives involved in computing the Weyl tensor lead to less smoothness in the numerical error. Overall, the two methods are competitive. [44]{} L. Lindblom, J. G. Baker and B. J. Owen, “Improved time-domain accuracy standard for modeled gravitational waveforms”, \[arXiv:gr-qc/1008.1803\] (2010). B. Abbott et al, (LIGO Scientific Collaboration), [Rept. Prog. Phys.]{} [72]{} 076901 (2009) F. Acernese et al, (Virgo Collaboration), [Class. Quant. Grav.]{} [25]{}, 114045 (2008). L. Lehner and O. M. Moreschi, “Dealing with delicate issues in waveform calculations”, [Phys.Rev.D]{} [76]{} 124040 (2007). M.C. Babiuc, N.T. Bishop, B. Szilágyi and J. Winicour, “Strategies for the characteristic extraction of gravitational waveforms”, [Phys. Rev. D]{} [79]{} 084011 (2009). N.T. Bishop, R. Gómez, L. Lehner, B. Szilágyi, J. Winicour and R. A. Isaacson, “Cauchy-Characteristic Matching”, in B Iyer and B Bhawal (Eds.), [Black Holes, Gravitational Radiation and the Universe]{}, Kluwer Academic Publishers, Dordrecht, 1998. N.T. Bishop, R. Gómez, L. Lehner, and J. Winicour, [Phys. Rev. D]{} [54]{} 6153 (1996). C. Reisswig, N. T. Bishop, D. Pollney and B. Szilágyi, “Unambiguous determination of gravitational waveforms from binary black hole mergers”, [Phys. Rev. Lett.]{} [95]{}, 221101 (2009). C. Reisswig, N. T. Bishop, D. Pollney and B. Szilágyi, “Characteristic extraction in numerical relativity: binary black hole merger waveforms at null infinity”, [Class. Quantum Grav.]{} [27]{}, 075014 (2010). C. Reisswig. C. D. Ott, U. Sperhake and E. Schnetter, “Gravitational wave extraction in simulations of rotating stellar core collapse”, \[arXiv:gr-qc/1012.0595\] (2010). C. Reisswig and D. Pollney. “Gravitational memory in binary black hole mergers”, \[arXiv:gr-qc/1004.4209\] (2010). S. Hahn and R. W. Lindquist, [Ann. Phys. NY]{} [ 29]{}, 304 (1964). L. Smarr, A. Cadez, B. de Witt and K. R. Eppley, [Phys. Rev. D]{} [14]{}, 2443 (1976). R. Matzner, H. E. Seidel, S. L. Shapiro, L. Smarr, W. -M. Suen, S. A. Teukolsky and J. Winicour (The Binary Black Hole Grand Challenge Alliance), [Science]{} [270]{}, 885 (1995). R. Arnowitt, S. Deser and C. W. Misner, [Gravitation: an introduction to current research]{}, ed. L. Witten (Wiley, New York, 1962). F. Pretorius,“Evolution of binary black-hole spacetimes”, [Phys. Rev. Lett.]{}, [95]{}, 121101(2005). “M. Campanelli, C. O. Lousto, P. Marronetti and Y. Zlochower”, “Accurate evolutions of orbiting black-hole binaries without excision”, [Phys. Rev. Lett.]{}, [96]{}, 111101 (2006). J. G. Baker, J. Centrella, D.-I. Choi, M. Koppitz and J. van Meter, “Gravitational-wave extraction from an inspiraling configuration of merging black holes", [Phys. Rev. Lett.]{}, [96]{}, 111102 (2006). M. Shibata and T. Nakamura, “Evolution of three-dimensional gravitational waves: Harmonic slicing case”, [Phys. Rev. D]{} [52]{}, 5428 (1995). T. Baumgarte and S. L. Shapiro, “On the numerical integration of Einstein’s field equations”, [Phys. Rev. D]{} [59]{}, 024007 (1999). H. Bondi, M.J.G. van der Burg and A.W.K. Metzner, [Proc. R. Soc. A]{} [269]{} 21 (1962). R.K. Sachs, [Proc. R. Soc. A]{} [270]{} 103 (1962). R. Penrose, [Phys. Rev. Letters]{} [10]{} 66 (1963). L.A. Tamburino and J. Winicour, [Phys. Rev.]{} [150]{} 1039 (1966). R.A. Isaacson, J.S. Welling and J. Winicour, [J. Math. Phys.]{} [24]{} 1824 (1983). N. T. Bishop, R. Gómez, L. Lehner, M. Maharaj and J. Winicour, [Phys. Rev. D]{} [56]{} 6298 (1997). M. Babiuc, B. Szilágyi, I. Hawke, and Y. Zlochower, “Gravitational wave extraction based on Cauchy-characteristic extraction and characteristic evolution”, [Class. Quantum Grav.]{} [22]{} 5089 (2005) R. Gómez, et al (The Binary Black Hole Grand Challenge Alliance), “Stable characteristic evolution of generic three-dimensional single-black-hole spacetimes”, [Phys. Rev. Lett.]{} [80]{}, 3915 (1998). R. G[ó]{}mez, L. Lehner, L. R. L.Marsa and J. Winicour, “Moving black holes in 3D”, [Phys. Rev. D]{} [57]{}, 4778 (1997). R. G[ó]{}mez, S. Husa, L. Lehner and J. Winicour, “Gravitational waves from a fissioning white hole”, [Phys. Rev. D]{} [66]{}, 064019 (2002). Y. Zlochower, R. Gómez, S. Husa, L. Lehner and J. Winicour, “Mode coupling in the nonlinear response of black holes,” [Phys. Rev. D]{} [68]{}, 084014 (2003). J. Winicour, “Characteristic evolution and matching”, [Living Rev. Relativity]{} [3]{}, (2009). F. Siebel, J. A. Font, E. M[ü]{}ller and P. Papadopoulos, “Axisymmetric core collapse simulations using characteristic numerical relativity”, [Phys. Rev. D]{} [67]{}, 124018 (2003). F. Linke and J. A. Font, H. T. Janka, E. M[ü]{}ller and P. Papadopoulos, “Spherical collapse of supermassive stars: Neutrino emission and gamma ray bursts”, [Astron. Astrophys.]{} [376]{}, 568 (2001). N. T. Bishop, R. Gómez, S. Husa, L. Lehner, J. Winicour “A numerical relativistic model of a massive particle in orbit near a Schwarzschild black hole”, [Phys.Rev. D]{} [68]{}, 084015 (2003). . [É]{}. Flanagan and S. A. Hughes, “Measuring gravitational waves from binary black hole coalescences. I. Signal to noise ratio for inspiral, merger and ringdown”, [Phys. Rev. D]{} [57]{}, 4535 (1998). L. Lindblom, B. J. Owen and D. A. Brown, “Model waveform accuracy standards for gravitational wave data analysis”, [Phys. Rev. D]{} [78]{}, 124020 (2008). L. Lindblom, “Optimal calibration accuracy for gravitational-wave detectors”, [Phys. Rev. D]{} [80]{}, 042005 (2009). E.T. Newman and R. Penrose, [J. Math. Phys.]{} [7]{}, 863 (1966). R. Gómez, L. Lehner, P. Papadopoulos and J. Winicour, “The eth formalism in numerical relativity”, [Class. Quantum Grav.]{} [14]{} 977, 1997. R. G[ó]{}mez, “Gravitational waveforms with controlled accuracy”, [Phys. Rev. D]{} [64]{}, 1–8 (2001). L. Lehner, “A dissipative algorithm for wave-like equations in the characteristic formulation”, [J. Comput. Phys.]{} [149]{}, 59–74 (1999). E.T. Newman and R. Penrose, [J. Math. Phys.]{} [3]{}, 566, 1962. J. Winicour, [Gen. Rel. and Grav.]{} [19]{} 281 (1987). N.T. Bishop, [Class. Quantum Grav.]{} [22]{} 2393 (2005). R.K. Sachs, [J. Math. Phys.]{} [3]{} 908 (1962). R. G[ó]{}mez and J. Winicour, “Asymptotics of gravitational collapse of gravitational waves”, [J. Math. Phys.]{} [33]{}, 1445 (1992). M. Campanelli, C. O. , P. Marronetti and Y. Zlochower, “Accurate evolutions of orbiting black-hole binaries without excision”, [Phys.Rev.Lett.]{} [96]{}, 111101 (2006). Y. Zlochower, J. G. Baker, M. Campanelli and C. O. Lousto, “Accurate black hole evolutions by fourth-order numerical relativity,” [Phys. Rev. D]{} [72]{}, 024021 (2005). . E. Schnetter, S. H. Hawley and I. Hawke, “Evolutions in 3D numerical relativity using fixed mesh refinement,” [Class. Quantum Grav.]{} [21]{}, 1465 (2004). J. G. Baker, J. Centrella, D.-I. Choi, M. Koppitz, and J. van Meter, “Gravitational wave extraction from an inspiraling configuration of merging black holes” [Phys.Rev.Lett.]{} [96]{}, 111102 (2006). M. Campanelli, C. O. Lousto and Y. Zlochower, “Gravitational radiation from spinning-black-hole binaries: The orbital hang up,” [Phys. Rev. D]{} [74]{}, 041501 (2006). J. Baker, M. Campanelli, C.O. Lousto and R. Takahashi “Modeling gravitational radiation from coalescing binary black holes”, [Phys.Rev. D]{}, [65]{}, 124012–124034 (2002). C. O. Lousto, H. Nakano, Y. Zlochower, and M. Campanelli, “Intermediate-mass-ratio black hole binaries: intertwining numerical and perturbative techniques,” arXiv\[gr-qc\] 1008.4360 (to appear in [Phys. Rev. D.]{}). M. Hannam, S. Husa, J. G. Baker, M. Boyle, B. Br[" u]{}gmann, T. Chu, N. Dorband, F. Hermann, I. Hinder, B. J. Kelly, L. E. Kidder, P. Laguna, K. D. Matthews, J. R. van Meter, H. P. Pfeiffer, D. Pollney, C. Reisswig, M. A. Scheel and D. Shoemaker, “Samurai project: Verifying the consistency of black-hole-binary waveforms for gravitational-wave detection”, [Phys. Rev. D]{} [79]{}, 084025 (2009). M. Boyle, D. A. Brown, L. E. Kidder, A. H. Mroue, H. P. Pfeiffer, M. A. Scheel, G. B. Cook, S. A. Teukolsky, “High-accuracy comparison of numerical relativity simulations with post-Newtonian expansions”, [Phys. Rev. D]{} [76]{}, 124038 (2007). L. Lindblom, “Use and abuse of the model waveform standards”, [Phys. Rev. D]{} [80]{}, 064019 (2009). M. Hannam, S. Husa, F. Ohme and P. Ajith, “Length requirements for numerical relativity waveforms”, [Phys. Rev. D]{} [82]{}, 124052 (2010). H. MacDonald, S. Nissake and H. P. Pfeiffer, “Suitability of post-Newtonian / numerical relativity hybrid waveforms for gravitational wave detectors”, [Class.Quantum. Grav.]{} [28]{}, 134002 (2011). H.-O. Kreiss and J. Winicour, “The well-posedness of the null-timelike boundary problem for quasilinear waves”, [Class. Quantum Grav.]{} [28]{},145020 (2011). B. Szil[á]{}gyi, and J. Winicour, “Well-posed initial-boundary evolution in general relativity” [Phys. Rev. D]{} [68]{}, 1 (2003). http://www.einsteintoolkit.org. G. L. Browning, J. J. Hack and P. N. Swarztrauber “A Comparison of Three Numerical Methods for Solving Differential Equations on the Sphere”, [Monthly Weather Review]{} [117]{}, 10582 (1989). C. Reisswig, N.T. Bishop, C.W. Lai, J. Thornburg and B. Szilágyi “Characteristic evolutions in numerical relativity using six angular patches”, [Class. Quantum Grav.]{} [24]{} S237-S339 (2007) R. Gómez, W. Barreto, and S. Frittelli, “A framework for large-scale relativistic simulations in the characteristic approach”, [Phys. Rev. D]{} [76]{} 124029–124050 (2007).	{ "pile_set_name": "ArXiv" }
--- abstract: \| Using supervised machine learning approaches to recognize human activities from on-body wearable accelerometers generally requires a large amount of labelled data. When ground truth information is not available, too expensive, time consuming or difficult to collect, one has to rely on unsupervised approaches. This paper presents a new unsupervised approach for human activity recognition from raw acceleration data measured using inertial wearable sensors. The proposed method is based upon joint segmentation of multidimensional time series using a Hidden Markov Model (HMM) in a multiple regression context. The model is learned in an unsupervised framework using the Expectation-Maximization (EM) algorithm where no activity labels are needed. The proposed method takes into account the sequential appearance of the data. It is therefore adapted for the temporal acceleration data to accurately detect the activities. It allows both segmentation and classification of the human activities. Experimental results are provided to demonstrate the efficiency of the proposed approach with respect to standard supervised and unsupervised classification approaches.\ author: - 'D. Trabelsi, S. Mohammed, F. Chamroukhi, L. Oukhellou, and Y. Amirat [^1]' title: An Unsupervised Approach for Automatic Activity Recognition based on Hidden Markov Model Regression --- Unsupervised learning, activity recognition, wearable computing, multivariate regression, Hidden Markov Model Introduction {#sec:Intro} ============ aging population has recently gained an increasing attention due to its socio-economic impact. By $2050$, the number of people in the European Union aged $65$ and above is expected to grow by $70\%$ and the number of people aged over $80$ by $170\%$[^2]. This demographic change poses increasing challenges for healthcare services and their adaptation to the needs of this aging population. Facing this problem or reducing its effect would have a great societal impact by improving the quality of life and regaining people independence to make them active in society. The aim is therefore to facilitate the daily activity lives of elderly or dependent people at home, to increase their autonomy and to improve their safety. In fact, most elderly prefer to stay at home in the so-called “aging in place" [@Kaluza]. The emergence of novel adapted technologies such as wearable and ubiquitous technologies is becoming a privileged solution to provide assistive services to humans, such as health monitoring, well being, security, etc. Among which, activity recognition has a wide range of promising applications in security monitoring as well as human machine interaction [@Lu2009]. A large amount of work has been done in this active topic over the past decades; nevertheless it is still an open and challenging problem [@Jinhui].\ Several techniques have been used to quantify these activities such as video-based sensors [@Brdiczka2009], wearable-based sensors, environmental sensors and object sensors (smart phones, RFID, etc.). Recently, the use of wearable-sensors based systems for activity recognition has gained more attention on a large number of technological fields such as navigation, monitoring and control of aircrafts [@MacKenzie; @Carminati], medical application [@Jovanov; @Wu], localization and robots [@Barshan; @Tan]. Among the inertial sensors used for activity recognition, the accelerometers are the most commonly used [@Jiayang]. They have shown satisfactory results to measure the human activities in both laboratory/clinical and free-living environment settings [@Mathie]. In addition, the latest advances in Micro-ElectroMechanical Systems (MEMS) technology have greatly promoted the use of accelerometers thanks to the considerable reduction in size, cost and energy consumption. Early studies in activity recognition used uniaxial accelerometers, while recent studies use mainly tri-axial accelerometers [@Noury; @Yang].\ Related work on human activity recognition ========================================== Regarding the human activity classification, one can make the distinction between supervised and unsupervised classification approaches. Supervised classification techniques consist in inferring a decision rule from labelled training data. The use of the supervised activity classification approaches has shown promising results [@Altun]. Some supervised approaches have enhanced the activity recognition process performances by using spatio-temporal information [@Chen]. Regarding the algorithms used in the supervised context, one can cite $k$-Nearest Neighbor ($k$-NN) algorithm [@Liu], multi-class Support Vector Machines (SVM) [@Qian] and Artificial Neural Networks (ANN) including both MultiLayer Perceptron (MLP) [@Dong; @Yang08] and Radial Basis Function (RBF) networks [@Faber]. Nevertheless, the collection of sufficient amounts of labelled data for a various and rich set of free-living activities may be sometimes difficult to achieve and computationally expensive [@Cvetkovic]. On the other hand, unsupervised classification techniques try to directly construct models from unlabelled data either by estimating the properties of their underlying probability density (called density estimation) or by discovering groups of similar examples (called clustering). The unsupervised learning techniques are of particular interest for an exploratory analysis of large amounts of unlabelled data. They can also consist in a preliminary task to further run a supervised classifier based on the obtained partition of the data. The use of an unsupervised approach may be needed in such a context of activity recognition when it is difficult to have labels for the data. Regarding the approaches used in the unsupervised context, one can cite the well-known $k$-Means algorithm [@Duda], the Gaussian Mixture Models (GMM) approach [@Allen] and the one based on Hidden Markov Model (HMM) [@Lin; @rabiner] or HMM with GMM emission probabilities [@Mannini]. Both the GMM and the HMM approaches use the EM algorithm [@dlr]. The HMM has shown good results in earlier exploratory studies thanks to their main advantage of suitability to model sequential data which is the case of monitoring human activities. Indeed, the acceleration data are measured over time during physical human activities of a person and are therefore sequential over time. The EM algorithm [@dlr] (also called Baum-Welch [@BaumWelch]) in the context of HMM is particularity adapted for unsupervised learning.\ In this study, an unsupervised approach for human activity recognition is proposed. It combines an HMM-based model with the use of acceleration data acquired during sequences of different human activities. More specifically, the proposed approach is based on a Hidden Markov Model in a multiple regression context and will be denoted by MHMMR. As the sequences of acceleration data consist in multidimensional time series where each dimension is an acceleration, the activity recognition problem is therefore formulated through the proposed MHMMR model as the one of joint segmentation of multidimensional time series, each segment is associated with an activity. In the proposed model, each activity is represented by a regression model and the switching from one activity to another is governed by a hidden Markov chain. The MHMMR parameters are learned in an unsupervised way from unlabelled raw acceleration data acquired during human activities. The most likely sequence of activities is then estimated using the Viterbi algorithm [@Viterbi]. The proposed technique is then evaluated on real-world acceleration data collected from three sensors placed at the chest, the right thigh and the left ankle of the subject.\ This study is an extension of the paper [@Trabelsi2012] where additional technical implementations are shown: Twelve activities and transitions are studied and performances of the proposed approach are evaluated and compared to those of some well known unsupervised and supervised techniques for activity recognition.\ This paper is organized as follows; section \[sec: experimental setup\] presents the experimental protocol and the data acquisition platform. Section \[sec: Seg with hidden Markov model regression\] presents the proposed model and its unsupervised parameter estimation technique from unlabelled acceleration data. In section \[sec: experimental study\], the performances of the proposed approach are evaluated and compared to those of some well known unsupervised and supervised techniques for activity recognition. Data Collection {#sec: experimental setup} =============== In this study, human activities are classified using three sensors placed at the chest, the right thigh and the left ankle respectively as shown in Figure \[fig\_MTX\]. Sensors placement is chosen to represent predominantly upper-body activities such as standing up, sitting down, etc. and predominantly lower body activities such as walking, stair ascent, stair descent, etc. The sensor’s placement guarantees at the same time less constraint and better comfort for the wearer. The attachment of the sensors to the human body should be well fitted and secured (Fig. \[fig\_MTX\]). These sensors consist of three MTx 3-DOF inertial trackers developed by Xsens Technologies [@Xsens]. Each MTx unit consists of a tri-axial accelerometer measuring the acceleration in the 3-D space (with a dynamic range of $\pm$5g where g represents the gravitational constant). Our experiences show also that the measured ankle-sensor accelerations during the different activities do not exceed the limit of $\pm$5g. The sampling frequency is set to 25 Hz, which is sufficient and larger than 20 Hz the required frequency to assess daily physical activity [@Bouten]. The sensors were fixed on the subject with the help of an assistant before the beginning of the measurement operation. Raw acceleration data are therefore collected over time when performing the activities. The MTx units are connected to a central unit called Xbus Master that is attached to the subject’s belt. Figure \[fig\_data\_gather\] shows the data gathering process from the Xbus-MTx acquisition system to the host pc. The Xbus Master is directly connected to the chest MTx unit while the remainig MTx units (thigh and ankle) are connected in series. Data transmission between the Xbus Master and the pc is carried out through a Bluetooth wireless link. ![MTx-Xbus inertial tracker and sensors placement[]{data-label="fig_MTX"}](MTX.eps){width="6.5cm" height="3.5cm"} ![Data ghathering from the MTx-Xbus acquisition system[]{data-label="fig_data_gather"}](data_gathering.eps){width="6.5cm" height="3cm"} The experiments were performed at the LISSI Lab/University of Paris-Est Créteil (UPEC) by six different healthy subjects of different ages (who are not the researchers) in the office environment. In order to gather various and rich dataset, the recruited volunteer subjects have been chosen in a given margin of age (25-30) and weight (55-70) kg. Twelve activities and transitions were studied and are listed as follows: Stairs down (A$1$) - Standing (A$2$) - Sitting down (A$3$) - Sitting (A$4$) - From sitting to sitting on the ground (A$5$) - Sitting on the ground (A$6$) - Lying down (A$7$) - Lying (A$8$) - From lying to sitting on the ground (A$9$) - Standing up (A${10}$) - Walking (A${11}$) - Stairs up (A${12}$). The activities were chosen to have an appropriate representation of everyday activities involving different parts of the body (fig. \[fig:Activities\]). The recognized activities and transition differ in duration and intensity level. Note that the activities $A_3$, $A_5$, $A_7$, $A_9$ and $A_{10}$ represent dynamic transitions between static activities. Each subject was asked to perform the twelve activities in his own style and was not restricted on how the activities should be performed but only with the sequential activities order. In addition, the duration of each activity is not restricted to be the same as it may vary from one subject to another. ![Examples of some considered activities: a) Climbing Stairs Down, b) Climbing Stairs Up, c) Walking, d) Sitting, e) Standing Up, f) Sitting on the ground.[]{data-label="fig:Activities"}](nouha.eps){width="8cm" height="3.5cm"} With three MTx sensor units, each one with a tri-axial accelerometer, a total of nine accelerations are therefore measured and recorded overtime for each activity. Since the goal is to recognize human activities from only the raw acceleration data, the acquired acceleration signals can be seen as multidimensional time series (of dimension 9) with regime changes due to the changes of activities over time. The activity recognition problem can therefore be formulated as a problem of multidimensional time series segmentation. Indeed, segmenting the time series according to different unknown regimes over time is equivalent to classifying the acceleration data into one set of activities; each activity being associated with a regime. This will be detailed in the next section that is dedicated to the proposed Hidden Markov Model Regression (HMMR) approach. Segmentation with Multiple Hidden Markov Model Regression - MHMMR {#sec: Seg with hidden Markov model regression} ================================================================== In this section, the problem of activity recognition (classification) is formulated as the one of joint segmentation of multidimensional time series. Indeed, the acceleration data are presented as multidimensional time series presenting various regime changes. In such context, the goal is to provide an automatic partition of the data into different segments (regimes), each segment being considered afterwards as an activity. Various modelling approaches of time series presenting regime changes have been proposed in literature. One can cite in particular the piecewise regression as one of the most adapted modelling approaches [@McGee; @Brailovsky]. The piecewise model has been applied in many domains including finance, engineering, economics, and bioinformatics [@Picard]. In the piecewise regression model [@Brailovsky], data are partitioned into several segments, each segment being characterized by its mean polynomial curve and its variance. However, the parameter estimation in such method requires the use of dynamic programming algorithm [@Bellman; @Stone] which may be computationally expensive especially for time series with large number of observations. Moreover, the standard piecewise regression model usually assumes that noise variance is uniform for all the segments (homoskedastic model). An alternative approach extended in this paper is based on Hidden Markov Model Regression [@Fridman]. This approach can be seen as an extension of the standard Hidden Markov Model (HMM) [@rabiner] to regression analysis. Each regime is described by a regression model rather than a simple constant mean over time, while preserving the Markov process modelling for the sequence of unknown (hidden) activities. Indeed, standard HMM-based approaches use simple Gaussian densities as density of observation. However, in the HMM regression context, each observation is assumed to be a noisy polynomial function to better model very structured data as the acceleration data. The approach we propose further extends the HMM model to a multiple regression setting. This is due to the fact the observed acceleration data is multidimensional. In the following, the Hidden Markov Regression Model for time series modelling is used by formulating its basic and multiple regression setting. In this framework, each observation, denoted by $\bsy_i$, represents the $i$th acceleration measurement while the associated state (class), denoted by $z_i$, represents its corresponding activity. General description of the Multiple Hidden Markov Model Regression {#ssec: MHMMR} ------------------------------------------------------------------ In Hidden Markov Model Regression (HMMR), each time series is represented as a sequence of observed univariate variables $(y_1, y_2, \hdots, y_n)$, where the observation $y_i$ at time $t_i$ is assumed to be generated by the following regression model [@Fridman]: $$y_i = \bsbeta^T_{z_i}\bt_i + \sigma_{z_i}\epsilon_i \quad ; \quad\epsilon_i \sim \N(0,1) , \quad (i=1,\ldots,n) \label{eq: HMM regression model}$$ where $z_i \in \{1,\ldots,K\}$ is a hidden discrete-valued variable. In this application case, $z_i$ represents the hidden class label (activity) of each acceleration data point and $K$ corresponds to the number of considered activities. The variable $z_i$ controls the switching from one polynomial regression model associated to one activity, to another of $K$ models at time $t_i$. The vector $\bsbeta_{z_i}=(\beta_{z_i0},\ldots,\beta_{z_i p})^T$ is the one of regression coefficients of the $p$-order polynomial regression model $z_i$ and $\sigma_{z_i}$ is its corresponding standard deviation, $\bt_i=(1, t_i,t_i^{2} \ldots, t_i^{p})^T$ is the $p+1$ dimensional covariate vector at time $t_i$ and the $\epsilon_i$’s are standard Gaussian variables representing an additive noise. The HMMR assumes that the hidden sequence $\bz=(z_1,\ldots,z_n)$ is a homogeneous Markov chain of first order parameterized by the initial state distribution $\pi$ and the transition matrix $\bA$. It can be shown that, conditionally on a regression model $k$ ($z_i=k$), $y_i$ has a Gaussian distribution with mean $\bbeta_k^T\bst_i$ and variance $\sigma_k^2$. Regarding the multiple regression case, the model can be formulated as follows: $$\begin{aligned} y^{(1)}_i &=& \bsbeta^{(1)T}_{z_i}\bt_i + \sigma^{(1)}_{z_i}\epsilon_i \nonumber \\ y^{(2)}_i &=& \bsbeta^{(2)T}_{z_i}\bt_i + \sigma^{(2)}_{z_i}\epsilon_i \nonumber \\ \vdots & & \vdots \nonumber \\ y^{(d)}_i &=& \bsbeta^{(d)T}_{z_i}\bt_i + \sigma^{(d)}_{z_i}\epsilon_i \label{eq: M-HMM regression model}\end{aligned}$$ where $d$ represents the dimension of the time series (sequence) and the latent process $\bz$ simultaneously governs all the univariate time series components. The model (\[eq: M-HMM regression model\]) can be rewritten in a matrix form as follows: $$\bsy_i = \bB_{z_i}^T \bt_i + \be_i \quad ; \quad\be_i \sim \N(\bO,\bsSigma_{z_i}) , \quad (i=1,\ldots,n) \label{eq: M-HMM regression model matrix form}$$ where $\bsy_i=(y^{(1)}_i,\ldots,y^{(d)}_i)^T$ is the $i$th observation of the time series in $\R^d$, $\bB_k = \left[\bsbeta^{(1)}_{k},\ldots, \bsbeta^{(d)}_{k} \right]$ is a $(p+1)\times d$ dimensional matrix of the multiple regression model parameters associated with the regime (class) $z_i=k$ and $\bsSigma_{z_i}$ its corresponding covariance matrix. The Multiple HMMR model is therefore fully parameterized by the parameter vector $\bstheta=(\bspi,\bA,\bB_1,\ldots,\bB_K,\bsSigma_1,\ldots,\bsSigma_K)$. The next sub-section gives the parameter estimation technique by maximizing the observed data likelihood through the Expectation-Maximization (EM) algorithm. The parameter vector $\bstheta$ is estimated using the well-known maximum likelihood method thanks to its very well-known attractive limiting properties of consistency, asymptotic normality and efficiency. Indeed, in our experiments, a considerable number of data points is acquired during time, which makes the sample size suitable to take advantage of the limiting properties of the maximum likelihood estimator. The log-likelihood to be maximized in this case is written as follows: [ ]{} Since this log-likelihood cannot be maximized directly, this can be performed using the EM algorithm [@dlr; @McLachlan], that is known as the Baum-Welch algorithm in the HMM context [@BaumWelch; @rabiner]. This algorithm alternates between the two following steps: #### E-step This step computes the conditional expectation of the complete-data log-likelihood given the observed data $\bY$, time $\bt$ and a current parameter estimation denoted by $\bstheta^{(q)}$: $$Q(\bstheta,\bstheta^{(q)})= \E\Big[\log p(\bY,\bz\|\bt;\bstheta)\|\bY,\bt;\bstheta^{(q)}\Big]\cdot \label{eq: Q-function for M-HMMR}$$ It can be easily shown that this step only requires the calculation of: - the posterior probability $$\tau^{(q)}_{ik} \!\!= p(z_i=k\|\bY,\bt;{\bm{\theta}^{(q)}}) \label{eq:the posterior probability}$$ $\forall$ $i=1,\ldots,n$ and $k=1,\ldots,K$ which is the posterior probability that $\bsy_i$ originates from the $k$th polynomial regression model given the whole observation sequence and the current parameter estimation $\bstheta^{(q)}$, - and the joint posterior probability of the state $k$ at time $i$ and the state $\ell$ at time $i-1$ given the whole observation sequence and the current parameter estimation $\bstheta^{(q)}$, that is $$\xi^{(q)}_{i\ell k} \!\! = \!\!p(z_{i}=k, z_{i-1}=\ell\|\bY,\bt;{\bm{\theta}^{(q)}}) \label{eq:joint posterior probability}$$ $\forall$ $i = 2,\ldots,n$ and $k,\ell=1,\ldots,K$. These posterior probabilities are computed by the forward-backward procedures in the same way as for a standard HMM [@rabiner]. More calculation details on this step can be found in [@rabiner]. #### M-step In this step, the value of the parameter $\bstheta$ is updated by computing the parameter $\bstheta^{(q+1)}$ that maximizes the conditional expectation (\[eq: Q-function for M-HMMR\]) with respect to $\bstheta$. It can be shown that this maximization leads to the following updating rules. The updates of the parameters governing the hidden Markov chain $\bz$ are the ones of a standard HMM and are given by: $$\pi^{(q+1)}_k = \tau^{(q)}_{1k} \label{eq: updating pi for the HMC}$$ $$\bA^{(q+1)}_{\ell k} = \frac{\sum_{i=2}^{n}\xi^{(q)}_{ik\ell}}{\sum_{i=2}^{n}\tau^{(q)}_{ik}} \label{eq: updating A for the HMC}$$ Updating the regression parameter consists in performing $K$ weighted multiple polynomial regressions. The regression parameter matrices updates are given by : $$\begin{aligned} {\bB}_k^{(q+1)} &=& \Big[\sum_{i=1}^{n}\tau^{(q)}_{ik} \bt_i\bt_i^T \Big]^{-1} \sum_{i=1}^{n}\tau^{(q)}_{ik} \bt_i \bsy^T_i \nonumber \\ &=& (\bX^T\bW_k^{(q)}\bX)^{-1}\bX^T\bW_k^{(q)}\bY, \label{eq: EM estimate of B_k for the MHMMR}\end{aligned}$$ where $\bW_k^{(q)}$ is a $n \times n$ diagonal matrix of weights whose diagonal elements are the posterior probabilities $(\tau_{1k}^{(q)},\ldots,\tau_{nk}^{(q)})$ and $\bX$ is the $n\times (p+1)$ regression matrix given by $(\bt_1,\ldots,\bt_n)^T$. The updating rule for the covariance matrices is written as a weighted variant of the estimation of a multivariate Gaussian density with the polynomial mean ${\bB}_k^{T(q+1)}\bt_i$ such as: [ ]{} Results and discussions {#sec: experimental study} ======================= This section presents experiments carried out to validate the two main ideas explored throughout this paper, i.e., the segmentation and the classification of the human activity from raw acceleration data using a MHMMR approach within an unsupervised learning framework[^3]. Series of experiments were conducted to evaluate the performance of the proposed approach and also to perform comparisons with well-known unsupervised and supervised classification approaches. Performance evaluation {#sec: Performance evaluation} ---------------------- Given a set of 9-dimensional acceleration data from three triaxial accelerometer modules mounted on the chest, right thigh and left ankle, the proposed approach allows both segmentation and classification of the twelve activities. Each obtained segment is indeed considered as an activity, achieving thus a classification task. We chose to take as a ground truth about the class of an activity, labeling obtained thanks to an expert. While the different subjects were performing the sequence of activities, an independent operator was asked to annotate the activities, thus providing a labeling of the dataset[^4]. The provided partition is indeed matched to the true labels (ground truth) by evaluating all the possible label switchings. The label switching leading to the minimum error rate is selected as the best class prediction. For the supervised classification approaches, data labels were used to both train and test the models. In this case, the performance was estimated through a 10-fold cross-validation procedure. Regarding the classification problem, confusion matrices between the annotated classes and the estimated classes for all the subjects in the database are computed. The criteria used to evaluate the performance of an approach are the correct classification rate and the prediction accuracy in terms of precision and recall. In the following, the results of the MHMMR approach obtained on real acceleration data of human activities are first detailed, then they are compared to those of standard unsupervised and supervised classification approaches. Classification performance of the MHMMR --------------------------------------- The following experiments were conducted to qualitatively assess the performances of the proposed approach in terms of automatic segmentation of human activity on the basis of raw acceleration signals. From the sequence of nine observed variables $\bsy_i = (y^{(1)}_i,\dots,y^{(9)}_i)$ at each time step $i$ for $i=1,\ldots,n$ corresponding to the 3-axis accelerations measured by the three sensors, the MHMMR is used to identify the latent sequence $\bz = (z_1, \dots, z_n)$ corresponding to the twelve activities. The number of classes $K$ is fixed to twelve and the order of regression $p$ is fixed empirically to three as it gives the best performance among several values of $p$. Model parameters are estimated from the data using the algorithm detailed in Section \[ssec: MHMMR\]. Figures \[fig:sequence1\] and \[fig:sequence2\] show the performance of the proposed method to segment the two following sequences: - Sequence 1: Standing - Sitting down - Sitting - From sitting to sitting on the ground - Sitting on the ground - Lying down - Lying, - Sequence 2: Standing - Walking - Climbing up stairs - Standing. These figures represent the evolution of the acceleration data and the corresponding posterior probabilities for the two different sequences. Note that the posterior probability is the probability that a sample $i$ will be generated by the regression model $k$ given the whole sequence of observations $(\bsy_1,\dots,\bsy_n)$. It can be observed that the obtained sequences are interesting and promising despite some confusion between activities such as (A${11}$, A${12}$). ![MHMRM segmentation for the sequence (Standing A$2$ - Sitting down A$3$ - Sitting A$4$ - From sitting to sitting on the ground A$5$ - Sitting on the ground A$6$ - Lying down A$7$- Lying A$8$) for the seven classes k=(1,…, 7)[]{data-label="fig:sequence1"}](sequence1.eps){width="8cm" height="6cm"} ![MHMMR segmentation for the sequence (Standing A$2$ - Walking A${11}$ - Climbing up stairs A${12}$ - Standing A$2$) for the three classes k= [1,…,3]{}[]{data-label="fig:sequence2"}](sequence2_bis.eps){width="8cm" height="6cm"} Table \[table:chest-thigh-right\] shows that the MHMMR gives $91.4\%$ as a mean correct classification rate averaged over all observations. It highlights the potential benefit of the proposed approach in terms of automatic segmentation and classification of human activity. Both the transitions and the stationary activities are well identified. Exhaustively, table \[table:Prec-recal-Act\] gives the percentage of precision and recall for each activity. Indeed, one can observe that static activities (A${2}$, A${4}$, A${6}$ and A${7}$) are easier to recognize than dynamic activities (A$1$, A${11}$, A${12}$). [\|\[1]{}[l\|]{} c\| c\| c\| c\| c\| c\| c\| c\| c\| c\| c\| c\|]{} Class & A$1$ & A$2$ & A$3$ & A$4$ & A$5$ & A$6$ & A$7$ & A$8$ & A$9$ & A${10}$ & A${11}$ & A${12}$\ Precision ($\%$)&71.9 &96.4 &78.4 &95.7 &92.3 &98.9 &97.6 &92.5 &82.6 &82.6 & 83.2 &95.6\ Recall ($\%$) &95 &87.8 &83.4 & 94 &97.3 &94.6 &95.4 &90.9 &98.5 &92.2 & 98.1 &82.3\ In order to focus on the efficiency ratio of the three sensors used for activity recognition, the MHMMR algorithm has been evaluated using data from only two sensors. The classification results, given in Table \[table:chest-thigh-right\], show as expected, that the percentage of correctly classified instances decreases with the number of data sources. The worst result is obtained when the sensor placed at the thigh is not taken into account. Sensors Percentage of correctly classified instances --------- ---------------------------------------------- $91.4\% \pm 1.65$ $83.9\% \pm 1.98$ $86.2\% \pm 2.03$ $84\% \pm 2.21$ : Effects of reducing the number of sensors when using MHMMR[]{data-label="table:chest-thigh-right"} Comparison with unsupervised and supervised classification approaches --------------------------------------------------------------------- Correct classification rates and the standard deviations obtained with standard unsupervised and supervised classification approaches as well as the MHMMR approach are given in Table \[table:class-prec-recall1\] and Table \[table:class-prec-recall2\]. [\|\[1]{}[l\|]{} c\|c\| c\|]{} & Correct Classification ($\%$)& Precision ($\%$) & Recall ($\%$)\ $k$-Means & $60.2 \pm 2.48$ & $60.4$ & $59.8$\ GMM & $72.3 \pm 2.05$ & $71.8$ & $73.5$\ HMM & $84.1 \pm 1.84$ & $83.8$ &$84$\ MHMMR & $91.4 \pm 1.65$ & $89$ &$95.6$\ Compared to standard unsupervised classifiers, the proposed MHMMR outperforms them since it provides a classification rate of 91.4 $\%$ while only 60 $\%$, $72\%$ and $84\%$ of instances are well classified with respectively the $k$-Means, the GMM and the standard HMM approaches. Notice that, the GMM and K-means approaches are not well suitable for this kind of longitudinal data. [\|\[1]{}[l\|]{} c\|c\| c\|]{} & Correct Classification & Precision & Recall\ & ($\%$) & ($\%$) & ($\%$)\ Naive Bayes & $80.6 \pm 0.91$ & $80.9$ & $80.6$\ MLP & $83.1 \pm 0.45$ & $82.8$ & $83.2$\ SVM & $88.1 \pm 1.32$ & $87.6$ & $88.3$\ $k$-NN & $95.8 \pm 0.32$ & $95.9$ & $95.9$\ Random Forest & $93.5 \pm 0.78$ & $93.5$ & $93.5$\ In Table \[table:class-prec-recall2\], it can be observed that the $k$-NN ($k=1$) gives the highest classification rates with 95.8$\%$, followed by the Random Forest with 93.5$\%$. Then, the SVM gives 88.1$\%$ and the MLP gives 83.1$\%$. However, the Naive Bayes gives the lowest classification rate with 80.6$\%$. Table \[table:class-prec-recall2\] shows also that the $k$-NN ($k=1$) has the best classification algorithm in terms of prediction accuracy since it achieves 95.9$\%$ of precision and recall. Compared to standard supervised classification techniques (using class labels), these results are very encouraging since the proposed approach performs in an unsupervised way. The main errors are due to the confusions located in transition segments. This is due to the fact that the transitions lengths are much shorter than the activities ones. Since the confusion matrix was computed using real labels supplied by a human expert, the obtained labels may not correspond perfectly to the expert labels, particularly, during transitions. Indeed, it is difficult to have the ground truth of the limit between an activity and a transition. Furthermore, the aforecited supervised classification approaches require a labelled collection of data to be trained. Besides, they do not explicit the temporal dependence in their model formulation as they assume an independent hypothesis for the data; the data are treated as several realizations in the multidimensional space ($\R^d$) without considering possible dependencies between the activities. Moreover, it can be noticed that assigning a new sample to a class using the $k$-NN approach requires the computation of as many distances as there are examples in the dataset, which may lead to a significant computation time. Using the proposed approach, classification needs the computation of the posterior probabilities, as many as there are activities. On the other hand, comparison with the unsupervised classification approaches ($k$-Means and the GMM) and the standard HMM shows that the proposed method gives relatively a high rate and better performances. Conclusion and future works =========================== In this paper, we presented a statistical approach based on hidden Markov models in a regression context for the joint segmentation of multivariate time series of human activities. It is based upon the use of raw accelerometer data acquired from body mounted inertial sensors in a health-monitoring context. The main advantage of the proposed approach comes from the fact that the statistical model explains the regime changes over time through the hidden Markov chain, each regime being interpreted as an activity (a class). Furthermore, learning with this statistical model is performed in an unsupervised way using unlabelled examples only; parameter estimates are computed by maximizing a likelihood criterion, using a dedicated EM algorithm. Considering human activity recognition within an unsupervised learning framework can be particularly interesting within an exploratory data-mining context in order to automatically cluster a large amount of unlabelled acceleration data into different groups of activity. The comparison with well-known supervised classification approaches shows that the proposed method is competitive even when performed in an unsupervised way. This work can be extended in several directions, namely integrating the model into a Bayesian context to better control the model complexity via choosing suitable prior distributions on the models parameters. Then, and perhaps more interestingly, another step to explore is to built a fully non Bayesian non-parametric model which will be useful for any kind of complex activities and in which the number of activities will not have to be fixed. In terms of application, a promising perspective in a rehabilitation context would be to use the proposed approach for recognizing in an unsupervised framework the undesirable compensatory physical behaviours observed with stroke and injury patients. [1]{} Baum, L.E. and Petrie, T. and Soules, G. and Weiss, N., “A maximization technique occurring in the statistical analysis of probabilistic functions”, [Annals of Mathematical Statistics*]{}, vol. 41, pp 164-171, 1970. Kaluza, B., Mirchevska, V., Dovgan, E., Lustrek, M., Gams M., “An Agent-based Approach to Care in Independent Living”, In Lecture Notes in Computer Science, International Joint Conference on Ambient Intelligence (AmI-2010), Malaga, Spain, 6439, 177–186, 2010. Lu, C.-H. and Fu L-C., “Robust Location-Aware Activity Recognition Using Wireless Sensor Network in an Attentive Home”, IEEE Transactions on Automation Science and Engineering, Vol.6, No.4, pp.598-609, 2009. Jinhui, Hu., and Nikolaos, V.B, “Fast human activity recognition based on structure and motion”, Pattern Recognition Letters, 32(14), 1814–1821, 2011. Brdiczka O., Langet M., Maisonnasse J., Crowley J. L., “Detecting human behavior models from multimodal observation in a smart home”, IEEE Transactions on Automation Science and Engineering, Vol. 6, No.4, pp 588-597, 2009. MacKenzie, D., “Inventing accuracy A Historical sociology of nuclear missile guidance”, MIT Press, Cambridge, MA, and London, 1991. Carminati, M., Ferrari, G., Sampietro, M., Grassetti, R., “Fault detection and isolation enhancement of an aircraft attitude and heading reference system based on MEMS inertial sensor”, Procedia Chemistry 1 (1), 509–512, 2009. Jovanov, E., Milenkovic, A., Otto, C., DeGroen P.C., “A wireless body area network of intelligent motion sensors for computer assisted physical rehabilitation”, J. Neuro Eng.Rehab. 2(6), 2005. Wu, W. H., Bui, A. A. T., Batalin, M. A., Liu, D., Kaiser, W. J., “Incremental diagnosis method for intelligent wearable sensor system”, IEEE Trans. Inf. Technol.B 11 (5), 553-562, 2007. Barshan, B., Durrant-Whyte, H.F., “Evaluation of a solid-state gyroscope for robotics applications”, IEEE Trans. Instrum. Meas. 44(1), 61-67, 1995. Tan, C.W., Park, S., "Design of accelerometer-based inertial navigation systems, IEEE Trans. Instrum. Meas. 54(6), 2520-2530, 2005. Jiayang, L., Lin, Z., Wickramasuriya, J. and Vasudevan, V., “uwave: Accelerometer-based personalized gesture recognition and its applications”, Pervasive and Mobile Computing, 657–675, 2009. Mathie, M. J., Celler, B. G., Lovell, N. H., Coster, A. C. F., “Classification of basic daily movements using a triaxal accelerometer”, Medical and Biological Engineering and Computing, 42, 679–687, 2004. Noury, N., Barralon, P., Virone, G., Boissy, P., Hamel, M., Rumeau, P., “A smart sensor based on rules and its evaluation in daily routines”, In Proceedings of the 25th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, (EMBS), Cancun, Mexico, 3286–3289, 2003. Yang, J. Y., Wang, J. S., Chen, Y. P., “Using acceleration measurements for activity recognition: An effective learning algorithm for constructing neural classifiers”, Pattern Recognition Letters, 29 (16), 2213–2220, 2008. Altun, K., Barshan, B., Tuncel, O., “Comparative study on classifying human activities with miniature inertial and magnetic sensors”, Pattern Recognition, 43 (10), 3605–3620, 2010. Khalili, A. H., and Aghajan, Ha., “Multiview activity recognition in smart homes with spatio-temporal features”, In Proceedings of the 4th ACM/IEEE ICDSC, 142–149,USA, 2010. Liu, C. L., Lee, C. H., and Lin P. M., “A fall detection system using k-nearest neighbor classifier”, Expert Systems with Applications, 37, 7174–7181, 2010. Qian, H., Mao, Y., Xiang, W., Wang, Z., “Recognition of human activities using SVM multi-class classifier”, Pattern Recognition Letters, 31 (2), 100–111, 2010. Dong, L. , Che, L., Sun, L., Wang, Y., “Effects of non-parallel combs on reliable operation conditions of capacitive inertial sensor for step and shock signals”, Sensors and Actuators A: Physical, 121 (2), 395–404, 2005. Yang, J. Y., Wang, J. S., and Chen, Y. P., “Using acceleration measurements for activity recognition: An effective learning algorithm for constructing neural classifiers”, Pattern Recognition Letters, 29, 342–350, 2008. Faber, G. S., Kingma, I., Bruijn, S. M., Van Dieen, J. H., “Optimal inertial sensor location for ambulatory measurement of trunk inclination”, Journal of Biomechanics, 42 (14), 2406–2409, 2009. Cvetkovic, B., Lustrek, M., Kaluza, B., and Gams, M., “Semi-supervised Learning for Adaptation of Human Activity Recognition Classifier to the User”, International Joint Conference on artificial Intelligence (STAMI), pp.24–29, 2011. Duda, R. O., Hart, P. E., and Stork, D. G., Pattern Classification (second ed.) A Wiley-Interscience Publication, John Wiley & Sons, 2000. Allen, F. R. , Ambikairajah, E., Lovell, N. H. , and Celler, B. G., “Classification of a known sequence of motions and postures from accelerometry data using adapted Gaussian mixture models”, Physiol. Meas., 27(10), 935–951, 2006. Lin, J. F. S., and Kulic, D., “Automatic human motion segmentation and identification using feature guided hmm for physical rehabilitation exercises”, In: Robotics for Neurology and Rehabilitation, Workshop at IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2011. Rabiner, L. R., A tutorial on hidden Markov models and selected applications in speech recognition, Proceedings of the IEEE, 77(2), 257–286, 1989. Mannini, A., and Sabatini, A., “Machine learning methods for classifying human physical activity from on-body accelerometers”, Sensors, 10, 1154–1175, 2010. Dempster, A. P., Laird, N. M. , and Rubin, D. B. , “Maximum likelihood from incomplete data via the [EM]{} algorithm”, Journal of The Royal Statistical Society, B, 39(1), pp.1-38, 1977. Viterbi, A. J., “Error bounds for convolutional codes and an asymptotically optimum decoding algorithm”, IEEE Transactions on Information Theory 13 (2), 260–269, 1967. Trabelsi D., Mohammed S., Oukhellou L., Amirat Y., “Activity Recognition Using Body Mounted Sensors: An Unsupervised Learning based Approach”, IEEE World Congress on Computational Intelligence - International Joint Conference on Neural Networks (IJCNN), June 10-15, Brisbane, Australia, 2012. http://www.xsens.com Bouten, C., Koekkoek, K., Verduin, M., Kodde, R., Janssen, J., “A triaxial accelerometer and portable data processing unit for the assessment of daily physical activity”, IEEE Transactions on Biomedical Engineering 44 (3), 136-47, 1997. McGee, V. E., Carleton, W. T., “Piecewise regression”, Journal of the American Statistical Association 65, 1109–1124, 1970. Brailovsky, V. L., Kempner, Y., “Application of piecewise regression to detecting internal structure of signal”, Pattern recognition 25 (11), 1361–1370, 1992. Picard, F., Robin, S., Lebarbier, E. , and Daudin, J. J., “A Segmentation/Clustering Model for the Analysis of Array CGH Data”, Biometrics, 63(3), 758–766, 2007. Bellman, R., “On the approximation of curves by line segments using dynamic programming”, Communications of the Association for Computing Machinery (CACM) 4 (6), 284, 1961. Stone, H., “Approximation of curves by line segments”, Mathematics of Computation 15 (73), 40–47, 1961. Fridman, M., Hidden markov model regression, Tech. rep., Institute of mathematics, University of Minnesota, 1993. McLachlan, G. J., Krishnan, T., The EM algorithm and extensions, New York: Wiley, 1997. Kuncheva, L. I., Combining Pattern Classifiers: Methods and Algorithms, Wiley-Interscience, 2004. [^1]: D. Trabelsi, S. Mohammed and Y. Amirat are with University Paris-Est Créteil (UPEC), LISSI, 122 rue Paul Armangot, 94400, Vitry-Sur-Seine, France e-mail: [email protected], [email protected], [email protected]. F. Chamroukhi is with University Sud Toulon-Var, LSIS Lab UMR CNRS 7296, email: [email protected]. L. Oukhellou is with University Paris-Est, IFSTTAR, GRETTIA, F-93166 Noisy-le-Grand, France, e-mail: [email protected] [^2]: http://ec.europa.eu/health-eu/my$\_$health/elderly/ [^3]: Note that, in this study, the raw acceleration data are directly used without any feature extraction. Indeed, in many area of application a feature extraction step is needed before running the classifier and may itself lead to an additional computational cost, which can be penalizing in real time applications. [^4]: Note that the labels were not used to train unsupervised models; they were only used afterwards for the evaluation of classification errors.	{ "pile_set_name": "ArXiv" }
--- bibliography: - 'aamnem99.bib' - 'aabib.bib' - 'proceedings.bib' --- The bulk composition of exo-planets Boris Gänsicke (University of Warwick), John Debes (STScI), Patrick Dufour (University of Montreal), Jay Farihi (UCL), Michael Jura (UCLA), Mukremin Kilic (University of Oklahoma), Carl Melis (UCSD), Dimitri Veras (University of Warwick), Siyi Xu (ESO), Ben Zuckerman (UCLA) [Motivation]{} Priorities in exo-planet research are rapidly moving from finding planets to characterizing their physical properties. Of key importance is their chemical composition, which feeds back into our understanding of planet formation. Mass and radius measurements of transiting planets yield bulk densities, from which interior structures and compositions can be deduced [@valenciaetal10-1]. However, those results are model-dependent and subject to degeneracies [@rogers+seager10-1; @dornetal15-1]. Transmission spectroscopy can provide insight into the atmospheric compositions [@singetal13-1; @demingetal13-1], though cloud decks detected in a number of super earths systematically limit the use of this method [@kreidbergetal14-1]. For the foreseeable future, far-ultraviolet spectroscopy of white dwarfs accreting planetary debris remains the only way to directly and accurately measure the bulk abundances of exo-planetary bodies. The exploitation of this method is limited by the sensitivity of HST, and significant progress will require a large-aperture space telescope with a high-throughput ultraviolet spectrograph. [Evolved planetary systems]{} Practically all known planet host stars, including the Sun, will evolve into white dwarfs, and many of their planets will survive [@verasetal13-1; @villaveretal14-1]. Observational evidence for such evolved planetary systems includes the detection of trace metals in the white dwarf photospheres [@koesteretal97-1], and infrared and optical emission from circumstellar debris disks [@zuckerman+becklin87-1; @gaensickeetal06-3; @farihietal09-1]. The generally accepted model explaining these observation is the tidal disruption of asteroids, minor planets, or planets [@jura03-1; @debesetal12-1; @verasetal14-1] perturbed onto star-crossing orbits by dynamical interactions with planets [@debesetal02-1; @frewen+hansen14-1; @veras+gaensicke15-1]. Spectroscopic surveys now unambiguously demonstrate that 25-50% of white dwarfs host evolved planetary systems [@zuckermanetal03-1; @zuckermanetal10-1; @koesteretal14-1]. [Debris-polluted white dwarfs as tracers of exo-planet bulk abundances]{} In a pioneering paper, @zuckermanetal07-1 showed that measuring the photospheric abundances of debris-polluted white dwarfs provides an unrivaled window into the bulk composition of exo-planetary material for planetary bodies with masses of $10^{20}-10^{25}$g [@girvenetal12-1], i.e. ranging from several 10km-sized asteroids to nearly the mass of Pluto. The ultraviolet wavelength range is fundamental for this work, as it contains strong transitions of the rock-forming elements (Si, Fe, Mg, O), refractory lithophiles (Ca, Al, Ti), and in particular of volatile elements (C, N, P, S) that trace the formation region of the planetary material relative to the snow line. We have led ten HST/COS programs that demonstrated the diagnostic potential of extra-solar cosmochemistry using white dwarfs, corroborating the rocky, volatile-depleted nature of the planetesimals [@juraetal12-1; @xuetal13-1], and detecting a variety in bulk compositions similar to, if not exceeding, that seen among solar-system bodies [@gaensickeetal12-1 see Fig.1]. Noticeably, we have discovered water-rich planetesimals [@farihietal13-2], which provide the potential for delivering water to planets in the habitable zone. The measured planetary debris abundances provide important input into our understanding of planet formation. Of particular importance for the properties of planetary systems are the C/O and Mg/Si ratios. C/O ratios $>0.8$ would result in a radically different setup from the solar system, with O-chemistry replaced by C-chemistry, which is discussed abundantly in the literature [e.g. @moriartyetal14-1]. The Mg/Si ratio determines the exact composition of silicates, which in turn has implications for planetary processes such as plate tectonics. Furthermore, the relative abundances of Fe and siderophiles (Cr, Mn, S, Ni), and of refractory lithophiles (Al, Ca, Ti) provides insight into the core and crust formation, respectively [@jura+young14-1; @melisetal11-1; @dufouretal12-1]. In contrast to indirect measurements, such as abundance studies of planet host stars [e.g. @delgado-menaetal10-1], far-ultraviolet spectroscopy of debris-polluted white dwarfs provides a direct measure of those ratios. The results from our published studies of have already informed recent models of planet formation [@carter-bondetal12-1]. However, global insight into the chemistry of planetary systems will only be possible from the detailed photospheric abundance studies of a substantial number of white dwarfs. The current roster of planetary debris abundance studies with at least five detected elements stands at $\simeq15$ [@jura+young14-1]. With HST, we may double, maybe triple this number over the next couple of years, but beyond that, the aperture of Hubble is too small to make significant progress. ![image](wd0843_1250.ps){width="8cm"} ![image](mm_mm.ps){width="8cm"} [Road-map for the next two decades, and the need for a large UV mission]{} Over the next four years, Gaia will identify $\simeq200\,000$ white dwarfs brighter than 20th magnitude. This sample will be volume-limited out to 300pc for white dwarfs with cooling ages of up to 600Myr, i.e. sufficiently hot to have significant ultraviolet flux. Ground-based spectroscopic follow-up of this sample (DESI/WEAVE/4MOST) will identify 1000s of strongly polluted white dwarfs, but typically only provide abundances for Ca and/or Mg. In addition, cross-correlation of the Gaia white dwarfs with the EUCLID and WFIRST surveys will result in the detection of 100s, possibly 1000s of debris discs. The detection of circumstellar debris is a direct proxy for metal-pollution at ultraviolet wavelengths. A subset of these debris discs will be bright enough to be followed-up with JWST/MIRI, providing detailed mineralogy. Those systems are particularly valuable, as the dust mineralogy can be directly compared to atomic abundances obtained from far-ultraviolet spectroscopy of the debris-polluted white dwarfs. In other words, the known sample of evolved planetary systems will explode, similar to the dramatic increase in the number of planets around main-sequence stars we witness thanks to missions like Kepler, TESS, and PLATO. To fully exploit the potential of evolved planetary systems for a detailed, large-scale statistical study of the bulk abundances of exo-planetary systems requires a large-aperture ultraviolet mission that can provide follow-up spectroscopy for several hundred debris-polluted white dwarfs. [Instrumental requirements]{} Assuming a factor 30 increase in sensitivity compared to COS ($\times$15 for a 10m aperture, and $\times2$ from improved optics, and improved orbital visibility) will increase the available volume for detailed abundance studies by a factor $\simeq150$ compared to what can be reached with HST, sufficient to include $>1000$ potential targets for high-quality ultraviolet spectroscopic follow-up. - Spectral resolution. A resolution of at least 20000, better 50000 is necessary to resolve photospheric and interstellar features, and to avoid blending of lines. - Spectral range. The “traditional” far-ultraviolet range 1100 to 1800Å contains most of the relevant atomic transitions. Extending coverage to 950Å, i.e. including Ly$\beta$ and Ly$\gamma$, would greatly improve the atmospheric parameters, $T_\mathrm{eff}$ and $\log g$, which, in turn, result in more accurate diffusion velocities, and finally abundances. The shorter-wavelength range also contains a number of higher-ionization lines detected in hotter, younger white dwarfs. - Signal-to-noise ratio. The abundance measurements require a minimum S/N of 40 to model the strongest absorption lines. Detection of trace species not observed so far (e.g. N, rare earth elements) will need higher S/N, of up to 100, which is currently difficult to obtain with COS because of the limited telescope aperture, and fixed-noise patterns in the detectors.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The nonthermal particle content of galaxy clusters should in part have a cosmological component generated during the early starburst phase of the member galaxies. This is reviewed in the framework of a simple cluster formation model suggested previously. It implies a nonthermal energy fraction of about 10 percent for the Intracluster gas. We also propose a mechanism for the early generation of Intracluster magnetic fields in terms of Galactic Winds. It results in typical field strengths of $10^{-7}$ Gauss. Such comparatively weak fields are consistent with an inverse Compton origin of the excess EUV and hard X-ray emission of the Coma cluster, given the radio synchrotron emission. The required relativistic electrons must have been accelerated rather recently, less than a few billion years ago, presumably in cluster accretion shocks. This is in contrast to the hadronic nonthermal component which accumulates on cosmological time scales, and whose $\pi^0$-decay TeV $\gamma$-ray emission is expected to be larger, or of the same order as the inverse Compton TeV emission. This $\gamma$-radiation characterizes the energetic history of cluster formation and should be observable with future arrays of imaging atmospheric Cherenkov telescopes.' address: - 'Max-Planck-Institut für Kernphysik, D-69029 Heidelberg, Germany' - 'Yerevan Physics Institute, 375036 Yerevan, Armenia' author: - 'H. J. Völk$^{\rm a}$ and A. M. Atoyan$^{\rm a, b}$' title: 'Clusters of Galaxies: magnetic fields and nonthermal emission' --- Introduction ============ Clusters of galaxies are the largest gravitationally bound structures in the Universe and may confine a representative fraction of its mass. Therefore the study of their dynamical properties and radiation content should allow, amongst other things, conclusions on the relative amounts of baryonic and nonbaryonic matter in cosmology (e.g. White and Fabian, 1995, and references therein). Another basic characteristic, also due to confinement, is the ratio of thermal to nonthermal energy in these objects. To a significant extent that ratio is established during the epoch of galaxy formation and thus preserves the energetic history of cluster formation. We shall review this topic here. The confinement of nonthermal particle components is intimately related to the existence of strong and chaotic magnetic fields in the intracluster medium (ICM), and we shall propose a mechanism for their early generation. This is followed by a discussion of the present-day nonthermal radiation from clusters in various wavelength ranges, in particular at very high $\gamma$-ray energies. [Rich Clusters]{} Rich clusters, i.e. conglomerates with typically more than 100 member galaxies, have typical radii $R_{\rm cl} \sim$ few Mpc and baryonic masses $M_{\rm cl} \sim 10^{14} \,{\rm to}\, 10^{15}\,{\rm M}_{\odot}$. The closest large cluster is the Virgo cluster at a distance of $d \sim 20 \, {\rm Mpc}$; all the others are at a distance $d$ of hundred Mpc and beyond. Examples for bright and relatively nearby clusters are the Perseus and the Coma clusters with $d \simeq 100\,\rm Mpc$. The Perseus cluster is the brightest cluster in soft X-rays (see Fig. 1). The large X-ray luminosity is due to the very hot ($T \sim 10^7 \,\rm to \, 10^8\,$K), massive ($M_{\rm gas} \sim {\rm few} \times \sum M_{\rm gal}$), and metal-rich ($[{\rm Fe}]_{\rm cl} \simeq 0.35 [{\rm Fe}]_{\odot}$) ICM gas (e.g. Böhringer, 1996). Apart from their primary cosmological interest, galaxy clusters also serve as extragalactic distance poles (Sunyaev-Zeldovich effect) and as gravitational telescopes for still more distant objects. This makes them important “instruments” for observational astronomy. [Cluster Formation]{} Most rich clusters - apart from objects still in an early formation phase, like the Virgo cluster - are characterized by a predominance of early type galaxies, i.e. elliptical (E) and lenticular (S0) galaxies which have little interstellar gas. Gas-rich spiral and irregular galaxies represent a minority. This situation is opposite to the one in the field, the space outside clusters, where the fraction of spiral galaxies is at least 75 percent. For rich clusters the metallicity of the ICM gas, for instance in the form of the fractional ICM iron mass, is also correlated with the optical luminosity in E and S0 galaxies (Arnaud et al., 1992). The correlation supports the qualitative view that early starbursts due to galaxy-galaxy interactions of protospirals have produced a large number of Supernovae (SNe) that heated the originally present interstellar gas and generated violent Galactic Winds, leaving gas-poor E and S0 galaxies behind. This mass loss should have led ultimately to the observed strong chemical enrichment of the ICM gas. We also conjecture that the radiation and the winds from these early galaxy mergers strongly heated the remaining primordial ICM gas and thus prevented further galaxy formation. This is perhaps the physical explanation for the observed inefficiency of galaxy formation which manifests itself in the remarkable preponderance of diffuse ICM gas mass over stellar mass in clusters of galaxies. A quantitative discussion of the dynamical prerequisites for Galactic Winds and the total number of SNe in clusters is contained in the paper by Völk et al. (1996, hereafter referred to as Paper I) which we shall summarize below. The total number of SNe since galaxy formation in the cluster, roughly a Hubble time $T_{\rm H} \simeq 1.5 \times 10^{10} {\rm yr}$ ago, is given by $$N_{\rm SN}=\int_{-T_{\rm H}}^{0}dt \times \nu_{\rm SN}(t)=\frac{0.35 \,{[Fe]}_{\odot} \times M_{\rm cl}}{\delta M_{Fe}}\;,$$ where $\delta M_{Fe}$ is the amount of iron produced per event. In such starbursts we dominantly expect core collapse SNII from massive progenitor stars to occur, with $\delta M_{Fe}\simeq 0.1 M_{\odot}$ on average. For the Perseus cluster this implies $N_{\rm SN}^{\rm Perseus} \sim 3 \times 10^{12}$. The corresponding total energy input into the interstellar medium is $N_{\rm SN}E_{\rm SN} \sim 3 \times 10^{63} E_{51}\,\rm erg$, where $E_{51}=10^{51} \,\rm erg$ is the average hydrodynamic energy release per SN. Assuming the early starbursts to occur at a typical redshift of $z\sim 2$ due to the merging of protospirals in the overdense protocluster environment (Steinmetz, 1993), with a duration of $T_{\rm SB} \leq 10^9 \,\rm yr$, we obtain $$\frac{(N_{\rm SN}^{\rm Perseus}/N_{\rm gal}^{\rm Perseus})}{T_{\rm SB}} \geq 100 \times \nu_{\rm SN}^{\rm Milky\, Way} .$$ Here $N_{\rm gal}^{\rm Perseus} \!\simeq \! 500$ denotes the number of galaxies in the cluster. As an example we can compare to the archetypical contemporary starburst galaxy $M82$. It has a current SN rate $\nu_{\rm SN}^{M82} \sim 10 \times \nu_{\rm SN}^{\rm Milky\, Way}$, a wind velocity $v_{\rm wind} \sim 2300\, {\rm km/sec}$, and a mass-loss rate of $\dot{M} \sim 0.8 M_{\odot}/{\rm yr}$ (Breitschwerdt, 1994). The starburst nucleus of $M82$ is characterized by the following values for the interstellar gas temperature $T$, gas density $n$, and thermal gas pressure $p$ : $T_{\rm base} \sim 10^8 \,{\rm K}$, $n_{\rm base} \sim 0.3\, {\rm cm}^{-3}$, and $p_{\rm gas}/k_{B} \sim 10^7\, {\rm K \,cm^{-3}}$ (Schaaf et al. 1989). Since the thermal ICM gas pressure in the Perseus cluster is $p_{\rm cl}^{\rm Perseus}/k_{B} \sim 10^4 \,{\rm K}\, {\rm cm}^{-3}$, it is clear that an object like M82 could readily drive a wind even against the [present-day]{} ICM pressure. At the galaxy formation epoch the ICM pressure was certainly much smaller than this value. In the expanding wind flow the SN-heated gas cools adiabatically to quite small temperatures. However it is reheated in the termination shock, where the ram pressure of the wind adjusts to the ICM pressure. Beyond this point the ejected galactic gas is mixed with the unprocessed ICM gas. [Particle Acceleration]{} Cluster formation also implies the production of a strong nonthermal component of relativistic particles. They will be accelerated during the early phase - and possibly also in later events - and confined in the turbulent ICM medium. The confinement time generally exceeds the cluster lifetime. Thus the energy spectrum of the energetic particles is the same as that generated by their sources. In Cosmic Ray parlance these are [cosmological]{} Cosmic Rays (CRs). During the early starburst particle acceleration will occur initially at the outer shocks of the Supernova Remnants (SNRs). However, like the thermal gas, this first generation of nonthermal particles will loose its energy almost completely by adiabatic cooling in the ensuing Galactic Wind. Fresh particle acceleration will occur with high efficiency at the strong wind termination shock, at distances $\sim \! 100 \,\rm kpc$. The wind magnetic field will still intersect the shock at an angle $\sim\! 10^\circ$ (e.g. Zirakashvili et al. 1996), so that the standard process of diffusive shock acceleration approximately still works there (Fig. 2a). Assuming an overall acceleration efficiency of 10 to 30 percent, one gets a total gas internal energy $E_{\rm gas}^{\rm GW} \sim {\rm few}\, 10^{62}\,$erg and a nonthermal energy $E_{\rm CR}^{\rm GW} \sim 10^{62}$ erg for a system like the Perseus cluster, ultimately from star formation and subsequent SN explosions. Since the galaxies are distributed across the cluster quasi-uniformly, this will also be true for the [nonthermal particle population and the radiation]{} they emit. The continuing gravitational contraction/accretion of the cluster will subsequently energize CRs and thermal gas at least adiabatically, or shock accelerate/heat both components, so that finally the total energy $E_{CR}$ of energetic particles reaches $10^{63}\, {\rm erg}\! \sim 0.1\, E_{\rm gas}$ in the cluster; $E_{\rm gas}$ now denotes the total internal energy of the ICM gas (Paper I). The resulting nonthermal energy density of some tenths of $\rm eV/cm^{3}$ happens to be roughly equal to that in the interstellar medium of our Galaxy. It is instructive to compare the expected nonthermal energy with the thermal energy content of the cluster galaxies. Assuming the stars internally to be in virial equilibrium and, for purposes of estimate, all of them to have a solar mass and radius, then $E_{\rm th}^{\rm star}\! \sim (3/10) GM_{\odot}^{2}/R_{\odot} \simeq 10^{48}\,\rm erg$. For a total mass of about $10^{14} M_{\odot}$ contained in the [galaxies]{} of the Perseus cluster this gives a total thermal energy in stars $ \sim\! 10^{62}\,\rm erg$, and thus $E_{\rm CR} > \sum_{\rm gal}\sum_{\rm stars} E_{\rm th}^{\rm star}$. This means that the nonthermal ICM energy is larger than the total thermal energy of all the stars in all the galaxies contained in the cluster! It has been argued more recently that, apart from star formation and overall gravitational contraction, also individual giant radio galaxies should have injected large and in fact comparable amounts of nonthermal particles during the life time of a cluster (En[ß]{}lin et al. 1997; Berezinsky et al. 1997). This is no doubt an important additional possibility. A weakness of this argument consists in the fact that per se it is predicated on statistical knowledge about the luminosity function for active galaxies in clusters in general, and not on direct observations of the individual cluster to which it is applied. [Particle Confinement]{} The large-scale magnetic field in the ICM gas may be quite chaotic and not well connected over distances exceeding typical intergalactic distances (see section 2). Thus energetic particles may not readily escape from the cluster due to such topological characteristics. However, already pure pitch angle diffusion along magnetic field lines with superposed turbulent fluctuations gives important insights into the confinement properties of galaxy clusters. Standard quasilinear theory yields a spatial diffusion coefficient $\kappa_{\parallel}$ along the large scale field $\bf B$ due to a power spectrum $P(k)$ of magnetic field fluctuations with wavelength $\lambda=2 \pi /k$ as $$\kappa_{\parallel}= (1/3) c r_{\rm g}(p) \frac{{\bf B}^2}{\int_k^{\infty} {\rm d}k' P(k')},$$ where $p$ denotes particle momentum, $kr_{\rm g}(p)\simeq 1$, and $k$ denotes wavenumber of the field fluctuations. Let us assume a relative fluctuation field strength of order unity at the inter-galaxy distance $1/k_{0}$, i.e. a totally turbulent field $P(k_0) \times k_0 \sim {\bf B}^2$ on this scale, and a power law form of $P(k)=P(k_0) (k/k_0)^{-n}$. Then the diffusion time across the cluster $T_{\rm esc} \sim R_{\rm cl}^2 / \kappa_{\parallel} > T_{H}$ for $(cp)_{\rm protons} \leq 10^{17}\, {\rm eV \; and} \leq10^{15}\,{\rm eV}$, for $n=3/2$ and $n=5/3$, respectively (Paper I). Also $t_{\rm loss}^{\rm protons}\gg T_{H}$ for nuclear collisions in the ICM gas. Therefore (except at subrelativistic energies with their prevailing Coulomb losses), up to these energies CR hadrons accumulate in the cluster since the galaxy formation epoch, and that is what we called [cosmological]{} CRs before. The situation is different for relativistic electrons, which suffer radiative losses: energetic electrons observed now must either be secondaries or be rather recently accelerated. Intracluster Magnetic Fields ============================= The large magnetic field strengths of $B \sim 1 \,\mu \rm G$ in the IC medium of rich clusters, in particular as observed by Faraday rotation measurements (e.g. Kronberg 1994), are not easily explained by a contemporary mechanism because present day turbulent dynamo effects in such a large-scale system should be extremely slow. Therefore we suggest here a field configuration that is due to the early formation history of galaxy clusters and that should be essentially preserved to this day. It should even be still in a state of development at the present epoch. The argument derives from the violent early Galactic Winds which accompany the starbursts responsible for the predominance of the early type galaxies in rich clusters. We assume first of all that the protospirals, whose mergers constitute the building blocks for the E and S0 galaxies, had already generated galactic magnetic fields of $\mu$G strength. This should indeed be possible within about $10^8$ yr, i.e. of the order of a rotation period of our Galaxy, from turbulent dynamo action that invokes boyancy effects from CRs and magnetic reconnection on spatial scales of $O(100 {\rm pc})$ (Parker, 1992); such a time scale and the resulting field strengths correspond to generally accepted numbers. In the second stage the ensuing Galactic Winds extend these fields from the interacting galaxies to almost intergalactic distances. In the final and by far longest stage, that lasts until now, the fields are recompressed by the contraction of the cluster to its present size. The ICM fields do not reconnect on the intergalactic scale in a Hubble time. Consequently there is no need for a continuous regeneration of these fields since their formation. However, this also implies that a topologically connected overall ICM field will on average not be formed either, and that the ICM field is chaotic on a scale smaller or equal to the present intergalactic distance. In detail we draw on arguments we have in the past used for the field configuration in a Galactic Wind from our own Galaxy (Zirakashvili et al. 1996; Ptuskin et al. 1997). They are based on estimates of the relative amount of field line reconnection vs. the extension of galactic field lines by a wind to “infinity” (Breitschwerdt et al. 1993). The basic result was that the rates of reconnection - and thus of the formation of “Parker bubbles” leaving the galaxy by their boyancy and allowing the generation of the disk magnetic field - and of extension of this field into the galactic Halo by the pressure forces of the wind are roughly equal. Thus both effects occur, and in the cluster galaxies, magnetic energy can be generated on the large scale of the wind at the expense of the thermal and nonthermal enthalpies produced in the starburst. The geometry of the field should roughly correspond to straight field lines out to radial distances $r$ of the order of the starburst (SB) radius, $r_{\rm gal}^{\rm SB} \sim 1$ kpc, and spherically diverging field lines beyond that. The slow rotation of the system should then lead to an azimuthal field component $\propto 1/r$ which dominates at large distances over any radial component. However, in contrast to the familiar situation in the Solar Wind equatorial plane, the axis of rotation is rather parallel than perpendicular to the flow at the base of the wind, and thus the dominance of the azimuthal field component is by no means as drastic as in the case of a stellar wind (Fig. 3). Assuming the ICM pressure $p_{\rm cl}(z=2)$ at the formation stage of the early type galaxies to be roughly a factor of $10^{-2}$ smaller than it is at present (after gravitational compression of the ICM gas), the termination shock distance $r_{\rm sh}$ is found from $\rho(r_{\rm sh})\, u^2 /k_{B} \sim 10^{-2} p_{\rm cl}(z=0)/k_{B} \sim 10^2 \,\rm K/cm^{-3}$. For M82 analogs, but with $r_{\rm gal}^{\rm SB} \sim 1$ kpc, we have $p_{\rm gas}/k_{B} \sim 10^7 \,{\rm K}{\rm cm}^{-3}$. Thus $r_{\rm sh}/r_{\rm gal}^{\rm SB} \sim p_{\rm gas}/ [10^{-2}p_{\rm cl}(0)] \simeq 300$, and therefore $$\frac{r_{\rm sh}}{d_{\rm gal}^{\rm field}(0)/(1+z)} \simeq \frac{300 \,{\rm kpc}}{(2\,{\rm Mpc}/3)} \simeq 0.5$$ The Wind Bubble containing the shock-heated wind gas will have a radius still exceeding $r_{\rm sh}$. Thus, even though the volume of hydrodynamically unaffected ICM gas may be large enough so that the ICM gas mass exceeds the mass associated with galaxies by a factor of a few - as observed - there may be that some Wind Bubbles touch. However, at the scale of $r_{\rm sh}$, reconnection with a speed between 1 and 10 percent of the Alfv$\acute{e}$n velocity is too slow to occur over a Hubble time, even in a present-day ICM magnetic field as high as $10^{-6}$ G. Therefore, on average, the field structure should remain topologically disconnected until today. The field strength $B_{\rm cl}(z=2)$ in the Wind Bubbles should be of the order of $$B_{\rm cl}(z=2) \simeq 4 B_{\rm gal} r_{\rm gal}^{\rm SB}/r_{\rm sh} \sim 10^{-2}B_{\rm gal} \sim 10^{-8} {\rm G}$$ or somewhat larger, if the field in the bubble increases in the decelerating postshock flow. The ongoing cluster contraction/accretion compresses the field to lowest order isotropically $\propto l^2$, with the scale factor $$l \simeq [n_{\rm cl}(0)/n_{\rm bar}(0)]^{1/3}/(1+z),$$ where $n_{\rm cl}(0)\sim (10^{-3}\,{\rm to}\, 10^{-4})\,{\rm cm}^{-3}$ and $z=2$. Choosing for the present mean baryon number density the value $n_{\rm bar} \sim 3 \times 10^{-7}\,{\rm cm}^{-3}$, we obtain $l \sim (2.3\,\rm to\, 5)$. Finally then, we obtain for the present-day ICM field: $B_{\rm cl}(z=0)/B_{\rm cl}(z=2) \sim (5 \,\rm to \,25)$. Therefore the present-day ICM magnetic field should have a mean strength of the order of $10^{-7}$G, from “primordial” seed fields, randomly directed on an intergalactic scale. Although smaller by about one order of magnitude than estimated from Farady rotation measurements, such fields need not necessarily be unrealistic, considering that observations might emphasize regions of high magnetic fields. In addition the increase of the field in the Galactic Wind bubbles beyond their postshock value might be more than a factor of unity as assumed above. Thus we cannot exclude $\mu$G fields although they certainly are at the upper limit our estimate permits. The interpretation of recent UV data also points to small field values (see section 3). In conclusion, there is hardly any need for a contemporary “turbulent IC dynamo”. Nevertheless, the estimated CR enthalpy of $\sim 0.5 \,{\rm erg/cm^3}$ is essentially a free energy reservoir for a future increase in $B$ towards a mean strength of a few $\mu$G. In the next section we consider the nonthermal radiation from clusters. In particular we shall discuss the implications of the broad-band observations for the nonthermal energy content and ICM magnetic fields. Nonthermal emission =================== From the Coma cluster radio fluxes are measured from 10.3MHz to 2.7GHz (Bridle and Purton, 1968; Henning 1989; Kim et al. 1990; Giovannini et al. 1993), and at 4.85GHz an upper flux limit was reported by Kim et al. (1990). The energy fluxes $J_{\nu} \propto \nu^{-\alpha_{\rm r}}$ between 30.9 MHz and 1.4 GHz are well fitted with a power-law index $\alpha_{\rm r} =1.16$ (Bowyer & Berghöfer 1998). The data at 2.7 and 4.85 GHz fall below this extrapolation, which may indicate a steepening in the electron spectrum (Schlickeiser et al. 1987) for Lorentz-factors $\gamma\geq 5\times 10^4$, but can be explained also as an instrumental effect (Deiss 1997). Below we use an index $\alpha_{\rm r} =1.16$. In the extreme ultraviolet (EUV) region, diffuse radiation between 65 and 245 eV is observed which appears in excess of the thermal fluxes of the X-ray emitting gas with $T\sim 2\times 10^6\,\rm K$, and this radiation was initially interpreted in terms of thermal emission from a gas of lower temperature, $T\sim 8\times 10^5\,\rm K$ (Lieu et al. 1996). Subsequently, the excess EUV emission, observed in a number of clusters (e.g. Mittaz et al. 1998), was suggested to represent inverse Compton (IC) radiation (Hwang 1997; Sarazin and Lieu 1998) of low energy electrons, $\gamma\sim 300$, on the 2.7 K microwave background radiation. Also En[ß]{}lin and Biermann (1998) used this possibility to estimate lower limits to the magnetic field strength in the Coma cluster (see also Lieu et al. 1998). Bowyer and Berghöfer (1998) have shown that the size of diffuse emission of the Coma cluster is significantly larger in radio than in EUV light. In addition, the spectral index of the energy flux $J(E)$ of the EUV emission is $\alpha_{\rm uv}\approx 0.75$ which is significantly different from that in the radio domain. Based on these arguments, Bowyer and Berghöfer (1998) suggested that two different populations of relativistic electrons should be responsible for the radio synchrotron and the EUV IC fluxes. We shall show below that the differences in these spectral indices are quite naturally explained in terms of a single population of electrons producing both the EUV excess IC and the radio synchrotron emission which avoids the need for a 2-component model. The characteristic synchrotron frequency produced by electrons with Lorentz factor $\gamma$ in the magnetic field $B$ is $\nu \simeq (B/1\,\mu {\rm G}) \gamma^2\,\rm Hz$ (e.g. Ginzburg 1979). Thus in a magnetic field of $B\sim 10^{-7}\,\rm G$ the production of synchrotron radiation with $\alpha_{\rm r} =1.16$ in the region from 30 MHz to 1.4 GHz requires the energy distribution of the electrons $N(\gamma)\propto \gamma^{-\alpha_{\rm e}}$ to extend with an index $\alpha_{\rm e}=1+2\alpha_{\rm r}\approx 3.3$ from $\gamma_1 \simeq 1.7\times 10^4$ to $\gamma_2 \simeq 1.4\times 10^5$. These Lorentz factors are smaller by a factor of 3 if $B\sim 10^{-6}\,\rm G$. For the IC process the mean energy of photons is $\epsilon \simeq (4/3) \epsilon_0 \,\gamma^2$, where $\epsilon_0 $ is the energy of target photons. For the 2.7K background radiation $\epsilon_0 = 6.5\times 10^{-4}\,\rm eV$. For the production of Coma’s EUV spectrum from 65 to 250 eV one therefore needs a spectral index of electrons $\alpha_{\rm e}\simeq 2.5$ only in a rather narrow energy region around $\gamma_{\rm ic} \simeq 300$. This energy is much smaller than the energy $\gamma_1$ necessary for the radio emitting electrons. Thus a scenario in which radiative losses induce a break in $\alpha_{\rm e}$ for a single population of electrons somewhere in between $\gamma_{\rm ic}$ and $\gamma_1$ can readily account for the observed indices both in the radio and the EUV. Indeed, assuming that the spectral index of the source function of electrons in the ICM is $\alpha_{\rm inj}=2.3$, the IC emission of those electrons at low energies will give $\alpha_{\rm uv}=0.65$, which is rather close to the value 0.75, in a rather narrow EUV band. After the break, the electron distribution steepens to $\alpha_{\rm e}=1+\alpha_{\rm inj}=3.3$, required for the radio emission of Coma. Such a break in the energy spectrum is necessarily produced by radiative losses of electrons on time scales $\geq 10^9\,\rm yr$. Indeed, the characteristic energy loss time in the 2.7K background radiation field as well as due to synchrotron emission can be written as $$t_{\rm rad}=2.4\times 10^{12} [C(\gamma)+0.1 B_{\mu \rm G}^{2}]^{-1} \,\gamma^{-1} \; \rm yr \; ,$$ where $B_{\mu \rm G}=B/1\,\mu \rm G$. The coefficient $C(\gamma)$ takes into account corrections for the Klein-Nishina effect in the IC energy losses of ultrarelativistic electrons, which is important for $B\leq 3\,\rm \mu G$. $C(\gamma)=1$ for $\gamma\ll 10^8$ (IC losses in the Thompson limit), but $C(10^8)\simeq 1/2$, $C(2\times 10^8)\simeq 1/3$, and $C(5\times 10^8)\approx 1/6$. For a magnetic field $\leq 3 \,\mu \rm G$ and timescales $\simeq (1-3)\times 10^{9} \,\rm yr$, equation (1) predicts the radiative cutoff energy $\gamma_{\rm br} \sim 10^3 $. It is important to note that relativistic electrons need to be produced (accelerated) in the ICM [continuously]{} during all these last years, because in the case of an ‘impulsive’ injection of the electrons the radiative losses would remove all particles with energies above $\gamma_{\rm br}$, and then the radio spectra cannot be explained. This is seen in Fig.3 which shows the results of calculations that assume both continuous and impulsive injection of electrons with a spectrum $$Q(\gamma)\propto \gamma^{-\alpha_{\rm inj}} \exp (-\gamma/\gamma_{\rm c}),$$ where $\gamma_{\rm c}$ defines the assumed characteristic maximum energies of accelerated particles. The fluxes for continuous injection are normalized to the radio flux $2\,\rm Jy$ observed at $400\,\rm MHz$. The IC radiation for continuous injection during the last $t=10^{9}\,\rm yrs$ (heavy solid line) has the spectral shape of the EUV flux, and for a slightly smaller magnetic field (cf. Fig.4), or an injection time larger by a factor of 2, it can also explain the absolute EUV flux. However in the case of relativistic electron production on much larger time scales, $\sim 10^{10}\,\rm yr$, the explanation of the shape of the EUV radiation becomes problematic even for continuos injection. Moreover, the agreement with observed spectra becomes worse than shown in Fig.3 if we take into account that for cosmological timescales the energy density of the microwave background increases with redshift as $w_{\rm mbr}\propto T^{4}(z)\propto (1+z)^4$. Thus, the electrons responsible for the excess EUV radiation, if indeed it has an IC origin, must have been produced continuously during the recent (1-3)Gyrs. This would be the characteristic age of the accretion shocks in the cluster which seem to be the most probable accelerators for the radio emitting electrons. An IC origin of the excess EUV flux imposes a strong constraint on the magnetic field in Coma. The variation of the IC flux for different magnetic fields are shown in Fig.4 which demonstrates that the lowest consistent magnetic field strength in Coma is equal to $B=7.5\times 10^{-8}\,\rm G$. This field could explain also the excess flux observed by [Beppo-SAX]{} in hard X-rays beyond 25 keV (Fusco-Femiano et al. 1998), that might be due to IC radiation of GeV electrons. The dot-dashed curve in Fig. 4 falls below both the X-ray and the EUV data. However, the assumption of injection for about 5 Gyr shifts the position of the radiative cutoff energy by factor of 5 and increases by the same factor the number of low energy electrons. The number of electrons with $\gamma\sim 300$ is also increased for a steeper injection spectrum, $\alpha_{\rm inj} = 2.5$. Both these options could marginally explain the radio synchrotron and EUV IC fluxes. Thus an IC origin of the excess EUV radiation restricts the cluster magnetic field to a narrow range around $0.1\,\rm \mu G$. For magnetic fields $B\simeq 10^{-7}\,\rm G$ the observed radio fluxes require injection of accelerated particles with a luminosity $L_{\rm inj}\simeq 3\times 10^{45}\,\rm erg/s$. In a thermal gas with density $n\sim 10^{-3}\,\rm cm^{-3}$, a [secondary]{} origin of the relativistic electrons, i.e. their production in $pp$ interactions, would require an enormously large total energy in relativistic protons. Indeed, the luminosity in the $\pi^{\pm}-\mu^{\pm}$-decay electrons can be estimated as $$L_{\pm}\simeq 7.7\times 10^{40}\,(n_{\rm p}/10^{-3}\,{\rm cm^{-3}})^{-3} (E_{\rm CR}/10^{60}\,\rm erg)\; erg/s.$$ The fluxes of IC gamma-rays to be expected at TeV energies are shown in Fig.5. for the case of $B=10^{-7}\,\rm G$ and acceleration of electrons up to an exponential cutoff energy $E_{\rm c}=100\,\rm TeV$. For magnetic fields $\sim 0.1\,\rm \mu G$ the diffusive shock acceleration mechanism in the Bohm limit still allows such high values for accelerated electrons. In this limit the acceleration time, defined by the rate ${\rm d}\gamma/{\rm d}t \simeq \gamma u^2/D(\gamma)$ where $u$ is the shock speed and $D(\gamma)$ is the diffusion coefficient, is estimated as $$t_{\rm acc}\simeq 5.4\times 10^{-5} \,u_{3}^{-2} B_{\rm \mu G}^{-1} \,\gamma \;\rm yr\;,$$ where $u_3 = u/10^3\,\rm km/s$. Equating $t_{\rm acc}$ with $t_{\rm rad}$ shows that values $\gamma\geq 10^8$ are possible. The solid line in Fig.5 corresponds to the solid curve in Fig.3. The TeV fluxes of IC $\gamma$-rays in that case correspond to several per cent of the TeV emission from the Crab Nebula. In the case of single power law injection with $\alpha_{\rm inj} =2.3$ a further increase of the IC fluxes at TeV energies are impossible because of the EGRET upper limit above 100 MeV (Sreekumar et al 1996). Note, however, that for a more realistic modelling of Coma as a spatially nonuniform source, the power law index $\alpha_{\rm inj}$ could be somewhat smaller than 2.3. We do not consider here this possibility, which should address also the question of different angular sizes of Coma in EUV and radio, as well as the flattening of the radio spectra observed towards the core of the cluster (Giovannini et al. 1993; Deiss et al. 1997). In Fig.5 the dashed line shows the maximum of the IC radiation fluxes to be expected in the case of $\alpha_{\rm inj} =2.1$, which is then at the 0.1 Crab level. The dots correspond to the $\pi^0$-decay fluxes produced by hadronic CRs with $\alpha_{\rm cr}=2.1$, assuming $E_{\rm CR}=3\times 10^{62}\,\rm erg$. The size of the radio emission produced by high-energy electrons in Coma is about half a degree. Therefore we can expect in that case a similarly large size for the TeV emission. Detection of such extended and weak ($\leq\,$0.1 Crab) TeV fluxes by current instruments is problematic, but they are quite accessible for the future HESS and VERITAS arrays. These results for the IC fluxes relate to the case of ICM magnetic fields $B\simeq 10^{-7}\,\rm G$. However, the estimates for the magnetic fields in Coma range from 0.1 to several $\mu \rm G$. In particular, Kim et al. (1990) and Feretti et al. (1995) deduced magnetic fields $B\simeq 1.7\,\rm \mu G$ and $B\simeq 6.0\,\rm \mu G$, respectively, from Faraday rotation measurements on background radio sources. If this is so, then an IC origin of the excess EUV fluxes is absolutely excluded. The only other possibility left for an explanation of this radiation in terms of nonthermal radiation is synchrotron production by very high energy electrons. The interpretation of the steep radio fluxes then requires a second component of nonthermal electrons. For a production of synchrotron photons with energies 200eV ($\nu \sim 5\times 10^{16}\,\rm Hz$) in magnetic fields $B\sim 1\,\rm \mu G$ one needs electrons with $\gamma \geq 2\times 10^8$. From Eq.(1), the radiative loss times of these electrons is very short, $t_{\rm rad}\leq 10^4 \,\rm yr$. Even assuming a rectilinear propagation of such electrons, the maximum possible distance from the acceleration sites would be less than 3kpc. Thus, the acceleration sites of the electrons should be rather smoothly distributed in the ICM in order to result in a smooth distribution of the synchrotron EUV radiation. This seems quite reasonable for accretion shocks, but not for isolated point sources like active galaxies. Another requirement is that the spectrum of accelerated particles should be very hard, with $\alpha_{\rm inj}\simeq 1.5$. Then the spectrum of electrons steepens to $\alpha_{\rm e} \simeq 2.5$ which is required for the production of synchrotron radiation with $\alpha_{\rm uv} \simeq 0.75$. In Fig.6 we show the spectra of the synchrotron and IC radiation generated in a magnetic field $B=2\,\rm \mu G$ by two different populations of relativistic electrons. The dashed curves are produced, as in previous figures, by electrons continuously injected into the ICM during recent epochs up to $t\geq 10^9\,\rm yr$ with $\alpha_{\rm inj}=2.3$ and luminosity $L_{\rm inj}=3.5\times 10^{42}\,\rm erg/s$. For this luminosity, these electrons can be well explained as [secondaries]{} produced in the interactions of hadronic CRs, of total energy $E_{\rm CR}=4.5\times 10^{61}\,\rm erg/s$, with an ICM gas of density $n_{\rm p}=10^{-3}\,\rm cm^{-3}$. They can generate the observed radio emission. On the other hand, the EUV fluxes in Figs. 3 and 4 are produced by the second component with $\gamma_{\rm c}=6\times 10^8$, and a hard injection spectrum with $\alpha_{\rm inj}=1$ (solid and dot-dashed curves) and $\alpha_{\rm inj}=1.5$ (three-dot–dashed). The curves show that continuous injection of this second component over times exceeding $3\times 10^7\,\rm yr$ from now, without reacceleration, would lead to an accumulation of electrons at energies smaller than $\gamma\sim10^5$. This would be quite sufficient to produce a radio synchrotron flux above the observed level. Moreover, in the case of $\alpha_{\rm inj}=1.5$ one has to assume a cutoff in the injection spectrum below $\gamma_{\rm low}\simeq 10^5$. Otherwise the number of low energy electrons in the injection spectrum would be unacceptably high. Thus, the scenario with high magnetic fields and nonthermal origin of the excess EUV radiation requires that the second electron component enters a reacceleration cycle at least once in $3\times 10^7$ year. In addition, this component should not contain electrons with low energies. Therefore it cannot be produced by reacceleration of secondary electrons. In a speculative vain this component could be due to run-away pulsars born with kick velocities $\sim 1000\,\rm km/s$. Another possibility for the pulsars to appear in the ICM would be that they could be dragged there in the process of galaxy-galaxy collisions. Pulsars produce, at the pulsar wind termination shocks, an electron distribution with a strong deficit of low-energy particles (e.g. Arons 1996). These electrons could be kept at those high energies by entering into frequent reacceleration processes on the accretion shocks in the ICM, and could be thus responsible for the second component. Reacceleration would not affect the radio electrons if it happened in the central region (see Bowyer and Berghöfer 1998). The signature of the second component would be a rising spectrum of IC $\gamma$-rays. Fields below $1\,\rm \mu G$ would even imply uncomfortably large TeV fluxes from the center of the Coma cluster. Thus the alternative between low and high magnetic fields in the Coma cluster would be given by the shape of the TeV $\gamma$-ray spectrum. [Acknowledgements]{} The authors thank F.A.Aharonian and E.N.Parker for illuminating discussions. They also thank H.Böhringer for the permission to use Fig.1 here. [999]{} M. Arnaud, R. Rothenflug, O. Boulade, L. Vigroux, E. Vangioni-Flam, [Astron. Astrophys.]{} [254]{} (1992) 49. J. Arons, [Space Sci. Reviews]{} [75]{} (1996) 235. H. Böhringer, in: R. Ekers et al., eds. [Extragalactic Radio Sources]{} 1996 IAU, 357. V.S. Berezinsky, P. Blasi, V.S. Ptuskin, [Astrophys. J.]{} [487]{} (1997) 529. S. Bowyer, T.W. Berghöfer, [Astrophys. J.]{} [505]{} (1998) 502. D. Breitschwerdt, J.F. McKenzie, H.J. Völk, [Astron. Astrophys.]{} [269 ]{} (1993) 54. D. Breitschwerdt, Habilitationschrift, Univ. Heidelberg (1994) 158. A.H. Bridle, C.R. Purton, [Astron. J.]{} [73]{} (1968) 8. B.M. Deiss, [Astron. Astrophys.]{} [325]{} (1997) 74. T.A. En[ß]{}lin, P.L. Biermann, [Astron. Astrophys.]{} [330]{} (1998) 90. T.A. En[ß]{}lin, P.L. Biermann, P.P. Kronberg, X.-P. Wu, [Astrophys. J.]{} [477]{} (1997) 560. L. Feretti, D. Dallacasa, G. Giovannini, A. Tagliani, [Astron. Astrophys.]{} [302]{} (1995) 680. R. Fusco-Femiano, D. Dal Fiume, L. Feretti, G. Giovannini, G. Matt, S. Molendi, [astro-ph/9808012]{} (1998). V.L. Ginzburg, [Theoretical Physics and Astrophysics]{}. Pergamon Press, Oxford, 1979. G. Giovannini, et al., [Astrophys. J.]{} [406]{} (1993) 399. P.A. Henning, [Astron. J.]{} [97]{} (1989) 1561. C.-Y. Hwang, [Science]{} [278]{} (1997) 1917. K.-T. Kim, P.P. Kronberg, P.E. Dewdney, T.L. Landecker, [ApJ]{} [355]{} (1990) 29. P.P. Kronberg, [Rep. Prog. Phys.]{} [57]{} (1994) 325. R. Lieu, W.-H. Ip, W.I. Axford, M. Bonamente, [astro-ph/9809175]{} (1998). J.P.D. Mittaz, R. Lieu, F.J. Lockman, [Astrophys. J.]{} [498]{} (1997) L17. E.N. Parker, [Astrophys. J.]{} [401 ]{} (1992) 137. V.S. Ptuskin, H.J. Völk, V.N. Zirakashvili, D. Breitschwerdt, [Astron. Astrophys.]{} [321]{} (1997) 434. R. Schaaf, W. Pietsch, P.L. Biermann, P.P. Kronberg, T. Schmutzler, [Astrophys. J.]{} [336]{} (1989) 722. R. Schlickeiser, A. Sievers, H. Thiemann, [Astron. Astrophys.]{} [182]{} (1987) 21. P. Sreekumar, et al., [Astrophys. J.]{} [464]{} (1996) 628. M. Steinmetz, PhD thesis, Technische Universität München (1993). H.J. Völk, F.A. Aharonian, D. Breitschwerdt, [Space Sci. Rev.]{} [75]{} (1996) 279. (Paper I) D.A. White, A.C. Fabian, [Mon. Not. R. Astron. Soc.]{} [273]{} (1995) 72. V.N. Zirakashvili, D. Breitschwerdt, V.S. Ptuskin, H.J. Völk, [Astron. Astrophys.]{} [311]{} (1996) 113.	{ "pile_set_name": "ArXiv" }
--- author: - Mikołaj Morzy - Tomasz Kajdanowicz - Przemysław Kazienko - Grzegorz Miebs - Arkadiusz Rusin bibliography: - 'priorityrank.bib' title: 'Priority Attachment: a Comprehensive Mechanism for Generating Networks' --- Introduction {#sec:introduction .unnumbered} ============ Motivation {#subsec:motivation .unnumbered} ---------- Many generative models of network formation have been proposed in the scientific literature [@chakrabarti2006graph], and some of them have gained significant notoriety, for instance the random network model of Erdös and Rényi [@erd6s1960evolution], the small world model of Watts and Strogatz [@watts1998collective], the cumulative advantage model of de Solla Price [@price1976general], the scale-free model of Albert and Barabási [@barabasi1999emergence], stochastic block model [@BlockModel_First_Steps], [@wang1987stochastic] or the forest fire model of Leskovec [@leskovec2005graphs]. Each of these generative network models is based on some phenomenon which (as is often claimed) explains the underlying process of network formation. For instance, in the case of the small-world model, the alleged phenomenon is the tendency of many systems to form tightly connected groups (small worlds) with incidental connections between groups serving as long distance bridges. In the case of the preferential attachment model the phenomenon which purportedly fuels the network formation process is the strong preference of vertices to connect to already well-connected vertices. In the stochastic block model, which is is a probabilistic model of the interactions between pairs of vertices, each vertex belongs to one of the groups or ’blocks’, and each edge exists with a certain probability depending only on the vertices’ membership to the particular group. Fluctuations in the same block are stochastically equivalent, indicating their equivalent role in generating of the network structure. This model is popular because it can generate a wide range of large-scale networking patterns and can learn social structure in unweighted and weighted networks [@BlockModel_weighted], [@BlockModel_structure_learning],[@guimera2009stochastic]. Some of the network generators do not attempt to model real-world processes directly, but they rely on some mathematical formalism, like the sequence of Kronecker products applied to a small seed set of networks [@leskovec2010kronecker]. Recently, an interesting proposal has been formulated to model complex networks using a stochastic sequence of predefined base actions [@arora2017action] and turning the re-creation of an empirical network into an optimization problem. For a long time, we have suspected that these individual phenomena are specialized instances of a more general mechanism of network creation. The main reason for this forefeeling was the fact that the generative network models seemed to be narrowly defined and each of them covered only a specific class of possible network topologies. A closer inspection of the generative network models further revealed that all of them were using, sometimes inadvertently, some type of prioritization while choosing target vertices during edge formation. The mechanism of prioritization using priority queues has a very long presence in almost all disciplines of science. In psychology priority queues are believed to be the driving mechanisms of human attention [@yantis1990mechanisms]. In systems science priority queues play an important role in controlling and scheduling [@cobham1954priority; @moon2000scalable] or in recommender systems for evolving networks[@liao2017ranking]. In sociology there are strong clues that the stochastic queuing theory can provide valuable insights into human dynamics [@walraevens2012stochastic; @jo2012time; @alexander2012priority]. In ecology prioritization influences the inter-species interactions [@louette2007predation; @almany2003priority; @blaustein1996priority]. In the domain of complex networks, the idea of network growth based on rankings was introduced by Fortunato, Flammini, and Menczer[@fortunato2006scale]. They have shown that substituting vertex feature distributions with global rankings[^1] of these features (both topological and non-topological) as the basis for preferential attachment, one always obtains a scale-free degree distribution in the resulting network. In addition, they have proven that it is sufficient to use ranking samples in the network growth process, as long as these samples maintain the original partial order of vertices. However, the approach presented in this paper differs from [@fortunato2006scale] in that we consider not only global rankings of vertices, but we allow local rankings as well. This is a fundamental difference since our model is not limited to scale-free networks, but can produce arbitrary network topologies. In addition, our model provides means for computational re-creation of empirical networks resulting in many instances of networks with nearly identical network profiles. It is a unique and promising property in the field of complex networks. In order to verify whether the priority attachment can explain processes of network formation, we have developed Priority Rank, a simple generative network model. This model uses the priority attachment phenomenon to drive the network formation. Our experiments show that Priority Rank is capable of generating a very wide spectrum of networks. We were able to successfully re-create network topologies generated by the most popular generative network models, which suggests that the priority attachment mechanism is indeed the common denominator for these models. More importantly, examining the prioritization scheme, which resulted in the best re-generation of the original network, we could provide viable interpretations of the latent network generative process. This unique feature of the Priority Rank model is best pronounced in empirical networks. Problems addressed {#subsec:problems.addressed.in.the.paper .unnumbered} ------------------ Research on complex networks is constantly struggling with a serious methodological obstacle. Despite the vast availability of networks that can be analyzed, each individual network has its distinctive characteristics and presents to a researcher a single data point. Any attempt to generalize the results obtained on one network to other networks strives with inherent vulnerabilities. The choice of target networks is arbitrary and the transfer of discovered patterns has to be selective. Oftentimes, unspoken assumptions come into play, for instance, a phenomenon discovered in a few networks of a particular class is presumed to be applicable to all networks of this class. These assumptions are much more questionable than usually assumed and they significantly weaken the scientific appeal of a discovery. Let us consider the phenomenon of the shrinking diameter of a network [@leskovec2005graphs]. Is it fair to conclude that this phenomenon should apply to all complex networks? Or maybe it is valid only for scale-free networks characterized by the preferential attachment mechanism of edge formation? Unfortunately, it is impossible to conduct a proper statistical inference since only one realization of each network is available. Furthermore, this problem cannot be addressed by sampling of networks, because sampling distorts the topology of networks to the point where sub-networks of a network loose their identifying properties [@Stumpf22032005]. Another problem addressed in this paper is the inherent incompatibility of networks. Trying to analyze different networks and to make quantitative comparisons between them is very difficult, because in the realm of complex networks even minuscule perturbations of local topology can lead to significant differences in global network properties. This is true even for synthetic networks which have been generated from the same generative model and with slightly different parameters. Distributions of centrality measures, such as degree, betweenness, closeness or clustering coefficient, vary dramatically in response to minor changes of initial configurations [@e18090320]. As the result, statistical comparison of seemingly similar networks suggests a much greater variability and difference between networks, which makes any machine learning extremely difficult in the domain of complex networks. Without the access to diversified training, testing, and validation sets of networks, all patterns discovered in networks are bound to overfit and not generalize well. Also, the relationships between network generation processes proposed in the literature are very unclear. Is the preferential attachment an alternative to the small world phenomenon, or is it merely a supplementation? We note that there is no common language at the level of network generating processes, making the comparisons between different networks even more challenging. Main findings {#subsec:main.results .unnumbered} ------------- The main contribution presented in this paper is the priority attachment mechanism of network formation. We show how priority queues built on local rankings can lead to very diversified network topologies. The function, which generates these local rankings, yields itself to interpretation. As it turns out, popular generative network models can be easily mimicked by the priority attachment, and some of these models (preferential attachment, cumulative advantage) are a special case of the priority attachment. The phenomenon of the priority attachment can be readily incorporated into a generative Priority Rank network model, which is a simple procedure of network creation. The Priority Rank model can produce networks with very different topologies using the same principle of priority attachment. Instead of using dedicated network models one can successfully force Priority Rank to produce random networks, preferential attachment networks, small world networks, and many more. The second main feature of Priority Rank is its applicability to empirical networks. Due to the priority attachment mechanism, which forms the basis of the model, it is possible to adjust Priority Rank to deliver networks with topology very similar to a given empirical network. The Priority Rank model has only to learn the distance function used in the priority attachment. We demonstrate Priority Rank’s ability to re-create empirical networks while preserving the distributions of centrality measures and other network characteristics. This ability allows for generating whole families of similar networks, thus, for conducting statistical inference on families of networks, etc. There are four main reasons, which make the Priority Rank model valuable: - Priority Rank is a comprehensive generative network model which can produce a very wide spectrum of network topologies, including the most popular network models. - Priority Rank offers insights into the generative processes behind modeled networks. Classical machine learning algorithms can be used to find the most fitting distance function for the priority attachment mechanism. In most cases such a distance function is easily interpretable and provides explanations for the latent network formation process. - Priority Rank allows us to generate multiple instances of networks with the same characteristics and distributions, because the model discovers the main generative process of network formation. It should be noted that this is incomparable with network sampling which oftentimes distorts the profile of sampled networks. Instead, the Priority Rank model allows to multiply networks for the purpose of A/B testing, statistical inference, simulations, etc. - Priority Rank does not require any hyper-parameters to be set a priori such as edge creation probability in the random network model or edge rewiring probability in the small world network. This feature of the Priority Rank model is very important because, contrary to popular belief, generative network models are very sensitive to the initialization of these parameters. Methods {#sec:methods .unnumbered} ======= Priority attachment {#subsec:priority.attachment .unnumbered} ------------------- The idea behind the priority attachment mechanism is fairly simple. Consider a new vertex which joins a network. The primary issue is the selection of target vertices to which the new vertex creates edges. Previously, several mechanisms have been proposed to model this selection process. For instance, in the Erdös-Rényi model the vertex selects target vertices randomly using the uniform distribution. In the Albert-Barabási model the vertex selects target vertices with the probability proportional to the current degree of each vertex. According to the priority attachment mechanism each vertex has a local ranking which arranges all possible target vertices by their “importance” from the point of view of the new vertex. The new vertex selects target vertices with the probability proportional to their position in the local ranking. One should regard this local ranking as the priority queue which orders all of the vertices of the network from the point of view of a single vertex. This means that each vertex uses its own local ranking while creating edges. In other words, the main mechanism of network formation is the attachment of vertices driven by their individual perception of priority of other vertices, hence the name “priority attachment”. The power of the priority attachment mechanism stems from the fact that local rankings can be computed by arbitrarily complex distance functions which can either model real-world phenomena, or model adjacency matrices of empirical networks. [0.45]{} ![Different network topologies and their degree distributions generated by the priority attachment mechanism[]{data-label="fig:network.topologies"}](random "fig:"){width="\textwidth"} [0.45]{} ![Different network topologies and their degree distributions generated by the priority attachment mechanism[]{data-label="fig:network.topologies"}](small-world "fig:"){width="\textwidth"} [0.45]{} ![Different network topologies and their degree distributions generated by the priority attachment mechanism[]{data-label="fig:network.topologies"}](preferential-attachment "fig:"){width="\textwidth"} [0.45]{} ![Different network topologies and their degree distributions generated by the priority attachment mechanism[]{data-label="fig:network.topologies"}](cosine "fig:"){width="\textwidth"} [0.45]{} ![Different network topologies and their degree distributions generated by the priority attachment mechanism[]{data-label="fig:network.topologies"}](disassortative "fig:"){width="\textwidth"} [0.45]{} ![Different network topologies and their degree distributions generated by the priority attachment mechanism[]{data-label="fig:network.topologies"}](hierarchical "fig:"){width="\textwidth"} Interestingly, the topology of the generated network depends almost exclusively on the properties of the function $D$ which is used to generate local rankings. Figure \[fig:network.topologies\] presents six different networks, each consisting of $n=50$ vertices, generated by the priority attachment mechanism. Random network is generated for $D(v_i,v_j) \sim N(\mu, \sigma)$, i.e., when local rankings are random permutations of the set of vertices. Small-world network is generated for $D(v_i,v_j) = \left\| a(v_i) - a(v_j) \right\|$ for an attribute which value is randomly chosen from a uniform distribution, $a(v) \sim U(0,1)$. In other words, local rankings arrange vertices by the distance defined by the vertex’s attribute $a(v)$, resulting in a strong preference for vertices in the local neighborhood. Preferential attachment network is generated for $D(v_i,v_j) = \frac{1}{\mathit{deg}(v_j) + \epsilon}$, i.e., when local rankings simply represent the global ranking of vertex degrees[@fortunato2006scale]. Cosine similarity network is generated if $D(v_i,v_j) = 1 - \frac{v_i \circ v_j}{\\|v_i\\|\\|v_j\\|}$, where $v_i=\left[a^1_i, a^2_i,\ldots,a^m_i\right]$ is a vector of numeric values, effectively linking vertices which are close in the vertex vector space. But priority attachment mechanism can be used to produce networks with more complex topologies. For instance, the function $D(v_i,v_j) = \mathit{deg}^* - \frac{\mathit{deg}(v_i)}{\mathit{deg}(v_j)}$, where $\mathit{deg}^$ is the maximum degree in a given network, produces a regular disassortative network with the degree assortativity coefficient $r = -0.28$. Another example is the function $D(v_i,v_j) = \alpha D_E(v_i,v_j) + (1-\alpha) D_H(v_i, v_j)$, where $D_E(v_i, v_j)$ is the Euclidean distance between vertices $v_i$, $v_j$ and $D_H(v_i, v_j)$ is the hierarchical distance between these two vertices (we assume that vertices belong to disjoint classes and that there is a total ordering between classes). When applied to the Priority Rank model, such distance function produces complex hierarchical structures. As can be seen in Figure \[fig:network.topologies\], networks created upon various functions $D$ have different topologies and different degree distributions. The main advantage of these simple distance function definitions is the fact that they are easily interpretable. For instance, when modeling a network of disease spreading it is likely that disease vectors would infect vertices in their close physical proximity. Thus, a distance function based on the physical distance between vertices (which generates small-world structures) would be a sound choice for a simple model of disease spreading. In addition, if one would like to differentiate the probability of edge formation based on additional factors (e.g., the transfer of a sexually transmitted disease is more likely between vectors of similar age), the incorporation of this factor into the distance function would be trivial. Similarly, when trying to model semantic relationships between words embedded in multidimensional space (a standard tool in the contemporary NLP), it is reasonable to assume that the proximity of word embeddings is an indication of some semantic relatedness of the words. A simple way to model these relationships would be to use cosine distance to define the distance function $D$. Let us now formally define the priority attachment mechanism and present the Priority Rank model. Further, we will use the following notation. - $G = \left< V, E \right>$ is a network with the set of $n$ vertices $V = \{ v_1, \ldots, v_n \}$ and the set of edges $E = \{ (v_i, v_j) : v_i, v_j \in V ; i \neq j\}$ - $\forall v_i \in V : v_i = \left[ a^1_i, a^2_i, \ldots, a^m_i \right]$, vertices are vectors of attributes, - $D(v_i, v_j): V \times V \rightarrow \mathbb{R}$ is the generic distance function which computes the distance between vertices $v_i$ and $v_j$, such that $D(v_i,v_j) > 0 \iff i \neq j$ and $D(v_i,v_i)=0$. Distance function $D$ does not have to be symmetrical. - $R_i = \left< v_i^1, v_i^2, \ldots, v_i^{n-1} \right>$ is a permutation of $V \setminus \{v_i\}$ representing the local ranking of vertices for the vertex $v_i$ According to the priority attachment mechanism, the probability of selecting a vertex $v_j$ as the target vertex for an edge originating from the vertex $v_i$ is inversely proportional to the position of the vertex $v_j$ in the ranking of vertices for $v_i$. The probability mass function of selecting the ith element of the ranking is given by $$P(i) = \frac{1}{\sum\limits_{k=1}^{n-1} \frac{1}{k}} \frac{1}{i} = \frac{1}{H_{n-1} i} \label{eq:pmf}$$ where $H_n$ is the nth harmonic number, serving as the normalizing constant so that Equation \[eq:pmf\] presents a proper probability mass function, i.e., $\sum_{i=1}^n P(i) = 1$. We will use Euler’s formula to approximate the nth harmonic number as $H_n \approx ln(n) + \frac{1}{2n} + \gamma$, where $\gamma$ is the Euler-Mascheroni constant, $\gamma = 0.57722$. Algorithm \[alg:network.generation\] presents the pseudo-code for generating networks using the Priority Rank model. Procedure $\mathit{sample}(\left[1,\ldots,m\right],P)$ samples an integer from the range $\langle 1,\ldots,m \rangle$ without replacement using the probability mass function from Equation \[eq:pmf\]. We assume a constant out-degree $k$ of vertices, but the number of edges each vertex creates can vary from vertex to vertex. In particular, when using Priority Rank model to re-create empirical networks, one can sample the out-degree distribution of the source network to obtain a particular value of $k$ for a given vertex. $D : V \times V \rightarrow \mathbb{R}$, distance function $V={v_1,\ldots,v_n}$, a set of vertices $k$, the number of edges each vertex creates compute the ranking $R_i$ using the distance function $D$ /\ priority evaluation \/ $t \leftarrow \mathit{sample}(\langle 1,\ldots,n \rangle, P)$ /\ random selection without replacement of the position in the ranking \/ $v_t \leftarrow R_i[t]$ /\ priority attachment \/ add edge $(v_i,v_t)$ In order to better illustrate the idea of priority attachment, let us consider an example of a simple network formation. Suppose that there are five people $v_1,..,v_5$ described by name, age, and sex. Let us also assume that the distance function $D$ is defined as follows: name age sex ------- ------- ----- -------- $v_1$ Alice 30 female $v_2$ Bob 40 male $v_3$ Cecil 25 male $v_4$ Diana 20 female $v_5$ Eve 35 female $$D(v_i,v_j) = \begin{cases} \|v_i[\mathit{age}]-v_j[\mathit{age}]\| ,& \text{if } v_i[\mathit{sex}] = v_j[\mathit{sex}] \\ \|v_i[\mathit{age}]-v_j[\mathit{age}]\| + 10,& \text{otherwise} \end{cases}$$ In other words, the social distance is defined in terms of the absolute difference of age, and the fact that two people share the same sex compensates for 10 years of age difference. In this example, people tend to form relationships with other people of similar age, and given two people of the same age, there is a preference to form relationships with people of the same sex. This model could be applicable, for instance, to the process of self-selecting students to form pairs in a large study group where the participants have no prior acquaintances. One possible instance of the network formation process driven by the priority attachment is presented in Figure \[fig:priority.attachment\], assuming that each vertex always creates $k=2$ outgoing edges. Individual priority rankings for each vertex, along with the value of the distance function and the probability of creating an edge to a vertex (computed using Eq. \[eq:pmf\]), are presented in Table \[tab:example.priority.attachment.rankings\]. ![Priority attachment process. Edges created at a given step, i.e., by a given person, are marked with solid lines.[]{data-label="fig:priority.attachment"}](example){width="\textwidth"} The process starts with Alice calculating her distance to all other vertices. The most similar vertex to Alice is Eve and she occupies the first position in the local ranking for Alice. Analogously, the most dissimilar vertex to Alice is Bob and he is placed at the end of the ranking. Inserting ranking positions into Equation \[eq:pmf\] yields the final probabilities of selecting vertices as target vertices for newly created edges. For the sake of simplicity we have assumed that for each vertex the first two most probable targets have been randomly chosen while forming the network. If two or more vertices are equidistant from the given vertex, they receive the same position in the local ranking, which may contain gaps. As the result, probabilities of selecting vertices at certain positions of the local ranking may change (compare local rankings of Alice, Bob, and Cecil). [clcc]{}\ rank & name & dist & prob\ 1 & Eve & 5 & 48%\ 2 & Diane & 10 & 24%\ 3 & Cecil & 15 & 16%\ 4 & Bob & 20 & 12% [clcc]{}\ rank & name & dist & prob\ 1 & Cecil & 15 & 39%\ 1 & Eve & 15 & 39%\ 3 & Alice & 20 & 13%\ 4 & Diane & 30 & 9% [clcc]{}\ rank & name & dist & prob\ 1 & Alice & 15 & 31%\ 1 & Bob & 15 & 31%\ 1 & Diane & 15 & 31%\ 4 & Eve & 20 & 7% \ [clcc]{}\ rank & name & dist & prob\ 1 & Alice & 10 & 45%\ 2 & Cecil & 15 & 22%\ 2 & Eve & 15 & 22%\ 4 & Bob & 30 & 11% [clcc]{}\ rank & name & dist & prob\ 1 & Alice & 5 & 45%\ 2 & Bob & 15 & 22%\ 2 & Diane & 15 & 22%\ 4 & Cecil & 20 & 11% A visualization of the Priority Rank model mechanics is available online at <https://priorityattachment.ml>. Re-creating empirical networks {#subsec:recreating.networks .unnumbered} ------------------------------ Given an empirical network $G$, we are interested in finding the distance function $D$ such that this distance function, when used inside the Priority Rank model, generates a network which is “similar” to the empirical network $G$. The problem of defining a robust and flexible measure of network similarity has been studied for many years and several network similarity measures have been proposed in the literature [@aliakbary2015distance; @schieber2017quantification]. However, these measures tend to be computationally exhaustive and difficult to apply to really large networks. For this reason, we have decided to use a simple and well-understood network similarity measure. In order to measure the degree of network similarity we compare the distributions of centrality measures using the Kolmogorov-Smirnov (K-S) non-parametric two-sample test of the equality of continuous one-dimensional distributions. The K-S test computes the maximal distance between cumulative distribution functions and provides rejection thresholds for the null hypothesis that the compared samples are drawn from the same distribution. The question remains, how to find the distance function $D$ which produces networks that minimize the K-S statistic for centrality measure distributions. Let $\delta(v_i,v_j)$ be the set of shortest paths between vertices $v_i$ and $v_j$ in the network $G$, and let $\delta_k(v_i,v_j)$ be the set of shortest paths between vertices $v_i$ and $v_j$ which pass through vertex $v_k$. Finally, let $\Delta(v_i,v_j)$ denote the length of the shortest path between vertices $v_i$ and $v_j$. A centrality measure* is a function $\mathcal{C}:V\rightarrow \mathbb{R}$ which assigns to each vertex a value representing the “importance” of the vertex in the network $G$. Four the most popular centrality measures include[@freeman1978centrality]: - degree centrality $C_D(v_i)=d(v_i)$ simply measures the number of vertices adjacent to the vertex $v_i$. The assumption here is that a vertex is important if it is directly connected to many vertices in the network. - betweenness centrality $C_B(v_i)=\sum_{j,k \neq i} \left\| \delta_i(v_j,v_k) \right\|$ measures the number of shortest paths between any pair of vertices which pass through vertex $v_i$. This interpretation of importance highlights the influence of a vertex on communication pathways through the network. - closeness centrality $C_C(v_i)= \frac{1}{\left\| V \right\|} \sum_j \Delta(v_i,v_j)$ measures the average distance from the vertex $v_i$ to all other vertices in the network. According to this definition, a vertex is important if it can quickly communicate with all remaining vertices in the network. - page rank centrality $C_P(v_i)= \alpha \sum_{j:(v_j,v_i) \in E} \frac{C_P(v_j)}{d(v_j)} + \frac{1-\alpha}{n}$ measures the importance of a vertex as a recursive sum of importances of vertices adjacent to $v_i$. According to this definition, a vertex is important if it connects to other important vertices in the network. Sometimes it is possible to “guess” the distance function $D$ given the description of the empirical network. Table \[tab:distance.functions\] presents a list of simple distance functions that can be used when re-creating empirical networks. Some of these distance functions are self-explanatory (like degree, betweenness, closeness, and page rank distances which are simply the expressions of preferential attachment to vertices with high values of these centrality measures). If the empirical network consists of vertices with attributes, euclidean distance can be used to generate local rankings of vertex priority. For the sake of simplicity we include only one- and two-dimensional euclidean distance. If vertices are described by numerical vectors, cosine distance can be utilized. The aggregate distance computes the distance on each pair of attribute values of compared vertices $v_i, v_j$, and then applies weights $w_k$ to distances computed on each attribute. name formulation --------------------------------- --------------------------------------------------------------------------------------------- -- random distance $D(v_i,v_j) \sim N(\mu,\sigma)$ degree distance $D(v_i,v_j) = \frac{1}{C_D(v_j) + \epsilon}$ betweenness distance $D(v_i,v_j) = \frac{1}{C_B(v_j) + \epsilon}$ closeness distance $D(v_i,v_j) = \frac{1}{C_C(v_j) + \epsilon}$ page rank distance $D(v_i,v_j) = \frac{1}{C_P(v_j) + \epsilon}$ euclidean 1-D distance $D(v_i,v_j) = \left\| a^1_i - a^1_j \right\| $ euclidean 2-D distance $D(v_i,v_j) = \sqrt{(a^1_i - a^1_j)^2+ (a^2_i - a^2_j)^2} $ cosine distance $D(v_i,v_j) = 1 - \frac{v_i \circ v_j}{\\|v_i\\| \cdot \\| v_j \\|}$ aggregate distance $D(v_i,v_j) = \sum\limits_{k=1}^m w_k D(a^k_i,a^k_j)$ linear regression distance $D(v_i,v_j) = W_{ij}\beta$, $W_{ij}=(a^1_i, \ldots, a^p_i, a^1_j, \ldots, a^p_j, \epsilon)$ naive bayes classifier distance $D(v_i,v_j) = \frac{P(C=1\|W_{ij})}{P(C=0\|W_{ij}) + \epsilon}$ : Distance functions[]{data-label="tab:distance.functions"} However, in most cases it is impossible to approximate the generative process of a network using a single, simple distance function. Given an empirical network, it is often precisely the aim of a researcher to deduce the guiding generative principle of a network. The main advantage of the Priority Rank model is its ability to derive the proper distance function $D$ from the existing network, which, if applied to the model, would generate the network most similar to the original one. The learning task can be defined as follows. Consider a network $G=\left< V,E \right>$, and in particular, consider an edge $(v_i,v_j) \in E$. For the Priority Rank model to re-generate this edge, it is expected the distance $D(v_i,v_j)$ to be minimized, and at the same time, the distance $D(v_i,v_k)$ should be maximized $\forall v_k : (v_i,v_k) \notin E$. The network $G$ provides the training data for a machine learning algorithm under the form of the adjacency matrix, which can be interpreted as a function: $$A: V \times V \rightarrow \{0,1\} : \quad A(v_i,v_j) = \begin{cases} 1, & \text{if } (v_i, v_j) \in E \\ 0, & \text{otherwise} \end{cases}$$ The positive cases in the training set consist of tuples $(v^1_i, \ldots, v^m_i, v^1_j, \ldots, v^m_j)$ for all pairs of vertices $(v_i, v_j)$, which are adjacent in $G$, and the negative cases consist of tuples $(v^1_i, \ldots, v^m_i, v^1_k, \ldots, v^m_k)$ for all pairs of vertices $(v_i,v_k)$ which are not adjacent in $G$. The training set can be fed into a classification algorithm, such as logistic regression, naive Bayes classifier, or SVM, to find patterns in vertex attribute co-occurrences which influence the probability of edge’s presence or absence. The model resulting from a classification algorithm can be interpreted as a condensed representation of the underlying principle of network formation. The last two distance functions presented in Table \[tab:distance.functions\] represent the learning procedure. The linear regression distance uses the least-squares method to fit the linear regression equation to the training set, and the naive Bayes classifier distance uses the well-known naive Bayes classifier to predict the probability of existence of an edge between two vertices. Results {#sec:results .unnumbered} ======= The main result reported in this paper is the development of the attachment priority mechanism. In this section, we present the outcomes of conducted experiments. Their aim was to verify whether the priority attachment mechanism is capable of explaining the underlying network formation process. We have used four popular generative network models to produce synthetic networks, and we have collected 18 empirical networks from various domains to test the ability of the priority attachment to re-create these networks. The experimental protocol was as follows. Since we cannot guess which distance function will be able to best reproduce a given $G$, we have applied all distance functions presented in Table \[tab:distance.functions\]. Then, for the best three distance functions we have run the generation process 20 times, aggregating the results afterwards. Of course, most of the distance functions require values of attributes describing vertices. These values were not always available, for instance, in the case of synthetic networks generated from theoretical network models. In such cases, we have created synthetic attributes for each vertex, generating four attributes: one ordinal, one categorical, two continuous. These attributes were generated from four different distributions: normal, uniform, log-normal, and the exponential distribution. For empirical networks, we have generated synthetic attributes only when no vertex attributes were present in the data, otherwise we have used only the real features of vertices. To compare synthetic and empirical networks with networks generated by the Priority Rank model, we have tested the conformity of centrality measure distributions using the Kolmogorov-Smirnov two sample test. Recall that the null hypothesis of two sample K-S tests states that the compared samples are drawn from the same distribution. We reject the null hypothesis for p-values less than the significance level $\alpha=0.05$. In Tables \[tab:random.network.results\] and \[tab:empirical.network.results\] we mark the results which pass the K-S test in boldface, i.e., instances where the null hypothesis holds. Synthetic networks {#subsec:artificial.networks .unnumbered} ------------------ We have tested the ability of the Priority Rank model to re-create networks using both synthetic networks obtained from the theoretical network models, and empirical networks representing various domains. Generative network models used in our experiments include the following: - Erdös-Rényi random model [@erd6s1960evolution]: an empty network with $n=50$ vertices is created, and then, for each pair of vertices an edge is formed with the probability $p=0.4$. - Watts-Strogatz small world model [@watts1998collective]: initially, $n=50$ vertices are connected in a ring topology, with each vertex connecting to its $k=3$ neighbors, and then, each edge is randomly rewired with the probability $p=0.01$. - Albert-Barabási scale free model [@barabasi1999emergence]: initial topology of the network consists of $n_0$ vertices forming a complete graph $K_{n_0}$, remaining vertices are added to the network sequentially until the desired number $n=50$ of vertices is reached, and each newly added vertex creates $k=3$ edges to existing vertices, choosing target vertices with the probability proportional to their degrees, hence the alternative name of the model: preferential attachment model. The regular, linear preferential attachment is achieved for $\alpha=1$ . - Leskovec forest fire model [@leskovec2005graphs]: vertices are added sequentially to the network, each out of $n=50$ new vertex provides $k$ edges to uniformly selected targets, and then adds more edges to direct neighbors of selected targets with the burning probability $p=0.3$, upon successful creation of a new edge the process continues recursively. The model produces networks of low density and relatively large diameter, but with average shortest path lengths significantly lower than in the case of small world networks. - Dorogovtsev-Goltsev-Mendes hierarchical model [@PhysRevE.65.066122]: initially, the network consists of a single edge connecting two vertices, and in each step of the network growth an edge produces a new vertex which is immediately attached to both ends of the edge. Although the resulting network is not strictly a fractal, it has a strong hierarchical structure with power law distributions of vertex degree and local clustering coefficient. In addition, the network contains multiple loops and is far from a trivial tree-like structure. - Disassortative model: an empty network with $n=100$ vertices is created and the following procedure is iteratively repeated. Vertices are ordered by their identifiers, the first 10 vertices create between 30 and 40 edges to randomly selected vertices from the range $\left< 50,100 \right>$, vertices from the range $\left< 10,50 \right>$ create up to 15 edges to random vertices, and vertices from the range $\left<51,100\right>$ create between 0 and 2 edges to random vertices. After each iteration self-loops and multiple edges are removed. The procedure continues until the degree assortativity coefficient falls below $-0.4$; this particular threshold has been set arbitrarily. symbol meaning ---------------- ---------------------------------------------------------------------------------------------------- $p_D$ p-value of the Kolmogorov-Smirnov test comparing the degree distributions of two networks $p_B$ p-value of the Kolmogorov-Smirnov test comparing the betweenness distributions of two networks $p_C$ p-value of the Kolmogorov-Smirnov test comparing the closeness distributions of two networks $d$ network diameter: the length of the longest shortest path in the network $\rho$ network density: the ratio of the number of existing edges to the maximum number of possible edges $L$ the average length of the shortest paths in the network $\rho^2_{C_D}$ Freeman’s centralization of the degree distribution of the network : Metrics used to compare networks[]{data-label="tab:metrics"} network $p_D$ $p_B$ $p_C$ $\left\| V \right\|$ $\left\| E \right\|$ $d$ $\rho$ $L$ $\rho^2_{C_D}$ --------- ------- ------- ------- -------------------- -------------------- ---------- ---------- ---------- ---------------- ---------------------------- -- 2.00 0.40 1.59 0.11 original 2.17 0.42 1.59 0.12 random distance 11.00 0.06 4.54 0.01 original 14.33 0.06 4.84 0.05 euclidean distance 3.00 0.06 1.53 0.25 original 4.18 0.06 1.83 0.38 linear regression distance 7.00 0.04 2.16 0.07 original 7.09 0.05 2.37 0.08 euclidean distance 5.00 0.02 1.70 0.12 original 6.00 0.02 2.08 0.16 degree distance 8.00 0.05 2.85 0.11 original 7.00 0.05 2.88 0.25 degree distance : Re-creation of synthetic networks. Results that pass the K-S test as well as scalar network descriptors such as diameter $d$ or density $\rho$ that are regenerated within $\pm10\%$ margin of the original value are marked with bold font[]{data-label="tab:random.network.results"} Metrics used to compare synthetic networks and networks generated by the Priority Rank model are presented in Table \[tab:metrics\]. Synthetic network statistics and the corresponding statistics of networks generated by the Priority Rank model are shown in Table \[tab:random.network.results\]. For distributions of degree, betweenness, and closeness, a standard two sample Kolmogorov-Smirnov test at level $\alpha=0.05$ has been performed and cases when the null hypothesis cannot be rejected are marked with boldface, i.e., the null hypothesis states that the compared samples come from the same distribution. Please note, however, that even if the re-created value does not fulfill this condition, it is rarely significantly different from the original since the original network descriptor is very small. ### Erdös-Rényi random network {#erdös-rényi-random-network .unnumbered} The Priority Rank model is able to re-create Erdös-Rényi random network very well. As can be seen in Table \[tab:random.network.results\], the network generated by the Priority Rank model reproduces all three centrality distributions of degree, betweenness, and closeness. It also retains the average shortest path length, and Freeman’s centralization of degree distribution. The only network statistic which is not duplicated is diameter. Overall, edges are created randomly in the Erdös-Rényi model, so random distance function works very well because vertices have no preference for other vertices. ### Watts-Strogatz small world network {#watts-strogatz-small-world-network .unnumbered} The initial topology of the ring is formed in one dimension. None of the considered distance functions was able to re-create the degree distribution, although the euclidean distance function managed to generate networks with similar betweenness and closeness distributions. The standard deviation of the K-S statistic is quite large, which weakens the result. The euclidean distance function used two synthetic attributes generated from exponential and normal distributions. The failure to re-generate the degree distribution is caused by the fact that Priority Rank does not have anything akin to random edge re-wiring, which is essential for the small-world model. The ring structure can be very easily reproduced. However, the presence of a few randomly re-wired edges, which reduce average path lengths in the network, is very difficult to mimic using any distance function. Apart from this deficiency, the Priority Rank model generates networks, which are very similar to the original small world network in terms of network diameter, density, and average shortest path lengths, these characteristics are re-created almost flawlessly. As a matter of fact, Priority Rank is capable of producing small world networks. The main issue is with reproducing the topology of a particular implementation of the Watts-Strogatz model. The original Watts-Strogatz model places all vertices on the circumference of a circle and attaches neighboring nodes. Thus, the proximity of vertices is defined by the Euclidean distance in two dimensions, with vertices artificially placed on a circle. This is a highly contrived scenario which bares little resemblance to any practical application of the model. On the other hand, we have assumed in our experiment that each vertex has a single numerical attribute. The proximity of vertices is defined by the one dimensional distance in the space of this attribute. For instance, one can interpret this attribute as yearly income. The distance function reflects the tendency of people to socialize within levels of financial hierarchy. Another example could be the tendency of people to form social ties with people of similar age, i.e., age is the attribute. Any scenario when the social network is driven by strong homophily can be modeled using our distance function, provided that the driver of homophily can be expressed by a single numerical attribute. For $n=100$ vertices, neighborhood size of $m=4$ and the probability of random edge rewiring $p=5\%$ the Priority Rank model produces the network with the global clustering coefficient (also known as transitivity) of $\mathit{gcc}=0.41$. The native function `igraph::sample_smallworld()` from the `igraph\R` package with the same parameters delivers a network with $\mathit{gcc}=0.47$. It is, in turn, far away from the network produced by Priority Rank using another degree distance function (the output network is scale free): $\mathit{gcc}=0.10$ as well as from the network provided by the random distance function (that mimics the Erdös-Rényi network): $\mathit{gcc}=0.06$. In our opinion, the Priority Rank model still produces valid small world networks defined as networks, which have disproportionately large global clustering coefficient and relatively small average shortest path length. ### Albert-Barabási scale free network {#albert-barabási-scale-free-network .unnumbered} Priority Rank successfully re-creates degree and betweenness distributions, while struggling to preserve the distribution of the closeness centrality (the linear regression distance function barely manages to pass the K-S test). The Priority Rank model also builds networks with larger diameters than the original network, but the density and the centralization of the degree distribution are very close to original values. We also note that the networks generated based on Priority Rank, in which average shortest paths are slightly longer than in the original Albert-Barabási model. Nevertheless, we conclude that synthetic scale-free networks can be mimicked by the Priority Rank model sufficiently. ### Leskovec forest fire network {#leskovec-forest-fire-network .unnumbered} The Priority Rank model can re-create forest fire networks very precisely. The best results are obtained when using euclidean distance based on one attribute with values drawn randomly from the normal distribution. Such distance function delivers networks with similar centrality measure distributions (degree, betweenness, closeness), and with very similar diameter, density, average shortest path lengths and degree distribution. ### Dorogovtsev-Goltsev-Mendes hierarchical network {#dorogovtsev-goltsev-mendes-hierarchical-network .unnumbered} The Priority Rank model is capable of successful generating hierarchical network structures, such as pseudo-fractal topologies of the Dorogovtsev-Goltsev-Mendes network. The only centrality measure that has not been reproduced is the closeness. Network density is reproduced exactly whereas the remaining descriptors are rather close to their original values. ### Disassortative network {#disassortative-network .unnumbered} As can be expected, the degree distance function was the best choice for re-creating disassortative networks. Distributions of degree and betweenness are re-created very well together with density and the average shortest path length. The diameter of the network is almost reproduced, and the only network descriptor which significantly differs from the original is Freeman’s degree centralization. Overall, the Priority Rank model can easily mimic synthetic networks produced by popular network generative models. A simple substitution of the distance function allows the Priority Rank model to produce instances of random networks, small world, scale-free, forest fire, hierarchical, and disassortative networks. In addition, one can easily introduce different variations of these generative models by modifying the distance function used to compute local priority rankings. This flexibility and ability to generalize multiple models is a unique and practical feature of the Priority Rank model. Empirical networks {#subsec:empirical.networks .unnumbered} ------------------ type name description ------ ------------------------- ---------------------------------------------------------------------- -- -- american bison Dominance relations among a group of American bison bulls bighorn sheep Dominance interactions among a group of female bighorn sheep C.elegans Neural connections of the C.elegans nematode mouse visual cortex Neuron interactions in the mouse primary visual cortex enzyme 108 Enzyme interaction network cage5 Cage model of DNA electrophoresis political books Co-purchase of books about U.S. politics on Amazon primary school Contacts among students and teachers at a primary school in Lyon vickers 7th graders Friendships among seventh grade students in Victoria, Australia freeman researchers E-mail messages exchanged between researchers terrorists Social associations of 9/11 hijackers karate club Friendships among members of a university karate club illinois high school Friendships among male students in a small high school in Illinois marseille high school Contacts among students in a high school in Marseilles CAG Integer matrices used in characteristic polynomials computations power network Symmetric structure of the standard test of the power system network football Foosball player contracts between countries st.marks ecosystem Carbon-flow among species in St.Marks National Wildlife Refuge : Empirical networks used in the experiments[]{data-label="tab:real.world"} Apart from popular generative network models, the Priority Rank model can generate networks with topologies not available through traditional models if provided with a suitable custom distance function. However, the most interesting and valuable property of the Priority Rank model is its ability to learn the generative processes of empirical networks and to generate multiple instances of these networks. The experimental evaluation of this feature is carried out on networks listed in Table \[tab:real.world\] and in the supplementary material. They are provided by The Colorado Index of Complex Networks [@icon] and The Network Repository [@nr]. Metrics used to compare empirical networks and networks generated by the Priority Rank model are presented in Table \[tab:metrics\]. Statistics of original networks along with corresponding ones for the networks generated by Priority Rank are shown in Table \[tab:empirical.network.results\]. network $p_D$ $p_B$ $p_C$ $\left\| V \right\|$ $\left\| E \right\|$ $d$ $\rho$ $L$ $r$ $\rho^2_{C_D}$ --------- ------- ------- ------- -------------------- -------------------- -------- ---------- ---------- ---------- ---------------- ---------------------------- 4 0.46 1.57 0.59 0.31 original 3 0.44 1.59 0.42 0.19 random distance 5 0.32 1.90 0.12 0.16 original 4 0.34 1.86 0.31 0.31 page rank distance 7 0.04 2.88 0.47 0.21 original 7 0.04 2.83 0.04 0.13 closeness distance 2 0.05 1.08 0.00 0.11 original 4 0.06 1.76 0.12 0.18 page rank distance 15 0.22 5.85 1.00 0.10 original 14 0.26 5.39 0.58 0.06 euclidean distance 6 0.18 2.22 0.56 0.31 original 7 0.18 2.10 0.23 0.40 degree distance 8 0.04 2.92 0.00 0.08 original 7 0.04 3.00 0.14 0.12 euclidean distance 4 0.29 1.77 0.00 0.14 original 4 0.30 2.03 0.30 0.22 euclidean distance 3 0.45 1.60 0.67 0.32 original 3 0.48 1.59 0.42 0.32 closeness distance 3 0.45 1.56 0.79 0.55 original 3 0.45 1.58 0.51 0.52 betweenness distance 5 0.16 2.95 1.00 0.28 original 8 0.14 2.84 0.07 0.28 degree distance 3 0.13 1.27 0.00 0.19 original 4 0.14 1.68 0.04 0.19 closeness distance 12 0.07 3.97 0.50 0.09 original 11 0.08 3.91 0.35 0.06 aggregate distance 6 0.11 2.03 0.00 0.11 original 6 0.11 2.40 0.10 0.31 linear regression distance 6 0.18 2.22 0.56 0.31 original 7 0.11 2.55 0.12 0.23 betweenness distance 11 0.06 4.39 0.00 0.04 original 13 0.08 4.77 0.00 0.03 random distance 3 0.10 1.30 0.00 0.19 original 4 0.11 1.88 0.13 0.31 degree distance 7 0.12 2.87 0.02 0.34 original 7 0.11 2.55 0.12 0.23 betweenness distance : Re-creation of empirical networks from Table \[tab:real.world\]. Results that pass the K-S test and scalar descriptors from within $\pm 10\%$ margin of the source network are in bold[]{data-label="tab:empirical.network.results"} - American bisons: The Priority Rank model correctly captures the underlying network structure and can re-create the network to a sufficient extent. All centrality measure distributions are retained in the generated networks, and these networks exhibit densities, and average shortest path lengths similar to the original network. The Priority Rank model slightly underestimates the diameter and the centralization of the degree distribution. Since the original network resembles the Erdös-Rényi random network, no wonder that the random distance best matches recreation process. Since Priority Rank is capable of re-creating the American bison network, this means that one could generate multiple instances of this network and belief that the generated instances reflect the same principle guiding the formation of the original network. These new networks could be interpreted as observations of another herd of bisons, or snapshots of the same group for different periods. This may be very useful for simulation purposes while analyzing transmission of a disease between individual animals living in the wild. Network instances generated with Priority Rank could be used as multiple alternative scenarios. - Bighorn sheep: The Priority Rank model very convincingly re-creates the topology of the original network maintaining distributions of all centrality measures, density and the average shortest path lengths. The generated networks have slightly smaller diameters than the original one, and the reciprocity is over-estimated. This experiment, however, supports our general claim that the Priority Rank model can discover the latent generative principle of an empirical network. The page rank distance aggregates the importance of all animals in the herd. Intuitively page rank reflects the true importance of animals given the history of dominance relationships. For the sake of brevity, we report only on the best fitting distance function for each network. Nevertheless, we have analyzed several different distance functions for each network. In this case, very similar results were obtained for two distance functions, which both made use of the age attribute. Naive Bayes classifier distance and linear regression distance functions are machine learning algorithms which map the relationship between two animals’ ages onto the probability of the existence of the dominance relationship between these animals. These two distance functions were close competitors of the page rank distance function. Thus, the interpretation of these distance functions allows us to extract a general principle of this particular network formation. - C.elegans: The Priority Rank model generates networks which are very similar in terms of degree and betweenness distributions to the C.elegans network, but we were unable to re-create a similar distribution of the closeness centrality measure (this could be due to a large number of vertices in the network). Also the distance, the density and the average path lengths are faithfully re-created with Priority Rank. One missing characteristic of the original network is the reciprocity. Our model cannot capture this feature using the closeness distance function. We note, however, that adding reciprocity to the distance function is straightforward. It is sufficient to include a component that would diminish the estimated distance between vertices $v_i$ and $v_j$, if an edge $(v_j,v_i)$ is already present in the network. - Mouse visual cortex: This small network is very difficult to re-create, most probably due to the fact that its generative process may be complex. Unfortunately, the source network does not contain any additional attributes and we have to rely only on topological features of the network. In this case, the best result is obtained for the page rank distance which creates networks similar to the scale free model. The Priority Rank model is able to produce networks with very similar degree and betweenness distributions. - Enzyme 108: The Priority Rank model using euclidean distance can re-generate all centrality measure distributions and obtain very similar values of diameter, density, and the average shortest paths length. The euclidean distance measure uses two synthetic attributes drawn from the uniform distribution. A distance function tends to produce networks similar to the small world model, and the enzyme 108 network definitely belongs to this network family. This small-worldliness of enzyme 108 is best manifested in the density and the average shortest path length, which are the largest among all analyzed networks. Interestingly, although Priority Rank had problems re-creating the synthetic small-world network, it manages to approximate the real world example of a small-world network very faithfully. Look at Illinois high school to see the same Priority Rank behavior. This might indicate that the original small-world model proposed by Watts and Strogatz over-simplifies the reality and the priority attachment based on the similarity of attributes is a better representation of real network formation phenomena. - Cage5: The degree distance function reproduces distributions of degree and betweenness, but cannot re-create the original distribution of the closeness centrality. As for the topology of the generated network, the Priority Rank model generates very similar networks in terms of diameter, density, average shortest path length and degree centralization. Our model underestimates the reciprocity of the network, most probably due to the fact that the degree distance function alone cannot account for the increased probability of reciprocal relationships. - Political books: This network is easily re-created with the Priority Rank model using the euclidean distance function based on a single discrete attribute, which reduces the function to a simple binary flag comparison. This result agrees very well with our intuition. One can interpret the synthetic discrete attribute as an indicator of a broad book category, and two books are purchased together if they belong to the same category. The Priority Rank model re-creates degree and betweenness distributions, but fails to retain the original distribution of closeness. Also the diameter, the density, the average shortest path length, and the centralization of degree are very similar to the original network. The difference in the reciprocity estimation results from lack of direction of edges in the original network. The Priority Rank model produces inherently directed networks. This example illustrates well the ability of the Priority Rank model to discover the latent process of network formation and provide simple, interpretable explanations of the underlying network structure. - Primary school: The main challenge that this network poses is the density of social interactions and very short average shortest path lengths. The Priority Rank model re-creates this network reasonably well using the euclidean distance function on a single discrete attribute. Similarly to the political books network, a simple comparison of a discrete attribute is sufficient to produce a good approximation of the original network. Although the Priority Rank model cannot reproduce original distributions of betweenness and centrality, the degree distribution is very well preserved. The remaining topological characteristics of the network are also retained, except for the reciprocity — for the same reasons as for political books network. The interpretation of the distance function is similar to the political books network, namely, the discrete attribute serves as a label of a coherent social group and students have a strong preference to create relationships within the social group they belong to. - Vickers 7th graders: The Priority Rank model re-creates this network perfectly using the closeness distance, where the closeness of the vertex is computed on incoming edges. With the exception of the reciprocity (slightly underestimated), all the remaining network profile is re-created precisely. One reason for this result is the fact that multiple types of edges have been flattened to a single layer in the original network, making the guessing of edge existence a bit easier. Incoming closeness seems to be a very reasonable explanation for edge presence. Popular students who have been nominated by many peers have indeed high value of inbound closeness centrality. Using this measure to generate ranking lists in the Priority Rank model yields good approximations of the original network. Again, we see how the Priority Rank model provides a simple and interpretable explanation of the latent network generation process. - Freeman researchers: An interesting feature of this network is the fact that it represents communication patterns in the pre-internet era. As expected, the network is characterized by very high reciprocity, low diameter and average shortest path length. The best distance function uses betweenness centrality. Our model not only re-generates identical distributions of centrality measures, but also captures all the remaining network statistics. Again, only the reciprocity has not been matched, due to the fact that the betweenness distance function is not capable of taking the existence of an edge into consideration. Since the network represents the flow of communication between people, it is not surprising that the best distance function uses the betweenness centrality. It ranks network vertices according to their impact on the communication pathways through the network. We find this result to be a strong indication that the Priority Rank model really can provide a viable explanation for the phenomenon driving the network formation. - 9/11 terrorists: Similarly to the network of researchers discussed above, the Priority Rank model re-creates the 9/11 terrorists network precisely. Although the distribution of closeness centrality is not preserved, the remaining network characteristics are almost identical to the original network. The choice of the distance function is also obvious. The network has been gathered post hoc in a way to outline the associations of a selected group of terrorists, and, as the result, these terrorists possess prominent degrees in the network. Again, the interpretation of the distance function is straightforward and the result supports our claim about the explanatory power of the Priority Rank model. - Karate club: Priority Rank is capable of re-creating the structures of the original network, with the notable exception of the closeness distribution and the overestimated average shortest path length. The best distance function and perfect network explanation is the closeness distance. Recall that the network represents social contacts within the karate club community at a university. The community has split in half due to a conflict between the two principal members, the administrator and instructor. Almost all of the members of the original club have links to one of the leaders of the club. Thus, one may conclude that the members had a strong preference to prioritize social contacts to these leaders, who are the two vertices with the shortest average paths to all the remaining vertices. - Illinois high school: The most distinguishing feature of this network is its very large diameter and relatively high average shortest path length. Theoretically, this network tends to follow the small world network model of Watts and Strogatz. The Priority Rank model re-generates this network almost perfectly, retaining the distributions of centrality measures and all other network descriptors. As expected, the best fitting is obtained for the aggregate distance function. This function compares the values of two attributes randomly drawn from the normal distribution. As a result, vertices with many similar attribute values tend to form communities (small worlds), with very few edges interconnecting these communities, hence the large network diameter. One may conclude that the evolution of these types of small world networks is primarily driven by the homogeneity of vertices constituting communities. We find it surprising that the Priority Rank model struggled to re-create the synthetic small world network, whereas it re-creates real world examples of small world networks very convincingly. In our opinion, this indicates that the assumptions behind the small world model of Watts and Strogatz are not fully supported by empirical networks. - Marseille high school: Surprisingly, the Priority Rank model has struggled to reconstruct this network, not being able to re-create distributions of betweenness and closeness. However, despite the problem with retaining the original centrality measure distributions, the networks generated by Priority Rank are very similar w.r.t. the remaining network characteristics. The best distance function is linear regression, which indicates that the attributes of vertices (class, sex) are crucial for inferring the existence of relationships. This is even more encouraging than previous examples where networks have been re-created using synthetic attributes. The linear regression distance function can be easily interpreted. It even provides quantitative estimations of the latent network generative process by means of regression coefficients. - CAG: The main focus of this experiment was to check whether Priority Rank is capable of capturing the generative process of highly atypical network induced by integer matrices used in computations of characteristic polynomials. Our model re-creates degree and betweenness distributions, as well as most of the descriptors of the original network, except reciprocity. Since the best fitting function is the degree distance, we are tempted to conclude that the CAG network is generated primarily by the preferential attachment mechanism. - Power network: This dataset representing the structure of power stations connections is challenging due to its very low density, large diameter and lack of degree centralization. Despite this, the Priority Rank model successfully reproduces the network, preserving all the distributions and measures. The best results are obtained for the random distance function, suggesting a haphazard setup of the network. - Football: Priority Rank produces networks very similar to the original network, with slightly over-estimated average shortest path length and degree centralization. The former is probably caused by the greater diameters than the original network. As expected, the best fitting function is the degree distance. It reflects the socio-economic reality of the process underlying the network formation: poorer countries are exporting best players to a few countries where very wealthy football clubs reside. Here we can see that the Priority Rank model provides an interpretation of the latent network formation mechanism, even if the mechanism involves complex socio-economical processes. - St.Marks ecosystem: In this experiment, we were particularly interested in discovering the underlying generative process of network formation. We can see that the Priority Rank model re-generates this network very accurately, preserving distributions of degree and betweenness and producing very similar diameters, densities, and average shortest path lengths. The betweenness distance turned out to be the best function. It is not surprising given the fact that the network focuses on modeling flows within the ecosystem. Again, we believe that the Priority Rank model discovers the main generative process of this network formation. Discussion {#sec:discussion .unnumbered} ========== In this paper, we have developed the priority attachment, a plausible principle of network formation. We have shown that priority attachment can generalize previously proposed mechanisms of network formation, such as the small world phenomenon or preferential attachment. The second major contribution of this paper is the introduction of the Priority Rank model, a comprehensive generative network model which utilizes the priority attachment principle. The Priority Rank model is capable of mimicking both synthetic networks produced by popular generative network models, as well as re-creating empirical networks. The priority attachment mechanism requires a distance function to compute local rankings for each vertex. The interpretation of the distance function allows us to infer the properties of generated networks in case of synthetic networks. More importantly, it captures the nature of the generative process imprinting formation of empirical networks. In our opinion this is the most valuable and handy feature of Priority Rank. The ability to discover the guiding principle of network formation facilitates generation of multiple similar but slightly different realizations for a given network, resulting in the whole population of networks. The stochastic nature of the priority attachment mechanism provides a small degree of randomization required to introduce variance into such population of networks, while preserving distributions of centrality measures and network profile at the same time. Additionally, the Priority Rank model and a suitable distance function can provide a network generation principle that better matches real networks than the other well-known generative models. The Enzyme 108 and Illinois high school networks are quite typical examples of small-world networks that are better represented by the Priority Rank model than by the Watts-Strogatz model [@watts1998collective]. Many different applications of the Priority Rank model can be enumerated. Apart from the obvious substitution of multiple generative network models by a single model, the Priority Rank model can be used for A/B testing of networks. Using our model, series of networks with a given profile may be generated. As a result, statistical inference on networks is possible owing to presence of many network instances belonging to the same family as the original one, which represents only a single data point. Priority Rank facilitates the simulation of various scenarios for network formation and growth. Given a model of network formation principle under the form of the distance function, we can predict evolution of networks. Priority Rank is an early implementation of the priority attachment principle, which is capable of creating or re-creating networks of varying topologies, but it should not be regarded as a final model. In fact, in the next section we outline several limitations of the current model and we sketch possible solutions. But despite some deficiencies of the Priority Rank model we see strong evidence in favor of our main claim, namely, that priority attachment is a generalization of previously proposed network formation principles and can successfully be used to explain the emergence of many network topologies. Limitations {#subsec:limitations .unnumbered} ----------- The main focus of this paper is the introduction of the priority attachment, an important and useful mechanism of network formation. Priority Rank, on the other hand, is a generative model which uses the priority attachment when generating networks. As such, Priority Rank has certain limitations which are unavoidable, since every network model has to make a compromise between accuracy and simplicity. Firstly, we note that the Priority Rank model does not always re-create empirical networks fully. It seems that the closeness distribution is notoriously difficult to reproduce. Also, we note the lack of a mechanism which would allow to easily account for the reciprocity. Of course, this can be introduced into the model by modifying the distance function to put more weight to existing edges. Another shortcoming of the current model is the fact, that while trying to re-create an empirical network, the search for the suitable distance function is performed via brute force approach. As of now, we are searching through the parameter grid of a fixed set of template distance functions and finding the distance function which best reproduces the input network. This is both time consuming and ineffective, as the search space is limited. Possible solution to this shortcoming would be to replace currently used distance functions by a single universal distance function discovery module. We believe that for larger networks, where adjacency matrices contain sufficient amount of information, the application of neural networks to form distance functions is a viable direction of future research. As a matter of fact, we are currently working on a deep neural network architecture that could extract several layers of topological features and learn the distance function. Finally, we note that the amount of randomness in the Priority Rank model is limited, which is most pronounced in the inability of the model to re-create small world networks (at least the networks produced by the Watts-Strogatz model). There are two places where randomness can be introduced into the Priority Rank model: either the distance function can produce random results for a subset of vertices, or one can modify the rank probabilities used by the model to select target vertices from local vertex rankings. The latter can be achieved, e.g., by flattening the distribution of probabilities of picking vertices at given positions in the ranking. This issue, will be another subject of future research. Acknowledgments {#acknowledgments .unnumbered} =============== We wish to thank F.Menczer, S.Fortunato, and A.Flammini for their valuable remarks. This work was partially supported by the National Science Centre, Poland, projects no. 2016/23/B/ST6/03962, 2016/21/B/ST6/01463 and 2016/21/D/ST6/02948; European Union’s Horizon 2020 research and innovation program under the Marie Skłodowska-Curie grant agreement No. 691152 (RENOIR); the Polish Ministry of Science and Higher Education fund for supporting internationally co-financed projects in 2016-2019 no. 3628/H2020/2016/2. The authors declare that there is no conflict of interest regarding the publication of this paper. Additional information {#additional-information .unnumbered} ====================== PriorityRank online generator is available at <https://priorityattachment.ml> Author information {#author-information .unnumbered} ================== Affiliations {#affiliations .unnumbered} ------------ Institute of Computing Science, Poznań University of Technology, Poland Mikołaj Morzy, Grzegorz Miebs, Arkadiusz Rusin Faculty of Computer Science & Management, Wroclaw University of Science and Technology, Poland Tomasz Kajdanowicz, Przemysław Kazienko Author contributions statement {#author-contributions-statement .unnumbered} ------------------------------ M.M., T.K., and P.K. devised the idea and designed experiments. A.R. constructed the prototype and conducted initial experiments on real-world networks. G.M. extended the prototype and conducted follow-up experiments on real-world data. M.M., T.K., P.K., and G.M. analyzed the data and experimental results. M.M., T.K., and P.K. edited the manuscript. Competing interests {#competing-interests .unnumbered} ------------------- The authors declare no competing financial interests. Corresponding author {#corresponding-author .unnumbered} -------------------- Correspondence to Mikołaj Morzy. [^1]: Within the scope of this paper we consider global ranking to be a universal ordering of vertices common to all vertices, whereas local ranking denotes an ordering of vertices as perceived by an individual vertex.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The proof of Anderson localization for the 1D Anderson model with arbitrary (e.g. Bernoulli) disorder, originally given by Carmona-Klein-Martinelli in 1987, is based in part on the multi-scale analysis. Later, in the 90s, it was realized that for one-dimensional models with positive Lyapunov exponents some parts of multi-scale analysis can be replaced by considerations involving subharmonicity and large deviation estimates for the corresponding cocycle, leading to nonperturbative proofs for 1D quasiperiodic models. In this paper we present a short proof along these lines, for the Anderson model. To prove dynamical localization we also develop a uniform version of Craig-Simon’s bound that works in high generality and may be of independent interest.' address: 'Department of Mathematics, University of California, Irvine' author: - 'Svetlana Jitomirskaya, Xiaowen Zhu' bibliography: - 'mybib1.bib' title: Large deviations of the Lyapunov exponent and localization for the 1D Anderson model --- Introduction ============ Anderson localization for the Anderson model can be proved in several different ways if the common distribution of the i.i.d.r.v’s is absolutely continuous. Without that condition (or at least some Hölder regularity) it remains an open question for $d\geq 2,$ and the number of approaches that work for $d=1$ also drops dramatically. Such is the situation, for example, for the Bernoulli-Anderson model. Anderson localization for [arbitrary]{} 1D disorder was first proved in [@ckm]. The approach was based on certain regularity of the Lyapunov exponents coming from the (analysis around) the Furstenberg theorem to obtain an analogue of Wegner’s lemma (automatic in the absolutely continuous case). After that the proof was reduced to multi-scale analysis, with initial scale coming again from the positive Lyapunov exponent. Another argument was later presented in [@svw], where an approach to positivity and regularity of the Lyapunov exponent using replica trick was given, again reducing the proof to multi-scale analysis. Multi-scale analysis is a method that allows to achieve Green’s function decay and ultimately localization from high probability of decay at the initial scale. It works in a variety of settings. Originally developed by Frohlich and Spencer [@fs], it was significantly simplified in [@von1989new] but remains somewhat involved. It should be noted that in the multidimensional case no shortcuts such as Furstenberg theorem or replica trick are available, and the multi-scale analysis is used to reach conclusions analogous to the positivity of the Lyapunov exponent simultaneously with the proof of localization. Yet in the one-dimensional case positivity of the Lyapunov exponent essentially provides the averaged decay statement, thus a large portion of the [conclusion]{} of the multi-scale analysis, making its machinery seem redundant. A method to effectively exploit positive Lyapunov exponent for a localization proof based on the analysis of the large deviation set for the Lyapunov exponent was first developed in [@j] for the almost Mathieu operator, initiating what was later called a non-perturbative approach, in contrast with earlier proofs based on some form of multi-scale analysis [@fsw; @sin]. A robust method based on subharmonic function theory and the theory of semianalytic sets was then developed in [@bg] and other papers summarized in [@bbook], to conclude localization from positive Lyapunov exponents for analytic quasiperiodic and some other deterministic potentials. The fact that those ideas can be applicable also to the Anderson model was mentioned in some talks by one of the authors circa 2000, but the details were never developed. One goal of this paper is to obtain a proof of Anderson localization for the 1D Anderson model in the spirit of [@j] but with appropriate simplifications due to randomness. Another proof, also based on large deviations and also avoiding multi-scale analysis was recently developed in [@7]. The proof of [@7] is based on deterministic ideas close to the ones in [@bs], which we believe may be somewhat more complicated than needed for the random case. We mention that yet another, purely dynamical, proof of localization for the 1D Anderson model appears in [@gk]. One ingredient in our simple argument for spectral localization, Theorem \[CS\], is Craig-Simon’s upper bound based on subharmonicity of the Lyapunov exponent [@cs], a statement that holds for any ergodic potential. In order to prove dynamical localization we need a uniform in energy and quantitative version of this statement, that we prove for general ergodic potentials satisfying certain large deviation bounds, a result that could be of independent interest. We note that our proof does not explicitly use subharmonicity. The rest of this paper is organized as follows. Section \[pre\] contains the preliminaries, the statement of the spectral localization result, Theorem \[thm1\], and its quick reduction to Theorem \[thm2\]. We then prove the preparatory Lemmas \[lemma1\], \[omega1\], \[omega3\], and Corollary \[omega2\] in Section \[alllemma\]. Then we complete the proof of Theorem \[thm2\] in Section \[pf\]. Our proof effectively establishes a more precise result, Theorem \[thm22\], which in turn immediately implies the Lyapunov behavior at all eigenvalues, Theorem \[CS2\]. We formulate and prove the general uniform Craig-Simon-type statement in Section \[uniformcs\], and use it in Section \[dyn\] to prove dynamical localization. Preliminaries {#pre} ============= The one dimensional Anderson model is given by a discrete Schrödinger operators $H_{\omega}$ $$(H_{\omega}\Psi)(n)=\Psi(n+1)+\Psi(n-1)+\omega_n\Psi(n),$$ where $\omega_n\in\mathbb{R}$ are independent identically distributed random variables with common Borel probability distribution $\mu.$ We will assume that $S\subset\mathbb{R}$, the topological support of $\mu,$ is compact , and contains at least two points. We will denote the probability space $\Omega=S^{\mathbb{Z}},$ with elements $\{\omega_n\}_{n\in\mathbb{Z}}\in\Omega$. Denote $\mu^{\mathbb{Z}}$ as $\mathbb{P}$. Let $\mathbb{P}_{[a,b]}$ be $\mu^{[a,b]\cap \mathbb{Z}}$ on $S^{[a,b]\cap \mathbb{Z}}$. Aldo let $T$ be the shift $T\omega_i=\omega_{i-1}$. Finally, we denote Lebesgue measure on $\mathbb{R}$ by $m$. We say that $H_\omega$ has spectral localization in $I$ if for a.e. $\omega$, $H_\omega$ has only pure point spectrum in $I$ and its eigenfunctions $\Psi(n)$ decay exponentially in $n$. We call $E$ a generalized eigenvalue ($g.e.$), if there exists a nonzero polynomially bounded function $\Psi(n)$ such that $H_\omega\Psi=E\Psi$. We call $\Psi(n)$ a generalized eigenfunction. Since the set of g.e. supports the spectral measure of $H_\omega$ (e.g. [@cycon]), we only need to show: \[thm1\] For a.e. $\omega$, for every g.e. $E,$ the corresponding generalized eigenfunction $\Psi_{\omega,E}(n)$ decays exponentially in $n$. For $[a,b]$ an interval, $a,b\in\mathbb{Z}$, define $H_{[a,b],\omega}$ to be operator $H_\omega$ resticted to $[a,b]$ with zero boundary conditions outside $[a,b]$. Note that it can be expressed as a “$b-a+1$”-dimensional matrix. The Green’s function for $H_\omega$ restricted to $[a,b]$ with energy $E\notin\sigma_{[a,b],\omega}$ is $$G_{[a,b],E,\omega}=(H_{[a,b],\omega}-E)^{-1}$$ Note that this can also be expressed as a “$b-a+1$”-dimensional matrix. Denote its $(x,y)$ entry as $G_{[a,b],E,\omega}(x,y)$. It is well known that $$\label{possion} \Psi(x)=-G_{[a,b],E,\omega}(x,a)\Psi(a-1)-G_{[a,b],E,\omega}(x,b)\Psi(b+1),\quad x\in[a,b]$$ and we have $$\label{sigma} \sigma:=\sigma(H_{\omega})=[-2,2]+S\quad a.e. \omega.$$ For $c>0, n\in\mathbb{Z}$, we say $x\in\mathbb{Z}~$ is $(c,n,E,\omega)$-regular, if $$G_{[x-n,x+n],E,\omega}(x,x-n) \leq e^{-cn}$$ $$G_{[x-n,x+n],E,\omega}(x,x+n) \leq e^{-cn}$$ Otherwise, we call it $(c,n,E,\omega)$-singular. By and definition 2, Theorem \[thm1\] follows from \[thm2\] There exists $ \Omega_0$ with $\mathbb{P}(\Omega_0)=1$, such that for every $ \tilde{\omega}\in\Omega_0$, for any g.e. $\tilde{E}$ of $H_{\tilde{\omega}}$, there exist $ N=N(\tilde{E},\tilde{\omega}),C=C(\tilde{E})$, such that for every $ n>N$, $2n,~2n+1$ are $(C,n,\tilde{E},\tilde{\omega})$-regular. Some other standard basic settings are below. Denote $$P_{[a,b],E,\omega}=det(H_{[a,b],E,\omega}-E), a\leq b$$ If $a>b$, let $ P_{[a,b],E,\omega}=1$. Then $$\label{A} \left\vert G_{[a,b],E,\omega}(x,y)\right\vert=\frac{\left\vert P_{[a,x-1],E,\omega}P_{[y+1,b],E,\omega}\right\vert}{\left\vert P_{[a,b],E,\omega}\right\vert},\quad x\leq y$$ If we denote the transfer matrix $T_{[a,b],E,\omega}$ as the matrix such that $$\left( \begin{array}{c} \Psi(b)\\ \Psi(b-1)\\ \end{array} \right) =T_{[a,b],E,\omega} \left( \begin{array}{c} \Psi(a)\\ \Psi(a-1)\\ \end{array} \right)$$ where $\Psi$ solves $H_\omega\Psi=E\Psi,$ then $$T_{[a,b],E,\omega}=\left( \begin{array}{cc} P_{[a,b],E,\omega} & -P_{[a+1,b],E,\omega}\\ P_{[a,b-1],E,\omega} & -P_{[a+1,b-1],E,\omega}\\ \end{array} \right)$$ The Lyapunov exponent exists by Kingman’s subadditive ergodic theorem and is given by $$\label{gamma} \gamma(E)=\lim_{n\to\infty}\frac{1}{n}\int_0^1 \log\Vert T_{[0,n],E,\omega}\Vert d\mathbb{P}(\omega)=\lim_{n\rightarrow\infty}\frac{1}{n} \log\Vert T_{[0,n],E,\omega}\Vert, \quad a.e. \omega.$$ Let $\nu=\inf\limits_{E\in \sigma}\gamma(E)$. By the Furstenberg’s theorem $\nu>0.$ It follows from (\[A\]) that the desired exponential decay of the Green’s function can be achieved if all the $P_{[a,b]}$ in (\[A\]) behave as $e^{(b-a)\gamma(E)}$, thus leading to the study of deviations of $\ln P_{[a,b]} $ from its mean. In fact, the key estimates underlying the analysis of [@ckm] are precisely large deviation bounds for the Lyapunov exponent due to Le Page [@lepage]. Here we will use a corresponding statement for the matrix elements [@tsay1999some] \[ldt lemma\] For any $\epsilon>0$, there exists $\eta=\eta(\epsilon)>0$ such that, there exists $N_0=N_0(\epsilon)$, such that for every $ b-a>N_0$, and any $E$ in a compact set, $$\label{eqLDT} \mathbb{P}\left\{ \omega:\left\vert \frac{1}{b-a+1} \log\Vert P_{[a,b],E,\omega}\Vert-\gamma(E) \right\vert\geq\epsilon \right\} \leq e^{-\eta (b-a+1)}$$ It will also be convenient to use the general subharmonicity upper bound due to Craig-Simon [@cs] \[CS1\] For a.e. $\omega$ [for all]{} $E$, we have $$ \varlimsup_{n\to\infty} \frac{\log\Vert T_{[0,n],E,\omega}\Vert}{n+1}\leq\gamma(E)$$ Main lemmas {#alllemma} =========== Denote $$\label{B+} B_{[a,b],\epsilon}^{+} =\left\{(E,\omega): \|P_{[a,b],E,\omega}\|\geq e^{(\gamma(E)+\epsilon)(b-a+1)}\right\}$$ $$\label{B-} B_{[a,b],\epsilon}^{-} =\left\{(E,\omega): \|P_{[a,b],E,\omega}\|\leq e^{(\gamma(E)-\epsilon)(b-a+1)}\right\}$$ and denote $B_{[a,b],\epsilon,E}^{\pm}=\{\omega:(E,\omega)\in B_{[a,b],\epsilon}^{\pm}\}$, $B_{[a,b],\epsilon,\omega}^{\pm}=\{E:(E,\omega)\in B_{[a,b],\epsilon}^{\pm}\}$, $B_{[a,b],}=B_{[a,b],}^+\cup B_{[a,b],}^-$. Let $E_{j,(\omega_a,\cdots,\omega_b)}$ be eigenvalues of $H_{[a,b],\omega}$ with $\omega\|_{[a,b]}=(\omega_a,\cdots,\omega_b)$. Large deviation theorem gives us the estimate that for all $E, a, b,\epsilon$ $$\label{ldt} \mathbb{P}(B_{[a,b],\epsilon,E}^\pm)\leq e^{-\eta(b-a+1)}$$ Assume $\epsilon=\epsilon_0<\frac{1}{8}\nu$ is fixed for now, so we omit it from the notations until Lemma \[omega3\]. Let $\eta_0=\eta(\epsilon_0)$ be the corresponding parameter from Lemma \[ldt lemma\] \[lemma1\] For $n \geq 2$, if $x$ is $(\gamma(E)-8\epsilon_0,n,E,\omega)$-singular, then $$(E,\omega)\in B_{[x-n,x+n]}^-\cup B_{[x-n,x]}^+\cup B_{[x,x+n]}^+$$ Note that from , for all $E,x,n\geq 2$, $$\mathbb{P}(B_{[x-n,x+n],E}^-\cup B_{[x-n,x],E}^+\cup B_{[x,x+n],E}^+)\leq 3e^{-\eta_0 (n+1)}$$ Follows imediately from the definition of singularity and . Now we will use the following three lemmas to find the proper $\Omega_0$ for Theorem \[thm2\]. \[omega1\] Let $0<\delta_0<\eta_0$. For a.e. $\omega$ (we denote this set as $\Omega_1$), there exists $ N_1=N_1(\omega)$, such that for every $ n>N_1$, $$\max\{m(B_{[n+1,3n+1],\omega}^-),m(B_{[-n,n],\omega}^-)\}\leq e^{-(\eta_0-\delta_0)(2n+1)}$$ By , $$m\times\mathbb{P}(B_{[n+1,3n+1]}^-)\leq m(\sigma)e^{-\eta_0(2n+1)}$$ $$m\times\mathbb{P}(B_{[-n,n]}^-)\leq m(\sigma)e^{-\eta_0(2n+1)}$$ If we denote $$\Omega_{\delta_0,n,+}=\left\{\omega:m(B_{[n+1,3n+1],\omega}^-)\leq e^{-(\eta_0-\delta_0)(2n+1)}\right\}$$ $$\Omega_{\delta_0,n,-}=\left\{\omega:m(B_{[-n,n],\omega}^-)\leq e^{-(\eta_0-\delta_0)(2n+1)}\right\},$$ We have by Tchebyshev, $$\label{A5} \mathbb{P}(\Omega_{\delta_0,n,\pm}^c) \leq m(\sigma)e^{-\delta_0(2n+1)}.$$ By Borel-Cantelli lemma, we get for $a.e.~\omega$, $$\max\{m(B_{[n+1,3n+1],\omega}^-),m(B_{[-n,n],\omega}^-)\}\leq e^{-(\eta_0-\delta_0)(2n+1)},$$ for $n>N_1(\omega)$. \[N1\] Note that we can actually shift the operator and use center point $l$ instead of $0$. Then we will get $\Omega_1(l)$ instead of $\Omega_1$, $N_1(l,\omega)$ instead of $N_1(\omega)$. And if we pick $N_1(l,\omega)$ in the theorem as the smallest integer satisfying the conclusion, we can estimate when we will have $N_1(l,\omega)\leq \ln^2 \|l\|$, which is very useful in the proof for dynamical localization in section 6. The next results follows from : \[CS\] For a.e. $\omega$(we denote this set as $\Omega_2$), for all $E$, we have $$\label{C} \max\left\{\varlimsup_{n\to\infty} \frac{ \log\Vert T_{[-n,0],E,\omega}\Vert}{n+1}, \varlimsup_{n\to\infty} \frac{\log\Vert T_{[0,n],E,\omega}\Vert}{n+1}\right\}\leq\gamma(E)$$ $$\label{D} \max\left\{\varlimsup_{n\to\infty} \frac{\log\Vert T_{[n+1,2n+1],E,\omega}\Vert}{n+1}, \varlimsup_{n\to\infty} \frac{\log\Vert T_{[2n+1,3n+1],E,\omega}\Vert}{n+1}\right\}\leq\gamma(E)$$ is a direct reformulation of the result of [@cs], Theorem \[CS1\], while follows by exactly the same proof. \[omega2\] For every $ \omega\in\Omega_2$, for every $E$, there exists $N_2=N_2(\omega,E)$, such that for every $ n>N_2$, $$\begin{aligned} &\max\{\Vert T_{[-n,0],E,\omega}\Vert, \Vert T_{[0,n],E,\omega}\Vert\}<e^{(\gamma(E)+\epsilon)(n+1)}\\ &\max\{\Vert T_{[n+1,2n+1],E,\omega}\Vert, \Vert T_{[2n+1,3n+1],E,\omega}\Vert\}<e^{(\gamma(E)+\epsilon)(n+1)} \end{aligned}$$ \[omega3\] Let $\epsilon>0,K>1$, For a.e. $\omega$(we denote this set as $\Omega_3=\Omega_3(\epsilon,K)$), there exists $ N_3=N_3(\omega)$, so that for every $ n>N_3$, for every $ E_{j,(\omega_{n+1},\cdots,\omega_{3n+1})}$, for every $ y_1,y_2$ satisfying $-n\leq y_1\leq y_2\leq n$, ${\lvert-n-y_1\rvert}\geq\frac{n}{K}$, and ${\lvertn-y_2\rvert}\geq\frac{n}{K}$, we have $E_{j,(\omega_{n+1},\cdots,\omega_{3n+1})}\notin B_{[-n,y_1],\epsilon,\omega}\cup B_{[y_2,n],\epsilon,\omega}$. Note that $\epsilon$ and $K$ are not fixed yet, we’re going to determine them later in section \[pf\]. Let $\bar {\mathbb{P}}$ be the probability that there are some $y_1, y_2, j$ with $$E_{j,(\omega_{n+1},\cdots,\omega_{3n+1})}\in B_{[-n,y_1],\epsilon,\omega}\cup B_{[y_2,n],\epsilon,\omega}.$$ Note that for any fixed $\omega_c,\cdots,\omega_d $, with $[c,d]\cap[a,b]=\emptyset$, by independence, $$\mathbb{P}(B_{[a,b],\epsilon,E_{j,(\omega_c,\cdots,\omega_d)}})=\mathbb{P}_{[a,b]}(B_{[a,b],\epsilon,E_{j,(\omega_c,\cdots,\omega_d)}})\leq e^{-\eta_0(b-a+1)}$$ Applying to $[a,b]=[-n,y_1]$ or $[y_2,n]$, $[c,d]=[n+1,3n+1]$ and integrating over $\omega_{-n},\cdots,\omega_{y_1}$ or $\omega_{y_2},\cdots,\omega_{n}$, we get $$\mathbb{P}(B_{[-n,y_1],\epsilon,E_{j,(\omega_{n+1},\cdots,\omega_{3n+1})}}\cup B_{[y_2,n],\epsilon,E_{j,(\omega_{n+1},\cdots,\omega_{3n+1})}}) \leq 2e^{-\eta_0(\frac{n}{K}+1)},$$ so $$\label{B5} \bar{\mathbb{P}}\leq(2n+1)^3 2e^{-\eta_0(\frac{n}{K}+1)}$$ Thus by Borel-Cantelli, we get the result. \[N3\] Similar to remark \[N1\], we can get $\Omega_3(l)$, $N_3(l,\omega)$ for an operator shifted by $\ell$ instead, and get the result that for $a.e. \omega$ (we denote this set as $\Omega_{N_3}$), there exists $L_3(\omega)$, such that for any $\|l\|>L_3$, $N_3(l,\omega)\leq \ln^2 \|l\|.$ This will be of use in section 6 for proving dynamical localization. Proof of Theorem 2.2 {#pf} ==================== We will only provide a proof that $2n+1$ is $(c, n, E,\omega)$-regular, the argument for $2n$ being similar. Let $\epsilon$ be small enough such that $$\label{epsilon1} \epsilon<\min\{(\eta_0-\delta_0)/3,\nu\}.$$ Now let $$L:=e^{(\eta_0-\delta_0-\epsilon)}>1,$$ and note that since $S$ is bounded, by we have there exists $ M>0$, such that $$\|P_{[a,b],E,\omega}\|<M^{(b-a+1)},\quad \forall E\in\sigma,\omega$$ Pick $K$ big enough such that $$M^{\frac{1}{K}}<L$$ Let $\sigma>0$ be such that $$\label{K} M^{\frac{1}{K}}\leq L-\sigma<L$$ Let $\Omega_0=\Omega_1\cap\Omega_2\cap\Omega_3(\epsilon,K)$. Pick $\tilde{\omega}\in\Omega_0$, and take $\tilde{E}$ a $g.e.$ for $H_{\tilde{\omega}}$, with $\Psi$ the corresponding generalized eigenfunction. Without loss of generality assume $\Psi(0)\neq 0$. Then there exists $ N_4$, such that for every $ n>N_4$, 0 is $(\gamma(\tilde{E})-8\epsilon_0,n,\tilde{E},\tilde{\omega})$-singular. For $n>N_0=\max\{N_1(\tilde{\omega}),N_2(\tilde{\omega},\tilde{E}),N_3(\tilde{\omega}),N_4(\tilde{\omega},\tilde{E})\}$, assume $2n+1$ is $(\gamma(\tilde{E})-8\epsilon_0,n,\tilde{E},\tilde{\omega})$-singular. Then both $0$ and $2n+1$ is $(\gamma(\tilde{E})-8\epsilon_0,n,\tilde{E},\tilde{\omega})$-singular. So by Lemma \[lemma1\], $\tilde{E}\in B_{[n+1,3n+1],\epsilon_0,\tilde{\omega}}^-\cup B_{[n+1,2n+1],\epsilon_0,\tilde{\omega}}^+ \cup B_{[2n+1,3n+1],\epsilon_0,\tilde{\omega}}^+ $. By Corollary \[omega2\] and , $\tilde{E}\notin B_{[n+1,2n+1],\epsilon_0,\tilde{\omega}}^+\cup B_{[2n+1,3n+1],\epsilon_0,\tilde{\omega}}^+$, so it can only lie in $B_{[n+1,3n+1],\epsilon_0,\tilde{\omega}}^-$. Note that in , $P_{[n+1,3n+1],E,\tilde{\omega}}$ is a polynomial in $E$ that has $2n+1$ real zeros (eigenvalues of $H_{[n+1,3n+1],\tilde{\omega}}$), which are all in $B=B^{-}_{[n+1,3n+1],\epsilon,\tilde{\omega}}$. Thus $B$ consists of less than or equal to $2n+1$ intervals around the eigenvalues. $\tilde{E}$ should lie in one of them. By Lemma \[omega1\], $m(B)\leq Ce^{-(\eta_0-\delta_0)(2n+1)}$. So there is some e.v. $E_{j,[n+1,3n+1],\tilde{\omega}}$ of $H_{[n+1,3n+1],\omega}$ such that $$\vert\tilde{E}-E_{j,[n+1,3n+1],\tilde{\omega}}\vert\leq e^{-(\eta_0-\delta_0)(2n+1)}$$ By the same argument, there exists $ E_{i,[-n,n],\tilde{\omega}}$, such that $$\vert\tilde{E}-E_{i,[-n,n],\tilde{\omega}}\vert\leq e^{-(\eta_0-\delta_0)(2n+1)}$$ Thus $\vert E_{i,[-n,n],\tilde{\omega}}-E_{j,[n+1,3n+1],\tilde{\omega}}\vert\leq 2e^{-(\eta_0-\delta_0)(2n+1)}$. However, by Theorem \[omega3\], one has $E_{j,[n+1,3n+1],\tilde{\omega}}\notin B_{[-n,n],\epsilon,\tilde{\omega}}$, while $E_{i,[-n,n],\tilde{\omega}}\in B_{[-n,n],\epsilon,\tilde{\omega}}$ This will give us a contradiction below.\ Since $\vert E_{i,[-n,n],\tilde{\omega}}-E_{j,[n+1,3n+1],\tilde{\omega}}\vert\leq 2e^{-(\eta_0-\delta_0)(2n+1)}$ and $E_{i,[-n,n],\tilde{\omega}}$ is the e.v. of $H_{[-n,n],\tilde{\omega}}$, $$\left\Vert G_{[-n,n],E_{j,[n+1,3n+1],\tilde{\omega}},\tilde{\omega}}\right\Vert\geq \frac{1}{2}e^{(\eta_0-\delta_0)(2n+1)}$$ Thus there exist $ y_{1},y_{2}\in [-n,n]$ and such that $$\left\vert G_{[-n,n],E_{j,[n+1,3n+1],\tilde{\omega}},\tilde{\omega}}(y_{1},y_{2})\right\vert\geq \frac{1}{2n}e^{(\eta_0-\delta_0)(2n+1)}$$ Let $E_j=E_{j,[n+1,3n+1],\tilde{\omega}}$. We have $E_j\notin B_{[-n,n],\epsilon,\tilde{\omega}}$, thus $$\vert P_{[-n,n],\epsilon,E_j,\tilde{\omega}}\vert\geq e^{(\gamma(E_j)-\epsilon)(2n+1)}$$ so by , $$\label{last} \left\Vert P_{[-n,y_{1}],\epsilon,E_jï¼Œ\tilde{\omega}}P_{[y_{2},n],\epsilon,E_j,\tilde{\omega}}\right\Vert\geq\frac{1}{2n}e^{(\eta_0-\delta_0)(2n+1)}e^{(\gamma(E_j)-\epsilon)(2n+1)}$$ Then for the left hand side of , there are three cases: 1. both $\|-n-y_{1}\|>\frac{n}{K}$ and $\|n-y_{2}\|>\frac{n}{K}$ 2. one of them is large, say $\|-n-y_{1}\|>\frac{n}{K}$ while $\|n-y_{2}\|\leq\frac{n}{K}$ 3. both small. For $(1)$, $$\frac{1}{2n}e^{(\eta_0-\delta_0+\gamma(E_j)-\epsilon)(2n+1)}\leq e^{2n(\gamma(E_j)+\epsilon)}$$ Since by our choice , $\eta_0-\delta_0+\gamma(E_j)-\epsilon>\gamma(E_j)+\epsilon$, for $n$ large enough, we get a contradiction. For $(2)$, $$ \frac{1}{2n}e^{(\eta_0-\delta_0+\gamma(E_j)-\epsilon)(2n+1)} \leq e^{(\gamma(E_j)+\epsilon)(2n+1)}(M)^{\frac{n}{K}} $$ is in contradiction with and For $(3)$, with and $$ \frac{1}{2n}e^{(\eta_0-\delta_0+\gamma(E_j)-\epsilon)(2n+1)}\leq M^{\frac{2n}{K}}\leq (L-\sigma)^{2n}\leq(e^{(\eta_0-\delta_0+\gamma(E_j)-\epsilon)}-\sigma)^{2n}, $$ also a contradiction. Thus our assumption that $2n+1$ is not $(\gamma(\tilde{E})-8\epsilon_0,n,\tilde{E},\tilde{\omega})$-regular is false. Theorem \[thm2\] follows. Note that we have established the following more precise version of Theorem \[thm2\] \[thm22\] There exists $ \Omega_0$ with $\mathbb{P}(\Omega_0)=1$, such that for every $ \tilde{\omega}\in\Omega_0$, for any g.e. $\tilde{E}$ of $H_{\tilde{\omega}}$, and $\epsilon>0,$ there exists $ N=N(\tilde{E},\tilde{\omega}, \epsilon)$, such that for every $ n>N$, $2n,~2n+1$ are $(\gamma(E)-\epsilon,n,\tilde{E},\tilde{\omega})$-regular. It is a standard patching argument (e.g. proof of Theorem 3 in [@j]) that this implies $\|\Psi_E(n)\|\leq C_{E,\epsilon}e^{-(\gamma(E) -\epsilon)n}$ for any $\epsilon>0.$ Combined with Theorem \[CS1\], this immediately implies that we have Lyapunov behavior at every generalized eigenvalue. \[CS2\] For a.e. $\omega$ for all generalized eigenvalues $E$, we have $$\label{ld} \lim_{n\to\infty} \frac{\log\Vert T_{[0,n],E,\omega}\Vert}{n+1}=\gamma(E)$$ Uniform and Quantitative Craig-Simon {#uniformcs} ==================================== Craig-Simon theorem \[CS1\] implies that for a.e. $\omega$ and every $E\in\sigma$ there exists $N(\omega,E)$ such that for $n>N,$ $\Vert T_{[0,n],E,\omega}\Vert\leq e^{(n+1)(\gamma(E)+\epsilon)}.$ For the proof of dynamical localization one however needs a statement of this type with $N$ uniform in $E.$ Such a statement is the goal of this section. We will show that it holds for any ergodic dynamical system satisfying the uniform LDT (Large Deviation Type) condition: Lemma \[ldt lemma\]. Thus this result has more general nature than the rest of the paper and may be of independent interest. In particular, it is applicable to quasiperiodic dynamics with Diophantine frequencies and analytic sampling functions. We note that uniform LDT condition can also be replaced by a combination of a pointwise LDT condition and continuity of the Lyapunov exponent. We have: \[QCS\] Let the ergodic family $H_\omega$ satisfy Lemma \[ldt lemma\]. Fix $\epsilon_0>0$. For a.e. $\omega$ (we denote this set as $\Omega_2=\Omega_2(\epsilon_0)$), there exists $N_2(\omega)$, such that for any $n>N_2(\omega)$, $E\in \sigma $, $$\|P_{[0,n],E,\omega}\|\leq e^{(\gamma(E)+\epsilon_0)(n+1)}$$ An immediate corollary is Let $H_\omega,\epsilon_0$ be as above. Then there exists $\Omega_2$ with $\mathbb{P}(\Omega_2)=1$, such that for $\omega\in\Omega_2$, there exists $N_2(\omega)$ such that $$\max\left\{\|P_{[0,n],E,\omega}\|,\|P_{[-n,0],E,\omega}\|,\|P_{[n+1,2n+1],E,\omega}\|,\|P_{[2n+1,3n+1],E,\omega}\|\right\}\leq e^{(\gamma(E)+3\epsilon_0)(n+1)}.$$ Thus we can replace Corollary \[omega2\] with this uniform version. We start with the following \[elementary\] Let $Q(x)$ be a polynomial of degree $n-1$. Let $x_i=\cos{\frac{2\pi(i+\theta)}{n}}$, $0<\theta<1/2$, $i=1,2,\cdots,n$. If $Q(x_i)\leq a^n$, for all $i$, then $Q(x)\leq Cna^n$, for all $x\in[-1,1]$, where $C=C(\theta)$ is a constant. By Lagrange interpolation, we have $$Q(x)=\sum_{i=1}^{n}Q(x_i)\prod_{j\neq i}\frac{x-x_j}{x_i-x_j}$$ Note that $$\sum_{j\neq i}\ln\|x_i-x_j\|=\sum_{j\neq i}\left\{\ln\left\|\sin\frac{\pi(i+j+2\theta)}{n}\right\|+\ln\left\|\sin\frac{\pi(i-j)}{n}\right\|+\ln 2\right\}=: A+B+(n-1)\ln 2.$$ We will use the following lemma without giving a proof. \[9.6\] Let $p$ and $q$ be relatively prime. Let $1\leq k_0\leq q$ be such that $$\vert \sin 2\pi(x+k_0p/(2q))\vert =\min _{1\leq k\leq q}\vert \sin 2\pi(x+kp/(2q))\vert.$$ Then $$\ln q+\ln (2/\pi)<\sum_{\substack{k=1\\ k\neq k_0}}^{q}\ln\vert\sin 2\pi(x+kp/(2q))\vert+(q-1)\ln 2\leq \ln q.$$ For $B$, we take $p=1,~q=n,~x=-i/(2n),~k=j$. Then $k_0=i$, and we get $$B\geq \ln n+\ln (2/\pi)-(n-1)\ln 2.$$ For $A$, we estimate by Lemma \[9.6\] with $p=1,~q=n,~x=(i+2\theta)/2n,~k=j$. If $k_0=j_0$ is the minimum term of $\ln\|\sin\frac{\pi(i+j+2\theta)}{n}\|$, then $$A\geq \ln n+\ln (2/\pi) -(n-1)\ln 2-\ln\left\|\sin\frac{\pi(2i+2\theta)}{n}\right\|+\ln\left\|\sin\frac{\pi(i+j_0+2\theta)}{n}\right\|$$ For $0<\theta<1/4$, we have $$\frac{\|\sin\frac{\pi(2i+2\theta)}{n}\|}{\|\sin\frac{\pi(i+j_0+2\theta)}{n}\|} = \frac{\|\sin\frac{\pi(2i+2\theta)}{n}\|}{\|\sin\frac{\pi\cdot 2\theta}{n}\|}\leq \frac{1}{\|\sin\frac{\pi\cdot 2\theta}{n}\|}= O(n)$$ Thus $$\sum_{j\neq i}\ln\|x_i-x_j\|\geq -(n-1)\ln 2+\ln n+C$$ Writing $x=\cos\frac{2\pi a}{n}$, by Lemma \[9.6\], we get $$\sum_{j\neq i}\ln \|x-x_j\|\leq -(n-1)\ln 2+ 2\ln n+ C$$ Thus $$\prod_{j\neq i}\frac{x-x_j}{x_i-x_j}\leq Cn$$ and we have $$Q(x)\leq Cna^n$$ Now we can finish the proof of Theorem \[QCS\]. We know that $\sigma$ is compact, so contained in some bounded closed interval. Assume we are dealing with $[a,a+A]$. Unifrom LDT implies that $\gamma$ is a continuous function of $E$ [@DK]. Since $\gamma(E)$ is uniformly continuous, for any $\epsilon_0$, there exists $\delta_0$ such that $$\label{holder} \|\gamma(E_x)-\gamma(E_y)\|\leq \epsilon_0,\quad if~ \|E_x-E_y\|\leq \delta_0.$$ Divide the interval $[a,a+A]$ into length-$\delta_0$ sub-intervals. There are $K=[A/\delta_0]+1$ of them (the last one may be shorter). Denote them as $I_k$, for $k=1,\cdots, K$. For $I_k=[E_{k,n},E_{k+1,n}]$, let $E_{k1,n},\cdots,E_{kn,n}$ be distributed as in Lemma 5.3. Namely, set $ E_{ki,n}=E_{k,n}+(x_i+1)\delta_0/2$, where $x_i$ are as in Lemma 5.3, $0<\theta<1/2$. Note that for any $E_x$, $E_y\in [E_{k1,n},E_{kn,n}]$, $\|\gamma(E_x)-\gamma(E_y)\|\leq \epsilon_0$. Since by the uniform-LDT condition $$\mathbb{P}\left(\left\{\omega: \exists i=1,\cdots,n,~s.t.~\|P_{[0,n],E_{ki,n},\omega}\|\geq e^{(\gamma(E_{ki,n})+\epsilon_0)(n+1)} \right\}\right)\leq ne^{-\eta_0(n+1)},$$ by Borel-Cantelli, for a.e. $\omega$, (we denote this set as $\Omega(k)$), there exists $N(k, \omega)$, such that for all $n>N(k,\omega)$, $$\|P_{[0,n],E_{ki,n},\omega}\|\leq e^{(\gamma(E_{ki,n})+\epsilon_0)(n+1)},\quad \forall i=1,\cdots,n.$$ If we denote $\gamma_{k,n}=\inf_{E\in [E_{k1,n},E_{kn,n}]}{\gamma(E)}$, then by $$\|P_{[0,n],E_{ki,n},\omega}\|\leq e^{(\gamma(E_{ki,n})+\epsilon_0)(n+1)}\leq e^{(\gamma_{k,n}+2\epsilon_0)(n+1)}, \quad \forall i=1,\cdots,n.$$ Let $M$ be big enough such that, for any $n>M$, $n^c\leq e^{\epsilon_0(n+1)}$. Thus by Lemma \[elementary\], applied to $Q(x)=P(E_{k,n}+\frac{(x+1)\delta_0}{2})$, for $E\in[E_{k,n},E_{k+1,n}]$, $n>\max\{N(k,\omega),M\}$, $$\|P_{[0,n],E,\omega}\|\leq n^ce^{(\gamma_{k,n}+2\epsilon_0)(n+1)}\leq n^ce^{(\gamma(E)+2\epsilon_0)(n+1)}\leq e^{(\gamma(E)+3\epsilon_0)(n+1)}$$ Let $\Omega_2=\bigcap\limits_k \Omega(k)$, $\tilde{N}(\omega)=\max_k\{N(k,\omega),M\}$. Then for any $n>\tilde{N}(\omega)$, $$\|P_{[0,n],E,\omega}\|\leq e^{(\gamma(E)+3\epsilon_0)(n+1)},\quad \forall E\in [a,a+A]$$ This allows us to also obtain a quantitative version of Theorem \[QCS\]. Assume the $N_2(\omega)$ in Theorem \[QCS\] is chosen to be the smallest satisfying the condition. Let $l\in\mathbb{Z}$, $N_2(l,\omega)=N_2(T^l\omega)$. Let $\bar{\Omega}_2=\bigcap_{l\in\mathbb{Z}}T^l\Omega_2$. \[5.5\] For a.e. $\omega$ (we denote this set as $\tilde{\Omega}_2$), there exists $L_2=L_2(\omega)$, such that for all $\vert l\vert>L_2$, $N_2(l,\omega)\leq \ln^2\vert l\vert$. In particular, if $n>\ln^2 \vert l\vert$, then $$\vert P_{[l,l+n],E,\omega}\vert\leq e^{(\gamma(E)+\epsilon_0)(n+1)},~for~all~E\in\sigma$$ Let $\omega\in\bar{\Omega}_2$, $l\in\mathbb{Z}$, $k\in\mathbb{N}$. By Theorem \[QCS\], $\bar{\Omega}$ has full measure. We have $$\begin{aligned} \mathbb{P}\{\omega:N_2(l,\omega)\geq k\} & \leq &\sum_{n=k}^{\infty}\mathbb{P}\{\omega:N_2(l,\omega)=n\}\leq \sum_{n=k}^{\infty}\mathbb{P}(B^+_{[l,l+n-1],E})\\ & \leq &\sum_{n=k}^{\infty}Ce^{-(\gamma(E)+\epsilon_0)n}\leq Ce^{-(\gamma(E)+\epsilon_0)k}\end{aligned}$$ Thus $$\mathbb{P}\{\omega:N_2(l,\omega)\geq \ln^2\vert l\vert\}\leq Ce^{-(\gamma(E)+\epsilon_0)(\ln^2\vert l\vert)}$$ By Borel-Cantelli lemma, we get the result and the corresponding $\tilde{\Omega}_2$. Dynamical Localization {#dyn} ====================== Now we have established the spectral localization for 1-d Anderson model. With some more effort, we can get the dynamical localization. We say that $H_\omega$ exhibits dynamical localization if for $a.e.~\omega$, for any $\epsilon>0$, there exists $\alpha=\alpha(\omega)>0$, $C=C(\epsilon,\omega)$, such that for all $x,y\in\mathbb{Z}$: $$\sup_t \|\langle\delta_x,e^{-itH_{\omega}}\delta_y\rangle\|\leq C_\epsilon e^{\epsilon\|y\|}e^{-\alpha\|x-y\|}$$ According to [@del1996operators], we only need to prove that for a.e. $\omega$, $H_\omega$ has SULE (Semi-Uniformly Localized Eigenfunction). We say $H$ has SULE if $H$ has a complete set $\{\varphi_E\}$ of orthonormal eigenfunctions, such that there is $\alpha>0$, and for each $\epsilon>0$, a $C_\epsilon$ such that for any eigenvalue $E$, there exists $l=l_E\in\mathbb{Z}$, such that $$\|\varphi_E(x)\|\leq C_\epsilon e^{\epsilon\|l_E\|}e^{-\alpha\|x-l_E\|},\quad x\in\mathbb{Z}$$ In fact, we will prove that $\vert \varphi_E(x)\vert\leq C_\epsilon e^{C\ln^2(1+\vert l_E\vert)}e^{-\alpha\vert x-l_E\vert}$, see , . In order to do this, we need to modify Lemma \[omega1\], Lemma \[omega3\] using the same method as in Lemma \[5.5\]. Assume the $N_i(\omega),~i=1,3$ in Lemmas \[omega1\], \[omega3\] are chosen to be the smallest parameters satisfying the condition. Let $l\in\mathbb{Z}$, $N_i(l,\omega)=N_i(T^l\omega)$. Let $\bar{\Omega}_i=\bigcap_{l\in\mathbb{Z}}T^l\Omega_i$, $i=1,3$. \[lem\] For a.e. $\omega$ (we denote this set as $\tilde{\Omega}_{1,3}$), there are $L_1(\omega),L_3(\omega)$ such that for any $\vert l\vert>\max\{L_1,L_3\}$, $$\max\{N_1(l,\omega),N_3(l,\omega)\}\leq\ln^2\vert l\vert$$ Let $\omega\in\bar{\Omega}_1$, $l\in\mathbb{Z}$, $k\in \mathbb{N}$, then by $$\mathbb{P}\{\omega:N_1(l,\omega)>k\}\leq \sum_{n= k}^\infty\mathbb{P}(\Omega_{\delta,n,\pm})\leq\sum_{n=k}^\infty 2m(\sigma)e^{-\delta_0(2n+1)}\leq Ce^{-\delta_0(2k+1)}$$ Thus $$\mathbb{P}\{\omega:N_1(l,\omega)>\ln^2 \vert l\vert\}\leq Ce^{-\delta_0(2\ln^2\vert l\vert+2)}$$ By Borel-Cantelli lemma, we can get the result. The same argument works for $N_3$. Then we rebuild the criteria for regularity around a singular point $l$. \[last cor\] For a.e. $\omega$ (we denote this set as $\tilde{\Omega}$), for any $l$, there exists $N(l,\omega)$, such that for any $n>N(l,\omega)$ and for all $E\in \sigma$ either $l$ or $l+2n+1$, and either $l$ or $l-2n-1$ are $(\gamma(E)-8\epsilon_0,n,E,\omega)$-regular. In section 4, we proved that either $0$ or $2n+1$ is $(\gamma(E)-8\epsilon_0,n,E,\omega)$-singular for all $n>N(\omega)$, with $N(\omega)=\max\{N_1(\omega),N_2(\omega),N_3(\omega)\}$. Here we set $N(l,\omega)=\max\{N(T^l\omega),N(T^{-l}\omega)\}$, and modify $\tilde{\Omega}$ accordingly. Now, take $\tilde{\Omega}=\tilde{\Omega}_2\cup\tilde{\Omega}_{1,3}$ and fix $\omega\in\tilde{\Omega}$. We omit $\omega$ from notations from now on. By Lemma \[lem\] and Lemma \[5.5\], there exist $L_1$, $L_2$, $L_3$ such that for all $\|l\|>\max\{L_1,L_2,L_3\}$, $$N_i(l)\leq \ln^2 \|l\|,\quad \forall i=1,2,3$$ for all $E\in\sigma$. Let $l_E$ be a position of the maximum point of $\varphi_E$. Take $L_4$ with $\ln^2L_4\geq[\frac{\ln 2}{\gamma(E)-8\epsilon_0}]+1$. For any $n\geq \ln^2L_4$, and any e.v. $E$, $l_E$ is naturally $(\mu-8\epsilon_0,n,E)$-singular by . Let $L=\max\{L_1, L_2, L_3, L_4\}$, $N(l):=\max\{N_1(l),N_2(l),N_3(l),\frac{\ln 2}{\gamma(E)-8\epsilon_0}\}$. Then for any $\|l\|>L$, $$\label{Nl} N(l)\leq \ln^2 \|l\|$$ If $\|l_E\|>L$, then for any $n\geq N(l_E)$, $l_E$ is $(\gamma(E)-8\epsilon_0,n,E)$-singular, so $x=l_E\pm(2n+1)$ is $(\gamma(E)-8\epsilon_0,n,E)$-regular. By , for any $\|x-l_E\|\geq N(l_E)$ $$\label{n>} \|\varphi_E(x)\|\leq 2e^{-(\gamma(E)-8\epsilon_0)\|x-l_E\|}$$ Since $\varphi_E$ is normalized, in fact for all $x$, $$\|\varphi_E(x)\|\leq 2e^{(\gamma(E)-8\epsilon_0)N(l_E)}e^{-(\gamma(E)-8\epsilon_0)\|x-l_E\|}$$ By , for any $\epsilon$, $$\label{1} \|\varphi_E(x)\|\leq 2e^{(\gamma(E)-8\epsilon_0)\ln^2 (1+\|l_E\|)} e^{-(\gamma(E)-8\epsilon_0)\|x-l_E\|}$$ If $\|l_E\|\leq L$, for any $\epsilon$, for $n\geq N(l_E)$, we use the same argument as and get $$\label{3} \|\varphi_E(x)\|\leq 2e^{-(\gamma(E) -8\epsilon_0)\|x-l_E\|}\leq 2e^{\epsilon \ln^2 (1+\|l_E\|)}e^{-(\gamma(E)-8\epsilon_0)\|x-l_E\|}$$ While for $n\leq N_{l_E}$, set $M_{2\epsilon}=\min_{k\in[-L,L],~\|x-k\|<N(k)} \{e^{\epsilon \ln^2(1+\|k\|)} e^{-(\gamma(E)-8\epsilon_0)\|x-k\|}\}$ and $C_{2\epsilon}=M_{2\epsilon}^{-1}$. Then for all $\|x-l_E\|< N(l_E)$, $$\label{2} \|\varphi_E(x)\|\leq 1\leq C_{2\epsilon} e^{\epsilon \ln^2 (1+\|l_E\|)} e^{-(\gamma(E)-8\epsilon_0)\|x-l_E\|}$$ Thus for $C_\epsilon=\max\{2,C_{2\epsilon}\}$, and provide SULE. Acknowledgments {#acknowledgments .unnumbered} =============== This research was partially supported by the NSF DMS-1401204. X. Z. is grateful to Wencai Liu for inspiring thoughts and comments for Sec. 5. We also thank Barry Simon for his encouragement.	{ "pile_set_name": "ArXiv" }
--- abstract: \| In this paper we conduct a data survey searching for well-defined streamer wave events observed by the Large Angle and Spectrometric Coronagraph (LASCO) on-board the Solar and Heliospheric Observatory (SOHO) throughout Solar Cycle 23. As a result, 8 candidate events are found and presented here. We compare different events and find that in most of them the driving CMEs ejecta are characterized by a high speed and a wide angular span, and the CME-streamer interactions occur generally along the flank of the streamer structure at an altitude no higher than the bottom of the field of view of LASCO C2. In addition, all front-side CMEs have accompanying flares. These common observational features shed light on the excitation conditions of streamer wave events. We also conduct a further analysis on one specific streamer wave event on 5 June 2003. The heliocentric distances of 4 wave troughs/crests at various exposure times are determined; they are then used to deduce the wave properties like period, wavelength, and phase speeds. It is found that both the period and wavelength increase gradually with the wave propagation along the streamer plasma sheet, and the phase speed of the preceding wave is generally faster than that of the trailing ones. The associated coronal seismological study yields the radial profiles of the Alfvén speed and magnetic field strength in the region surrounding the streamer plasma sheet. Both quantities show a general declining trend with time. This is interpreted as an observational manifestation of the recovering process of the CME-disturbed corona. It is also found that the Alfvénic critical point is at about 10 R$_\odot$ where the flow speed, which equals the Alfvén speed, is $\sim$ 200 km s$^{-1}$. author: - 'S.W. $^{1,2,3}$, Y. $^{3}$, B. $^{3}$, H.Q. $^{3}$, X.L. $^{3}$, L.D. $^{3}$, X.S. $^{1}$' title: Streamer Wave Events Observed in Solar Cycle 23 --- Introduction {#S-Introduction} ============ Streamer waves can be excited by the interaction of a rapidly moving and expanding CME ejecta with a nearby streamer structure, representing one of the largest wave phenomena ever observed in the solar corona (Chen et al., 2010, Paper [slowromancap1@]{}). We interpret the waves as the fast kink body mode (Edwin and Roberts, 1982) carried by and propagating outwards along the streamer plasma sheet structure in the wake of the CME-caused streamer deflections (e.g., Hundhausen, Holzer, and Low, 1987; Sheeley, Hakala, and Wang, 2000; Tripathi and Raouafi, 2007; Filippov and Srivastava, 2010). Using the LASCO coronagraph data, Paper [slowromancap1@]{} obtains the wave properties like the period, wavelength, and propagation phase speed for the streamer wave event on 6 July 2004. In the follow-up study by Chen et al. (2011, Paper [slowromancap2@]{}), a coronal seismological method was developed to diagnose the values of the Alfvén speed and magnetic field strength in the region surrounding the plasma sheet structure, with the assumption that the streamer wave is the fast kink mode propagating along the plasma sheet. Interesting results concerning the temporal evolution of the physical conditions of the CME-disturbed corona are found. However, only one specific streamer wave event which started on 6 July 2004 was investigated in these studies. This gives rise to natural questions asking about how many other similar events exist and, if they do, how the waves get excited and how their properties compare with the well-studied 2004 event. To address these questions, a complete data survey of LASCO observations during Solar Cycle 23 was conducted. It was found that in only about 8 cases there exist well-defined wavy motions along the streamer stalk among innumerable CME-streamer interaction events. Therefore, it is apparent that the generation of streamer waves requires certain strict excitation conditions to be satisfied. The conditions may relate to the large-scale coronal magnetic field topology, CME dynamics and morphological evolution, and physical details of the CME impact on the streamer structure. One main purpose of this paper is to provide more clues on the wave excitation conditions by studying the 8 candidate streamer wave events observed in Solar Cycle 23 and collecting common observational features of these events. The paper is organized as follows. In Section 2, we first present observations of the 8 events and discuss the physical factors that may play a role in the wave formation. In Section 3, we conduct further detailed investigations on a specific event on 5 June 2003 to deduce the evolution of the wavelength, period, and propagation phase speed, we also present a coronal seismological study to diagnose the distribution of the Alfvén speed and magnetic field strength in the plasma sheet region. A summary is provided in the final section of this paper. Candidate streamer wave events observed in Solar Cycle 23 {#S-general} ========================================================= CDAW (Coordinated Data Analysis Workshops) data center provides a detailed catalog of the CME events observed by LASCO (Gopalswamy et al., 2009), including the associated white light and running difference images (RDIs) as well as some deduced physical parameters. Our data survey focusing on the streamer wave events was conducted taking advantage of the information gathered by the data center. All CME events in the CDAW data base during Solar Cycle 23 were viewed to find those in which the streamer stalk presented snakelike wavy motions after being hit by a nearby CME. The RDIs were further examined to confirm the presence of one to a few pairs of bright-dark and dark-bright (BD-DB) patches along the streamer stalk, which could be a manifestation of the wavy motions. Eight candidate streamer wave events, including the event on 6 July 2004, studied previously, were found. In the following text, the events will be named after the date on which the associated CME is first observed. For example, the just-mentioned event will be called the 20040706 event. Some relevant physical parameters of the CMEs are presented in the first to sixth columns of Table 1, including the appearance time (UT) of the CME ejecta in the C2 field of view (FOV), the CME type, the central position angle (CPA), the linear speed, the class of the accompanying flares according to the X-ray flux recorded by the GOES satellite. The last column of Table 1 gives the CPA of the associated streamer structure measured at 5 R$_\odot$. In 3 events, it is difficult to rule out the possibility of the wavy motions being humps pushed by the CME trailing material or magnetic structures adjacent to the streamer. That is to say in these events it is possible that both processes, including the action of the magnetic restoring force as a result of the streamer deflection and the direct interaction with nearby structures trailing the CME, may play a role in forming the streamer wavy motions. The 3 events have been indicated by the symbol “$\ast$” after the observation dates given in the first column. From the table, we see that most events take place in 2003 and 2004, years of high level of solar activity, and few events occur in the rising or declining phase of solar activity. This is easy to understand since statistically CMEs and streamers interact much more frequently and energetically during times of higher than lower solar activity. We also find that there are 6 halo CMEs, in all except one event the apparent angular widths exceed 100$^{\circ}$, and the average width is about 290$^{\circ}$. The lowest linear speed is 964 km s$^{-1}$ for the 20030527 event and for all the other events it is larger than 1000 km s$^{-1}$ with the largest speed being 2861 km s$^{-1}$ for the 20050115 event, and the average speed being 1580 km s$^{-1}$. In addition, accompanying flares are observed for the four front-side events. The flare class varies from C8.8 to X3.6. The other four events, indicated by the symbol “$\setminus$”, are mostly back-side events according to EIT observations (Delaboudinière et al., 1995). Considering that all front-side events have accompanying flares, it is highly possible that this may also be the case for the back-side CMEs. In the following section, we will continue to describe some of these observational features and discuss their relevance to the excitation conditions of streamer waves. Individual events ----------------- Among the candidate streamer wave events listed in Table 1, events 20040705 and 20040706 have been already reported with the latter carefully studied in Paper [slowromancap1@]{} and [slowromancap2@]{}. In the following text, we will first present a brief introduction to these two events, and then describe relevant observations of the other events in the order of their occurrence. ### Events 20040706 and 20040705 According to Table 1, the 20040706 event results from the interaction between a fast brightness-asymmetric (BA) halo CME with a linear speed of 1307 km s$^{-1}$ and a streamer structure with a CPA of 225$^\circ$ as measured at 5 R$_\odot$. The interaction starts at about 20:30 UT as recorded by C2, causing an obvious deflection of the streamer structure away from its original position. The deflection is seen at the bottom of the C2 FOV indicating that the interaction takes place at an even lower height. At 20:58 UT, the CME ejecta already left the C2 FOV, and the streamer starts to move backwards. In Paper [slowromancap1@]{}, it was suggested that the wavy motion is controlled by the magnetic restoring force given by the deflection of the streamer structure. The wavy motion is indicated by the BD-DB pairs observed in the RDIs. The presence of these difference structures facilitates greatly the extraction of the wave profiles and further quantitative measurements of the heliocentric distances of selected wave phases, which are given by subsequent wave crests and troughs observable in the event. The distances of 5 phases, marked as P1 to P5, are measured for event 20040706 in Paper [slowromancap1@]{}. In other words, about two wavelengths of the streamer wave are observable in the event, with the distance between P1 to P3 representing the first wavelength and P3 to P5 the second one. The wavelength is about 2 - 3 R$_\odot$, and the period is about 1 hour. It is interesting to find that both parameters increase with the outward propagation of the wave. More quantitative results and discussions will be presented in Section 3 together with that for event 20030605, to facilitate the comparison between events. Studies in Paper [slowromancap1@]{} also examined the magnetic field topology given by the extrapolation of the measured photospheric magnetic field with the Potential Field Source Surface (PFSS) model (Schatten, Wilcox, and Ness, 1969; Schrijver and De Rosa, 2003). It was concluded that the CME source lies at the flank of the closed loop systems corresponding to the streamer. Such geometry makes it possible that the expanding CME ejecta hits the streamer from the flank, which has been considered as a physical condition favoring the formation of streamer wave. The other event 20040705 results from the interaction between the same streamer structure and a previous fast halo CME. The CME, taking place, 23 hours before, possibly originates from the same source region and has a linear speed of 1444 km s$^{-1}$. The RDIs of this event have been shown in Paper [slowromancap1@]{} and will not be presented here. The morphological evolution of the two events look similar. However, from the RDIs we find that there exist loop-like trailing eruptive structures moving together with the streamer kinks. This makes it difficult for us to rule out the possibility that the wavy motions are streamer humps pushed by the nearby CME structures. In the following, we shall discuss relevant observations of the other events listed in Table 1. For each event, we will pick out several white light images obtained by C2 to illustrate the CME-streamer interaction process and the morphology of the streamer wavy motion. For some events, we will also show several RDIs for better visualizations of the motion. ### Event 20010420 The CME ejecta appears in the C2 FOV at 10:06 UT with a CPA of 65$^\circ$, an angular width of 127$^{\circ}$, and a linear speed of 1160 km s$^{-1}$. The CME hits a nearby streamer structure with a CPA of about 15$^{\circ}$. From the first two white light images of Figure 1, we observe the deviation of the streamer structure away from its previous position. This results in the left (east)-dark and right (west)-bright feature appearing at the bottom of the first RDI. At 11:30 UT, we see that the CME front has left the C2 FOV; till 13:31 UT, the streamer stalk presents a wave-like motion. The wavy motion can be easily related to the BD-DB pairs in the RDIs of the corresponding intervals. It should be noted that trailing material continues to flow outwards even after the CME left, as seen from the corresponding white light and difference images. The motion of this material produces the complex difference structures below the CME ejecta. Some of them are rather close to the streamer structure, and thus may get mixed up with the bright-dark pairs associated with the streamer wavy motion, which makes our judgement of the driving mechanisms of the wavy motions nontrivial. However, with a careful examination of the white light images, we prefer the explanation that the motions are driven by the streamer inherent magnetic restoring force, as an aftermath of the streamer deflection caused by the CME. Examining the RDIs, we find that a left-bright and right-dark feature start to emerge from the bottom of the C2 FOV between 10:54 and 11:06 UT. This color pattern is distinct from what emerged earlier in association with the CME-streamer deflection, indicating that the streamer starts to swing backwards (eastwards), probably under the action of the magnetic restoring force. From the 11:06 - 11:54 UT RDIs shown in Figures 1k and 1l, we see that the streamer starts to move westwards again. In the meantime, the left-bright and right-dark feature observed in Figure 1j already moves upwards to the location marked by the plus sign, which gives the first wave crest of the streamer wave. During the interval of 12:06 - 12:30 UT, we observe again the eastward motion of the streamer structure. Therefore, the streamer structure completes one period of the wavy motion from 11:06 UT to 12:30 UT. The period is thus estimated to be about 1 - 1.5 hours. In the last three RDIs, the locations of the following wave trough and crest are marked by the stars and squares. It is roughly estimated that the wavelength of the streamer wave is about 2 - 3 R$_\odot$ and the phase speed is about 300 - 400 km s$^{-1}$. ### Event 20030527 At 23:50 UT on May 27, LASCO recorded a halo CME with a linear speed of 964 km s$^{-1}$ accompanied by an X1.5 solar flare. At 00:06 UT on the following day, another BA halo CME with a linear speed of 1366 km s$^{-1}$ accompanied by an X3.6 flare is observed. According to the corresponding EIT data, we find that both CMEs probably originate from the same active region. During the eruptions, both CMEs trigger wavy motions of a nearby streamer with a CPA being about 185$^{\circ}$, as shown in Figure 2. The most interesting observational feature seen from this figure is that the streamer exhibits a waving and tangling morphology from 00:50 to 01:50 UT. Possible reasons accounting for the tangling morphology can be deduced by examining the Wilcox Solar Observatory Source Surface Synoptic Charts (http://wso.stanford.edu/synsourcel.html) for CR2003. It is found that the current sheet structure relevant to the streamer in question lies in between 50$^{\circ}$ - 60$^{\circ}$ southern latitude and extends over a rather wide longitudinal range from 200$^{\circ}$ - 300$^{\circ}$. Thus, it is suggested that distinctive wavy motions can be supported by different longitudinal parts of the plasma sheet, whose projections onto the sky plane form the observed tangling morphology. During the event, the CME structures associated with both CMEs flow outwards persistently, and we are unable to tell the exact mechanism driving the wavy motions. ### Event 20030605 From Table 1, we see that the CME appears in the C2 FOV at 20:06 UT with a CPA of 230$^{\circ}$, an angular width of 310$^{\circ}$, and a linear speed of 1458 km s$^{-1}$. At 20:30 UT, the CME hits a streamer structure with a CPA being about 316$^{\circ}$ from the bottom of the C2 FOV, as seen from Figure 3a. This results in an obvious leftward deflection, and a subsequent rightward motion of the streamer structure. It can be seen that the CME front leaves the C2 FOV very rapidly within about 1.5 hours after its first appearance. The majority of the CME ejecta is present on the right side of the streamer. On the other side, there exist structures with relatively weak brightness moving outwards along with the CME front. The material trailing these weak frontal structures are looking faint from the white light observations on the left side of the streamer. We therefore take this event as a streamer wave event suggesting that the wavy motion is supported by the inherent magnetic restoring force of the deflected structure. The morphology of the wavy motion can be observed from the last two images taken at 21:30 UT and 21:54 UT, when most part of the CME already leaves the C2 FOV. It is found that the event is simple and clear with more than three wave phases observable. This allows us to conduct a further investigation on the wave properties, which will be done as we proceed. ### Event 20031118 The CME driving this event appears in the C2 FOV at 8:50 UT with an accompanying M3.9 flare and a linear speed of 1660 km s$^{-1}$, whose image observed at 9:06 UT is given in Figure 4a. At 10:26 UT, another CME with a CPA of 95$^{\circ}$ and a linear speed of 1824 km s$^{-1}$ is observed by C2 and shown in Figure 4b. The ejecta of the first CME interacts with a streamer structure with a CPA of about 319$^{\circ}$. From Figure 4a, we see that the interaction starts from the bottom of the C2 FOV or probably at an even lower height, which results in a leftward deflection of the streamer structure. The RDIs in the streamer region rotated by 41$^{\circ}$ counterclockwise are shown in Figures 4c - 4f. From Figure 4c, it is easy to see that the CME-caused deflection, corresponding to the left-bright and right-dark feature in the upper part of the figure, is followed by a rightward motion as indicated by the right-bright and left-dark feature in the lower part, while no obvious trailing material of the CME is present on the left side of the streamer. Therefore, it is suggested that the above rightward motion is supported by the inherent magnetic restoring force of the streamer, and thus gives another example of streamer wave event. According to the timing and the enhanced deflection of the streamer structure as seen from the second right-bright and left-dark feature at the bottom of Figures 4e and 4f, we suggest that this feature represents the streamer deflection caused by the second CME. It is found that the associated ejecta has no direct contact with the streamer structure. So it is possible that the impact of the CME on the streamer is achieved by expelling the surrounding coronal magnetic field and/or propagating disturbances. Examining the data from Wind/WAVES (Bougeret, et al., 1995), we find that there exists a type [slowromancap2@]{} radio burst at 10:10 UT, corresponding temporally to the presence of the lower right-bright and left-dark feature of Figure 4e. Therefore, it is possible that the feature is a result of the streamer deflection caused by the associated CME shock (see, e.g., Sheeley, Hakala, and Wang, 2000). The subsequent observation indicates that this deflection does not grow into a streamer wave, yet terminates the development of the wave event caused by the previous eruption. ### Event 20050115 At 23:06 UT, a halo CME appears in the C2 FOV with a linear speed as large as 2860 km s$^{-1}$ and an accompanying X2.6 solar flare. According to Figures 5a and 5b, we see that the eruption results in a global perturbation with all the streamers observed in the plane of the sky deflected. Among the streamers, the one with a CPA of 240$^{\circ}$ at 5 R$_\odot$ presents the largest amplitude of deflection, along which the concerned wavy motion develops. In Figures 5c - 5f, we show the corresponding RDIs in the streamer region within the range of 4 R$_\odot$ to 12 R$_\odot$ taken from the C3 observations. The right-bright and left-dark feature in the upper part, i.e., father away from the Sun, of Figure 5c is associated with the rightward deflection of the streamer structure in response to the CME impact. The feature with different brightness distribution in the lower part of this figure is believed to be caused by the inherent magnetic restoring force of the streamer since no obvious trailing structures are observed on the right side of the streamer. In the last three panels of RDIs, we observe the outward propagation of the bouncing motion. At the lower part of Figures 5e and 5f we can discern the presence of another rightward wavy motion of the streamer structure. Thus, only one pair of the BD-DB feature is observable in this event. ### Event 20061106 The CME driving this wave event is present in the C2 FOV at 17:54 UT with a CPA of about 80$^{\circ}$, an angular width of about 80$^{\circ}$, and a linear speed of 1994 km s$^{-1}$. The streamer with a CPA of about 128$^{\circ}$ is deflected by the ejecta, as seen from Figures 6a and 6b. From Figures 6b and 6d, we see that the CME front already leaves the C2 FOV at 18:54 UT. The streamer starts to wave rightwards from the CME deflection producing the left-dark and right-bright feature in the streamer region at the bottom of Figure 6d. From Figures 6e and 6f, we see that the feature continues to propagate outwards along the streamer stalk to the outer edge of the FOV. There are no obvious trailing structures observed on the left side of the streamer. So this event is suggested to be a streamer wave event. It is also found that there is no direct contact of the CME ejecta with the streamer structure indicating that the interaction is possibly achieved by the CME expelling the surrounding coronal magnetic field and/or relevant disturbances. The Wind/WAVES data show that there exists a type [slowromancap2@]{} decametric radio burst at about 18:00 UT. Therefore, it is also possible that the streamer wave is related to the shock disturbance driven by the eruption. Discussion on the excitation conditions of streamer waves {#S-labels} --------------------------------------------------------- As mentioned previously, the excitation or formation of streamer wave requires certain strict physical conditions to be satisfied. The conditions may be related to the large-scale magnetic geometry of the corona, the CME dynamics and morphology, as well as the details of the CME-streamer interaction. To provide clues on these conditions, we first summarize the common observational features among the candidate events of streamer waves discussed in the previous subsection. First, all the driving CMEs are fast and wide with an average linear speed of 1580 km s$^{-1}$ and an average apparent angular width of 290$^{\circ}$. Second, all the front-side CMEs have accompanying flares indicating the occurrence of magnetic reconnections during the process. Third, in most events CMEs hit the streamer from the flank, and the sites where the impact takes place are no higher than the bottom of the C2 FOV, i.e., lower than about 2 R$_{\odot}$ in heliocentric distance. Keeping these observational common characteristics in mind, in the following we explain why the above conditions can be taken as necessary conditions for the excitation of streamer waves. We start by discussing the physical origin of the associated restoring force. It is generally believed that the dynamical equilibrium between the expansion of hot coronal plasmas and the confinement of closed magnetic arcades gives rise to coronal streamers consisting of closed field arcades rooted on the photosphere and a cusp atop of them. Upon the impact of the CME ejecta, the streamer structure below the cusp gets deflected away from its equilibrium position. Due to the photospheric line-tying effect of the deflected field lines, the streamer structure responds to the CME impact with a magnetic restoring force, which may be given by both the magnetic tension and pressure. Thus, to produce such a restoring force, the closed field arcades in the streamer should be deflected. This requires that the site where the initial interaction between the ejecta and the streamer structure takes place be low enough, in agreement with the observations. Furthermore, to have a restoring force strong enough to produce observable effects on the streamer structure, the following conditions should also be satisfied. First, the deflection amplitude should be large enough. This requires that the CME ejecta be close enough to the streamer, and the CME expands fast enough, also consistent with the observations of most of our events. Second, the duration of the CME interacting with the streamer structure below the cusp, say from 1.5 $-$ 2.5 R$_\odot$, should be short in comparison with the acting time scale of the inherent restoring force. For a fast-moving CME with a speed no less than 1000 km s$^{-1}$, the ejecta can travel more than 5 R$_\odot$ in 1 hour. In this case, the duration of the CME-streamer interaction in the region between about 1.5 $-$ 2.5 R$_\odot$ is appreciably short compared to the time scale of the restoring force which is estimated to be about 1 hour according to our observations. Therefore, a fast CME is also believed to be one necessary excitation condition for the growth of streamer waves. In addition, to make space for the bouncing motion of the streamer structure following the deflection, some of the magnetic field lines stretched outward by the ejecta should close back. This requires the occurrence of magnetic reconnections along with the eruptions, consistent with the presence of solar flares with all front-side CMEs. We note that the above conditions are only necessary ones, other conditions including the details of the CME-streamer interacting process may be also required. On the other hand, it is true that whether the streamer wave, if already formed, is observable with the coronagraph still depends on the observational view angle, presence of interfering bright structures in the fore- and background corona, and existence of disturbing features of the present eruption and/or other nearby CMEs. We note that the work on excitation conditions of streamer waves is expected to improve by the analysis of more events in the future, as well as magnetohydrodynamic (MHD) modeling efforts designed to study the physics of CME-streamer interaction. Wave properties and relevant seismological study on the 20030605 event {#S-features} ====================================================================== This is suggested to be a simple and clear streamer wave event, which allows us to conduct a further analysis of the wave properties. In the following subsection, we will first determine the heliocentric distances of four wave crests/troughs based on the LASCO observations; then, we deduce the radial profiles of the wave period and wavelength, as well as the propagating phase speeds. In the second subsection, following the seismological approach developed by Paper [slowromancap2@]{} we estimate the radial profiles of the Alfvén speed and magnetic field strength in the region surrounding the plasma sheet. The results will be compared with the 20040706 event studied previously. Wave properties {#S-equations} --------------- Four white light images of this event have been presented in Figure 3. In Figure 7, we show all the relevant RDIs given by LASCO C2 (lower panels) and C3 (upper) observations. From the RDIs, it is easy to recognize the dark-bright pairs corresponding to the streamer deflection and the subsequent wavy motion. Using these difference images, we delineate the profiles of the streamer wave, and determine four specific wave phases at which the wavy motion reaches the maximum amplitude. The phases are called as P1, P2, P3, and P4, and marked with plus signs, asterisks, squares, and triangles. In comparison with the first three phases, P4 appears later associated with a smaller wave amplitude and a lower number of observations and less accuracy of distance and speed measurements as a result. Therefore, in the following text we will focus on the behavior of the first three phases. The variations of the heliocentric distances of the four phases are shown in Figure 8a with corresponding symbols. The solid lines are given by the second order polynomial fittings to the measurements, which are then used to deduce the radial and temporal profiles of the propagation speed of a certain phase. Before presenting the speed profiles, we first discuss the variation of the wave period and wavelength deduced from Figure 8a. For streamer waves, the wavelength is defined as the distance between two adjacent wave crests or wave troughs, e.g., the distance between P1 and P3 at a fixed time; the wave period is defined as the time difference of P1 and P3 passing the same altitude. With these definitions, radial evolutions of the two parameters can be deduced from Figure 8a and are shown in Figures 8c and 8d, respectively. Results for the 20040706 event are also plotted with dashed lines. We can see that, for the 20030605 (20040706) event, the period and wavelength increase from 2.4 R$_\odot$ and 70 minutes (2.0 R$_\odot$ and 50 minutes) at 4 R$_\odot$ to 2.8 R$_\odot$ and 85 minutes (2.6 R$_\odot$ and 85 minutes) at 8 R$_\odot$. Both parameters exhibit similar increasing trend with distances. It should be pointed out that, the statement regarding the wavelength increasing with distance has been explained with the positive difference of speeds of preceding and trailing wave phases in Paper [slowromancap1@]{}. For example, P1 moves faster than P2 and P3, as observed, this makes the distance between P1 and P3, i.e., the first wavelength, increases with the wave propagation. Nevertheless, the increase of period with the wave propagation is not mentioned in previous studies. Here, we provide an explanation of this observation still making use of the difference of phase speeds. From the definition of wave period given above, the time required by P1 to pass through a certain distance range, say from 4 to 6 R$_\odot$, is shorter than the time used by P3 since at a fixed distance P1 moves faster than P3 in general. It can be found that in this case the wave period increases with distance. The phase speed variations involved here have both the spatial and temporal contributions, which are likely associated with the recovering process of the CME-disturbed corona. We will further discuss this process in the following subsection. The phase speeds deduced from the distance-time fittings are plotted as solid lines in Figure 8b. From this figure, we see that the speeds of both P1 and P2 decrease from 435 and 400 km s$^{-1}$ at 3 R$_\odot$ to 425 and 390 (415 and 382) km s$^{-1}$ at 5 (7) R$_\odot$, while the speed of P3 keeps almost constant at about 371 km s$^{-1}$. In addition, the speed averages of the four phases are 402, 376, 371, and 350 km s$^{-1}$, which basically decrease with the order of their presence. Comparing the results of events 20030605 and 20040706, we find that the values and variation trends of the speeds associated with individual phases are essentially similar. Relevant seismological study {#S-simple-equations} ---------------------------- Using measurements of the 20040706 event, Paper [slowromancap2@]{} conducted a seismological study to diagnose the distribution of the Alfvén speed and magnetic field strength in the region surrounding the streamer plasma sheet, interpreting the streamer wave as the fast kink body mode propagating along the sheet structure. The measured phase speed $v_p$ has two contributions, one is the speed of the background solar wind $v_{sw}$, the other is the phase speed of the mode at the plasma rest frame $v_k$. The possible variation range of the former parameter is constrained by the statistical results of the speeds of blobs flowing together with the wind along the plasma sheet (Wang et al., 2000). The latter is connected to the Alfvén speed in the region outside of the plasma sheet ($v_{Ae}$) with an approximation deduced from a parameter study of the corresponding dispersion relation given by Edwin and Roberts (1982). We note that the situation considered by them, where a slab is embedded in an otherwise uniform environment, is different from the present case where an electric current sheet exists inside the high density slab, as suggested by the PFSS extrapolation results for all streamers studied here. The magnetic field reverses its direction across this sheet, thereby presenting a further transverse structuring. However, it can be shown that this current sheet is transparent to the two-dimensional perturbations we consider, as long as it is infinitely thin. Details of our deductions are given in the Appendix. Thus, the radial profiles of $v_{Ae}$ can be deduced for individual phases. With the electron density distribution limited by the inversion of the pB (polarized brightness) data (van de Hulst, 1950; Hayes, Vourlidas, and Howard, 2001) recorded by LASCO, the magnetic field strengths associated with individual phases are then evaluated. It should be pointed out that these diagnoses have large errors and uncertainties. Measurement errors of the phase speeds are estimated to be about $\pm$10$\%$ (or a total of 20$\%$) in Paper [slowromancap1@]{}, and factors contributing to the uncertainties of our seismological studies are discussed in Paper [slowromancap2@]{} and will not be repeated here. Following the approach developed in Paper [slowromancap2@]{}, we conduct a seismological study using the streamer wave observations in the 20030605 event. For simplicity, only the average speed of the blob measurements is used in the study, which is plotted as the dot-dashed line in Figure 9a. It can be seen that the solar wind speed increases gradually from 50 km s$^{-1}$ at 3 R$_\odot$ to 110 (150) km s$^{-1}$ at 5 (7) R$_\odot$. Subtracting the above values of $v_{sw}$, we get the phase speed of the kink mode at the plasma rest frame $v_k$. According to the parameter study on the dispersion relation carried out in Paper [slowromancap2@]{}, $v_k$ is related to $v_{Ae}$ with $v_k \approx \alpha v_{Ae}$, where $\alpha \approx 0.92$. Note that this value of $\alpha$ is determined with the electron density distribution for the 20040706 event, which is basically similar to that for the present event as seen from the results of the pB inversion plotted in the following figure. Therefore, we make use of the same value of $\alpha$ in this study. The obtained radial profiles of $v_{Ae}$ associated with phases P1, P2, and P3 are shown in Figure 9a with solid, dotted, and dashed lines, respectively. From this figure, we see that values of $v_{Ae}$ are 408 (P1), 373 (P2), and 343 (P3) km s$^{-1}$ at 3 R$_\odot$, and decrease to 341 (P1), 304 (P2), and 283 (P3) km s$^{-1}$ at 5 R$_\odot$ and 287 (P1), 249 (P2), and 237 (P3) km s$^{-1}$ at 7 R$_\odot$. It can also be seen that in the region surrounding the plasma sheet the Alfvénic critical point, where $v_{Ae}=v_{sw}$, lies at $\sim$ 10 R$_\odot$ and the Alfvén speed is $\sim$ 200 km s$^{-1}$ there. This result is in agreement with the previous deduction using data from the Helios spacecraft (Pizzo et al., 1983). The radial profiles of the electron density below 5 R$_\odot$ can be obtained with the pB inversion method given in the SolarSoft package, and that beyond 5 R$_\odot$ are given by assuming an $r^{-2}$ dependence. The pB data on 5 June 2003 were recorded at 21:05 UT when the streamer exhibits significant snakelike motion and, therefore, are not appropriate for the density inversion along a specific radial direction. There are no data obtained on the previous day, so we make use of the pB data recorded at 21:00 on 6 June for the required density deduction. This is equivalent to assume that the density along the streamer structure does not change appreciably during the day after the CME eruption. The densities along position angles of 315$^\circ$ and 300$^\circ$, representing different directions along and away from the streamer, are plotted in Figure 10b with dotted and dashed lines, respectively. The average of the two sets of densities is delineated with the solid line with values of $6.5 \times 10^{5}$ $(8.6 \times 10^{4}$, $4.4 \times 10^{4})$ cm$^{-3}$ at 3 (5, 7) R$_\odot$, this is substituted in the deduced profiles of $v_{Ae}$ to diagnose the magnetic field strengths $B_e$ in the corresponding region. The obtained field strengths are shown in Figure 9b for phases P1 (red-solid), P2 (blue-dotted), and P3 (yellow-dashed). From the figure, we see that $B_e$ is 0.145 (0.045, 0.028) G for P1, 0.139 (0.041, 0.024) G for P2, and 0.129 (0.038, 0.023) G for P3 at 3 (5, 7) R$_\odot$, respectively. Basically $B_e$ decreases according to the $r^{-2}$ dependence. In Figure 9b, we also show other estimates on the magnetic field strength in the corona with various symbols, including results from our previous study employing the streamer wave method to the 20040706 event presented as shadow areas, and several other diagnoses obtained employing various radio methods. To be specific, the strength-distance relationship in the heliocentric range of 1.02 - 10 $R_\odot$ above active regions (dot-dashed line) given by Dulk and McLean (1978) is mainly based on radio burst observations, the results of Vršnak et al. (2002) and Cho et al. (2007) presented as crosses and diamonds are deduced using the band-splitting phenomenon of type [slowromancap2@]{} radio bursts, the results from the Faraday-rotation measurement of radio signals emitted from the Helios spacecraft and extragalactic radio sources are included as open circles with error bars (Pätzold et al., 1987), triangles (Spangler, 2005), and squares (Ingleby et al., 2007). The latest results obtained by Ramesh et al. (2010) employing the low-frequency circularly polarized radio emission inside a streamer structure are given as solid inverse triangles. We note that the above list of diagnoses of the coronal field strength is incomplete, and there exist many other estimates (see e.g., references in Vršnak et al., 2002). From this figure, it can be seen that the magnitude and variation trend of the magnetic field strength in the region surrounding the plasma sheet are basically consistent with the results for event 20040706 and other estimates. In addition, we also find that both the Alfvén speed and magnetic field strength have a general decline trend with time at a fixed distance. According to the explanation provided in Paper [slowromancap2@]{}, this trend of variation of the two quantities is a result of the recovering process of the CME-disturbed corona. From Figure 9, it can be seen that both the magnetic field strength and Alfvén speed decline by about 15% during about 60-90 minutes, i.e., in about one wave period. Of course, the coronal field can not decrease unlimitedly. However, at this time it is not possible to assess how the magnetic field strength may evolve in such a short interval in the absence of a well-observed streamer wave phenomenon. Finally, we note that the differences between the diagnostic results for the three wave phases are possibly not significant considering the errors and uncertainties associated with our diagnostics. Conclusions =========== In this paper, we present 8 candidate streamer wave events found via a data survey through LASCO observations in Solar Cycle 23. We find the following three common observational features in these events: (1) the driving CMEs are wide and fast with a linear speed no less than $\sim$ 1000 km s$^{-1}$; (2) all front-side CMEs have accompanying flares; (3) in most events the bright CME and streamer structures have direct contact with each other, and the interaction starts at a height no higher than the bottom of the C2 FOV, i.e., lower than $\sim$ 2 R$_\odot$. These common features shed light on the excitation conditions of streamer waves. Nevertheless, more similar events, when available in the future, should be analyzed and MHD models should be developed for a better understanding on the physics of CME-streamer interaction. A further study on the event dated on 5 June 2003 gives radial and temporal evolution of wave properties including the period, wavelength, and propagation speeds of four observable wave crests/troughs. It is found that both the period and wavelength increase gradually with the wave propagation along the streamer plasma sheet, and the phase speed of the preceding wave phase is generally faster than that of the trailing ones. The associated coronal seismological study yields the radial profiles of the Alfvén speed and magnetic field strength in the region surrounding the streamer plasma sheet. It is found that the Alfvénic critical point is at about 10 R$_\odot$ where the flow speed, which equals the Alfvén speed, is $\sim$ 200 km s$^{-1}$. The magnetic field strengths corresponding to the first three wave phases are 0.145, 0.139, and 0.129 G at 3 R$_\odot$, and decrease generally according to the $r^{-2}$ dependence to 0.045, 0.041, and 0.038 G at 5 R$_\odot$, and to 0.028, 0.024, and 0.023 G at 7 R$_\odot$, respectively. The obtained results are generally consistent with that of another well-studied event on 6 July 2004. The SOHO/LASCO data used here are produced by a consortium of the Naval Research Laboratory (USA), Max-Planck-Institut für Aeronomie (Germany), Laboratoire d’Astronomie Spatiale (France), and the University of Birmingham (UK). SOHO is a project of international cooperation between ESA and NASA. We thank Dr. A. Vourlidas for helping us analyze the LASCO pB data. This work was supported by grants NNSFC 40825014, 40890162, A Foundation for the Author of National Excellent Doctoral Dissertation of PR China (2007B24), and the Specialized Research Fund for State Key Laboratory of Space Weather in China. B Li is supported by the grant NNSFC 40904047, LD Xia by 40974105, and HQ Song by Natural Science Foundation of Shandong Province ZR2010DQ016 and Independent Innovation Foundation of Shandong University 2010ZRYB001. We note that the situation considered by Edwin and Roberts (1982, ER82 from now on), where a slab is embedded in an otherwise uniform environment, is different from the present case where an electric current sheet (CS) exists inside the high density slab. The magnetic field reverses its direction across this sheet, thereby presenting a further transverse structuring. However, it can be shown that this CS is transparent to the two-dimensional perturbations we consider, as long as it is infinitely thin. To illustrate this, let us start with Figure 1 in ER82 where the vertical lines $x=\pm x_0$ separate the slab from its environment. The dispersion relation (Equation (11) in ER82) then follows from the linearized ideal MHD equations, the ansatz that any small-amplitude time-dependent perturbation $f(x,z;t)$ is in the form $\hat{f}(x)\exp[i(\omega t + kz)]$, the condition that the perturbations are not identically zero, as well as the requirements that the perturbed total pressure (gas plus magnetic) and transverse velocity $v_x$ be continuous across the interfaces $x=\pm x_0$. With the above ansatz in mind, a pair of $(\omega, k)$ that satisfies the dispersion relation will give a $\hat{v}_x$ continuous throughout the entire $x-z$ plane. To be specific, Equation (10) in ER82) expresses both kink and sausage modes, with the former satisfying $\beta_e=-\alpha_e=\beta_0\sinh(m_0 x_0)$, and the latter obeying $\beta_e=\alpha_e=\alpha_0\cosh(m_0 x_0)$. Here $\alpha_0$ or $\beta_0$ is arbitrary. Among the rest of perturbations pertaining to these modes, $\hat{v}_z$, $\hat{p}_T$ are given by Equations (17) and (18) in Roberts (1981), while $\hat{b}_x$ and $\hat{b}_z$ can be found via Equation(10) in Roberts (1981), $\hat{p}$ via Equation (9) and $\hat{\rho}$ via Equation (6a). Here $\vec{v}=(v_x, 0, v_z)$ and $\vec{b}=(b_x, 0, b_z)$ are the velocity and magnetic field perturbations, respectively. Moreover, $p, \rho$ and $p_T$ represent the perturbed pressure, density and total pressure. The hat $\hat{}$ refers to the Fourier amplitudes. Let the solution set $(\hat{\vec{v}}, \hat{\rho}, \hat{p}, \hat{\vec{b}})$ be denoted by $S^{\mathrm{ER82}}$. Now suppose the sign of the magnetic field in equilibrium in the right half of the plane ($x>0$) turns negative. It turns out that $S^{\mathrm{ER82}}$ is also a solution of the linearized MHD equations in this case, the only modification being that $\hat{\vec{b}}$ should be in the opposite direction to $\hat{\vec{b}}$ found in ER82 for $x>0$. Across the (perturbed) interface initially located at $x=0$ in equilibrium, the normal magnetic field, transverse displacement, and the perturbed total pressure $p_T$ are all continuous. (That $p_T$ is continuous follows from the fact that $p$ and $B_0 b_z$ are both continuous, and one can easily see that the latter does not change if $B_0$ and $b_z$ changes their signs simultaneously.) This allows us to see the CS as transparent within the framework of linear, ideal MHD. In particular, the dispersion relation derived in ER82 also applies. Bougeret, J.L., Kaiser, M.L., Kellogg, P.J., Manning, R., Goetz, K., Monson, S.J., Monge, N., Friel, L., Meetre, C.A., Perche, C., et al.: 1995, Space Sci. Rev. 71, 231 Brueckner, G.E., Howard, R.A., Koomen, M.J., Korendyke, C.M., Michels, D.J., Moses, J.D., Socker, D.G., Dere, K.P., Lamy, P.L., Llebaria, A., et al.: 1995, Solar Phys. 162, 357. Chen, Y., Song, H.Q., Li, B., Xia, L.D., Wu, Z., Fu, H., Li, X.: 2010, Astrophys. J. 714, 644. Chen, Y., Feng, S.W., Li, B., Song, H.Q., Xia, L.D., Kong, X.L., Li, X.: 2011, Astrophys. J. 728, 147. Cho, K.S., Lee, J., Gary, D.E., Moon, Y.J., Park, Y.D.: 2007, Astrophys. J. 665, 799. Dulk, G.A., Mclean, D.J.: 1978, Solar Phys. 57, 279. Delaboudinière, J.P., Artzner, G.E., Brunaud, J., Gabriel, A.H., Hochedez, J.F., Millier, F., Song, X.Y., Au, B., Dere, K.P., Howard, R.A., et al.: 1995, Solar Phys. 162, 291. Edwin, P.M., Roberts, B.: 1982, Solar Phys. 76, 239. Filippov, B., Srivastava, A.K.: 2010, Solar Phys. 266, 123. Gopalswamy, N., Yashiro, S., Michalek, G., Stenborg, G., Vourlidas, A., Freeland, S., Howard, R.: 2009, Earth, Moon, and Planets, 104, 295. Hayes, A.P., Vourlidas, A., Howard, R.A.: 2001, Astrophys. J. 548, 1081. Hundhausen, A.J., Holzer, T.E., Low, B.C.: 1987, J. Geophys. Res. 92, 11173. Ingleby, L.D., Spängler, S. R., Whiting, C.A.: 2007, Astrophys. J. 668, 520. Pätzold, M., Bird, M.K., Volland, H., Levy, G.S., Seidel, B.L., Stelzried, C.T.: 1987, Solar Phys. 109, 91. Pizzo, V., Schwenn, R., Marsch, E., Rosenbauer, H., Muehlhaeuser, K.H., Neubauer, F.M.: 1983, Astrophys. J. 271, 335. Ramesh, R., Kathiravan, C., Sastry, Ch. V.: 2010, Astrophys. J. 711, 1029. Roberts, B.: 1981, Solar Phys. 69, 27. Schatten, K.H., Wilcox, J.M., Ness, N.F.: 1969, Solar Phys. 6, 442. Schrijver, C.J., De Rosa, M.L.: 2003, Solar Phys. 212, 165. Sheeley, N.R., Hakala, W.N., Wang, Y.M.: 2000, J. Geophys. Res. 105, 5081. Spangler, S.R.: 2005, Space Sci. Rev. 121, 189. Tripathi, D., Raouafi, N.E.: 2007, Astron. Astrophys. 473, 951. van de Hulst, H.C.: 1950, Bull. Astron. Inst. Neth. (11), 135. Vršnak, B., Magdalenic, J., Aurass, H., Mann, G.: 2002, Astron. Astrophys. 396, 673. Wang, Y.M., Sheeley, N.R., Socker, D.J., Howard, R.A., Rich, N.B.: 2000, Geophys. Res. Lett. 105, 25133. -------------- ------------ ----------- ------- --------------- ------------- ------- Streamer Date Time CPA Width Speed Flares CPA (yyyy/mm/dd) (UT) (deg) (deg) (km s$^{-1}$) (deg) 20010420 10:06:05 65 127 1160 $\setminus$ 15 20030527\* 23:50:05 Halo/S 360 964 X1.3 185 (20030528) (00:50:05) (Halo/BA) (360) (1366) (X3.6) (185) 20030605 20:06:05 230 239 1458 $\setminus$ 316 20031118\* 08:50:05 Halo/BA) 360 1660 M3.9 319 20040705\* 23:06:05 Halo 360 1444 $\setminus$ 225 20040706 20:06:06 Halo(BA) 360 1307 $\setminus$ 225 20050115 23:06:50 Halo(BA) 360 2861 X2.6 240 20061106 17:54:04 80 80 1994 C8.8 128 -------------- ------------ ----------- ------- --------------- ------------- ------- : Some relevant physical parameters of the CMEs and streamers of the 8 candidate streamer wave events. The first to sixth columns present the appearance date and time (UT) of the CME ejecta in the C2 field of view (FOV), the CME type, the central position angle (CPA), the linear speed, and the importance of the accompanying flares according to the X-ray flux recorded by the GOES satellite. The last column gives the CPA, measured at 5 R$_\odot$, of the streamer structure with wavy motion. ![White light images and RDIs in the streamer region observed by LASCO C2 for the 20010420 event. The plus signs, asterisks, and squares mark the location of phases P1, P2, and P3. See text for details.[]{data-label="f1"}](fig1.eps){width="100.00000%"} ![White light images of the disturbed streamer structure observed by C2 on 28 May 2003. Arrows indicate the waving and tangling features of the streamer stalk. []{data-label="f2"}](fig2.eps){width="100.00000%"} ![White light images of the CME-streamer interaction event observed by LASCO C2 on 5 June 2003.[]{data-label="f3"}](fig3.eps){width="100.00000%"} ![White light images (a, b) of the two CME eruptions, and RDIs (c-f) of the streamer region for the 20031118 event. The RDIs have been rotated counterclockwise by 41$^{\circ}$.[]{data-label="f4"}](fig4.eps){width="100.00000%"} ![A white light image (a) and RDIs (b-f) for the CME-streamer wavy event observed by LASCO C2 on 15 January 2005. The CPA of the wavy streamer is 240$^{\circ}$. The RDIs in the lower panels are in the range of heliocentric distances from 4 R$_\odot$ to 12 R$_\odot$ taken from the C3 observations. []{data-label="f5"}](fig5.eps){width="100.00000%"} ![White light images (a, b) and RDIs (e-f) of the CME-streamer interaction event observed by LASCO C2 on 6 November 2006. The CPA of the wavy streamer is 128$^{\circ}$. The RDIs in the lower panels have been rotated clockwise by 128$^{\circ}$. []{data-label="f6"}](fig6.eps){width="100.00000%"} ![RDIs in the streamer region for the 20030605 event observed by LASCO C2 (lower panels) and C3 (upper). All images have been rotated counterclockwise by 35$^{\circ}$. The four selected wave phases (P1 - P4) are marked by plus signs, asterisks, squares, and triangles, respectively.[]{data-label="f7"}](fig7.eps){width="100.00000%"} ![Properties of the streamer wave event 20030605. (a) Radial variations of heliocentric distances of the four phases (P1 - P4) indicated by plus signs, asterisks, squares, and triangles, respectively. The solid lines are given by the second order polynomial fittings to the measurements. (b) The fitted profiles of phase speeds for the first three wave phases. The deduced radial evolutions of the wavelengths (c) and periods (d) for the 20030605 (solid) and the 20040706 (dashed) events.[]{data-label="f8"}](fig8.eps){width="100.00000%"} ![Radial profiles of the estimated Alfvén speed (a) and magnetic field strength (b) in the region surrounding the streamer plasma sheet for the streamer wave event 20030605. Profiles corresponding to different wave phases are given by the red-solid (P1), blue-dotted (P2), and yellow-dashed (P3) lines. The black dot-dashed line in (a) presents the average blob speed taken from Wang et al. (2000). Various symbols in (b) represent other estimates of the coronal magnetic field strength. []{data-label="f9"}](fig9.eps){width="100.00000%"} ![(a) The pB intensity distribution observed by LASCO on 6 June 2003, the dotted (dashed) line denotes a specific position angle of 315 (300)$^{\circ}$. (b) Dotted and dashed lines are the corresponding electron number density profiles given by the pB inversion method ($\le$ 5 R$_\odot$) and the r$^{-2}$ dependence ($>$ 5 R$_\odot$) along the two position angles. The solid line is given by their average.[]{data-label="f10"}](fig10.eps){width="100.00000%"}	{ "pile_set_name": "ArXiv" }
--- abstract: 'Singlet fission has emerged as one of the most exciting phenomena known to improve the efficiencies of different types of solar cells and has found uses in diverse optoelectronic applications. The range of available singlet fission molecules is, however, limited as to undergo singlet fission, molecules have to satisfy certain energy conditions. Recent advances in material search using inverse design has enabled the prediction of materials for a wide range of applications and has emerged as one of the most efficient methods in the discovery of suitable materials. It is particularly helpful in manipulating large datasets, uncovering hidden information from the molecular dataset and generating new structures. However, we seldom encounter large datasets in structure prediction problems in material science. In our work, we put forward inverse design of possible singlet fission molecules using a transfer learning based approach where we make use of a much larger ChEMBL dataset of structurally similar molecules to transfer the learned characteristics to the singlet fission dataset.' author: - \| Akshay Subramanian^1^, Utkarsh Saha^2^, Tejasvini Sharma^2^,\ Naveen K. Tailor^2^, Soumitra Satapathi^2[\]{}^ bibliography: - 'publication.bib' title: Inverse Design of Potential Singlet Fission Molecules using a Transfer Learning Based Approach --- 1\. Department of Metallurgical and Materials Engineering, Indian Institute of Technology Roorkee, Roorkee, India\ 2. Department of Physics, Indian Institute of Technology Roorkee, Roorkee, India\ [\]{}corresponding author: Soumitra Satapathi ([email protected]) Introduction {#introduction .unnumbered} ============ Singlet fission (SF) is a process by which a high-energy singlet exciton, resultant from the absorption of a photon, is converted into two triplet excitons, each carrying about half the energy. [@1; @2; @3; @4; @5] As the process is spin allowed, it can occur very rapidly (on a picosecond or femtosecond timescale) and out-compete radiative decay (that generally occurs on a nanosecond timescale) thereby producing two triplets with very high efficiency. [@1; @2] For a molecule to undergo SF, certain conditions need to be satisfied. One of the main requirements is that the energy of the singlet state (E\[S1\]) should be more than double the energy of the triplet state (E\[T1\]), i.e., E\[S1\]/E\[T1\] $\geq$ 2. [@1; @2; @3] As an ideal case, E\[S1\]/E\[T1\] $\approx$ 2 for fast formation of the spin-coupled triplet state and too much exoergicity can result in unwanted heat generation. The other requirement for singlet fission is that the energy of the next higher triplet level should be greater than or equal to the energy of the two low-lying triplets, E\[T2\] $\geq$ 2$\times$E\[T1\]. [@1; @2; @3] SF offers the possibility of overcoming thermalisation losses in PVs, as every photon absorbed above the bandgap leads to the formation of two electron-hole pairs. Consequently, there is a growing scope of interest in this area as SF can overcome the theoretical Shockley–Queisser limit for the power conversion efficiency of a single junction solar cell, which is about 32%. [@6] However, the realization of SF-based solar cells is hindered by dearth of suitable materials. Due to the high computational cost of excited-state quantum mechanical calculations, predictive descriptors that are fast to evaluate must be found in order to explore the chemical space in search of new SF materials.\ \ Previously, quantum mechanical methods and time dependent density functional theory (TD-DFT) have been used to the study of singlet fission process. Berkelbach et al. applied Redfield theory to investigate the charge transfer states in the SF dynamics of pentacene. [@4; @5; @7] Tamura et al. utilized the time-dependent Hartree quantum mechanical approach to study the SF mechanism in a pentacene derivative and in rubrene. [@8] Krishnapriya et.al. employed spin density distribution to encode SF in a series of pentacene dimers using phenyl-, thienyl- and selenyl- flanked diketopyrrolopyrrole (DPP) derivatives. [@9] Several research groups have also applied machine learning methods to study SF dynamics. Chen et.al. have reported that deep learning can be used for describing nonadiabatic excited state dynamics for SF. [@10] Schröder et.al. have used tree-tensor network state simulations to compute the real-time dynamics of exponentially large vibronic states of SF molecules. [@11] Although these interesting studies have explored the energetics and charge carrier dynamics of different existing SF molecules, these methods cannot predict new potential materials for SF. Generally, for the development of new materials, the stepwise procedure of selection of materials, prediction of material properties, chemical synthesis, and experimental validation is usually repeated until satisfactory performance is achieved. As this experimental approach is tedious and expensive, progressively predictive techniques, for example, high-throughput computational screening (HTCS) have gained popularity in recent years. [@12; @13; @14; @15] In HTCS, extensive property prediction using DFT calculations [@16] or machine learning [@17; @18] is carried out after screening out suitable materials from molecular libraries and open databases. This permits highly efficient categorization of potential candidates for subsequent experimental verification. This procedure has been increasingly used in materials science for various applications such as learning the chemistry of materials using only elemental composition [@35] , crystal structure prediction [@32] and target property prediction [@33; @34] . But for DFT calculations, the computational complexity increases with the number of atoms and achieving the results up to a certain accuracy can be computationally quite costly. Predicting material properties using machine learning requires the selection of suitable features and feature selection is intuitive as it cannot be reasoned out why some features work better than the others. Moreover, there is no assurance regarding whether the correct chemical space is being investigated.\ \ To overcome the above-mentioned issues, inverse design of materials using deep learning has emerged as one of the most promising methods in recent years for predicting potential materials with specific target properties for an application. [@26; @27; @28; @29; @30] Inverse design aims to design materials that are expected to meet the given target properties in a direct manner, whereas the conventional approach designs materials first and predicts their properties subsequently. The inverse design approach extracts the molecular design knowledge hidden in the molecular database and generates new molecules on the basis of its own knowledge, thereby allowing systematic materials exploration without the need for researcher experience or intuition. Deep learning has recently found significant use in the inverse design of molecules, especially in the area of drug discovery. Gomez-Bombarelli et al. [@19] have shown that deep generative models could be utilized for the inverse design of potential drug molecules by optimizing certain properties of importance. They used a Variational Auto-Encoder (VAE) [@20] , a robust generative model for this purpose, the architecture and specifics of which have been elaborated in later sections. Similarly, Popova et al. [@21] experimented with stack-augmented RNNs to design drug molecules, while using Reinforcement Learning approaches to tune properties such as solubility.\ \ Our primary goal was to investigate the possibilities of inverse design of singlet fission molecules. The search for singlet fission molecules is hindered by the fact that there are too few molecules undergoing singlet fission. The dataset we used comprised of about 1000 screened singlet fission molecules from previous literature [@31; @1; @3; @5; @37; @38; @39; @40; @41].\ \ In our work, we propose an inverse design model for the generation of new potential singlet fission molecules. Initially, we had applied a deep generative model to the 1000 molecule extract for the inverse design of singlet fission molecules. But, the number of molecules was not sufficient to successfully carry out inverse design using traditional deep learning techniques. This is because deep learning techniques try to represent the problem in the form of complex functions and therefore require large amounts of data to accurately describe the function. To solve this problem of limited data availability, we develop a new approach that makes use of a transfer learning approach. The transfer learning [@22] approach has primarily been applied to image classification and detection tasks in the past. Recently, this approach has started to gain recognition in molecular discovery field. Segler et al. [@23] used transfer-learning to first train the RNN on a whole dataset of molecules and later fine-tune the model towards the generation of drug molecules with physico-chemical properties of interest. Gupta et al. [@24] applied the transfer learning approach to grow drug molecules from fragments. In our approach, we initially train a deep network on a much larger dataset (ChEMBL - 100,000 molecules) [@25] to teach the model to encode and decode Simplified Molecular-Input Line-Entry System(SMILES) strings and understand the intricacies in their grammar. We then fine-tune this model on our singlet-fission dataset to transfer the learned characteristics to our task of importance. This fine-tuned model was then used to generate novel molecules that could possibly exhibit singlet fission property. The ChEMBL and singlet-fission datasets contained molecules that were structurally quite similar. This prompted us to choose the ChEMBL dataset for the pretraining task. Results {#results .unnumbered} ======= Computational Approach {#computational-approach .unnumbered} ---------------------- Building the Generative Model {#building-the-generative-model .unnumbered} ----------------------------- Our model primarily comprised of 3 networks : a) A Variational Autoencoder (VAE), b) A Multi Layer Perceptron (MLP) and c) A classifier. The VAE consists of an encoder that is responsible for encoding the input SMILES string into a 1-D vector, and a decoder that is responsible for decoding this latent vector back into a SMILES string. We call the 1-D encoded vector the latent representation of the molecule. A more detailed explanation of the autoencoder architecture is described in the Supplementary Information. For the generation of new structures, the chemical structures encoded in the continuous representation have to be connected with the target properties that we are seeking to optimize. We trained the VAE on a reconstruction task and the MLP on a regression task with the target property as E\[S1\] - 2 $\times$ E\[T1\], which we wish to optimize. This results in the formation of a property-wise distribution of the latent space, which means that molecules with similar values of the target property will lie close to each other (measured by Euclidean distance) on the latent space. This then allows us to easily maneuver and explore the latent space, giving us flexibility while generating new molecular structures. A diagrammatic representation of the latent space is shown in Fig \[fig:latent\]. ![Latent space optimization. After the target property prediction of molecules based on their latent representation, we can optimize the values to find new latent representations expected to have better values of the desired properties.[]{data-label="fig:latent"}](1.jpg){width="80.00000%"} To generate promising new singlet fission molecules, the latent vector of an encoded molecule is taken and then we advance in the direction anticipated to improve the desired target property. A schematic flowchart depicting the autoencoder along with a property prediction network is shown in Fig \[fig:auto\_flowchart\]. ![Flowchart for generating new structures using the inverse design model. The encoder converts the SMILES input into a 1-D vector which is decoded back to a SMILES string by the decoder. The formation of the property wise distribution of the latent space allows us to take the latent vector of an encoded molecule and move in a direction most likely to improve the target property, which in this case, is the first SF condition. For the second SF condition, a neural network is used as a binary classifier.[]{data-label="fig:auto_flowchart"}](2.jpg){width="100.00000%"} The resulting new candidate vectors can then be decoded into corresponding molecules. Then, for the second singlet fission condition E\[T2\] - 2$\times$E\[T1\] $>$ 0 to be satisfied, we train another neural network for classification of the generated structures. Applying this to SF dataset {#applying-this-to-sf-dataset .unnumbered} --------------------------- Since the singlet fission dataset has only about 1000 molecules, accurate reconstruction of new molecules is not feasible using deep inverse design, since deep learning approaches often require large amounts of data to perform well. When we tried training our deep network on the singlet fission dataset, the model was not able to learn the intricate and complex SMILES grammar and hence, generated many invalid SMILES strings. This is a major hurdle in applying modern inverse design approaches to tasks that contain less amounts of data. As a result, we propose a new approach for the design of new structures using transfer learning which allowed our model to perform far better and generate a substantial number of valid SMILES strings. Transfer Learning Approach {#transfer-learning-approach .unnumbered} -------------------------- For the transfer learning approach, we need a database which contains similar structures as there are in the singlet fission. We choose the ChEMBL database of bioactive molecules which are very similar in structure with the molecules in the singlet fission dataset after which we extract around 100,000 molecules for the model. Then, we train the VAE on the 100,000 molecule extract of CHEMBL database to teach the model to learn basic SMILES grammar. We followed the following steps to transfer the learned characteristics from the ChEMBL dataset to the singlet fission dataset: 1) Pretrained the VAE on ChEMBL extract on only reconstruction of input SMILES strings, 2) Joint training of pretrained VAE and newly initialized MLP on Reconstruction + Regression (Property Prediction) tasks, 3) Froze VAE layers and only trained MLP on regression task (Model started overfitting on reconstruction task earlier than on the regression task during joint training). The loss curves obtained while training the model on the singlet fission dataset are shown in Fig \[fig:plots\]. [.9]{} ![ Training plots on the singlet fission dataset. Blue line indicates loss on the validation dataset and red line indicates loss on the training dataset[]{data-label="fig:plots"}](images/plot1.jpg "fig:"){width=".9\textwidth"} [.9]{} ![ Training plots on the singlet fission dataset. Blue line indicates loss on the validation dataset and red line indicates loss on the training dataset[]{data-label="fig:plots"}](images/plot2.jpg "fig:"){width=".9\textwidth"} Our VAE contains Dense (Fully connected) layers not only as the final layer but also in some positions in the middle of the network. These dense layers are input size dependent and so we had to follow identical preprocessing methods in both the ChEMBL and Singlet Fission datasets so as to have identical input sizes in both cases. The presence of these size dependent layers was also the key reason that we chose to use the combined vocabulary of both datasets to define our fully connected layer sizes. Generating novel structures {#generating-novel-structures .unnumbered} --------------------------- To generate potential singlet fission molecules using our model, we decoded points in the latent space close (measured by Euclidean distance) to molecules already known to satisfy the Singlet Fission condition (E\[S1\] - 2 $\times$ E\[T1\] $>$ 0). The VAE being probabilistic in nature, generates many invalid SMILES strings. To account for this stochasticity, we performed 200 decoding attempts per molecule so as to produce valid strings. From these decoding attempts, we usually got a prominent molecule and many others appeared with lower frequencies. By optimizing the Singlet Fission property (E\[S1\] - 2 $\times$ E\[T1\]), we were able to generate multiple potential molecules satisfying the above criterion. After that, we input the generated molecules through a neural network classifier for satisfying the second singlet fission condition E\[T2\] $\geq$ 2$\times$E\[T1\]. We show four of the resultant structures in Fig \[fig:mol\_joint\]. ![Generated molecular structures[]{data-label="fig:mol_joint"}](images/mol_joint.jpg){width="100.00000%"} We see that although all four structures satisfy the singlet fission criteria, the third and fourth structures seem unstable due to the presence of 9 and 10 membered rings respectively. We can therefore shortlist the first two structures for further experimental validation or DFT analysis. Discussion {#discussion .unnumbered} ========== From the above results, we see that transfer learning can be a powerful tool in applying deep generative neural networks for the inverse design of materials, especially on small sized datasets. This approach is generalizable to a large number of materials science problems owing to the fairly common issue of small sized datasets. To the best of our knowledge, this is the first time that a transfer learning based approach is being applied to the inverse design of materials while optimizing properties of interest. This method is easily extendable to many other materials science problems with small dataset constraints. We can screen out the final potential molecules for further study and analysis. Consequently, this formulation of a transfer learning based inverse design framework is expected to minimize the efforts in the computational screening of feasible molecules in a boundless search space. Methods {#methods .unnumbered} ======= Data Collection and Preprocessing {#data-collection-and-preprocessing .unnumbered} --------------------------------- Pretraining was done on a 100,000 molecule extract of the ChEMBL database. The learned features were then transferred to the singlet fission dataset [@31]. First, we take the molecules from the datasets and convert them into their SMILES representation by using RDKIT. Then, for feeding into the neural network, we generate one-hot encoded vector representations of the SMILES strings with a vocabulary (set of unique characters) of length 52. We chose this length by calculating the vocabulary of the combined singlet fission dataset + ChEMBL extract. For computational ease, we encoded strings up to a maximum length of 120 characters and padded shorter strings with spaces. We chose the above form of representation to allow for easy encoding of SMILES to vectors and decoding of vectors back to SMILES or in other words, a two-way mapping between the strings and their corresponding vectors. Model Architecture {#model-architecture .unnumbered} ------------------ We made many architectural design choices for our VAE based on work by Gomez-Bombarelli et al. The following was the construction of our VAE model: The encoder consisted of four 1-D convolutional layers of size 9, 9, 11 and 11 followed by two fully connected (Dense) layers of sizes 435 and 292 respectively. Our latent representation was therefore a 1-D vector of length 292. The decoder consisted of three Gated Recurrent Unit (GRU) layers of hidden size 501 followed by a fully connected layer of size 52 whose outputs were passed to a softmax activation function. Our decoder therefore produced a probability distribution over each character in our vocabulary for each of the 120 characters of our output string. Rectified Linear Unit activation function was applied after all hidden layers except for the pre-final layer of the encoder and the first layer of the decoder where Scaled Exponential Linear Unit (SELU) activation function was used. Adam was the optimizer used during the VAE training. Pretraining of the VAE on the reconstruction task on the ChEMBL extract required a total of 136 epochs to train and finetuning of VAE + MLP on joint reconstruction + regression (property prediction) tasks was carried out for 42 epochs after which our model started overfitting on the reconstruction task. This called for the third step of our training procedure which was to freeze all VAE layers and only train the MLP on the property prediction task for another 50 epochs to achieve the best possible results on both tasks. We used a learning rate of 1e-3 during pretraining of the VAE and a learning rate of 1e-4 during finetuning of the VAE + MLP model. All our experiments were carried out using the PyTorch [@36] framework. Data Availability {#data-availability .unnumbered} ================= Data for training the model is available upon request. Code Availability {#code-availability .unnumbered} ================= An implementation of the algorithm described in the paper and pretrained model weights are available at https://github.com/aksub99/FissionNet. Author Information {#author-information .unnumbered} ================== Affiliations {#affiliations .unnumbered} ------------ Department of Physics, Indian Institute of Technology Roorkee, Roorkee, Uttarakhand, 247667, India\ Utkarsh Saha, Tejasvini Sharma, Naveen K. Tailor & Soumitra Satapathi\ \ Department of Metallurgical and Materials Engineering, Indian Institute of Technology Roorkee, Roorkee, Uttarakhand, 247667, India\ Akshay Subramanian Authors’ contributions {#authors-contributions .unnumbered} ====================== S.S., A.S. and U.S. conceived the project, A.S. wrote the computer software and carried out simulations and experiments with contributions from U.S.; U.S., A.S., S.S., T.S. and N.K.T. wrote the manuscript. A.S. and U.S. contributed equally on this work. Competing interests {#competing-interests .unnumbered} =================== The authors declare no competing interests.	{ "pile_set_name": "ArXiv" }
--- --- [The Frobenius problem for homomorphic\ embeddings of languages into the integers]{} [Michel Dekking]{} [DIAM, Delft University of Technology, December 12, 2017 ]{} Introduction ============ The Frobenius problem is also known as the ‘coin problem’. Since the value of a coin can only be positive, we will consider exclusively embeddings into the natural numbers $\mathbb{N}=\{1,2,3,\dots\}$. Let ${{\mathcal{L}}}$ be a language, i.e., a sub-semigroup of the free semigroup generated by a finite alphabet under the concatenation operation. A homomorphism of ${{\mathcal{L}}}$ into the natural numbers is a map ${{\rm S}}:{{\mathcal{L}}}\rightarrow {{\mathbb{N}}}$ satisfying $${{\rm S}}(vw)={{\rm S}}(v)+{{\rm S}}(w), \quad {\rm for\: all }\: v,w \in {{\mathcal{L}}}.$$ The two main questions to be asked about the image set ${{\rm S}}({{\mathcal{L}}})$ are (Q1) Is the complement ${{\mathbb{N}}}\setminus {{\rm S}}({{\mathcal{L}}})$ finite or infinite? (Q2) If the complement of ${{\rm S}}({{\mathcal{L}}})$ is finite, then what is the largest element in this set? These two questions are known as the Frobenius problem in the special case that ${{\mathcal{L}}}$ is the full language consisting of all words over a finite alphabet. In this case they have been posed as a problem (with solution) for an alphabet $\{a,b\}$ of cardinality 2 by James Joseph Sylvester in 1884 [@Sylvester]: ${{\mathbb{N}}}\setminus{{\rm S}}({{\mathcal{L}}})$ is finite, and its largest element is $${{\rm S}}(a){{\rm S}}(b)-{{\rm S}}(a)-{{\rm S}}(b).$$ In this paper we will also restrict ourselves to the two symbol case: alphabet $\{a,b\}$. In Section \[sec:gold\] we prove that for the golden mean language (“no $bb$") the set ${{\mathbb{N}}}\setminus{{\rm S}}({{\mathcal{L}}})$ is finite, with largest element $${{\rm S}}(a)^2+{{\rm S}}(a){{\rm S}}(b)-3{{\rm S}}(a)-{{\rm S}}(b).$$ Our main interest is however not in sofic languages[^1], but in languages with low complexity, where the complement of ${{\rm S}}({{\mathcal{L}}})$ can be infinite. In Section \[sec:Sturm\] we analyse the case of Sturmian languages, and show that for the Fibonacci language a $0\!-\!\infty$ law holds: either the complement is empty or it has infinite cardinality. In Section \[sec:TM\] we show that for any homomorphism ${{\rm S}}$ the image of the Thue-Morse language will consist of a union of 5 arithmetic sequences. In Section \[sec:twodim\] we consider two-dimensional embeddings, which behave quite differently. We usually suppose that $\gcd({{\rm S}}(a),{{\rm S}}(b))=1$. First of all this is not a big loss since automatically the complement will have infinite cardinality in this case. Secondly, if $r$ divides both ${{\rm S}}_1(a)$ and ${{\rm S}}_1(b)$ for some homomorphism ${{\rm S}}_1$, then $${{\rm S}}_1({{\mathcal{L}}}^n)=r^n{{\rm S}}_2({{\mathcal{L}}}^n), \quad {\rm for}\: n=1,2,\dots,\: {\rm where}\; {{\rm S}}_2(a)=\frac{{{\rm S}}_1(a)}{r},\, {{\rm S}}_2(b)=\frac{{{\rm S}}_1(b)}{r}.$$ Our work is related to the work on abelian complexity, see, e.g., [@abeliancomp], [@RSZ], [@Karhumaki-S-K]. See Lemma \[lem:compl\] for such a connection. Our work is also related to the notion of additive complexity, see [@Sahas] and [@Ardal]. The additive complexity of an infinite word $w$ over a finite set of integers (see [@Ardal]) is the function $n\rightarrow\phi^+(w, n)$ that counts the number of distinct sums obtained by summing $n$ consecutive symbols of $w$. In general we write ${{\mathcal{L}}}^n$ for the set of words of length $n$ in a language ${{\mathcal{L}}}$. Let ${{\mathcal{L}}}_w$ be the language of all words occurring in the infinite word $w$. Then the additive complexity is $\phi^+(w, n)={\rm Card} \{{{\rm S}}(u): u\in {{\mathcal{L}}}_w^n\}$, where ${{\rm S}}$ is the identity map on the alphabet of $w$. We finally mention that homomorphisms ${{\rm S}}$ from a language to the natural numbers already occur in the 1972 paper [@Car-Sco-Hog Section 6] in the context of the Fibonacci language, where they are called weights. Homomorphic images of the golden-mean language {#sec:gold} ============================================== The golden mean language is the language ${{\mathcal{L}}}_{ \rm GM}$ consisting of all words over $\{a,b\}$ in which $bb$ does not occur as a subword. Now if ${{\rm S}}$ satisfies ${{\rm S}}(a)=1$ or ${{\rm S}}(b)=1$, then it is easily seen that ${{\rm S}}({{\mathcal{L}}}_{ \rm GM})={{\mathbb{N}}}$, so for these homomorphisms the golden mean and the full language both map to $\mathbb{N}$. One could say they both have Frobenius number $0$. In general however, the Frobenius number will increase substantially. If we take ${{\rm S}}$ defined by $${{\rm S}}(a)=100,\; {{\rm S}}(b)=3,$$ then the Frobenius number of the full language under ${{\rm S}}$ is $300-100-3=197$, and the Frobenius number of ${{\rm S}}({{\mathcal{L}}}_{ \rm GM})$ is equal to 9997. For arbitrary homomorphisms the solution of the Frobenius problem for the golden mean language is given by the following, where we write ${{\rm S}}_a:={{\rm S}}(a),\, {{\rm S}}_b:={{\rm S}}(b)$. \[th:Goldmean\] Let ${{\rm S}}: {{\mathcal{L}}}_{ \rm GM}\rightarrow {{\mathbb{N}}}$ be a homomorphism. Suppose $\gcd({{\rm S}}_a,{{\rm S}}_b)=1$, and both ${{\rm S}}_a>1$ and ${{\rm S}}_b>1$. Then the Frobenius number of ${{\rm S}}({{\mathcal{L}}}_{ \rm GM})$ is equal to $$\max\: {{\mathbb{N}}}\setminus {{\rm S}}({{\mathcal{L}}}_{ \rm GM})= {{\rm S}}_a({{\rm S}}_a-3) +{{\rm S}}_b({{\rm S}}_a-1).$$ Proof: Let an ${{\rm S}}_a$-$point$ be defined as a multiple $n{{\rm S}}_a$, $n=0,1,\dots$, and an ${{\rm S}}_a$-$interval$ as the set of numbers between two consecutive ${{\rm S}}_a$-points. We also consider ${{\rm S}}_b$-$chains$, defined for $n\ge0$ by $$C(n)=\{n{{\rm S}}_a+{{\rm S}}_b,\, n{{\rm S}}_a+2{{\rm S}}_b,\dots, n{{\rm S}}_a+(n+1){{\rm S}}_b\}.$$ Note that the union of the ${{\rm S}}_a$-points and the ${{\rm S}}_b$-chains will give ${{\mathcal{L}}}_{ \rm GM}$. The key observation is that the ${{\rm S}}_b$-chain $C({{\rm S}}_a-2)$ has ${{\rm S}}_a-1$ elements, which are all different modulo ${{\rm S}}_a$. This is a consequence of $\gcd({{\rm S}}_a,{{\rm S}}_b)=1$. It follows that the ${{\rm S}}_b$-chains fill in more and more points of the ${{\rm S}}_a$-intervals. The last point to be filled in is modulo ${{\rm S}}_a$ equal to ${{\rm S}}_a-{{\rm S}}_b$, produced by the last element of the chain $C({{\rm S}}_a-2)$. This is the number $$P:=({{\rm S}}_a-2){{\rm S}}_a+({{\rm S}}_a-1){{\rm S}}_b.$$ But then the largest number in the complement of ${{\mathcal{L}}}_{ \rm GM}$ is $P-{{\rm S}}_a$, which is the number as claimed in the theorem. In this argument we used that if a point in an ${{\rm S}}_a$-interval is filled in, then the corresponding points modulo ${{\rm S}}_a$ in all later intervals will also be filled in, simply because the later chains will be extensions of the earlier ones. $\Box$ ![Example with ${{\rm S}}(a) = 7, {{\rm S}}(b) = 3$: row $n$ shows the ${{\rm S}}_a$-points in blue, and the ${{\rm S}}_b$-chain $C(n-1)$ in yellow and green, for $n = 1,\dots,8$ (truncated at 56).[]{data-label="gold"}](Golden-mean-language-TikZ-37.pdf){height="3.5cm"} Sturmian languages {#sec:Sturm} ================== Sturmian words are infinite words over a two letter alphabet that haven exactly $n+1$ subwords for each $n=1,2,\dots$. We call the collection of these subwords a Sturmian language. There is a surprising characterization of Sturmian words: $s$ is Sturmian if and only if $s$ is irrational mechanical, which means that there exists an irrational number $\alpha\in (0,1)$ and a number $\rho$ such that $s=s_{\alpha,\rho}$, or $s=s'_{\alpha,\rho}$, where $$s_{\alpha,\rho}=\big([ (n+1)\,\alpha +\rho]-[ n\,\alpha +\rho]\big)_{n\ge 0},\; s'_{\alpha,\rho}=\big(\lceil (n+1)\,\alpha +\rho\rceil-\lceil n\,\alpha +\rho\rceil\big)_{n\ge 0}.$$ See, e.g., [@Lothaire Prop. 2.1.13]. Because of this representation, we will use the alphabet $\{0,1\}$ instead of $\{a,b\}$ in this section. Of special interest are the Sturmian words $s_\alpha:=s_{\alpha,0}$ and $s'_\alpha:=s'_{\alpha,0}$ of intercept 0. These have the property that they only differ in the first element: $$s_\alpha=0\,c_\alpha, \qquad s'_\alpha=1\,c_\alpha.$$ Here $c_\alpha:=s_{\alpha,\alpha}$ is called the characteristic word of $\alpha$. For $n\ge 0$ we have $$c_\alpha(n)=s_{\alpha,\alpha}(n)=[ (n+1)\,\alpha +\alpha]-[ n\,\alpha +\alpha] =[ (n+2)\,\alpha ]-[ (n+1)\,\alpha ].$$ The words $s_\alpha, s'_\alpha$ and $c_\alpha$ generate the same language ([@Lothaire Prop.2.1.18] ), which we denote ${{\mathcal{L}}}_\alpha$. Recall that ${{\mathcal{L}}}^n_\alpha$ is the set of words of length $n$ in ${{\mathcal{L}}}_\alpha$. \[lem:compl\] Let ${{\mathcal{L}}}_\alpha$ be a Sturmian language, and let ${{\rm S}}$ be a homomorphism with ${{\rm S}}(0)\ne{{\rm S}}(1)$. Then ${\rm Card} \,{{\rm S}}({{\mathcal{L}}}^n_\alpha)=2$ for all $n\ge 1$. Proof: This follows directly from the fact ([@Lothaire Th.2.1.5]) that Sturmian words are balanced, i.e., any two words of the same length can at most differ 1 in their number of ones.$\Box$ A sequence $([n\alpha])$, where $[.]$ denotes integer part, is called a Beatty sequence if $\alpha>1$, and a slow Beatty sequence if $0<\alpha<1$ (terminology from [@KimStol]). \[th:Sturm\] Let $\alpha$ be an irrational number from $(0,1)$. Let ${{\mathcal{L}}}_\alpha$ be the Sturmian language generated by $\alpha$, and let $(q_n)_{n\ge 0}$ be the slow Beatty sequence defined by $$q_n=[(n+1)\alpha].$$ Let ${{\rm S}}: {{\mathcal{L}}}_\alpha\rightarrow {{\mathbb{N}}}$ be a homomorphism. Define ${{\rm S}}_0={{\rm S}}(0), {{\rm S}}_1={{\rm S}}(1)$. Then $${{\rm S}}({{\mathcal{L}}}_\alpha) = \{({{\rm S}}_1-{{\rm S}}_0)q_n+n{{\rm S}}_0+{{\rm S}}_0: \; n=0,\dots\} \cup \{({{\rm S}}_1-{{\rm S}}_0)q_n+n{{\rm S}}_0+{{\rm S}}_1: \; n=0,\dots\}.$$ Proof: If ${{\rm S}}_0={{\rm S}}_1$ then this is certainly true, so suppose ${{\rm S}}_0\ne{{\rm S}}_1$ in the sequel. We denote $c_\alpha[i,j]:=c_\alpha(i)\dots c_\alpha(j)$ for integers $0\le i<j$. Let $N_{\ell}(w)$ denote the number of occurrences of the letter $\ell$ in a word $w$ for $\ell=0,1$. Then $$N_1(c_\alpha[0, n-1])=\sum_{k=0}^{n-1} c_\alpha(k)=[(n+1)\alpha]-[\alpha]=q_n, \quad N_0(c_\alpha[0,n-1])=n-q_n.$$ Of course all words $c_\alpha[0,n-1]$ are in the Sturmian language ${{\mathcal{L}}}_\alpha$, but ${{\mathcal{L}}}_\alpha$ also contains the words $0c_\alpha[0,n-1]$ and $1c_\alpha[0,n-1]$. It thus follows from Lemma \[lem:compl\] that ${{\rm S}}({{\mathcal{L}}}_\alpha)$ is given by the union of all images ${{\rm S}}(0c_\alpha[0,n-1])$ and ${{\rm S}}(1c_\alpha[0,n-1])$. Since $${{\rm S}}(0c_\alpha[0,n-1])={{\rm S}}_0+(n-q_n){{\rm S}}_0+q_n{{\rm S}}_1=({{\rm S}}_1-{{\rm S}}_0)q_n+n{{\rm S}}_0+{{\rm S}}_0,$$ the result follows.$\Box$ The Fibonacci language ----------------------- Let $\Phi=(\sqrt{5}+1)/2=1.61803\dots$ be the golden mean, and let $\alpha:=2-\Phi$. We have $$c_\alpha=([(n+1)\alpha]-[n\alpha])_{n\ge 1}= 0, 1, 0, 0, 1, 0, 1, 0, 0, 1, 0, 0, 1, 0, 1, 0, 0,\dots,$$ the infinite Fibonacci word. We write ${{\mathcal{L}}}_{\rm F}:={{\mathcal{L}}}_\alpha$. \[th:Fib\] Let ${{\rm S}}: {{\mathcal{L}}}_{\rm F}\rightarrow {{\mathbb{N}}}$ be a homomorphism. Then $${{\rm S}}({{\mathcal{L}}}_{\rm F})=\big(({{\rm S}}_0-{{\rm S}}_1)[n\Phi]+(2{{\rm S}}_1-{{\rm S}}_0)n+{{\rm S}}_0-{{\rm S}}_1\big)_{n\ge 1} \,\cup\, \big(({{\rm S}}_0-{{\rm S}}_1)[n\Phi]+(2{{\rm S}}_1-{{\rm S}}_0)n\big)_{n\ge 1}.$$ Proof: This is a corollary to Theorem \[th:Sturm\], using $[-x]=-[x]-1$ for non-integer $x$: $$\begin{aligned} ({{\rm S}}_1-{{\rm S}}_0)q_{n-1}+n{{\rm S}}_0 &=& ({{\rm S}}_1-{{\rm S}}_0)[n\alpha]+n{{\rm S}}_0= ({{\rm S}}_1-{{\rm S}}_0)[n(2-\Phi)]+n{{\rm S}}_0\\ &=& 2({{\rm S}}_1-{{\rm S}}_0)n+({{\rm S}}_1-{{\rm S}}_0)[-n\Phi]+n{{\rm S}}_0\\ &=& (2{{\rm S}}_1-{{\rm S}}_0)n+({{\rm S}}_1-{{\rm S}}_0)(-[n\Phi]-1)\\ &=& ({{\rm S}}_0-{{\rm S}}_1)[n\Phi]+(2{{\rm S}}_1-{{\rm S}}_0)n+{{\rm S}}_0-{{\rm S}}_1.\hspace{4cm}\Box\end{aligned}$$ \[lem:empty\] For ${{\rm S}}(0)=1,\, {{\rm S}}(1)\le 3$ or ${{\rm S}}(0)=2,\, {{\rm S}}(1)=1$ one has ${{\rm S}}({{\mathcal{L}}}_{\rm F})={{\mathbb{N}}}$. Proof:* Take $({{\rm S}}_0,{{\rm S}}_1)=$ (1,1). Then obviously ${{\rm S}}({{\mathcal{L}}}_{\rm F})={{\mathbb{N}}}$. Take $({{\rm S}}_0,{{\rm S}}_1)=$ (2,1). Then ${{\rm S}}({{\mathcal{L}}}_{\rm F})={{\mathbb{N}}}$, since by Theorem \[th:Fib\] ${{\rm S}}({{\mathcal{L}}}_{\rm F})$ is the union of $([n\Phi])$ and $([n\Phi]+1)$, where the difference of two consecutive terms in $([n\Phi])$ is never more than 2. Take $({{\rm S}}_0,{{\rm S}}_1)=$ (1,2). Then ${{\rm S}}({{\mathcal{L}}}_{\rm F})={{\mathbb{N}}}$, since ${{\rm S}}({{\mathcal{L}}}_{\rm F})$ is the union of $([n(3-\Phi)])$ and $([n(3-\Phi)])+1)$, where the difference of two consecutive terms in $([n(3-\Phi)])$ is never more than 2. Take $({{\rm S}}_0,{{\rm S}}_1)=$ (1,3). This case is more complicated. Let $u:=(-2[n\Phi]+5n-2)_{n\ge 1}$, and $v:=u+2$. Then according to Theorem \[th:Fib\], the union of the sets determined by $u$ and $v$ is ${{\rm S}}({{\mathcal{L}}}_{\rm F})$. Let $\Delta u$ be the difference sequence defined by $\Delta u_n= u_{n+1}-u_n$ for $n \ge 0$. It is easy to see that the difference sequences $\Delta v$ and $\Delta u$ are both equal to the Fibonacci sequence $1,3,1,1,3,1,\dots$ on the alphabet $\{1,3\}$ (cf. [@AllDekk]). We claim that if two consecutive numbers $m , m+1$ are missing in $u$, then these two do appear in $v$, implying that ${{\rm S}}({{\mathcal{L}}}_{\rm F})={{\mathbb{N}}}$. Indeed the two missing numbers are characterized by $u_{n+1}-u_n=3$ for some $n$, and the missing numbers are $m=u_n+1$ and $u_n+2$. The second number appears in $v$, simply because $v=u+2$. The first number appears because $u_{n+1}-u_n=3$ implies $u_n-u_{n-1}=1$ (no 33 in the 1-3-Fibonacci sequence), and so $v_{n-1}=v_n-1=u_n+1$. $\Box$ We define $\mathcal{E}:=\{(1,1), (1,2), (1,3), (2,1)\}$. \[th:card\] Let ${{\rm S}}: {{\mathcal{L}}}_{\rm F}\rightarrow {{\mathbb{N}}}$ be a homomorphism. Then ${{\mathbb{N}}}\setminus {{\rm S}}({{\mathcal{L}}}_{\rm F})$ has infinite cardinality, unless $({{\rm S}}(0),{{\rm S}}(1))\in \mathcal{E}$, in which case the complement is empty. Proof: According to Lemma \[lem:empty\] the complement of ${{\rm S}}({{\mathcal{L}}}_{\rm F})$ is empty for $({{\rm S}}_0,{{\rm S}}_1)\in \mathcal{E}$. The density of the set ${{\rm S}}({{\mathcal{L}}}_{\rm F})$ in the natural numbers exists, and equals $$\delta:=\frac2{({{\rm S}}_0-{{\rm S}}_1)\Phi+\,2{{\rm S}}_1-{{\rm S}}_0}.$$ The theorem will be proved if we show that $\delta<1$ for $({{\rm S}}_0,{{\rm S}}_1)$ not in $\mathcal{E}$. First we note that the denominator of $\delta$ is positive: $$({{\rm S}}_0-{{\rm S}}_1)(\Phi-1)+{{\rm S}}_1>-{{\rm S}}_1(\Phi-1)+{{\rm S}}_1={{\rm S}}_1(2-\Phi)>0,$$ where we used that $1<\Phi<2$. We now have $$\delta<1\;\Leftrightarrow\; ({{\rm S}}_0-{{\rm S}}_1)\Phi+\,2{{\rm S}}_1-{{\rm S}}_0 >2 \;\Leftrightarrow\; ({{\rm S}}_0-{{\rm S}}_1)\Phi > {{\rm S}}_0-{{\rm S}}_1 +2-{{\rm S}}_1.$$ If ${{\rm S}}_0>{{\rm S}}_1$, this is satisfied, since under this condition $(2-{{\rm S}}_1)/({{\rm S}}_0-{{\rm S}}_1)\le 0$, unless $({{\rm S}}_0,{{\rm S}}_1)=(2,1)\in \mathcal{E}$. If ${{\rm S}}_0<{{\rm S}}_1$, we have to see that $\Phi < 1 +(2-{{\rm S}}_1)/({{\rm S}}_0-{{\rm S}}_1)$. This holds for ${{\rm S}}_0\ge2$, since then $(2-{{\rm S}}_1)/({{\rm S}}_0-{{\rm S}}_1)\ge 1$. If ${{\rm S}}_0=1$, then this does not hold for ${{\rm S}}_1=1,2,3$, i.e., for pairs from $\mathcal{E}$, but it will hold for all ${{\rm S}}_1\ge 4$.$\Box$ For particular values of ${{\rm S}}(0)$ and ${{\rm S}}(1)$ the complement of the embedding of the language has a nice structure, as it can be expressed in the classical Beatty sequences $A(n)=[ n\Phi ]$ for $n\ge 1$, and $ B(n)=[ n\Phi^2]$ for $n\ge 1$. The sequences $A$ and $B$ are called the lower Wythoff sequence and upper Wythoff sequence; they are extremely well-studied. [Example 1]{}. Let ${{\rm S}}$ be given by ${{\rm S}}(0)=3$ and ${{\rm S}}(1)=2$. In the following we use the notation $pX+qY+r= (pX(n)+qY(n)+r)_{n\ge 1}$ for real numbers $p,q,r$ and functions $X,Y:{{\mathbb{N}}}\rightarrow{{\mathbb{N}}}$. Then ${{\rm S}}({{\mathcal{L}}}_{\rm F})= B({{\mathbb{N}}}) \:\cup\: B\!+\!1\,({{\mathbb{N}}}), \quad {{\mathbb{N}}}\!\setminus{{\rm S}}({{\mathcal{L}}}_{\rm F})= \{1,4, 9, 12, \dots\}= 2A+{ {\rm Id}}+1\,({{\mathbb{N}}}\cup \{0\}).$ The first statement follows directly from Theorem \[th:Fib\]. The second statement follows in a number of steps from the fact that $A$ and $B$ form a Beatty pair: $A({{\mathbb{N}}})\cap B({{\mathbb{N}}})=\emptyset$, and $A({{\mathbb{N}}})\cup B({{\mathbb{N}}})={{\mathbb{N}}}$. This implies that $A(A({{\mathbb{N}}}))\cup A(B({{\mathbb{N}}}))\cup B({{\mathbb{N}}})={{\mathbb{N}}}$, where the three sets are disjoint. But $AA=B-1$ (see, e.g., Formula (3.2) in [@Car-Sco-Hog]). Adding 1 to all three sequences it follows that $B({{\mathbb{N}}}) \:\cup\: B+1\,({{\mathbb{N}}}) \:\cup\: AB+1\,({{\mathbb{N}}}) = {{\mathbb{N}}}\setminus \{1\}.$ Moreover, according to [@Car-Sco-Hog Formula (3.5)] one has $ AB=A+B=2A+{ {\rm Id}}$. But then the three sequences $([ n\Phi ]+n)_{n\ge 1},\, ([ n\Phi ]+n+1)_{n\ge 1},\, (2[ n\Phi ]+n+1)_{n\ge 0}$, form a complementary triple, i.e., as sets they are disjoint, and their union is ${{\mathbb{N}}}$. A similar result holds for[^2] ${{\rm S}}(0)=4,\, {{\rm S}}(1)=3$. [Example 2]{}. Let ${{\rm S}}$ be given by ${{\rm S}}(0)=3$ and ${{\rm S}}(1)=1$, then by Theorem \[th:Fib\] ${{\rm S}}({{\mathcal{L}}}_{\rm F})= 2A-{ {\rm Id}}\,({{\mathbb{N}}}) \:\cup\: 2A-{ {\rm Id}}+2\,({{\mathbb{N}}})$. It is proved in [@AllDekk] that ${{\mathbb{N}}}\setminus{{\rm S}}({{\mathcal{L}}}_{\rm F})= \{2, 9, 20, 27, 38, 49, \dots\}= 4A+3{ {\rm Id}}+2\,(\mathbb{N} \cup \{0\}),$ and that the three sequences $(2[ n\Phi ]-n)_{n\ge 1},\, (2[ n\Phi ]-n+2)_{n\ge 1},\, (4[ n\Phi ]+3n+2)_{n\ge 0}$, form a complementary triple. The Thue-Morse language {#sec:TM} ======================= Let $\theta$ given by $\theta(a)=ab,\, \theta(b)=ba$ be the Thue-Morse morphism. Let ${{\mathcal{L}}}_{\rm TM}$ be the language generated by this morphism. Let $R_{r,s}=\{s,r+s,2r+s,\dots\}$ be the set determined by the arithmetic sequence with terms $rn+s$ for $n=0,1\dots$. \[th:TM\] Let ${{\rm S}}: {{\mathcal{L}}}_{\rm TM}\rightarrow {{\mathbb{N}}}$ be a homomorphism. Define $p={{\rm S}}(0), q={{\rm S}}(1)$. Then $${{\rm S}}({{\mathcal{L}}}_{\rm TM})=R_{p+q,0}\cup R_{p+q,p}\cup R_{p+q,q}\cup R_{p+q,2p}\cup R_{p+q,2q}.$$ Proof: Let ${{\mathcal{L}}}_{\rm TM}^n$ be the set of words of length $n$ in the Thue-Morse language. Put $r={{\rm S}}(ab)=p+q.$ It is clear (and for $p=0, q=1$ observed also in [@RSZ]) that since the Thue-Morse word is a non-periodic concatenation of $ab$ and $ba$ that for $n=1,2,\dots$ $${{\rm S}}({{\mathcal{L}}}_{\rm TM}^{2n})=\{rn, rn+q-p, rn+p-q\},\quad {{\rm S}}({{\mathcal{L}}}_{\rm TM}^{2n-1})=\{rn+p, rn+q\}.$$ This implies the statement of the theorem.$\Box$ \[th:cardTM\] Let ${{\rm S}}: {{\mathcal{L}}}_{\rm TM}\rightarrow {{\mathbb{N}}}$ be a homomorphism. Then ${{\mathbb{N}}}\setminus {{\rm S}}({{\mathcal{L}}}_{\rm F})$ has infinite cardinality if and only if ${{\rm S}}(a)+{{\rm S}}(b)\ge 6$. For ${{\rm S}}(a)+{{\rm S}}(b) < 6$, the complement is either empty or a singleton. Proof: This follows directly form Theorem \[th:TM\]. If ${{\rm S}}(a)+{{\rm S}}(b)\ge 6$, then the density of ${{\mathbb{N}}}\setminus{{\rm S}}({{\mathcal{L}}}_{\rm TM})$, is at least 1/6, so the set has infinite cardinality. The results for ${{\rm S}}(a)+{{\rm S}}(b) < 6$ follow also directly from the previous theorem. $\Box$ [Remark]{} Let $\sigma$ given by $\sigma(a)=ab, \sigma(b)=aa$ be the period-doubling or Toeplitz morphism. The difficulty—see [@Karhumaki-S-K Lemma 6]—of determining the abelian complexity of the period-doubling morphism already indicates that solving the Frobenius problem for the period-doubling language will be much more involved than for the Thue-Morse language. Two dimensional embeddings {#sec:twodim} ========================== Here we consider homomorphisms ${{\rm S}}: {{\mathcal{L}}}\rightarrow {{\mathbb{N}}}\times {{\mathbb{N}}}$ and ${{\rm S}}: {{\mathcal{L}}}\rightarrow {{\mathbb{Z}}}\times{{\mathbb{Z}}}$. The situation changes drastically for this ‘double-coin’ problem. \[prop:2D\] Let ${{\mathcal{L}}}$ be a language on the alphabet $\{a,b\}$, and let ${{\rm S}}: {{\mathcal{L}}}\rightarrow {{\mathbb{N}}}\times {{\mathbb{N}}}$ be a homomorphism. Then ${{\mathbb{N}}}\times{{\mathbb{N}}}\,\setminus\, {{\rm S}}({{\mathcal{L}}})$ has infinite cardinality for all pairs $\{{{\rm S}}(a), {{\rm S}}(b)\}$ which are not equal to the pair $\{{ \rm (0,1),(1,0)}\}$. Proof: It suffices to prove this for the full language ${{\mathcal{L}}}_{\rm full}$. The image under ${{\rm S}}$ is an integer lattice, with a complement of infinite cardinality, unless ${{\rm S}}(a)$ and ${{\rm S}}(b)$ are the unit vectors. $\Box$ We learn from this that the alphabet is ‘too small’, and that we should rather consider embeddings in ${{\mathbb{Z}}}\times{{\mathbb{Z}}}$ instead of ${{\mathbb{N}}}\times {{\mathbb{N}}}$. We focus again on low complexity languages, in particular on those generated by a primitive morphism $\varphi$ on an alphabet $A$. Such a morphism has a language ${{\mathcal{L}}}_\varphi$ associated to it, where each word $w\in {{\mathcal{L}}}_\phi$ has a measure $\mu_\varphi(w)$. For a given homomorphism ${{\rm S}}:{{\mathcal{L}}}_\varphi\rightarrow {{\mathbb{Z}}}\times{{\mathbb{Z}}}$ we call the average $$\Delta_\varphi({{\rm S}}):=\sum_{a\in A} \mu_\varphi(a){{\rm S}}(a)\vspace{-.5cm}$$ the drift* of ${{\rm S}}$. \[prop:drift\] Let ${{\mathcal{L}}}_\varphi$ be a language generated by primitive morphism on an alphabet $A$, and let ${{\rm S}}: {{\mathcal{L}}}_\varphi\rightarrow {{\mathbb{Z}}}\times {{\mathbb{Z}}}$ be a homomorphism. Then ${{\mathbb{Z}}}\times{{\mathbb{Z}}}\,\setminus\, {{\rm S}}({{\mathcal{L}}})$ has infinite cardinality if $\Delta_\varphi({{\rm S}})\ne (0,0)$. Proof: It is well-known that the measure $\mu_\varphi$ is strictly ergodic. Because of this, we have for words $w$ from ${{\mathcal{L}}}_\varphi$, where $\|w\|$ denotes the length of $w$, $$\frac1{\|w\|}{{\rm S}}(w)=\frac1{\|w\|}\sum_{a\in A} N_a(w){{\rm S}}(a)\rightarrow \sum_{a\in A} \mu_\varphi(a){{\rm S}}(a)=\Delta_\varphi({{\rm S}}) {\:\rm as}\: \|w\|\rightarrow\infty.$$ Thus for long words $w$ the images ${{\rm S}}(w)$ will be concentrated around the line in the direction of the drift of ${{\rm S}}$, and so the complement of ${{\rm S}}({{\mathcal{L}}}_\varphi)$ will have infinite cardinality if the drift is not $(0,0)$. $\Box$ Can we say something about the Frobenius problem for homomorphic images of morphic languages of an embedding with drift $(0,0)$? We shall give an infinite family of morphic languages ${{\mathcal{L}}}_\theta$ on an alphabet $A=\{a,b,c,d\}$ of four letters where for the homomorphism ${{\rm S}}^\oplus$ given by $${{\rm S}}^\oplus(a)=(1,0),\; {{\rm S}}^\oplus(b)=(0,1),\;{{\rm S}}^\oplus(c)=(-1,0),\; {{\rm S}}^\oplus(d)=(0,-1)$$ the homomorphic embedding is the whole $ {{\mathbb{Z}}}\times{{\mathbb{Z}}}$—and thus the complement is empty. We shall make use of the paperfolding morphisms introduced in [@Dek-paper]. Let $\sigma$ be the rotation morphism on the alphabet $\{a,b,c,d\}$ given by $\sigma(a)=b,\; \sigma(b)=c,\; \sigma(c)=d,\; \sigma(d)=a,$ and let $\tau$ be the anti-morphism given by $\tau(w_1\dots w_n)=w_n\dots w_1$. A morphism $\theta$ on $\{a,b,c,d\}$ is called a paperfolding morphism if 1\) $\sigma \tau \theta=\theta$, 2\) Letters from $\{a,c\}$ alternate[^3] with letters from $\{b,d\}$ in $\theta(a)$. A paperfolding morphism is called symmetric if $\sigma \theta=\theta$. It is clear that this happens if and only if the word $\theta(a)$ is a palindrome. Let $G$ be a (semi-) group with operation $+$ and unit $e$. In general an infinite word $x=(x_n)$ over an alphabet $A$ and a homomorphism ${{\rm S}}:A^\rightarrow G$ generate a walk* $Z=(Z_n)_{n\ge 0}$ by (cf. [@Dek-Marches]) $$Z_0=e, \qquad Z_{n+1}=Z_{n}+{{\rm S}}(x_n)={{\rm S}}(x_0\dots x_n),\:{\rm for}\: n\ge 0.$$ A paperfolding morphism $\theta$ with $\theta(a)=a...$ is called perfect if the four walks generated by the fixed point $x=\theta^\infty(a)$, and its three rotations over $\pi/2, \pi$ and $3\pi/2$ visit every integer point in the plane exactly twice (except the origin, which is visited 4 times). In [@Dek-paper] it is—not explicitly—proved that for any odd integer $N$ that is the sum of two squares there exists a perfect symmetric paperfolding morphism of length $N$. To make the proof explicit, one uses that according to the paragraph at the end of Section 7 in [@Dek-paper] there exists a symmetric planefilling and self-avoiding string for each such $N$, and then one observes that the construction of such a string in the proof of [@Dek-paper Theorem 4] always satisfies the perfectness criterion given in [@Dek-paper Theorem 5]. The smallest length is $N=5$, with morphism $\theta$ given by $$\theta(a)=abcba,\; \theta(b)=bcdcb, \; \theta(c)=cdadc, \; \theta(d)=dabad.$$ ![The four images of the words $\theta^4(a),\dots,\theta^4(d)$ under ${{\rm S}}^\oplus$, where $\theta$ is the perfect symmetric 5-folding morphism. The origin is not covered, but it is the image of the word $abcd\in {{\mathcal{L}}}_\theta$.[]{data-label="fold5"}](5-fold-points-ZXZ.jpg){width="6cm"} \[prop:drift\] Let ${{\mathcal{L}}}_\theta$ be the language generated by a perfect symmetric paperfolding morphism $\theta$. Then ${{\rm S}}^\oplus({{\mathcal{L}}}_\theta)= {{\mathbb{Z}}}\times {{\mathbb{Z}}}$. Proof: This follows directly from Theorem 5 in [@Dek-paper], using the observation above. $\Box$ [3]{} J.-P. Allouche and F. M. Dekking, Beatty sequences and complementary triples, In preparation. H. Ardal, T. Brown, V. Jungi´c, J. Sahasrabudhe, On additive and Abelian complexity in infinite Words, Integers 12 (2012). F. Blanchet-Sadri, Daniel Seita, and David Wise, Computing abelian complexity of binary uniform morphic words, Theoretical Computer Science 640 (2016) 41–-51. L. Carlitz, Richard Scoville, and V.E.Hogatt jr, Fibonacci representations, Fib. Quart. 10 (1972),1–28. F.M.Dekking, Marches automatiques, J. Théor. Nombres Bordeaux 5 (1993), 93–100. Paperfolding morphisms, planefilling curves, and fractal tiles, Theoretical Computer Science 414 (2012), 20–37. Morphisms, symbolic sequences, and their standard forms, Journal of Integer Sequences 19 (2016), Article 16.1.1, 1-–8. J. Karhumäki, A. Saarela, and L. Q. Zamboni. Variations of the [M]{}orse-[H]{}edlund Theorem for k-abelian equivalence. Developments in Language Theory, Proceedings, Ekaterinburg, Russia, 26–29 August 2014, Volume 8633 of the series Lecture Notes in Computer Science, 203–214. (http://arxiv.org/abs/1302.3783 , 2013.) C. Kimberling and K. B. Stolarsky, Slow Beatty sequences, devious convergence, and partitional divergence, Amer. Math. Monthly, 123 (No. 2, 2016), 267-273. M. Lothaire, Algebraic Combinatorics on Words, Encyclopedia of Mathematics and its Applications 90, Cambridge University Press, 2002. The On-Line Encyclopedia of Integer Sequences, founded by N.J.A Sloane, sequences A276885 and A276886. Gwénaël Richomme, Kalle Saari, and Luca Q. Zamboni, Abelian complexity of minimal subshifts, J. Lond. Math. Soc. (2) 83 (2011), no. 1, 79–95. MR 2763945, https://doi.org/10.1112/jlms/jdq063 Julian Sahasrabudhe, Sturmian words and constant additive complexity, Integers 15 (2015). Sylvester, James Joseph, “Question 7382". Mathematical Questions from the Educational Times 41 (1884), 21. [^1]: Languages defined by the labelling of infinite paths of an automaton. [^2]: In these two cases ${{\mathbb{N}}}\setminus{{\rm S}}({{\mathcal{L}}}_{\rm F})$ is given by sequences A276885, respectively A276886 in OEIS ([@OEIS-Fib]). It is easily seen that the definitions of these sequences in OEIS are equivalent to the way in which we obtain them. [^3]: This corrects an omission in [@Dek-paper Definition 1].	{ "pile_set_name": "ArXiv" }
--- abstract: 'We present in this talk the recent investigation of the effective weak chiral Lagrangian within the framework of the instanton-induced chiral quark model. The low energy constants in leading order $g_8$ and $g_{27}$, and their ratio are determined. The momentum-dependence of the constituent quark mass turns out to improve the low energy constants to the great extent, respecting the $\Delta T=1/2$ enhancement.' address: - ' Institute for Theoretical Physics II, Ruhr-University Bochum, D-44780 Bochum, Germany' - ' Department of Physics, Pusan National University, 609-735 Pusan, Republic of Korea' author: - 'Mario Franz [^1], Hyun-Chul Kim[^2][^3], and Klaus Goeke$^{\rm a}$ [^4]' title: \| Effective Weak Chiral Lagrangian from the Instanton Vacuum [^5]\ -4cm PNU-NTG-03/99\ RUB-TPII-10/99\ July 1999 1.5cm --- ------------------------------------------------------------------------ [1.]{} Understanding the $\Delta T = 1/2$ selection rule, best known as the fact that the isospin amplitude of the $K\rightarrow \pi\pi$ decay is about 22 times larger than the $T=2$ amplitude, remains still elusive in hadronic physics. Despite a great deal of effort this dominance of the $\Delta T=1/2$ channel over the $\Delta T=3/2$ one has not been explained in a satisfactory manner. While a part of the answer comes from perturbative gluons created by the evolution from a scale of $M_W\simeq 80$ GeV to around 1 GeV, another part of the answer is supposed to arise from the structure of the light hadrons, whose description at scales around 1 GeV requires a method of nonperturbative QCD. In chiral perturbation theory ($\chi$PT) [@Kamboretal; @Esposito; @EKW] the $\Delta T=1/2$ enhancement is represented by the ratio of the low energy constants $g_8$ and $g_{27}$: $\|g_8 +\frac15 g_{27}\|_{\rm exp}\simeq 5.1, \;\|g_{27}\|_{\rm exp}\simeq 0.16$ [@PichRafael]. Kambor [et al.]{} showed that the octet coupling constant is decreased by $30\%$ with the chiral loops considered. So, the ratio is reduced to about $22$. However, the inclusion of the chiral loops ${\cal O}(p^4)$ introduces so many low energy constants (LECs) that it is not possible to constrain the LECs to order ${\cal O}(p^4)$ based on the experimental data to date. One need to resort to effective QCD-inspired models in order to proceed without relying on experimental results. Recently, M. Franz [et al.]{} [@FranzKimGoeke] have investigated the effective $\Delta S=1,2$ weak chiral Lagrangian to order ${\cal O}(p^4)$ within the framework of the chiral quark model ($\chi$QM), focusing on determining the LECs in the effective weak chiral Lagrangian. However, the ratio of the $g_8/g_{27}$ obtained from the $\chi$QM are deviated from the phenomenological values from $\chi$PT. In this talk, we want to present the recent improvement of the former study [@FranzKimGoeke], considering the momentum dependence of the mass of the constituent quark which considered as a constant in the former work. The momentum-dependent dynamical quark mass arises from the picture of the instanton vacuum which pertains to nonperturbative QCD. The instanton vacuum elucidates one of the most important low-energy properties of QCD, [i.e.]{} the mechanism of spontaneous breaking of chiral symmetry [@DP], which brings about the momentum-dependent quark mass. [2.]{} The low-energy QCD partition function in Euclidean space can be written as $${\cal Z}\;=\; \int {\cal D} \psi {\cal D} \psi^\dagger {\cal D} \pi^a \exp \left[\int d^4 x \psi^{\dagger \alpha}_{f} \left(i\rlap{/}{\partial} + i\sqrt{M(-i\partial)} U^{\gamma_5} \sqrt{M(-i\partial)}\right)_{fg} \psi^{\alpha}_{g} \right], \label{Eq:Dirac}$$ The $M(k)$ is the momentum-dependent constituent quark mass expressed as follows: $$M(k) \;=\; M_0 F(k)^2$$ The $F(k)$ is normalized to unity at $k=0$. Thus, $M_0$ is the value of the $M(k)$ at $k=0$. The momentum-dependent constituent quark mass can be regarded as a UV regulator. The pion decay constant $f_\pi$ is obtained as follows: $$f_{\pi}^2 \;=\; 4N_c\int \frac{d^4 k}{(2\pi)^4} \frac{M^2(k) - \frac12 M(k) \frac{dM(k)}{dk}k + \frac14 \left(\frac{dM(k)}{dk}\right)^2 k^2}{(k^2 +M^2(k))^2} . \label{Eq:fpiq}$$ The $M(k)$ is contrained by reproducing the experimental value of $f_\pi=93$ MeV. The quark and gluon condensates are written, respectively, by $$\langle \bar{\psi}\psi\rangle\;=\; -4 N_c \int \frac{d^4 k}{(2\pi)^4}\frac{M(k)}{k^2 +M^2(k)},\;\;\; \left\langle\frac{\alpha_s}{4\pi}G^a_{\mu\nu}G^{a\mu\nu}\right\rangle \;=\; 8 N_c \int \frac{d^4 k}{(2\pi)^4}\frac{M^2(k)}{k^2 +M^2(k)}. \label{Eq:condensate}$$ The effective weak chiral Lagrangian $S^{\Delta S = 1}_{\rm eff}[\pi]$ in lowest order of $G_F$ can be obtained as follows: $${\cal L}^{\Delta S = 1}_{\rm eff} \;=\; -\frac{1}{\cal N} \int {\cal D} \psi {\cal D} \psi^\dagger {\cal H}^{\Delta S = 1}_{\rm eff} \exp \left[\int d^4 x \psi^\dagger \left( i\rlap{/}{\partial} + i\sqrt{M(-i\partial)} U^{\gamma_5} \sqrt{M(-i\partial)} \right)\psi\right]. \label{Eq:part1}$$ Here the effective weak quark Hamiltonian ${\cal H}^{\Delta S = 1}_{\rm eff}$ consists of ten four-quark operators among which only seven operators are independent: $${\cal H}^{\Delta S = 1}_{\rm eff} \;=\; -\frac{G_F}{\sqrt{2}} V_{ud} V^_{us} \sum_i c_i (\mu) {\cal Q}_i (\mu) + {\rm h.c.} .$$ The expression for the four-quark operators ${\cal Q}_i$ and Wilson coefficients can be found elsewhere [@Burasetal]. Integrating over the quark fields and using the derivative expansion, the effective weak chiral Lagrangian in leading order is then obtained as follows: $$\begin{aligned} {\cal L}^{\Delta S = 1,{\cal O} (p^2)}_{\rm eff}&=& -\frac{G_F}{\sqrt{2}} V_{ud} V^_{us} f^4_{\pi} \left[g_{8} \left\langle\lambda_{23} L_\mu L^\mu \right\rangle \right.\nonumber \\ && \;+\; \left.g_{27} \left(\frac23 \left\langle \lambda_{12} L_\mu\right\rangle \left\langle \lambda_{31} L^\mu \right\rangle + \left\langle \lambda_{32} L_\mu\right\rangle \left\langle \lambda_{11} L^\mu \right\rangle \right) \right] \;+\; {\rm h.c.} \end{aligned}$$ The coupling constants $g_{8}$ and $g_{27}$ can be extracted from the $K\rightarrow \pi\pi$ decay rate and the $\Delta T = 1/2$ enhancement is reflected in these constants. In our model those constants are expressed in terms of the Wilson coefficients and the dynamical coefficients ${\cal K}$, ${\cal M}$, ${\cal P}$, and ${\cal R}$ [@FranzKimGoeke2] which are functions of the momentum-dependent quark mass: $$\begin{aligned} g^{(1/2)}_{8} &=& \frac{16 N_c^{2}{\cal K}^2}{f^4_{\pi}} \left(-\frac25 c_1 + \frac35 c_2 + c_4 -\frac35 c_9 + \frac25 c_{10} \right) +\frac{64 N_{c}^{2} {\cal M} ({\cal P} + {\cal R})}{f_{\pi}^{4}}c_6, \nonumber \\ g^{(3/2)}_{27} &=& \frac{16 N_c^{2} {\cal K}^2}{f^4_{\pi}} \left(\frac35 c_1 + \frac35 c_2 + \frac{9}{10} c_9 +\frac{9}{10} c_{10}\right). \label{Eq:ratiol}\end{aligned}$$ [3.]{} In this talk, we will show the results with the original $M(k)$ [@DP] employed, which turns out to be the best for the present calculation. For the other types of the $M(k)$, we refer to the corresponding work [@FranzKimGoeke2]. In Figures 1 and 2 the quark and gluon condensates are drawn, respectively. ![The gluon condensate as a function of $M_0$.[]{data-label="fig2"}](oquark.ps){width="70mm" height="50mm"} ![The gluon condensate as a function of $M_0$.[]{data-label="fig2"}](ogluon.ps){width="70mm" height="50mm"} Both of the quark and gluon condensates increase as the $M_0$ decreases. In particular, they increase relatively faster in smaller values of the $M_0$. In Figure 3, we draw the low energy constants $g_8$ and $g_{27}$, respectively. While the $g_8$ changes drastically as the $M_0$ does, the $g_{27}$ midly decreases as the $M_0$ increases. It indicates that the $\Delta T=1/2$ enhancement is better explained as we choose smaller values of the $M_0$. Figure 4 shows the ratio $g_{8}/g_{27}$ of the LECS in leading order, compared with the former result in the shaded band for which the constant quark mass ($F^2(k)=1$) is used and the cut-off parameter $\Lambda$ is introduced to tame the divergence. Note that the quark condensate appears directly in the expression for the LECs in the former work. On the other hand, the quark condensate does not enter in the present result as seen in Eq.(\[Eq:ratiol\]). The coefficient ${\cal M}$ in Eq.(\[Eq:ratiol\]) becomes the quark condensate only when we turn off the momentum-dependence of the quark mass, which turns out to be identical to the former result. The improvement achieved by the momentum-dependent quark mass is prominent, compared to the former work [@FranzKimGoeke]. In particular, the improvement is remarkable at smaller values of the $M_0$, as we already noticed in Fig. 3. ![The ratio $g_{8}/g_{27}$ as a function of $M_0$. The solid curve draws the present result, while the shaded band designates the former one with the uncertainty from the range of input parameters.[]{data-label="fig4"}](olecs.ps){width="70mm" height="50mm"} ![The ratio $g_{8}/g_{27}$ as a function of $M_0$. The solid curve draws the present result, while the shaded band designates the former one with the uncertainty from the range of input parameters.[]{data-label="fig4"}](oratiob.ps){width="70mm" height="50mm"} For example, with the value of $M_0=200$ MeV, we obtain the ratio which is almost the same as the empirical result, allowing the quark condensate to be larger than usual. However, bearing in mind that there might be other sources of the $\Delta T=1/2$ enhancement such as the matching between the scale of the physical decay processes and that of the weak Hamiltonian [@Bijnens], or $1/N_c$ corrections ([e.g.]{} meson loops), we do not want to insist too much strongly smaller values of the $M_0$ so as to reproduce the empirical result of the ratio. The conclusion we draw from the present work is that the momentum-dependence of the constituent quark mass plays an essential role in improving the low energy constants to the great extent, respecting the $\Delta T =1/2$ enhancement. [9]{} J. Kambor, J. Missimer, and D. Wyler, [Nucl. Phys.]{} [B346]{} (1990) 17; [Phys. Lett.]{} [261B]{} (1991) 496 G. Esposito-Farèse, [Zeit. f. Phys.]{} [C50]{} (1991) 255. G. Ecker, J. Kambor, and D. Wyler, [Nucl. Phys.]{} [B394]{} (1993)101. A. Pich and E. de Rafael, [Nucl. Phys.]{} [B358]{} (1991) 311. M. Franz, H.-C. Kim, and K. Goeke, hep-ph/9903275 (1999). D. Diakonov and V. Petrov, [Phys. Lett.]{} [B147]{} (1984) 351; [Nucl. Phys.]{} [B272]{}(1986) 457;“[ Spontaneous breaking of chiral symmetry in the instanton vacuum]{}”, in Hadron matter under extreme conditions, Kiew (1986) p. 192. G. Buchalla, A.J. Buras, M.E. Lautenbacher, [Rev. Mod. Phys.]{} [68]{} (1996) 1125 (and references therein). M. Franz, H.-C. Kim, and K. Goeke, PNU-NTG-01/99, RUB-TPII-09/99 (1999). J. Bijnens, in this proceeding (1999). [^1]: email:[email protected] [^2]: Research supported by the Korea Research Foundation made in the program year of 1998 [^3]: email:[email protected] [^4]: email:[email protected] [^5]: Talk given by HCK at PANIC ‘99, Uppsala Sweden, June 10, 1999.	{ "pile_set_name": "ArXiv" }
--- abstract: \| The effect of the lattice periodic potential on superconductivity which was ignored by BCS theory has been investigated. According to the effective mass approximation of band theory, the effect of lattice periodic potential can be embodied in the effective mass $m^{\ast }$ of the Ginzburg-Landau (GL) equations. A special property of $m^{\ast }$ is that it can be negative. Negative effective mass leads to many unusual phenomena. The superconducting order parameter shows the period distribution. Its modulate wavevector is proportional to the condensed carrier density, which explains the linear relation between the magnetic peaks displacement $\epsilon $ and $x$ for La$_{2-x}$Sr$_x$CuO$_4$. The superconducting phase is always local and separated originally and evolves into global superconducting phase at the certain pairs concentration, which explains why the cuprate superconductors must be insulator at low doped. The doped concentrations of insulator to superconductor transition for La$_{2-x}$Sr$_x$CuO$% _4 $ is consistent with the experiment results. The relation of the superconducting gap (SG) and the pseudogap (PG) was discussed.\ Keywords: Stripes; Ginzburg-Landau theory; Superconductivity; Charge fluctuations. author: - Zhongqing Wu$^1$ and Xuanjia Zhang$^2$ title: Charge fluctuations in cuprate superconductors --- Since the high temperature superconductor (HTS) was found in 1986, HTS has been one of the most attractive field in condensed matter physics. Almost any physical properties about HTS have been intensely studied and many anomalous properties have been found. which make many people believe that the BCS theory and Fermi liquid theory are no longer fit for HTS. The most striking of all anomalous properties are the stripe phase and pseudogap (PG). Unlike conventional metals in which the charge distribution is homogenous, the charge carriers are segregated into one-dimensional stripe in HTS, which was predicted by t-j model[@Zan1] and then supported by many experiment[@Moo; @Dai; @Nod; @Yam; @Mat; @Tran; @Tranq]. The relations of many physical properties such as Knight Shifts[@Wal], NMR relaxation rates[@Ish], DC conductivity[@Mom], with the temperature in normal state of HTS do not conform to Fermi liquid theory, which has been valid in the conventional metal. The temperature at which these physical properties begin to depart from Fermi liquid theory are almost the same. Some experiments such as angle-resolved photoemission, tunnelling spectroscopy indicate that the density of states (DOS) in Fermi surface begin to decrease and an PG open at this temperature. The PG is of the same size and k dependence as superconducting gap (SG). The microscope mechanism of PG is still unclear. In superconducting states of HTS, anomalous properties have also been found. For example, the ratio of the energy gap in 0K $\Delta(0)$ with the superconducting phase transition temperature $T_c$ ($2\Delta(0)/k _{B}T _{c}\approx 8\sim9$) far more than BCS theory’ prediction ($2\Delta(0)/k _{B}T _{c}=3.53$) Some people believe that these anomalous properties result from the correlation effect between electrons, which has been overlook by BCS and Fermi liquid theory. Many model have been proposed for the study of correlation effects. However up to now, there is a lack of the breakthrough in this aspect. Another important effect that has also been overlook by BCS theory is the lattice periodic potential. The band theory that describes the effect forms the basis of the modern theory of electron in solids. According to the effective mass approximation of band theory, the effect can be included in BCS theory if we substitute the effective mass $m^{\ast }$ for mass of bare electron. $m^{\ast }$ should be negative for HTS, as will be discussed in the end of our paper. Since Ginzburg-Landau (GL) equations can be derived from the BCS theory, the mass in equation should be also negative for HTS. Negative effective mass will lead to extraordinary spatial distribution of the order parameter, which can well illuminate many anomalous properties. Localized holes organize into one-dimensional structures, which has been observed in HTS in many experiments[Tran,Bor,Nod,Tranq,Sha]{}. The spin modulation also shows one-dimensional properties. Neutron scattering studies reveal that there are two types of twin domains and spin modulation is one dimensional in each domain for La$% _{2-x}$Sr$_{x}$CuO$_{4}$ and YBa$_2$Cu$_3$O$_6.6$[@Tra; @Moo]. We think that these one-dimensional properities result from the one-directional modulation of order parameter. Then with no magnetic field considered, the first equation of GL can be written as $$\alpha \psi +\beta {\|\psi \|}^{2}\psi -\frac{\hbar ^{2}}{2m_{l}^{\ast }}% \frac{\partial^{2} \psi}{\partial l^{2}} =0\text{,}\eqno{(1)}$$where $m_{l}^{\ast }$ is the component of the effective mass in $l$ direction, $\psi =\sqrt{n_{s}(r)}e^{i\phi (r)}$ is the order parameter, which only vary along $l$ and is equal in direction perpendicular to $l$, where $n_{s}(r)$ is the local condensed carrier density and $\phi $ is the phase of the effective wave function. Taking $f=\frac{\psi }{\psi _{0}}$ as the dimensionless effective wave function where $\|\psi _{0}\|^{2}=-\frac{\alpha }{\beta }$, under the condition $m_{l}^{\ast }<0$, the equation (1) reduces to $$\xi ^{2}\frac{\partial^{2}f}{\partial l{}^{2}}=f(1-f^{2})\text{,}\eqno{(2)}$$where$\ \xi ^{2}=\frac{\hbar ^{2}}{\|{2m_{l}^{\ast }\alpha }\|}$. The corresponding equation for $m_{l}^{\ast }>0$ can be obtain as long as the left side of equation (2) multiply by -1. As will be shown later, just above difference between equation for $m_{l}^{\ast }>0$ and $m_{l}^{\ast }<0$ leads to that they have very different spatial distribution of the order parameter. Since the coefficient of each term in equation (2) does not include the variable $l$, the origin can be arbitrary for the infinity system. Therefore we choose the origin where $f$ has the minimum. Then the boundary condition is $$f(0)=f_{0}\text{,}$$ $$\left.\frac{\partial f}{\partial l}\right\| _{l=0}=0\text{.}\eqno{(3)}$$ ![$f^{2}=\frac{\|\protect\psi \|^{2}}{\|\protect\psi _{0}\|^{2}}$ as a function of space coordinate $l$ (in units of $\xi$) for $f_{0}^{2}=0.01,0.36,0.81$[]{data-label="fig:1"}](fig1.eps){width="8.5cm"} The dependence of $f^{2}$ on the space coordinate $l$ is shown in Fig. 1. The spatial distribution of the order parameter is sensitive to $f_{0}^{2}$. When $f_{0}^{2}$ is close to 1, the condensed carrier density fluctuation is similar to the charge density wave. When $f_{0}^{2}$ is close to 0, the phase separation is obvious. The condensed carriers get together and form periodically charged stripes. Between the charged stripes are regions where few condensed carrier can be found and the microscope composition is similar with the parent compounds of HTS. We call these regions antiferromagnetic (AF) stripes. It is shown that AF stripes are wider than charge stripes when $f_{0}^{2}=0.01$ in Fig. 1, which is more distinct when further decreasing $f_{0}^{2}$. The stripes have been observed by many experiments in HTS and their structure change with the doped concentration. The low-energy neutron-scattering studies performed on La$_{2-x-y}$Nd$_{y}$Sr$_{x}$CuO$_{4}$ by Yamada$\ {\it {et}}$ $% {\it {al}}$[@Yam] have demonstrated that the charge modulation wavevector $\epsilon $ initially increases linearly with $x$ before saturating for $% x>1/8$. In order to compare with the results of the low-energy neutron-scattering studies, we investigated the relation of the average condensed carrier density with the order parameter modulation wavevector. The average condensed carrier density is defined as $$\rho =\frac{1}{\tau }\int_{-\frac{\tau }{2}}^{\frac{\tau }{2}}{\|\psi \|}% ^{2}dl=\frac{1}{\tau }{\|\psi _{0}\|}^{2}\int_{-\frac{\tau }{2}}^{\frac{\tau }{% 2}}\|f\|^{2}dl,\eqno{(4)}$$where $\tau $ is the order parameter modulation period. The dependence of $\rho $ on $\frac{1}{\tau }$ is shown in Fig. 2. $\rho $ varies linearly with $\frac{1}{\tau }$ as $$\rho =\frac{\xi \|\psi _{0}\|^{2}}{k}\frac{1}{\tau }\eqno{(5)}$$for $\rho <0.6{\|\psi _{0}\|}^{2}$, where $k=\frac{4}{15}$ is the slope of the broken line. Because that the modulation wavevector $\epsilon $ is proportional to the inverse of the charge modulation period and the average condensed carrier density $\rho $ is proportional to Sr concentration $x$ for La$_{2-x}$Sr$_{x}$CuO$_{4}$ if all of carrier condensed, the equation (5) denotes $$\epsilon \propto x\eqno{(6)}$$when $\rho <0.6{\|\psi _{0}\|}^{2}$. Furthermore, the modulation wavevector $\epsilon $ reaches the maximum when the order parameter is homogenous distribution. These results accord well with the neutron scattering measurements on La$_{2-x-y}$Nd$_{y}$Sr$_{x}$CuO$_{4}$[@Yam; @Mat]. ![The dependence of inverse of charge modulation period $\frac{1}{% \protect\tau }$(in units of $\frac{1}{\xi}$) on the condensed carrier density $\protect\rho$ ( in units of $\|\psi _{0}\|^{2}$).[]{data-label="fig:2"}](fig2.eps){width="8.5cm"} Fig. 1 indicates that there are two kinds of the spatial distribution of the condensed carriers density: One is the global distribution, the other shows that the condensed carriers is local and separate each other. The kind of the distribution can be determined by whether $f_{0}^{2}$ is close to 0 and should be close correlative with the average condensed carrier density $\rho $. The relation of $f_{0}^{2}$ with $\rho $ is shown in Fig. 3. When $\rho $ is smaller than $% \rho _{c}=0.35{\|\psi _{0}\|}^{2}$, $f_{0}^{2}$ is almost 0 and the superconducting phase is local and separated each other, which leads to that the system does not show global superconductivity. It is obvious the temperature of the global superconductivity $T_c$ is determined by the value of $f_0^2$ at ground state. Expanding $T_c$ with the value of $f_{0}^{2}$ at ground state by Thaler Formula, we have that $T_{c}$ is proportional to $f_{0}^{2}$ when $f_{0}^{2}<<1$. Because all of carrier condensed at ground state, so $\rho \propto x$. Therefore, in fact, Fig. 3 reveals the dependence of $T_{c}$ on Sr concentration $x$. $\rho _{c}$ just corresponds to the Sr concentration $x_{c}$ of non-superconductor to superconductor transition. According to the fact that modulated wave vector $\epsilon $ has maximum at $\rho ={% \|\psi _{0}\|}^{2}$ in our result and is saturating at $x=1/8$ by the experiment [@Yam], we evaluated $\ x_{c}$ to be about $0.044$ for La$_{2-x}$Sr$_{x}$CuO$_{4}$. ![$f_{0}^{2}$ as a function of the condensed carrier density $\protect\rho $ ( in units of $\|\psi _{0}\|^{2}$)[]{data-label="fig:3"}](fig3.eps){width="8.5cm"} It is well known that the free energy of system must decrease with increasing the condensed carrier density. Obviously, we should make sure whether our free energy meets the above demand. The free energy density of the system $$F_{s}=\frac{1}{\tau }\int_{-\frac{\tau }{2}}^{\frac{\tau }{2}}[\alpha \|\psi \|^{2}+\frac{\beta }{2}\|\psi \|^{4}+\frac{1}{2m_{l}^{\ast }}\|(-i\hbar \nabla )\psi \|^{2}]dl,\eqno{(7)}$$where the free energy density in the normal state is taken as zero. The relation of $F_{s}$ (in units of $\alpha \|\psi _{0}\|^{2}$) and $\rho $ is shown in Fig. 4. $F_{s}$ decrease monotonously with increasing the condensed carrier density $\rho $(note: $\alpha $ is negative). It implies that more and more carriers condensed with decreasing the temperature. An intriguing phenomenon is that $F_{s}$ is weak-dependence on $\rho $ when $\rho \approx 1$, which means that the electron pairs can be excited without energy. Namely, the energy gap has node. At the conventional superconductor, the energy gap is non-zero and manifests itself in exponentially activated temperature dependence of a wide variety of dynamic and thermodynamic properties at low temperature. At HTS the exponentially activated temperature dependence is disappear. Cavity perturbation measurements[@Zha] and muon spin rotation study[@Har] on high quality YBa$_2$Cu$_3$O$_7$ crystals have revealed the linear temperature dependence in the penetration depth $\lambda$ below 30 K. A linear term in the low temperature thermal conductivity has also been found[@Kri]. Above linear temperature dependence of physical properties indicates non-exponentially activated temperature dependence of the DOS of the quasiparticle excitation, which can appear only when there is node in the energy gap. The presence of a zero-bias conductance peak in tunneling spectroscopy also supports that the energy gap has node[@Hu]. ![The dependence of the free energy density $F_{s}$ (in units of $\alpha \|\psi _{0}\|^{2}$ ) on the condensed carrier density $\protect\rho $ ( in units of $\|\psi _{0}\|^{2}$).[]{data-label="fig:4"}](fig4.eps){width="8.5cm"} The electrons begin to pair below the superconducting phase transition temperature. But the system does not show the zero-resistance properties because the superconducting phase is local at low pairs concentration. The condensation of electrons leads to the decrease of density of states in Fermi surface, which results in the anomalous of normal state transport properties and the open of the pseudogap (PG). The normal state PG turns into the superconducting gap (SG) when the local superconducting phase evolves into the global one. Therefore it is natural that many experiment, such as angle-resolved photoemission, tunnelling spectroscopy, nuclear magnetic resonance and neutron scattering, reveal that SG emerges from the normal state PG and is of the same size and k dependence as PG[@Tim]. Even all of carriers transform into electron pairs, the superconducting phase is still local and separate for $x<x_{c}$ ($=0.044 $ for La$_{2-x}$Sr$_{x}$CuO$_{4}$). The system does not show global superconductivity at any temperature. They become insulators because the local superconducting phase consumes all of carrier at low temperature. For $x>x_{c}$, the local superconducting phase can always evolve into the global one at certain temperature. Therefore the $x_{c}$ is the concentration of insulator to superconductor transition. Since the transition concentration is determined by extending the experiment data of $T_{c}\neq 0K$, it may be arbitrary a value between $\rho _{c}$ and $\rho _{f}$. This gives out a likely explanation why the insulator-superconductor transition concentration reported by different experiment groups is obvious disagreement. $\rho _{c}=0.34\|\psi _{0}\|^{2}$ and $\rho _{f}=0.45\|\psi _{0}\|^{2}$ respectively correspond to $x=0.044$ and $x=0.056$ for La$_{2-x}$Sr$_{x}$CuO$_{4}$, which is well consistent with the experiment result $x\approx 0.05\sim {0.06}$[@Tak; @ell]. Contrary with the common sense, the superconducting phase appears also in insulator. Furthermore the modulate wave vector $\epsilon $ in insulator and in superconductor has the same dependence on doped concentration, $\epsilon \propto x$, which is verified by neutron scattering experiment on La$_{2-x}$Sr$_{x}$CuO$% _{4}$[@Mat]. The Josephon effect experiment by Decca ${\it {et}}$ ${\it {% al}}$[@Dec] indicates the existence of an anomalously large proximity effect in underdoped insulating YBa$_{2}$Cu$_{3}$O$_{6+x}$. Our results that insulator has local superconducting phase will be helpful to understand above anomalously large proximity effect. It is generally perceived that the superconducting phase transition temperature is the temperature $T_C$ at which the resistance disappear. However, according to our result, this view only suits the conventional metals. The genuine phase transition temperature for HTS is the PG temperature $T_p$ where the order parameter has appeared. The zero-resistance temperature $T_C$ is just the temperature at which the local superconducting phase is translated into the globe superconducting phase. It is well known that the gap should vanish and the breaking of the symmetry will happen at the phase transition temperature. However, these phenomena do not occur at the zero-resistance temperature $T_C$. On the contrary, a breaking of time-reversal symmetry at the PG temperature $T_p$ is recently found by angle-resolved photoemission study[@Kam]. Therefore, at the region between $T_p$ and $T_C$, the system that has been taken for normal state is actually in superconducting state. Then it is natural that many physical properties at this region do not conform to Fermi liquid theory that describes the normal-state metal. Noting that the physical properties above the PG temperature accord well with Fermi liquid theory, We are sure that Fermi liquid theory is still fit for HTS. The opinion that the PG temperature is superconducting phase transition temperature is helpful to clarify the perplexity that predict of BCS is invalid for HTS. With the PG temperature $T_p$ given by the experiments, the value of $2\Delta(0)/k _{B}T _{p}$ is close to the predict of BCS theory. Although the negative effective mass $m^{\ast }$ is general concept in Solid State Physics, it is still necessary to discuss whether the effective mass may be negative in GL theory. GL theory is a phenomenological theory. Considering that GL theory can be derived from the microscopic BCS theory and the meaning of electron mass in BCS theory is clearer than that in GL theory, we first study the electron mass of BCS theory. Ashcroft had put forward that in its simplest form the BCS theory makes a gross oversimplification in the basic Hamitanian that describes the conduction electrons [@Ash]. The conduction electrons are treated in the free electron approximation and the effect of lattice periodic potential (band structure effect) is ignored. The oversimplifications may seem surprising because the band theory that describes the effect forms the basis of the modern theory of electrons in solids. Then the questions why BCS theory that ignores the effect can still explain the conventional superconductivity and whether the effect can be neglected in HTS rise. To answer these questions, we must consider the effect of the lattice periodic potential. The effective mass approximation of band theory shows that the electrons nearby the band gap can be treated as free electrons with the effective mass. The effect of lattice periodic potential is embodied in the effective mass. According to the approximation, if Fermi surface is nearby the band gap, the conduction electrons can be taken as free electron. Then the lattice periodic potential is considered as long as we replace the mass of bare electron with the effective mass $m^{\ast }$. The replacement does not influence the BCS’major equilibrium predictions. It partly explains why BCS theory that ignores the effect of lattice periodic potential can still explain successfully superconductivity. Comparing with bare electron mass, $m^{\ast }$ has s special property. It may be negative. For example, the electrons in top of band have the negative $m^{\ast }$, which is a very important conclusion of band theory. GL theory can be derived from the microscopic BCS theory. Therefore its mass may also be negative. The negative $m^{\ast }$ only holds true near the top of the band, which means that $\|-i\hbar \nabla \psi \|^{2}$ should be limited. We need not be afraid that the minimum of the GL free energy will be unbound from below. Although the hole description is always introduced when $% m^{\ast }$ is negative, its application to superconductor is questioned since the hole is the collective behavior of the whole band, whereas electron pairing is only involved in a thin layer of electrons nearby Fermi surface. Now we can affirm that the lattice periodic potential is very important to HTS. The anomalous properties about HTS are primarily due to this effect. Three striking anomalous properties: the stripe phase, the pseudogap and the energy gap nodes, can been deduce from GL equation if the lattice periodic potential is considered. BCS theory and Fermi liquid theory are still fit for HTS. Some opinions should change. The genuine superconducting phase transition temperature for HTS is the PG temperature $T_p$ where the order parameter has appeared. The zero-resistance temperature $T_C$ is just the temperature at which the local superconducting phase is translated into the globe superconducting phase. The superconducting phase transition happens also in insulator. [99]{} J. Zannen, and O. Gunnarsson, Phys. Rev. B ${\bf 40}$, 7391(1989). P. Dai ${\it {et}}$ ${\it {al}}$., Science ${\bf 284}$, 265(1999). T. Noda, H. Eisaki, and S. Uchida, Science ${\bf 286}$, 265(1999). H. A. Mook, D. Pengcheng, F. Dogan, and R. D. Hunt, Nature ${\bf 404}$, 729(2000). K. Yamada ${\it {et}}$ ${\it {al}}$., Physica C. ${\bf 282-287% }$, 85(1997). M. Matsuda ${\it {et}}$ ${\it {al}}$., Phys. Rev. B. ${\bf 62}$, 9148([2000)]{}. J. M. Tranquada ${\it {et}}$ ${\it {al}}$., Phys. Rev. B. $% {\bf 54}$, 7489(1996). J. M. Tranquada, B. J. Sternlieb, J. D. Axe, Y. Nakamura, and S. Uchida, Nature, ${\bf 375}$, 561(1995). Walstedt R.E, et. al.,Physical Review B-Condensed Matter,.41,(1990).9574. K. Ishida, et. al., Phys. Rev. B 58, (1998)R5960 N. Momono, et.al., Physica C 317, (1999)603. E. Borsa ${\it {et}}$ ${\it {al}}$., Phys. Rev. B. ${\bf 52,}$ 7334(1995). R. P. Sharma ${\it {et}}$ ${\it {al}}$., Nature, ${\bf 404}$, 736(2000). J. M. Tranquada, Physica B ${\bf 241-243}$, 745(1998). K. Zhang, D.A. Bonn, et. al., Phys. Rev. Lett 73, 2484(1994). W. N. Hardy, D.A. Bonn, D. C. Morgan, R. X. Liang and K. Zhang, Phys. Rev. Lett 70, 3999(1993) K. Krishana, N. P. Ong, et. al., Science 277, 83(1997) C. R. Hu, Phys. Rev. Lett. 72, 1526(1994). T. Timusk and B.Statt, Rep. Prog. Phys. ${\bf 62}$, 61(1999). H.Takagi, T. Ido, S. Ishibashi, M. Uota, and S. Uchida, Phys. Rev. B. [ ]{}${\bf 40}$, 2254(1989). B. Ellman ${\it {et}}$ ${\it {al}}$., Phys. Rev. B.[ ]{}$% {\bf 39}$, 9012(1989). R. S. Decca, H. D. Drew, E. Osquiguil, B. Maiorov, and J. Guimpel, Phys. Rev. Lett. ${\bf 85}$, 3708(2000). N. Miyakawa $\it {et}$ $\it {al}$., Phys.Rev.Lett. $\bf{83}$, 1018 (1999). A. Kaminski, et.al., Nature 416, (2002)610. Neil. W. Ashcroft and N. David Mermin, Solid State Physics (Holt, Rinehart and Winston, New York, 1976).	{ "pile_set_name": "ArXiv" }
--- abstract: 'Autoscaling system can reconfigure cloud-based services and applications, through various configurations of cloud software and provisions of hardware resources, to adapt to the changing environment at runtime. Such a behavior offers the foundation for achieving elasticity in modern cloud computing paradigm. Given the dynamic and uncertain nature of the shared cloud infrastructure, cloud autoscaling system has been engineered as one of the most complex, sophisticated and intelligent artifacts created by human, aiming to achieve self-aware, self-adaptive and dependable runtime scaling. Yet, existing Self-aware and Self-adaptive Cloud Autoscaling System (SSCAS) is not mature to a state that it can be reliably exploited in the cloud. In this article, we survey the state-of-the-art research studies on SSCAS and provide a comprehensive taxonomy for this field. We present detailed analysis of the results and provide insights on open challenges, as well as the promising directions that are worth investigated in the future work of this area of research. Our survey and taxonomy contribute to the fundamentals of engineering more intelligent autoscaling systems in the cloud.' author: - Tao Chen - Rami Bahsoon - Xin Yao bibliography: - 'references.bib' title: 'A Survey and Taxonomy of Self-Aware and Self-Adaptive Cloud Autoscaling Systems' --- <ccs2012> <concept> <concept\_id>10011007.10010940.10010971.10011120.10003100</concept\_id> <concept\_desc>Software and its engineering Cloud computing</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10011007.10010940.10011003.10011002</concept\_id> <concept\_desc>Software and its engineering Software performance</concept\_desc> <concept\_significance>300</concept\_significance> </concept> </ccs2012> [^1] Introduction ============ Modern IT companies, from small business to large enterprises, increasingly leverage cloud computing to improve their profits and reduce the costs. Throughout all the Software-as-a-Service (SaaS), Platform-as-a-Service (PaaS) and Infrastructure-as-a-Service (IaaS) levels, one of the pronounced benefits of the cloud is referred to as elasticity, which reflects the extent to which a system can adapt to the workload fluctuations by adjusting configurations and resource provisioning close to the demand. In certain predictable scenarios where the environmental condition has strong and stable seasonality, the configurations and resources can be approximately specified by human experts in advance. Nevertheless, for many other cases, for examples, unexpected workload changes, elasticity can be only enabled by runtime automatic scaling, or simply autoscaling: a dynamic process, often operating on a Physical Machine (PM), that adapts software configurations (e.g., threads, connections and cache, etc) and hardware resources provisioning (e.g., CPU, memory, etc) on-demand, according to the time-varying environmental conditions. The ultimate goal of autoscaling is to continually optimize the non-functional Quality of Service (QoS) (e.g., response time and throughput) and cost objectives for all cloud-based services[^2]; thus their Service Level Agreement (SLA) and budget requirements can be better complied with. In particular, autoscaling systems help to realize elasticity by providing timely and elastic adaptation in scales, which is one of the key benefit of cloud computing that attracts a wide range of practitioners [@lorido2014review][@qu2016auto]. Comparing with the other cloud resources management in general, autoscaling system has been specifically designed for: (i) scaling cloud-based applications in response to dynamic changes in load, uncertainties in operations, handling multitenancy while ensuring Service Level Agreement compliance etc. In contrast, other resource management tasks, e.g., resource scheduling, often work on planned and deterministic sequence of resource demand. (ii) Adapting both the software configurations and hardware resources (and their interplays) that span over all SaaS, PaaS and IaaS levels, whereas most of the other resource management considers hardware resources and IaaS only. (iii) Taking the QoS for cloud-based applications/services at the centre of the concerned objectives (explicitly or implicitly) while the other resource management tasks often focus on resource utilization. From the literal meaning of the word autoscaling, it is obvious that the process is dynamic and requires the system to adapt itself subject to the dynamic and uncertain state of the services being managed and the environment. In such a way, the cloud-based services, running on a Virtual Machine (VM) or containers, can be expanded and shrink according to the environmental conditions at runtime. This characteristic has made autoscaling systems well-suited to the broad category of self-aware and self-adaptive systems [@Chen:2015:computer] [@roadmap] [@landscape]. However, given the unique characteristics of cloud, engineering Self-aware and Self-adaptive Cloud Autoscaling System (SSCAS) poses many challenges, including efficient autoscaling architectural styles, accurate model to predict the effects of autoscaling decision[^3] on the quality attributes, appropriate granularity of runtime control and effective trade-off decision making. In particular, a SSCAS should be able to handle various dimensions of QoS attributes, software configuration and hardware resources, in the presence of *QoS interference* [@fuzzy-vm-interference] [@qcloud] [@vm-fuzzy-MIMO] [@2014-decision-tree-software-CP-2014] where the quality of a single cloud-based service can be influenced by the dynamic behaviors of its neighbors on a PM, i.e., the other co-located services and co-hosted VMs or containers, under the sharing infrastructure of cloud. In this article, we provide a survey and taxonomy for the landscape of SSCAS research to better understand the state-of-the-arts and to identify the open challenges in the field. In particular, we focus on cloud autoscaling researches with respect to the well-known principles of self-awareness [@epics] and self-adaptively [@landscape] in computing systems, as well as the fundamental approaches and techniques that realize them. Broadly speaking, we aim to answer the following research questions: (i) What are the levels of self-awareness and self-adaptivity that have been captured in SSCAS? (ii) What are the architectural patterns used for engineering SSCAS? (iii) What are the approaches used to model the quality related to SSCAS? (iv) What is the granularity of control in SSCAS? (v) What are the approaches used for decision making in SSCAS? In a nutshell, our key findings are: - Stimulus-, time- and goal-awareness are the most commonly considered self-awareness levels in current SSCAS research while self-configuring and self-optimizing are more attractive than the others on self-adaptivity (Please refer to Section \[sec:s-aware\] and \[sec:s-adaptive\] for their definitions, respectively). - It was found that the general and simple feedback loop architectural pattern (and its variations) has been prominent for engineering SSCAS. - Analytical and machine learning based modeling have been the most widely used approaches for modeling the effects of autoscaling decision on quality attributes. But surprisingly, systematic selection of model’s input features and QoS interference are rarely considered for SSCAS. - The level of service/application is the most popular granularity of control for SSCAS. - Explicit optimization driven decision making is the most commonly used approach in SSCAS, but most of them have assumed single objective or using weighted sum aggregation of objectives. Further, we noted that QoS interference is again absent in many studies. In addition, we obtain the following insights on the open challenges for future research of the field: - There is a lack of considering required knowledge and its representation for SSCAS architecture with respect to the principles of self-awareness [@epics], thus urging further investigations. This can help to reason about and prevent improper design decisions, leading to better self-adaptivity. - Despite that QoS interference has been found to be an important issue [@fuzzy-vm-interference] [@qcloud] [@vm-fuzzy-MIMO] [@2014-decision-tree-software-CP-2014], it is often overlooked in both the QoS modeling and decision making aspects of SSCAS. Therefore, we call for novel and effective approach to manage and mitigate QoS interference in cloud. - Most studies attempt to scale hardware resources at the IaaS level only. However, a mature SSCAS should additionally consider the software configurations related to the cloud-based services and their interplay with the hardware resources, as found in recent studies [@2013-JRAO-most-closest-work-2013] [@software-RP-two-loops] [@2014-decision-tree-software-CP-2014]. - Instead of using fixed granularity of control (i.e., the boundary of decision making is on each service/application, VM/container, PM or cloud), future SSCAS should consider more flexible ones, e.g., dynamic and hybrid level, as discovered in recent studies [@MIMO-fuzzy-hill-climbing-2013] [@11icde_smartsla_full] [@Chen:2014:seams]. - The assumption of autoscaling bundles (e.g., the VM instances from Amazon EC2) has been made in a considerable amount of SSCAS studies. However, it is known that renting bundles cannot and does not reflect the interests of consumers and the actual demand of their cloud-based services [@rule-control-elasticity-cloud]. Thus, considering arbitrary and custom combinations of configurations and resources is an inevitable trend in the cloud computing paradigm. - Considering multi-objectivity in SSCAS is a must in order to create better diversity and possibly better trade-off quality without the needs of weights specification. In addition, how to achieve balanced trade-off over the set of non-dominated solutions is worth investigating [@Chen:2015:tsc-pending]. - More real world cases and scenarios are needed as this can be the only way to fully verify the potentials, effectiveness and impacts of SSCAS. This article is structured as follows: Section \[sec:back\] introduces the background and challenges. Section \[sec:related\] compares our work with the other reviews. A taxonomy and findings of the survey, with respect to different aspects of a SSCAS, are presented in Section \[sec:result\]. Section \[sec:diss\] discusses on the findings and presents the open challenges we learned. Section \[sec:con\] presents the conclusion. Backgound {#sec:back} ========= Autoscaling System in Cloud --------------------------- ![The conceptual design of an autoscaling system. (Note that this figure represents the conceptual design of an autoscaling system in cloud. Practical deployment of the autoscaling engine can be either centralized, or decentralized where there are different engines, each of which is running on a PM.[]{data-label="ch1:simple_arch"}](figures/simple-arch.pdf){width="7.3cm"} Given that it is almost impossible to access the low level details of cloud-based services (e.g., their codes and algorithms) at runtime, an autoscaling system often consist of two physical parts: a managing part containing the autoscaling engine and a manageable part encapsulating services and VMs/containers running in the cloud. The two physical parts are seamlessly and transparently connected for realizing the entire autoscaling process, known as external adaptation [@roadmap] [@landscape]. The external adaptation of an autoscaling system is shown in Figure \[ch1:simple\_arch\]. As we can see, the core of an autoscaling system in the cloud is the autoscaling engine, which can consist of multiple logical aspects. A typical example of autoscaling system is a feedback loop that covers monitoring and scaling: the former gathers the service’s or application’s current state while the latter utilizes the information to decide an action after being analyzed and reasoned about by the autoscaling engine. Given the multi-tenant nature of cloud, cloud-based services often come with different QoS objectives, SLA requirements and budget constraints, etc. The ultimate goal of an autoscaling system is to adapt those cloud-based services, through scaling the related software configurations and hardware resources, in such a way that their objectives are continually optimized. To execute an autoscaling decision, the scaling actions could be vertical (scale up/down where changes occurs on a VM/container), horizontal (scale in/out that adds/removes other VMs/containers), or both. Self-Adaptivity in Software Systems {#sec:s-adaptive} ----------------------------------- The broad category of automatic and adaptive systems aim to deal with the dynamics that the system exhibited without human intervention; but this does not necessarily involve uncertainty, i.e., there are changes related to the system but it is easy to know when they would occur and the extent of these changes. Self-adaptivity, being a sub-category, is a particular capability of the system to handle both dynamics and uncertainty. Here, self-adaptive systems refer to the systems that are capable to adapt their behaviors according to the perception of the uncertain environment and its own state. To date, self-adaptivity in software systems remain an important and challenging research field [@femosaa][@2014-kalman-pair-wised-coupling-on-tiers-2014][@icpechen]. According to the adaptive behaviors, self-adaptivity can be regarded as the following four properties, each of which covers a specific set of goals, as discussed by [@landscape]: - Self-configuring: “The capability of reconfiguring automatically and dynamically in response to change by installing, updating, integrating, and composing/decomposing software entities.” [@landscape] - Self-healing: “This is the capability of discovering, diagnosing, and reacting to disruptions. It can also anticipate potential problems, and accordingly take proper actions to prevent a failure.” [@landscape] - Self-optimizing: “This is also called self-tuning or self-adjusting, is the capability of managing performance and resource allocation in order to satisfy the requirements of different users, e.g., response time, throughput and utilization.” [@landscape] - Self-protecting: “This is the capability of detecting security breaches and recovering from their effects. It has two aspects, namely defending the system against malicious attacks, and anticipating problems and taking actions to avoid them or to mitigate their effects.” [@landscape] Self-Awareness in Software Systems {#sec:s-aware} ---------------------------------- In contrast, self-awareness is about the capability of a system to acquire knowledge about its current state and the environment. Such knowledge permits better reasoning about the system’s adaptive behaviors. Consequently, self-awareness is often seen as the lowest level of abstraction of self-adaptivity [@landscape], and thus it can improve the basic perceptions and self-adaptivity of a system [@2014_epics_handbook] [@7185305] [@epics_survey] [@Chen2016:book]. Inspired from the psychology domain, Becker et al. [@epics] have classified self-awareness of a computing system into the following general capabilities (they have used node to represent any conceptual part of a system being managed): - Stimulus-aware: “A node is stimulus-aware if it has knowledge of stimuli. The node is not able to distinguish between the sources of stimuli. It is a prerequisite for all other levels of self-awareness.” [@epics] - Interaction-aware: “A node is interaction-aware if it has knowledge that stimuli and its own actions form part of interactions with other nodes and the environment. It has knowledge via feedback loops that its actions can provoke, generate or cause specific reactions from the social or physical environment.” [@epics] - Time-aware: “A node is time-aware if it has knowledge of historical and/or likely future phenomena. Implementing time-awareness may involve the node possessing an explicit memory, capabilities of time series modeling and/or anticipation.” [@epics] - Goal-aware: “A node is goal-aware if it has knowledge of current goals, objectives, preferences and constraints. It is important to note that there is a difference between a goal existing implicitly in the design of a node, and the node having knowledge of that goal in such a way that it can reason about it. The former does not describe goal-awareness; the latter does.” [@epics] - Meta-self-aware: “A node is meta-self-aware if it has knowledge of its own capability(ies) of awareness and the degree of complexity with which the capability(ies) are exercised. Such awareness permits a node to reason about the benefits and costs of maintaining a certain capability of awareness.” [@epics] Comparison to Related Surveys {#sec:related} ============================= Research on cloud autoscaling systems and the related topics have been reviewed in some other surveys. For example, Manvi and Shyam [@manvi2014resource] present a review on resource management in the cloud, particularly at the IaaS level. They have provided a board survey on different issues related to managing cloud resources, e.g., resource adaptation, resource mapping and resource brokering etc. While resource management has some similarities to autoscaling, they lie in different levels of abstraction: the latter is more specific than the former. In other words, cloud autoscaling is one, but probably the most important part of the board cloud resource management. Another review from Mana [@mann2015allocation] is explicitly concerned with VM to PM mapping problem which is also belong to the cloud resource management category, but is often regarded as a fundamentally different issue to cloud autoscaling. Ardagna et al. [@ardagna2014quality] present a survey on QoS modeling and its application in the cloud. Indeed, QoS is the major concern for a cloud autoscaling system, but its management can be governed by various different approaches other than autoscaling, e.g., load balancing and admission control, which are also covered in [@ardagna2014quality]. In contrast to the above, our survey has explicitly focused on automatically scaling software configuration and hardware provisioning in the cloud in order to change the capacity of cloud-based services to handle the dynamic workloads. Al-Dhuraibi et al. [@al2017elasticity] present a review on elastic autoscaling approaches in the cloud, specifically focusing on the physical infrastructure level support, e.g., benchmarking and containerization techniques. Our survey, in contrast, is primarily concerned with the logical architecture and algorithmic level techniques for achieving different aspects of cloud autoscaling, e.g., modeling and decision making. Qu et al. [@qu2016auto] also survey autoscaling approaches for a special type of cloud application, i.e., web applications, with a coarse correlation to self-adaptivity, e.g., if an approach is self-adaptive or not; while our survey is application agnostic and we present finer correlation of an approach to different levels of self-adaptivity, e.g., self-optimization. The most related survey from the literature is probably the one by Lorido-Botran et al. [@lorido2014review], in which different category of algorithmic level techniques for QoS modeling and decision making in cloud autoscaling are reviewed. However, their survey differs from ours in the following three aspects: (i) they have not provided a comprehensive taxonomy on the cloud autoscaling problem; such a taxonomy (i.e., modeling, architecture, granularity and decision making), which we will present in the next section, is important as it clearly state the open problems and challenges related to different aspects of the cloud autoscaling domain, providing better clarifications and clearer directions for researchers on this research field. (ii) In addition, Lorido-Botran et al. [@lorido2014review] did not explicitly link the cloud autoscaling systems to different levels of self-awareness and self-adaptivity, which is one of the key contributions of our survey. (iii) Finally, we discuss the open problems and challenges of cloud autoscaling systems in a broader fashion. In summary, our survey differs from the other similar reviews in the following: - We present a focused survey on the logical architecture and algorithmic level techniques for cloud autoscaling which are application agnostic. - We explicitly correlate the reviewed approaches with different levels of self-awareness, self-adaptivity and the required knowledge in a fine grained manner. - We provide a clarified taxonomy that covers different logical aspects for engineering cloud autoscaling systems, and classify every study accordingly. - We discuss the open problems and challenges of cloud autoscaling systems in a broader fashion. Taxonomy and Survey Results for SSCAS {#sec:result} ===================================== In this section, we present a taxonomy and survey results for the state-of-the-art SSCAS research obtained from our review process. Review Process and Research Questions {#sec:rq} ------------------------------------- The review is intended to create a broad scope to cover the landscape of SSCAS research. Particularly, the following research questions serve as the main drivers of this review: - RQ1: What are the levels of self-awareness and self-adaptivity that have been captured in SSCAS? - RQ2: What are the architectural patterns used for engineering SSCAS? - RQ3: What are the approaches used to model the quality related to SSCAS? - RQ4: What is the granularity of control in SSCAS? - RQ5: What are the approaches used for decision making in SSCAS? The following prominent indexing services were used during the review: IEEE Xplore, ACM Digital Library, Science Direct, ISI Web of Knowledge, and Google Scholar. The search term was “Cloud computing” AND “Autoscaling” AND (“QoS modeling” OR “Performance modeling” OR “decision making” OR “optimization” OR “Architecture” OR “Interference” OR “Resource allocation” ). After applying inclusion (e.g., considering only journal, conference, and workshop papers) and exclusion criteria (e.g., removing duplicate entries and considering only the extended version) to the initial search result, the review has ended with the total of 109 studies. A Taxonomy of SSCAS Research ----------------------------- The overall taxonomy, concluded from the extracted studies, is given in Figure \[fig:tax\]. As we can see, current research on SSCAS often require sophisticated designs in different highest leveled logical aspects of the autoscaling engine, which we have classified and discussed as the following: ![A taxonomy of SSCAS research.[]{data-label="fig:tax"}](figures/global.pdf){width="7.3cm"} - *Self-Awareness:* This is concerned with the ability to acquire and maintain knowledge about the system’s own states and the environment, as specified in Section \[sec:s-adaptive\]. The key challenges here are which level(s) of knowledge is required for a SSCAS, what does it means for certain level in the problem context, (e.g., what does interaction refers to?) and what is the representation for different levels of knowledge, e.g., how do we represent goals in the SSCAS? - *Self-Adaptivity:* This is about the ability to change the system’s own behavior with specific goals in mind, as specified in Section \[sec:s-aware\]. Often, the required levels of self-adaptivity depending on the requirements, but they could be also related to the specific levels of knowledge that the SSCAS is able to capture, e.g., the SSCAS has to be goal-aware to achieve self-optimization. - *Architectural Pattern:* Autoscaling architecture is the most essential element of SSCAS. It describes the structure of the autoscaling process, the interaction between components and the modularization of the other important logical aspects in autoscaling. The challenge of architecting SSCAS is concerned with how to systematically capture different logical aspects (e.g., decision making) of SSCAS using a given architectural pattern. More importantly, how to encapsulate these aspects and the algorithms that realizes them into different components of the pattern. - *QoS Modeling:* While modeling the cost incurred by cloud-based services is straightforward, modeling the QoS is often much more complex and challenging. Here, the QoS modeling is concerned with the sensitivity of QoS with respect to the environment conditions (e.g., workload) and the control knobs (e.g., software configurations and hardware resources). The resulting model is a powerful tool to assist the autoscaling decision making process. Without loss of generality, in this article, we use *cloud primitives* to refer to both control knobs and environmental conditions in the cloud. In particular, we further decompose the notion of primitives into two categories, termed *control primitive* and *environmental primitive. Control primitives refer to the internal control knobs that can be either software or hardware. They are the fundamental features that can be controlled by the cloud providers to support QoS. Specifically, software control primitives are the key cloud configurations at the software level, e.g., the number of threads in the thread pool, the buffer size and the cache size, etc. In contrast, hardware control primitives refer to the computational resources, such as CPU, memory and bandwidth, etc. Typically, software control primitives exist on the PaaS layer while the hardware control primitives present on the IaaS layer. It is worth noting that considering software control primitives when autoscaling in the cloud is a non-trivial task, as they have been proved to be important features that can significantly influence the QoS [@2013-JRAO-most-closest-work-2013] [@software-RP-two-loops] [@2014-decision-tree-software-CP-2014]. The environmental primitives, on the other hand, refer to those external stimuli that is uncontrollable but can cause dynamics and uncertainties in the cloud. These, for example, can be the workload, incoming data, node failure, etc. The above examples of primitives listed above are not exhaustive, Ghanbari et al. [@rule-control-elasticity-cloud] have provided a more completed and detailed list of the possible control primitives in cloud. The challenges of QoS modeling include: (i) which primitives should be selected as model’s input features; (ii) how does the QoS change in conjunction with those primitives; (iii) how to incorporate the information of QoS interference into the model; (iv) whether the model is built offline or online; (v) and whether the model is dynamic, semi-dynamic or static. - Granularity of control:* Determining the granularity of control in the autoscaling engine is essential to ensure the benefits (e.g, QoS and cost objectives) for all cloud-based service. It is concerned with understanding whether certain objectives can be considered in isolation with some of the others, i.e., the boundary of decision making. This is because *objective-dependency* (i.e., conflicted or harmonic objectives) often exist in the decisions making process, which implies that the overall quality of autoscaling can be significantly affected by the inclusion of conflicted or harmonic objectives when making decision, hence rendering it as a complex task. This is especially important for the shared cloud infrastructure where objective-dependency exists for both intra- and inter-services. That is to say, objective-dependency is not only caused by the nature of objectives (intra-service), e.g, throughput and cost objective of a service; but also by the QoS interference (inter-services) due to the co-located services on a VM/container and co-hosted VMs/containers on a PM [@2013-JRAO-most-closest-work-2013] [@software-RP-two-loops] [@2014-decision-tree-software-CP-2014] [@qcloud]. Here, the challenges are which granularity of control to use, what is the basic entity to control (e.g., application or VM), and whether the control is in a centralized or decentralized manner. - *Decision making:* The final logical aspect in autoscaling logic is the dynamic decision making process that produces the optimal (or near-optimal) decision, which consists of the newly configured values of the related control primitives, for all the related objectives. In the presence of objective dependency, autoscaling decision making requires to resolve complex trade-offs, subject to the SLA and budget requirements. The trade-off decision can be then executed using either vertical (scale up/down) and/or horizontal scaling actions (scale in/out), which adapt the cloud-based services and/or VMs/containers correspondingly. The challenges of decision making in SSCAS include: (i) how to reason about and search for the effected adaptation decisions; (ii) what are the objectives, their representations and conflicting relations, if any; (iii) and which are the control primitives to tune. In the following, we present our detailed findings in regards to the taxonomy of SSCAS. The Levels of Self-Awareness and Self-Adaptivity in Cloud Autoscaling Systems {#sec:sscas} ----------------------------------------------------------------------------- \[fig:sscas\] [0.33]{} \[fig:self-aw\] [0.33]{} \[fig:self-ad\] [0.33]{} \[fig:knowledge\] RQ1: What are the levels of self-awareness and self-adaptivity that have been captured in SSCAS? It is worth noting that not all the studies have explicitly declared which levels of self-awareness/self-adaptivity that they have taken into account, therefore we identified this by looking at the studies in details with respect to the definitions of self-awareness/self-adaptivity. From Figure \[fig:self-aw\], we can see that stimulus-awareness, which is the fundamental level in self-awareness, has been considered in all the studies. The time- and goal-awareness have attracted relativity similar amount of attention. However, interaction- and meta-awareness has not been widely studied in recent SSCAS research. In Figure \[fig:self-ad\], we see that self-configuring and -optimizing have been predominately captured in the studies, whereas self-healing and -protecting receive little attentions. In particular, we found no study that explicit aims for self-protecting in SSCAS. Figure \[fig:knowledge\] illustrates whether the required knowledge representations at the SSCAS architecture level, e.g., knowledge of goal is required in the architecture, have been discussed, implicitly discussed or explicitly discussed in the studies. We can see that the majority of the studies surveyed do not attempt to declare what knowledge is required in SSCAS architecture, leaving only 33% of the studies have discussed the knowledge implicitly or explicitly. Architectural Pattern {#sec:arch} --------------------- RQ2: What are the architectural patterns used for engineering SSCAS? We classified the predominantly applied architectural patterns for SSCAS into three categories based on their basic form; these are Feedback Loop [@Brun:2009], Observe-Decide-Act (ODA) [@self-aware-ML-adaptive-control] and Monitor-Analysis-Plan-Execute (MAPE) [@ibm]. From Figure \[fig:arch\], we see that the generic feedback loop has been the predominant architecture pattern in SSCAS, following by the MAPE pattern. Particularly, as we can observe form Table \[tb:arch\], single and close feedback loop are widely exploited in SSCAS. In the following, we specify some representative studies under each category in details. ### Feedback Loop Feedback loop is the most general architectural pattern for controlling self-adaptive systems, including the autoscaling systems. It is usually a closed-form loop made up of the managing system itself and the path transmitting its origin (e.g., a sensor) to its destination (e.g., an actuator). Here, we further divide the pattern as single or multiple loops: - 4.4.1.1 Single Loop: Single loop is the simplest, yet the most commonly used pattern for SSCAS due to its flexibility. The most common practice with single loop is to build a closed feedback control where the core is the decision making component and an optional QoS modeling component, e.g., Ferretti et al. [@05557978] , CloudOpt [@fine-grained-servce-LQM-MIP-power], SmartSLA [@11icde_smartsla_full] and CLOUDFARM [@2014-linear-centralized-decision-making-2014], etc. Some other studies have included an additional component for workload or demand prediction based on either offline profiling, e.g., Jiang et al. [@typical-navie-scaling-VM-number] and Fernandez et al. [@2014-profiling-decision-tree-scaling-2014], or online learning, e.g., Kingfisher [@2011-cost-aware-v-h-scaling-2011] and PRESS [@signal-resource-trend-prediction]. Open feedback exists for single loop, as presented in Cloudine [@2013-elasticity-primitives-2013], where the scaling actions are partially triggered by user requests. In particular, they use a centralized Resource and Execution Manager to handle all the scaling actions. Apart from the general autoscaling architecture, other efforts are particularly designed upon specific cloud providers. For example, Zhang et al. [@MPC-price] as well as Kabir and Chiu [@cache-static-ANN-bi-obj-2012] propose to use a simple feedback loop for architecting autoscaling system, which is heavily tied to the properties of Amazon EC2. [\|P[2.4cm]{}\|P[0.8cm]{}\|P[1.1cm]{}\|X\|]{} *Architectural Pattern&Style&Open/Close&Representative Examples\ &&close&[@05557978], [@fine-grained-servce-LQM-MIP-power],[@11icde_smartsla_full], [@HPL-2008-123R1-mimo], [@2014-eplison-GA-weigh-h-scaling-2014], [@06119056], [@2014-linear-centralized-decision-making-2014], [@2014-kalman+rule-based-2014], [@typical-navie-scaling-VM-number], [@2014-profiling-decision-tree-scaling-2014], [@2011-cost-aware-v-h-scaling-2011], [@kriging-controller], [@2013-self-organising-map-2013], [@signal-resource-trend-prediction], [@2014-fuzzy-compare-to-JRao-2014], [@2013-software-CP-only-2013],[@acdc09], [@compare-to-Rao-fuzzy-2013], [@full-simulation-model], [@GA-full-simulation] [@CloudSim], [@CloudAnalysis], [@DCSim], [@sla-provision], [@parallel-RL-vertical-QoS-2013], [@MASCOTS11], [@E3-R-extended], [@wosp10sla], [@queue-VM-group], [@3-stages-game-theory], [@2013-rule-based-multi-elasiticity-2013], [@linear-mapping-CP-bundles-2012], [@qcloud], [@vm-fuzzy-MIMO], [@2014-decision-tree-software-CP-2014], [@dynamic-model-comparison], [@change-point], [@2013-single-learner-filter-wrapper-LR-2013], [@kalman-AR], [@kalman-clustering], [@ILP-cost-only-scaling-2013], [@2014-ARMA-single-agg-objective-2014], [@scale-rule-based], [@IWQOS11], [@2013-scaling-select-one-predictor-and-error-correlation-2013], [@2010-demand-pattern-match-2010], [@ensemble-prediction-VM-2014-full], [@seelam2015polyglot], [@souza2015using], [@farokhi2016hybrid], [@shariffdeen2016workload], [@gandhi2017model], [@qu2016reliable], [@sun2016roar], [@da2016autoelastic], [@rameshan2016augmenting], [@baresi2016discrete], [@zhang2016container], [@li2015rest], [@sun2017automated], [@awada2017improving]\ &&open& [@2013-elasticity-primitives-2013], [@MPC-price], [@cache-static-ANN-bi-obj-2012]\ &multiple&close&[@2014-kalman-pair-wised-coupling-on-tiers-2014], [@2012-app-VM-mapping-2012], [@EMA-CPU-memory-PD], [@MIMO-fuzzy-hill-climbing-2013], [@Arc-cover-all-controller], [@ICDCS2011], [@PCA-model-scaling-2013], [@2014-2-SVM-workload-type-2014], [@06032254], [@fuzzy-2-loop-control], [@MASCOTS11-bu-software-CP-full] [@2013-JRAO-most-closest-work-2013], [@TR-10-full-version], [@Chen:2013:iccs], [@Chen:2015:tsc-pending], [@Chen:2015:computer], [@multitier-resalloc-Cloud11], [@profit-cent-local-search], [@2014-navie-ANN-GA-2014], [@JSS_Kousiouris], [@Chen:2013:seams], [@MOCS2014-full], [@Chen:2014:ucc], [@Chen:2015:tse-pending], [@ma2016auto]\ OAD&single&close&[@van2015mnemos],[@herbein2016resource], [@HuBrKo2011-SEAMS-ResAlloc]\ &single&close&[@li-opt-clouds-4], [@cloud_computing_2010_5_20_50060-ext], [@Compsac_2010_I_Brandic], [@SEASS_2010_Michael_Maurer], [@tse1], [@fuzzy-vm-interference], [@queue-prediction], [@Emeakaroha_CloudComp2010], [@HPCS_IWCMC_Vincent], [@self-sla-and-resource], [@megahed2017stochastic], [@lakew2017kpi], [@aslanpour2017auto], [@gill2017chopper]\ &multiple&close&[@software-RP-two-loops], [@BRGA-resource],[@Chen:2014:seams]\ - 4.4.1.2 Multiple Loops:* It is possible to use multiple loops and controllers for autoscaling in the cloud. Here, multiple feedback loops operate in different levels of the architecture, e.g., one operates at the cloud level while the others operate on each VM. The benefit is that multiple loops provide low coupling in the design of the loops for SSCAS. Notably, multiple loop control can be used to separate global and local controls. Among others, Kalyvianaki et al. [@2014-kalman-pair-wised-coupling-on-tiers-2014] apply multiple decentralized feedback loops for autoscaling CPU in the cloud. Although it aims to exploit one loop per individual application, the controllers actually operate on each tier of an application. Chen and Bahsoon [@Chen:2015:tsc-pending] also leverage multiple feedback loop to auto-scale cloud services, where each PM maintains a loop. Unlike classic feedback loop where the adaptations occur only on the manageable part of SSCAS, their adaptations also happen on the manging part. Multiple loop control is also effective for isolating the logical aspects of autoscaling and management in the cloud. For instance, Wang, Xu and Zhao [@fuzzy-2-loop-control] propose a two layered feedback control for autoscaling in the cloud. The first layer, termed guest-to-host optimization, controls the hardware resources, e.g., CPU and memory. Subsequently, the host-to-guest optimization adapts the software configuration accordingly. ### Observe-Decide-Act Observe-Decide-Act (ODA) loop [@self-aware-ML-adaptive-control] is considered as an extended pattern of the generic feedback loop. As specified in the SEEC framework [@self-aware-ML-adaptive-control], ODA is unique in the sense that it decouples multiple loops to different roles (i.e., application developer, system developer, and the SEEC runtime decision infrastructure) in the development life-cycle, each role focuses on one or more steps in ODA. In such a way, ODA links the effects of human activities on the adaptive behaviors. Among others application of OAD in SSCAS, MNEMOS [@van2015mnemos] has relied on OAD to realize an integrated, datacenter-wide architecture for autoscaling resources in the cloud, in which the System Monitor acts as the observer, the Portfolio Scheduler acts as the decider, and the VM Manager acts as the executioner. Huber et al. [@HuBrKo2011-SEAMS-ResAlloc] also use ODA for self-aware autoscaling resources in the cloud. However, unlike traditional ODA loop, it has an additional Analysis step which is used to detect the type of problems that trigger adaptation. ### Monitor-Analyze-Plan-Execute Another pattern extended from the generic feedback loop, namely Monitor-Analyze-Plan-Execute (MAPE), is firstly proposed by IBM for architecting self-adaptive systems. In such pattern, the Decide step in OAD is further divided into two substeps, these are Analyze and Plan, where the former is particularly designed to determine the causes for adaptations, e.g., SLA violation; the latter, on the other hand, is responsible for reasoning about the possible actions for adaptation. MAPE sometime can be extended by a Knowledge component (a.k.a. MAPE-K) which maintains historical data and knowledge used by the system for better adaptation. MAPE can be also realized as either single or multiple loops: - 4.4.3.1 Single Loop: MAPE (or MAPE-K) is also widely applied for SSCAS. For example, the architecture of the FoSII project [@Compsac_2010_I_Brandic] [@SEASS_2010_Michael_Maurer] leverages single MAPE-K to realize the self-management interface, aiming to prevent SLA violations in cloud by devising the related actions. They also use the additional Knowledge (K) component to record cases and the related solutions, which can assist the autoscaling decision making. Chen and Bahsoon [@Chen:2014:seams] have realized MAPE as a single loop where the adaptations occur on both the managing and manageable parts of the SSCAS. - 4.4.3.2 Multiple Loops: Realizing multiple MAPE loops for SSCAS is also possible. Zhang et al. [@software-RP-two-loops] introduce an architecture for autoscaling using two nested MAPE loops. The first loop is responsible for adapting the software primitives while the other loop is used to change the hardware primitives. These two loops run sequentially upon autoscaling, that is, adapting the software control primitives before changing the hardware control primitives. Similarly, BRGA [@BRGA-resource] utilizes MAPE to realize a framework for autoscaling in the cloud. Such solution consists of both the local and global view of the cloud-based application. QoS Modeling {#sec:qos} ------------ RQ3: What are the approaches used to model the quality related to SSCAS? QoS modeling, or performance modeling, is a fundamental research theme in cloud computing and it can serve as useful foundations for addressing many research problems in the cloud [@lorido2014review], including autoscaling. The QoS models correlate the QoS attributes to various control primitives and environmental primitives. Clearly, these models are particularly important for SSCAS, since they are powerful tools that can assist the reasoning about the effects of adaptation on objectives in the autoscaling decision making process. Note that although QoS model can provide great helps to the decision making in SSCAS, not all of the studies have considered QoS modeling as part of their solutions. In fact, some of them rely on model-free solution, e.g., control theoretic approach, which we will review in Section \[sec:dm\]. Typically, QoS modeling consists of two phases: the primitives selection phase and the QoS model construction phase. More precisely, the primitives selection phase determines which and when the cloud primitives correlate with the QoS; while the QoS model construction phase identifies how these primitives correlate with the QoS, i.e., their magnitudes in the correlation. The QoS models might come as three forms: (i) static models where the models’ expression and their structure (e.g., the number of inputs and the coefficients) do not change over time; (ii) dynamic models which permits those changes; or (iii) semi-dynamic models in the sense that the expression (e.g., coefficients) could be dynamically updated but the input features do not. Further, those models can be built online at system runtime, or offline at the design phase of the system. In this section, we classified the studies mainly based on the modeling methods applied to the QoS model construction phase, since we found that the primitives selection is often conduced using manual and static approaches in the studies. As we can see in Figure \[fig:qos\], the majority of the studies has exploited analytical model (43%) and machine learning model (38%) to predict QoS. In contrast, simulation and hybrid model receives much less attention. From Table \[tb:qos\], we can obtain the following observations: [\|P[1.3cm]{}\|P[1.7cm]{}\|P[0.7cm]{}\|P[1.3cm]{}\|P[1.2cm]{}\|X\|]{} *Modeling Approach&Type&Built&QoS Interference&State&Concrete Models\ & Queuing model&offline&no&static& S-QUEUE (5):* [@MPC-price], [@06119056], [@E3-R-extended], [@queue-VM-group], [@typical-navie-scaling-VM-number] M-QUEUES (3): [@multitier-resalloc-Cloud11], [@queue-prediction], [@wosp10sla] LQN (4): [@3-stages-game-theory], [@fine-grained-servce-LQM-MIP-power], [@li-opt-clouds-4], [@sla-provision]\ &\[-0.05cm\]&online&no&semi& MDP (1): [@tse1] MODEL@RUNTIME (1): [@2014-eplison-GA-weigh-h-scaling-2014]\ &&offline&no&static& PCM (1): [@HuBrKo2011-SEAMS-ResAlloc] GRAPH (1): [@2013-rule-based-multi-elasiticity-2013]\ &&offline&no&dynamic&PCA (1): [@PCA-model-scaling-2013]\ &&offline&no&static& EMPIRICAL-MODEL (9): [@06032254], [@Compsac_2010_I_Brandic], [@2014-linear-centralized-decision-making-2014], [@Emeakaroha_CloudComp2010], [@HPCS_IWCMC_Vincent], [@profit-cent-local-search], [@cache-static-ANN-bi-obj-2012], [@BRGA-resource], [@shariffdeen2016workload], [@megahed2017stochastic], [@gill2017chopper], [@awada2017improving]\ &&offline&yes&static& EMPIRICAL-MODEL (1): [@ma2016auto]\ Simulation& &offline&no&semi& PROFILING (2): [@2014-profiling-decision-tree-scaling-2014], [@sun2016roar] SIMULATOR (4): [@full-simulation-model], [@CloudSim], [@CloudAnalysis], [@DCSim], [@GA-full-simulation]\ &&online&no&semi& LR (3): [@acdc09], [@software-RP-two-loops], [@linear-mapping-CP-bundles-2012] ARMA (1): [@HPL-2008-123R1-mimo], [@MIMO-fuzzy-hill-climbing-2013] KF (1): [@2014-kalman-pair-wised-coupling-on-tiers-2014]\ &&online&no&dynamic& LR (1): [@ISPASS07]\ &&online&yes&semi& MIMO (4): [@fuzzy-vm-interference], [@qcloud], [@vm-fuzzy-MIMO], [@2014-decision-tree-software-CP-2014]\ &&online&no&semi& KM (1): [@kriging-controller] RT (1): [@11icde_smartsla_full] ANN (4): [@dynamic-model-comparison] [@2014-navie-ANN-GA-2014] [@JSS_Kousiouris] [@Chen:2013:seams] SVM (2): [@2014-2-SVM-workload-type-2014] [@dynamic-model-comparison] CHANGE-POINT (1): [@change-point]\ &&online&yes&semi& SVM (1): [@rameshan2016augmenting]\ &&online&no&semi& ARMA+SVM (1): [@TR-10-full-version] ANN+ANN (1): [@MOCS2014-full]\ &&online&no&dynamic& KNN+LR:+RT (1): [@2013-single-learner-filter-wrapper-LR-2013]\ &&online&yes&dynamic& ARMA+ANN+RT (2): [@Chen:2014:ucc], [@Chen:2015:tse-pending]\ &&semi&no&semi& LQN+KF (2): [@2014-kalman+rule-based-2014], [@kalman-AR] LQN+KF+K-MEAN (1): [@kalman-clustering] S-QUEUE+ARMA (1): [@ICDCS2011] M-QUEUE+KF (1): [@gandhi2017model]\ &&offline&yes&static& EMPIRICAL-MODEL+PROFILING (1): [@sun2017automated]\ 1. Despite the high importance of QoS interference, it has not received much attention when modeling the QoS (only 13%). 2. Truly dynamic QoS modeling, i.e., changing both the input features and their coefficients, is still minority (7%). Other studies have merely considered changing coefficients of the model while ignoring the input features’ dynamics (54% semi-dynamic), or none at all (39% static). 3. The concrete modeling methods applied for machine learning model is more diverse than the methods in other categories. Additionally, from Table \[tb:qos-inout\], we can observe that: 1. Although most studies (65%) have only considered certain inputs/output during their experiments, they have claimed that their model is compatible with any given inputs (i.e., any cloud control primitives) and/or output (i.e., any QoS attributes). 2. The most widely considered input dimension is CPU while the most common output is response time (except the general one, i.e., QoS attribute). 3. The most explicitly modeled number of outputs is three, while the most explicitly considered number of inputs is four. In the following, we specify some representative studies on QoS modeling for SSCAS in details. [\|P[3cm]{}\|X\|]{} *Outputs&Cloud Primitives\ QoS attributes& CPU (1):* [@HuBrKo2011-SEAMS-ResAlloc] number of VM (1): [@2014-eplison-GA-weigh-h-scaling-2014], [@2014-profiling-decision-tree-scaling-2014] configurations (1): [@2014-2-SVM-workload-type-2014], [@2014-decision-tree-software-CP-2014] resources (2): [@2014-navie-ANN-GA-2014], [@qcloud] CPU and bandwidth (1): [@vm-fuzzy-MIMO] CPU and memory (4): [@kriging-controller], [@MIMO-fuzzy-hill-climbing-2013], [@fuzzy-vm-interference], [@GA-full-simulation] configurations and resources (6): [@PCA-model-scaling-2013], [@software-RP-two-loops], [@tse1], [@Chen:2014:ucc], [@Chen:2015:tse-pending], [@Chen:2013:seams] configurations, CPU and memory (1): [@TR-10-full-version] resources and workload (8): [@ISPASS07], [@JSS_Kousiouris], [@kalman-AR], [@full-simulation-model], [@CloudSim], [@CloudAnalysis], [@DCSim], [@gill2017chopper] CPU, memory and disk (2): [@2013-rule-based-multi-elasiticity-2013], [@HPL-2008-123R1-mimo] CPU, memory and bandwidth (3): [@2014-linear-centralized-decision-making-2014], [@dynamic-model-comparison], [@self-sla-and-resource] CPU, storage and bandwidth (2): [@06032254], [@Emeakaroha_CloudComp2010] configurations, resources and workload (1): [@2013-single-learner-filter-wrapper-LR-2013] CPU, bandwidth, storage and number of VM (1): [@Compsac_2010_I_Brandic] CPU, memory, workload and number of VM (1): [@11icde_smartsla_full] CPU, memory, workload and bandwidth: (1) [@linear-mapping-CP-bundles-2012] CPU, memory, storage and bandwidth (1): [@HPCS_IWCMC_Vincent] CPU, number of VMs and workload (1): [@ma2016auto] workload and interference index (1): [@rameshan2016augmenting]\ Response time& CPU (1): [@2014-kalman-pair-wised-coupling-on-tiers-2014] number of VM (1): [@multitier-resalloc-Cloud11] workload and number of VMs (4): [@3-stages-game-theory], [@queue-prediction], [@queue-VM-group], [@sla-provision] CPU and memory (2): [@E3-R-extended], [@MPC-price] workload and CPU (2): [@li-opt-clouds-4], [@ICDCS2011] thread and CPU (1): [@wosp10sla] CPU, memory, workload and number of VMs (2): [@fine-grained-servce-LQM-MIP-power], [@2014-kalman+rule-based-2014] CPU, workload and number of VMs (1): [@gandhi2017model]\ Response time and workload& number of VMs (1): [@cache-static-ANN-bi-obj-2012] CPU and memory (1): [@profit-cent-local-search] workload and number of VMs (1): [@typical-navie-scaling-VM-number] workload, number of VMs and VM type (1): [@sun2017automated]\ Response time and utilization & number of VMs and VM type (1): [@awada2017improving]\ Response time and throughput & CPU and memory (1): [@06119056] CPU, memory, number of VMs and VM type (1): [@sun2016roar]\ CPU utilization & workload and number of VM (1): [@acdc09]\ QoS attributes and hardware demand& configurations and resources (1): [@MOCS2014-full] CPU and workload (1): [@kalman-clustering]\ QoS attributes and workload& workload and number of PM (1): [@change-point]\ QoS attributes and overhead& resources (1): [@BRGA-resource]\ Cost& workload and number of VM (2): [@shariffdeen2016workload], [@megahed2017stochastic]\ ### Analytical Modeling Analytical modeling approaches rely on a closed-form structure to model the cloud-based service. These models are often built offline based on theoretical principles and assumptions. Next, we further divide the analytical modeling approach into queuing theory, dependability models and black box models. - 4.5.1.1 Queuing theory: Queuing model and queuing network are widely applied for QoS modeling in the cloud. They model the cloud-based services as a single queue or a collection of queues that are interacting through a mixture of request arrivals and completes. Specifically, a single queue has been used to model the correlation of response time (or throughput) to CPU, number of VM and workload. For example, depending on the assumption of the distribution on arrival and service rate, the model can be built as M/G/m queue[^4] by Zhang et al. [@MPC-price], M/G/m queue by Jiang et al. [@06119056], M/M/1 queue by E$^3$-R [@E3-R-extended] and JustSAT [@queue-VM-group], and M/M/m queue by Jiang et al. [@typical-navie-scaling-VM-number]. To create more detailed modeling with respect to the internal structure of cloud-based services, multiple queues can be used to create QoS models, for example, Goudarzi and Pedram [@multitier-resalloc-Cloud11] apply multiple queues to model the response time for cloud-based multi-tiered applications with respect to number of VM and workload. Their work calculates average response time for the queue in the forward direction throughout the tiers. Unlike classical queuing model and queuing network, the Layer Queuing Network (LQN) additionally model the dependencies presented in a complex workflow of requests to cloud-based services and applications. For instance, Zhu et al. [@sla-provision] have also used LQN where the authors employ a global M/M/m queue for the entire on-demand dispatcher and then a M/G/1 queue on each tier of an application. The former queue correlates the response time to number of VMs, while the latter queue models the relationship between response time and CPU of the VM that contains the corresponding tier. - 4.5.1.2 Dependability models: Dependability models focus on the modeling of various states for QoS attributes. For example, in QoSMOS [@tse1], the authors analytically solve the Markov Models (Discrete-Time Markov Chain and Markov Decision Process) to model the QoS for services in an application. The model correlates QoS attributes with hardware resources and workload. Huber et al. [@HuBrKo2011-SEAMS-ResAlloc] uses Palladio Component Model (PCM) as architecture-level QoS model since it permits to explicitly model different usage profiles and resource allocations. - 4.5.1.3 Black box models: Black-box models handle the QoS using empirical and historical domain knowledge. Among others, the CLOUDFARM framework [@2014-linear-centralized-decision-making-2014] uses a empirical QoS model where the correlation between certain QoS values and the required resource is captured (i.e., CPU). In particular, the authors assumed that the magnitudes of resources to the QoS values is known, as specified by the cloud service or application provider. Another study from Emeakaroha et al. [@Emeakaroha_CloudComp2010] [@HPCS_IWCMC_Vincent] propose an empirical model that maps the expected QoS values with CPU, memory, bandwidth and storage based on the assumptions of the system that being managed. ### Simulation Based Modeling QoS models can be also generated by various simulators; here, conducting simulations is usually a complex and expensive process and thus they are used in an offline manner. In practice, simulation is required to be setup by the domain experts, who will often need to analyze, interpret and profile the data collected after simulation runs. Specifically, Fernandez et al. [@2014-profiling-decision-tree-scaling-2014] have relied on a profiling approach that builds the QoS model for each bundle of VM offline. The process is similar to a simulation modeling approach. CDOSim [@full-simulation-model] is a framework that simulates the actual application in the cloud to restrict the search-space for autoscaling and to steer the exploration towards promising decisions. CloudSim [@CloudSim] is a simulation toolkit that models QoS attributes (of VM) with respect to resource allocation. It supports both single cloud and multiple clouds scenarios. As an extension of CloudSim, CloudAnalyst [@CloudAnalysis] allows the simulation of QoS attributes for the application deployed on geographically-distributed datacenters. Similarly, DCSim [@DCSim] simulates the overall quality of resource autoscaling for the entire cloud. ### Machine Learning Based Modeling The increasing complexity of cloud-based services has rendered the modeling process an extremely difficult task for human experts. To this end, recent studies have exploited the advances of machine learning algorithms and theory to create more reliable QoS models. In the following, we survey the key studies that apply machine learning approaches for QoS modeling in the cloud. In particular, we have further classified them into two categories, these are: linear and nonlinear modeling. - 4.5.3.1 Linear modeling: Learning algorithms based on linear models for QoS modeling in the cloud can handle linear correlation between a selected set of cloud primitives (e.g, CPU, memory, number of VM, workload etc) and output (i.e., QoS attributes), and they are sometime very efficient. Simple linear models most commonly rely on linear regression, where each primitive input is associated with a time-varying weight, e.g., Lim et al. [@acdc09], Zhang et al. [@software-RP-two-loops] and Collazo-Mojica et al. [@linear-mapping-CP-bundles-2012]. More advanced forms exist, e.g., Padala et al. [@HPL-2008-123R1-mimo] have used Auto-Regressive-Moving-Average (ARMA) model that is trained continually by Recursive Least Squares (RLS). The authors claim that the second-order linear ARMA model is easy to be estimated online and can simplify the corresponding controller design problem with adequate accuracy. We found that there are limited studies, which attempt to capture the information of QoS interference in the linear QoS model and they only focus on the VM-level [@fuzzy-vm-interference], [@qcloud], [@vm-fuzzy-MIMO], [@2014-decision-tree-software-CP-2014]. As an example, Q-Cloud [@qcloud] has explicitly considered QoS interference by using the hardware control primitives of all co-hosted VMs as inputs, rendering it in a Multi-Inputs-Multi-Output (MIMO) model, which is trained by Least Mean Square (LMS) method. - 4.5.3.2 Nonlinear modeling: Learning algorithms based on nonlinear models for QoS modeling in the cloud is able to capture complex and nonlinear correlation, in addition to the linear one. However, it can also produce relatively large overhead than the linear modeling. Here, existing studies often aim to model the correlation between hardware control primitives (e.g., CPU, memory and bandwidth) and QoS. The nonlinear modeling can be relied on kriging model [@kriging-controller], Regression Tree (RT)[@11icde_smartsla_full], Artificial Neural Network (ANN) [@dynamic-model-comparison] [@2014-navie-ANN-GA-2014] [@JSS_Kousiouris] [@Chen:2013:seams], Support Vector Machine (SVM) [@2014-2-SVM-workload-type-2014] [@dynamic-model-comparison], change-point detection [@change-point]. For example, SmartSLA [@11icde_smartsla_full] employs Regression Tree (RT) and boosting to model the QoS. RT partitions the parameter space in a top-down fashion, and organizes the regions into a tree style. The tree is then trained by M5P where the leaves are regression models. The study from Kunda et al. [@dynamic-model-comparison] presents sub-modeling based on ANN and SVM for correlating QoS with hardware control primitives in the cloud. Instead of building a single model for a QoS attribute, they train n sub-models, whereby n is determined by performing k-mean clustering based on the similarity between data values of QoS, creating more accurate and finer grained models. - 4.5.3.3 Ensemble modeling: Examples exist for cases where multiple linear and/or nonlinear machine learning algorithms are explored together. Among others, Chen, Bahsoon and Yao [@Chen:2014:ucc] [@Chen:2015:tse-pending] exploit a bucket of learning algorithms (both linear and non-linear models). The model accuracies are tracked continually at runtime, considering QoS interference. The best model for a given input values, according to both local and global errors, will be used to make prediction. - 4.5.3.4 Comparison of different learning algorithms: Given the various types of machine learning algorithms, it can be difficult to determine which one(s) are the appropriate algorithms for QoS modeling in the cloud, with respect to both accuracy and overhead. There are researches that have conducted empirical comparisons of different possible learning algorithms for QoS modeling in the cloud [@2013-offline-profiling-2013] [@determine-CP-compare-models] [@tp38]. ### Hybrid Modeling We discovered that linear machine learning algorithms are also commonly used with analytical approaches to form QoS models. Specifically, Grandhi et al. [@2014-kalman+rule-based-2014] and Zheng et al. [@kalman-AR] have proposed hybrid model: to model the multi-tiered application, they have relied on a modified LQN containing some time-varying coefficients. The authors then employ the Kalman filter as an online parameter estimator to continually estimate those coefficients. The approach proposed by Xiong et al. [@ICDCS2011] has relied on a combined model, where a M/G/1 queue is used to model the correlation of workload to response time; while ARMA is used to model the relationship between response time and CPU. ### Dynamic Primitives Selection We noticed that the majority of the aforementioned studies regard the primitives selection as a manual and offline process, most commonly, they have relied on empirical knowledge and heavy human analysis to select the important primitives as the input features of QoS models. Although not many, there are some studies that explicitly consider dynamic process in primitives selection, which tends to be more accurate and can be easily applied [@PCA-model-scaling-2013] [@ISPASS07] [@2013-single-learner-filter-wrapper-LR-2013] [@Chen:2015:tse-pending]. As an example, vPerfGuard [@2013-single-learner-filter-wrapper-LR-2013] is a framework that correlates QoS attributes with respect to software control primitives, hardware control primitives and environmental primitives. The authors achieve primitive selection based on both filter (relevance based correlation coefficient) and wrapper (i.e., hill-climbing comparison for different learning algorithms). Chen and Bahsoon [@Chen:2015:tse-pending] dynamically select primitives that maximize both information relevance (between a primitive and quality) and minimize redundancy (between already selected primitives). While explicitly modeling the effects of QoS interference, the authors propose a fully self-adaptive approach that selects primitives that improve prediction accuracy given a learning algorithm. ### Workload and Demand Modeling We found that some existing studies (e.g., [@2011-cost-aware-v-h-scaling-2011], [@ILP-cost-only-scaling-2013], [@JSS_Kousiouris]) attempts to model the workload and demand for assisting autoscaling decision making. In those cases, the modeling is reduced to a single dimension, where the core is to model the trend of the workload or demand using its historical data. However, unlike QoS modeling which is often multi-dimensional, the single dimension in workload or demand models do not offer the ability to reason about the effects of autoscaling decisions and the possible trade-offs. Granularity of Control {#sec:gra} ---------------------- RQ4: What is the granularity of control in SSCAS? The ultimate goal of autoscaling is to optimize the QoS and cost objectives, which are referred to as benefit, for all cloud-based services. To this end, the granularity of control in autoscaling plays an integral role, since it determines the boundary of decision making: which and how many objectives should be considered in a decision making process of autoscaling. In the following, we classify existing SSCAS studies depending on what level of granularity they operate at. As we can see from Figure \[fig:gra\], the predominant granularity of control is at the service/application level where the boundary of decisions making is grouped by each service/application. Notably, controlling at the cloud level tends to be the second most popular, leaving the other levels being minority. Generally, the finer granularity of control implies that it is harder to achieve globally-optimal benefit but likely to generate smaller overhead. On the other hand, globally-optimal benefit can be easier reached with large overhead if the granularity of control is coarser. According to Table \[tb:gra\], we can obtained the following observations: [\|P[1.6cm]{}\|P[1.3cm]{}\|P[1.6cm]{}\|X\|]{} *Granularity&Entity&Style&Representative Examples\ &Application&Decentralized& [@2011-cost-aware-v-h-scaling-2011], [@queue-VM-group], [@2014-profiling-decision-tree-scaling-2014], [@linear-mapping-CP-bundles-2012], [@TR-10-full-version], [@wosp10sla], [@06119056], [@cloud_computing_2010_5_20_50060-ext], [@signal-resource-trend-prediction], [@PCA-model-scaling-2013], [@acdc09], [@ILP-cost-only-scaling-2013], [@typical-navie-scaling-VM-number], [@multitier-resalloc-Cloud11], [@2014-decision-tree-software-CP-2014], [@2014-aco-app-to-VM-consolidation-2014], [@scale-rule-based], [@2014-kalman+rule-based-2014], [@ICDCS2011], [@IWQOS11], [@compare-to-Rao-fuzzy-2013], [@HPL-2008-123R1-mimo], [@souza2015using], [@farokhi2016hybrid], [@shariffdeen2016workload], [@megahed2017stochastic], [@lakew2017kpi], [@aslanpour2017auto]\ &Service&Decentralized& [@2013-rule-based-multi-elasiticity-2013], [@2014-ARMA-single-agg-objective-2014], [@Compsac_2010_I_Brandic], [@SEASS_2010_Michael_Maurer], [@kriging-controller], [@tse1], [@2014-eplison-GA-weigh-h-scaling-2014], [@E3-R-extended], [@GA-full-simulation], [@2013-scaling-select-one-predictor-and-error-correlation-2013], [@2010-demand-pattern-match-2010], [@cache-static-ANN-bi-obj-2012], [@HuBrKo2011-SEAMS-ResAlloc]\ & Application&Decentralized& [@fuzzy-2-loop-control], [@parallel-RL-vertical-QoS-2013], [@2013-software-CP-only-2013], [@software-RP-two-loops], [@2014-2-SVM-workload-type-2014], [@vm-fuzzy-MIMO], [@qcloud], [@gandhi2017model],\ &Application&Centralized& [@gill2017chopper], [@sun2017automated], [@awada2017improving]\ &VM&Decentralized& [@2014-fuzzy-compare-to-JRao-2014], [@2014-kalman-pair-wised-coupling-on-tiers-2014], [@ensemble-prediction-VM-2014-full]\ &VM&Centralized& [@qu2016reliable], [@da2016autoelastic], [@rameshan2016augmenting]\ &Container&Decentralized& [@baresi2016discrete], [@li2015rest]\ &Container&Centralized& [@zhang2016container]\ PM level&Application&Decentralized&[@MASCOTS11], [@MASCOTS11-bu-software-CP-full], [@2013-JRAO-most-closest-work-2013], [@2014-navie-ANN-GA-2014], [@fuzzy-vm-interference]\ &Application&Centralized&[@3-stages-game-theory], [@2012-app-VM-mapping-2012], [@2013-elasticity-primitives-2013], [@profit-cent-local-search], [@2014-linear-centralized-decision-making-2014], [@li-opt-clouds-4], [@sla-provision], [@06032254], [@fine-grained-servce-LQM-MIP-power], [@MPC-price], [@EMA-CPU-memory-PD], [@BRGA-resource], [@self-sla-and-resource], [@van2015mnemos], [@herbein2016resource], [@2013-self-organising-map-2013]\ &Application&Decentralized&[@05557978],[@Arc-cover-all-controller]\ &Application&Decentralized&[@MIMO-fuzzy-hill-climbing-2013], [@11icde_smartsla_full], [@ma2016auto]\ &Application&Centralized&[@seelam2015polyglot]\ &Service&Decentralized&[@Chen:2013:iccs], [@Chen:2014:seams], [@Chen:2015:tsc-pending], [@Chen:2015:computer]\ 1. Most of the studies (72%) see each application as the basic entity regardless to the granularity of control in SSCAS. 2. Decentralization (74%) is the most popular approach for all granularity of control, except the cloud level where centralized (or partially centralized) control is predominately exploited. In the following, we specify each granularity of control for SSCAS in details. ### Controlling at Service and Application Level Service/Application level is the finest level of control in a SSCAS. It is worth noting that by service, we refer to any conceptual part of the system being managed. As a result, control granularity at the service/application level may refer to independently controlling/scaling an application, a tier of an application or a cloud-based service. We found that most of the studies have focused on controlling each cloud-based application. These approaches have relied on controlling the QoS and/or cost for each individual application in isolation, and therefore, they sometimes regard an application as a service. Examples of such include: Lim et al. [@acdc09] control the application and its required VM, in which case an application is regarded as a service. Sedaghat et al. [@ILP-cost-only-scaling-2013] regard application as a service, and considered the required number of VMs and the fixed VM bundles for such service. There are studies that explicitly controls cloud-based service in general. Among others, Copli et al. [@2013-rule-based-multi-elasiticity-2013] control the QoS, cost and their elasticity for each service deployed in the cloud. Yang et al. [@2014-ARMA-single-agg-objective-2014] control the cost of individual cloud-based services. The FoSII project [@Compsac_2010_I_Brandic] controls individual cloud-based service, their QoS and cost. Gambi et al. [@kriging-controller] control at the service level, where the controller decides on the optimal autoscaling decision for cloud-based service in isolation. ### Controlling at Virtual Machine and Container Level VM hypervisor and container are two fundamental infrastructure that underpin cloud computing. In particular, VM and container differ in the sense that the former requires a full Operating System to be installed on a VM while the latter does not. This fact allows the container to set naively with the host PM, providing much faster creation and removal time of VM image. However, such benefit comes in the expenses of weaker security guarantees and potentially greater chances of interference, given that the container instances have less isolation. Despite such a difference, the two infrastructures are conceptually similar as they both aim to provide certain level of isolation on top of the hosting PM, and thus they can be regarded as the same granularity of control. VM level means that the control and decision making operate at each VM in the SSCAS. In particular, certain studies assumes a one-to-one mapping between application (or a tier) and VM and thus they can be categorized as either service level or VM level granularity. To better separate them from the pure service/application level granularity of control, these studies are regarded as VM level granularity. Specifically, FC2Q [@2014-fuzzy-compare-to-JRao-2014] regards application tier and VM interchangeably, therefore controlling each tier of an application is equivalent to control each individual VM. Similarly, Kalyvianaki et al. [@2014-kalman-pair-wised-coupling-on-tiers-2014] control a tier of an application that resides on a VM, and the authors only focus on CPU allocation of a VM. ### Controlling at Physical Machine Level Autoscaling decision making on each PM independently is referred to as PM level control in the SSCAS. The primary intention of PM level control is to manage the QoS interference caused by co-hosted VMs. Among the others, Xu et al. [@MASCOTS11] [@MASCOTS11-bu-software-CP-full] control the VMs collectively at the PM level, thus the autoscaling promotes better management of QoS interference at the VM level. ### Controlling at Cloud Level The most coarse level of control granularity is at the cloud level for SSCAS. The majority of the studies achieves autoscaling at the cloud level by using a centralized and global controller, with an aim to manage utility ([@3-stages-game-theory], [@2012-app-VM-mapping-2012], [@6008793], [@2013-elasticity-primitives-2013], [@profit-cent-local-search], [@2014-linear-centralized-decision-making-2014]), profits ([@li-opt-clouds-4], [@sla-provision]) and availability ([@06032254]). Among others, Ferretti et al. [@05557978] control the QoS for all cloud-based services in a global manner. However, the actual deployment can be either centralized or decentralized. Similarly, CRAMP [@EMA-CPU-memory-PD] uses a centralized and global controller, it controls the entire cloud for cost and QoS. CloudOpt [@fine-grained-servce-LQM-MIP-power] also controls the entire cloud using centralized control, as the considered optimization involves all the PM in the cloud. Some of the studies have relied on a decentralized manner where a consensus protocol is employed for controlling at the cloud granularity. For example, Wuhib et al. [@Arc-cover-all-controller] aim to control the entire cloud, and thus the QoS and the overall power consumption of cloud can be collectively managed. Further, they have relied on decentralized deployment, which can reduce the overhead of cloud-level control. ### Controlling at Hybrid Level We found that it is also possible for SSCAS to operate at multiple and hybrid levels, with an aim to better manage the overhead and global benefit. For example, Minarolli and Freisleben [@MIMO-fuzzy-hill-climbing-2013] combine both PM level and cloud level control, where the PM level is decentralized and the objective is to optimize the utility locally. Similarly, SmartSLA [@11icde_smartsla_full] aims to control the resource allocation for all the cloud-based services, therefore it utilizes a global, cloud-level control in addition to the decentralized local control on each VM. Different to the others, Chen and Bahsoon [@Chen:2014:seams] exploit a dynamic schema where the multiple simultaneously presented granularity are changed at runtime, according to the objective-dependency. Trade-off Decision Making {#sec:dm} ------------------------- RQ5: What are the approaches used for decision making in SSCAS?* The final important logical aspect in cloud autoscaling is the challenging decision making process, with the goal to optimize QoS and cost objectives. It is even harder to handle the trade-off between possibly conflicting objectives. Such decision making process is essentially a combinatorial optimization problem where the output is the optimal (or near-optimal) decision containing the newly configured values for all related control primitives. In the following, we survey the key studies on the decision making for SSCAS. In particular, we classify them into three categories, these are rule based control, control theoretic approach and search-based optimization. As we can see from Figure \[fig:dm\], while rule-based control and control theoretic approach share similar popularity in the studies, search-based optimization receives much more attentions than the other two. From Table \[tb:dm\], we can observe that: 1. Most of the studies (63%) in SSCAS do not attempt to consider the trade-off between objectives during the decision making, or they handle such trade-off in the way that different objectives are aggregated using weighted sum (29%), which essentially converting the multiple objectives into a single one. 2. Explicit consideration of QoS interference (16%) is still rare during the decision making. From Table \[tb:dm-inout1\] and \[tb:dm-inout2\] we can see that: 1. Most of the studies (78%) claim that they can work on any given objectives, thus the QoS attributes and cost being the most popular objectives to be improved during the decision making process of SSCAS. 2. Hardware resources, particularly CPU and memory, are the most commonly considered dimension of control primitives to be tuned in SSCAS. However, there are little studies (9%) that consider the interplay between software and hardware control primitives. 3. Considerable amount of studies (34%) has assumed bundles on the autoscaling decision, which will reduce the search space but might negatively constrain the quality of decision making. Observations from Table \[tb:dm-vh\] shows us that: 1. The majority of the studies (66%) has considered both vertical and horizontal scaling. 2. Horizontal scaling receives much more attentions than the vertical one. In the following, we specify some of the decision making approaches for SSCAS in details. [\|P[2cm]{}\|P[0.8cm]{}\|P[1.5cm]{}\|P[1.2cm]{}\|Xp[2.5cm]{}\|]{} *Decision Making Approach&Form&Trade-off&QoS Interference&Concrete Methods&\ &&None&No&\ &&None&Yes&\ &&None&No&PDC (3):* [@2012-app-VM-mapping-2012], [@acdc09], [@EMA-CPU-memory-PD] KC (2): [@2014-kalman-pair-wised-coupling-on-tiers-2014], [@2014-kalman+rule-based-2014] FC (3): [@2014-fuzzy-compare-to-JRao-2014], [@compare-to-Rao-fuzzy-2013], [@fuzzy-2-loop-control]&\ &&Weighted sum&No&FC (1): [@IWQOS11]&\ &&None&No&PIC+LA (1): [@ICDCS2011] MPC+MA (1): [@farokhi2016hybrid] PIC+ILP (1): [@baresi2016discrete] & ANN+FC (1): [@2013-software-CP-only-2013] MIMO(1): [@lakew2017kpi]\ &&Weighted sum&No&FC+QPS (1): [@MIMO-fuzzy-hill-climbing-2013] & PIDC+RL+ES (1): [@TR-10-full-version]\ &&Weighted sum&Yes& MPC+QP (1): [@fuzzy-vm-interference]&\ &&None&Yes& FUZZY-MIMO (1): [@vm-fuzzy-MIMO] & MIMO (1): [@qcloud]\ &&None&Yes&RL (3):[@MASCOTS11-bu-software-CP-full], [@MASCOTS11], [@2013-JRAO-most-closest-work-2013]&\ &&None&No&\ &&Single&No&HEURISTIC (2): [@2014-ARMA-single-agg-objective-2014], [@ma2016auto] DP (1): [@MPC-price] ES (7): [@PCA-model-scaling-2013], [@queue-VM-group], [@linear-mapping-CP-bundles-2012], [@HuBrKo2011-SEAMS-ResAlloc], [@shariffdeen2016workload], [@megahed2017stochastic], [@zhang2016container] ILP (2): [@2011-cost-aware-v-h-scaling-2011], [@ILP-cost-only-scaling-2013] & MIP (1): [@fine-grained-servce-LQM-MIP-power] ACO (1): [@2014-aco-app-to-VM-consolidation-2014] LA (1): [@2014-2-SVM-workload-type-2014]\ &&Single&Yes& HEURISTIC+DT (1): [@2014-decision-tree-software-CP-2014]&\ &&Weighted sum&Yes&ES (2): [@3-stages-game-theory], [@gandhi2017model] RS (1): [@Chen:2014:seams] & HEURISTIC (1): [@sun2017automated]\ &&Weighted sum&No&ES (6): [@tse1], [@kriging-controller], [@typical-navie-scaling-VM-number], [@Compsac_2010_I_Brandic], [@SEASS_2010_Michael_Maurer], [@awada2017improving] NFM (1): [@li-opt-clouds-4] FDS (1): [@multitier-resalloc-Cloud11] BS (1): [@cache-static-ANN-bi-obj-2012] DP (1): [@2014-linear-centralized-decision-making-2014] & LS (1): [@profit-cent-local-search] GS (1): [@11icde_smartsla_full] DT (1): [@2014-profiling-decision-tree-scaling-2014] QP (1): [@HPL-2008-123R1-mimo]\ &&Weighted sum&No&TS (1): [@sla-provision] GA (3): [@software-RP-two-loops], [@2014-navie-ANN-GA-2014], [@BRGA-resource] PSO (1): [@software-RP-two-loops]&\ &&Pareto&No&SMS-MOEA (1): [@wosp10sla] NSGA-II (3): [@2014-eplison-GA-weigh-h-scaling-2014], [@E3-R-extended], [@GA-full-simulation]&\ &&Pareto&Yes&MOACO (3): [@Chen:2013:iccs], [@Chen:2015:tsc-pending], [@Chen:2015:computer] &\ ### Rule Based Control Rule-based control is the most classic approaches for making decision in SSCAS. Commonly, one or more conditions are manually specified and mapped to a decision, e.g., increase CPU and memory by x if the throughput is lower than y. Therefore, the possible trade-off is often implicitly handled by the conditions and actions mapping. Specifically, Cloudline [@2013-elasticity-primitives-2013] allows programmable elasticity rules to drive autoscaling decisions. It is also possible to modify these rules at runtime as required by the users. Copil et al. [@2013-rule-based-multi-elasiticity-2013] handle the decision making process by specifying different condition-and-actions mapping for autoscaling in the cloud. In addition, the rules can be defined at different levels, e.g., PaaS and IaaS. Similarly, Ferretti et al. [@05557978] allow to setup mapping between QoS expectation and actions using XML like notations. [\|P[3cm]{}\|P[1.2cm]{}\|X\|]{} *Objective&Bundles&Control Primitives\ &Yes&number of VMs (3):* [@2014-aco-app-to-VM-consolidation-2014], [@shariffdeen2016workload], [@qu2016reliable] CPU and memory (2): [@2011-cost-aware-v-h-scaling-2011], [@MPC-price] CPU, memory and number of VM (1): [@ILP-cost-only-scaling-2013], [@fine-grained-servce-LQM-MIP-power]\ &No&number of VMs (1): [@megahed2017stochastic] configurations (1): [@2014-2-SVM-workload-type-2014] CPU (1): [@ICDCS2011]\ &Yes&number of VMs (3): [@3-stages-game-theory], [@typical-navie-scaling-VM-number], [@cache-static-ANN-bi-obj-2012] resources (1): [@2014-ARMA-single-agg-objective-2014]\ &No&number of VM (1): [@sla-provision] CPU, memory and bandwidth (1): [@scale-rule-based] CPU, memory and number of VM (1): [@2014-kalman+rule-based-2014]\ &Yes&number of VMs (1): [@2012-app-VM-mapping-2012] configurations and resources (1): [@PCA-model-scaling-2013] CPU and memory (1): [@2014-profiling-decision-tree-scaling-2014] CPU, memory and number of VMs (2): [@2013-elasticity-primitives-2013], [@EMA-CPU-memory-PD] CPU, memory and disk (1): [@2013-elasticity-primitives-2013] CPU, memory and bandwidth (2): [@2014-linear-centralized-decision-making-2014], [@linear-mapping-CP-bundles-2012]\ &No&CPU (4): [@2014-fuzzy-compare-to-JRao-2014], [@li-opt-clouds-4], [@IWQOS11], [@compare-to-Rao-fuzzy-2013] number of VM (2): [@multitier-resalloc-Cloud11], [@2014-eplison-GA-weigh-h-scaling-2014] resources (2): [@tse1], [@2014-navie-ANN-GA-2014] CPU and memory (6): [@kriging-controller], [@MIMO-fuzzy-hill-climbing-2013], [@parallel-RL-vertical-QoS-2013], [@MASCOTS11], [@fuzzy-vm-interference], [@GA-full-simulation] configurations and resources (4): [@Chen:2013:iccs], [@Chen:2015:tsc-pending], [@Chen:2015:computer], [@Chen:2014:seams] CPU, memory and disk (1): [@HPL-2008-123R1-mimo] CPU, memory and configurations (1): [@TR-10-full-version] CPU, storage and bandwidth (1): [@06032254] CPU, memory and thread (2): [@fuzzy-2-loop-control], [@MASCOTS11-bu-software-CP-full] CPU, bandwidth, storage and number of VM (1): [@Compsac_2010_I_Brandic] CPU, thread, session, buffer and memory (1): [@2013-JRAO-most-closest-work-2013].\ QoS attributes&No& configurations (1): [@2013-software-CP-only-2013] configurations and resources (1): [@software-RP-two-loops] CPU, memory and bandwidth (1): [@05557978] thread, CPU and memory (1): [@cloud_computing_2010_5_20_50060-ext] CPU and memory (1): [@lakew2017kpi] resources (1): [@gill2017chopper] number of VMs (1): [@rameshan2016augmenting]\ &No&CPU (1): [@2014-kalman-pair-wised-coupling-on-tiers-2014] memory (1): [@farokhi2016hybrid] CPU and bandwidth (2): [@vm-fuzzy-MIMO], [@qcloud] CPU and thread (1): [@wosp10sla] CPU and number of VMs (1): [@gandhi2017model] CPU, memory and number of VMs (1): [@li2015rest]\ &Yes& CPU, memory and number of VMs (1): [@baresi2016discrete]\ [\|P[5.8cm]{}\|P[1cm]{}\|X\|]{} *Objective&Bundles&Control Primitives\ CPU utilization&No&CPU and number of VMs (1):* [@acdc09]\ CPU utilization&Yes&number of VMs (1):[@da2016autoelastic] CPU and number of VMs (1): [@ma2016auto]\ Throughput&Yes&number of VMs (1): [@seelam2015polyglot]\ General utilization&No& configurations (1): [@2014-decision-tree-software-CP-2014] number of VMs (1): [@2013-scaling-select-one-predictor-and-error-correlation-2013] resources (1): [@signal-resource-trend-prediction] CPU (2): [@2010-demand-pattern-match-2010], [@herbein2016resource] CPU and memory (1): [@van2015mnemos]\ General utilization&Yes&number of VMs (1): [@ensemble-prediction-VM-2014-full] CPU, memory, number of VMs and bandwidth (1): [@zhang2016container]\ QoS attributes and power&Yes&CPU, memory and number of VM (1): [@Arc-cover-all-controller]\ Response time and utilization&Yes&number of VM (1): [@aslanpour2017auto] number of VM and VM type (1): [@awada2017improving]\ Response time, cost and availability&No&CPU and memory (1): [@profit-cent-local-search]\ VM consumption&Yes&number of VM (2): [@queue-prediction], [@queue-VM-group]\ SLA penalty&No&CPU, memory and number of VM (1): [@11icde_smartsla_full]\ Response time, throughput, CPU utilization and cost&No&CPU and memory (1): [@06119056]\ Response time, throughput, CPU utilization&Yes&number of VMs and VM type (1): [@sun2017automated]\ SLA and power&No&resources (1): [@2013-self-organising-map-2013]\ SLA and power&No&CPU (1): [@souza2015using]\ Response time, throughput and cost&No&CPU and memory (1): [@E3-R-extended]\ Benefits and overhead&No&CPU and memory (1): [@BRGA-resource]\ QoS attributes, utilization and cost&No&CPU (1): [@HuBrKo2011-SEAMS-ResAlloc]\ QoS attributes, utilization, number of actions&No&CPU, memory and bandwidth (1): [@self-sla-and-resource]\ [\|P[1.3cm]{}\|X\|]{} *Scaling&Decision Making Approaches\ Vertical& RULES (1):* [@souza2015using] CONTROL THEORY (4): [@2014-kalman-pair-wised-coupling-on-tiers-2014], [@2013-software-CP-only-2013], [@farokhi2016hybrid], [@lakew2017kpi] SEARCH-BASED OPTIMIZATION (6): [@parallel-RL-vertical-QoS-2013], [@tse1], [@wosp10sla], [@06119056], [@2010-demand-pattern-match-2010], [@2014-decision-tree-software-CP-2014]\ Horizontal& RULES (6): [@herbein2016resource], [@Arc-cover-all-controller], [@seelam2015polyglot], [@aslanpour2017auto], [@da2016autoelastic], [@rameshan2016augmenting] CONTROL THEORY (1): [@2012-app-VM-mapping-2012] SEARCH-BASED OPTIMIZATION (13): [@3-stages-game-theory], [@MPC-price], [@multitier-resalloc-Cloud11], [@queue-VM-group], [@sla-provision], [@typical-navie-scaling-VM-number], [@cache-static-ANN-bi-obj-2012], [@2014-aco-app-to-VM-consolidation-2014], [@2013-scaling-select-one-predictor-and-error-correlation-2013], [@ensemble-prediction-VM-2014-full], [@shariffdeen2016workload], [@megahed2017stochastic], [@sun2017automated]\ Both& RULES (12): [@2013-elasticity-primitives-2013], [@2013-rule-based-multi-elasiticity-2013], [@05557978], [@06032254], [@Compsac_2010_I_Brandic], [@scale-rule-based], [@self-sla-and-resource], [@cloud_computing_2010_5_20_50060-ext], [@van2015mnemos], [@qu2016reliable], [@gill2017chopper], [@li2015rest] CONTROL THEORY (14):[@2014-fuzzy-compare-to-JRao-2014], [@acdc09], [@EMA-CPU-memory-PD], [@fuzzy-2-loop-control], [@MIMO-fuzzy-hill-climbing-2013], [@2014-kalman+rule-based-2014], [@ICDCS2011], [@IWQOS11], [@TR-10-full-version], [@fuzzy-vm-interference], [@compare-to-Rao-fuzzy-2013], [@vm-fuzzy-MIMO], [@qcloud], [@baresi2016discrete] SEARCH-BASED OPTIMIZATION (35): [@MASCOTS11-bu-software-CP-full], [@MASCOTS11], [@2011-cost-aware-v-h-scaling-2011], [@2014-ARMA-single-agg-objective-2014], [@2014-linear-centralized-decision-making-2014], [@fine-grained-servce-LQM-MIP-power], [@kriging-controller], [@li-opt-clouds-4], [@PCA-model-scaling-2013], [@profit-cent-local-search], [@software-RP-two-loops], [@2014-2-SVM-workload-type-2014], [@2014-navie-ANN-GA-2014], [@2014-profiling-decision-tree-scaling-2014], [@HPL-2008-123R1-mimo], [@linear-mapping-CP-bundles-2012], [@2014-eplison-GA-weigh-h-scaling-2014], [@BRGA-resource], [@E3-R-extended], [@GA-full-simulation], [@HuBrKo2011-SEAMS-ResAlloc], [@2013-JRAO-most-closest-work-2013], [@signal-resource-trend-prediction], [@2013-self-organising-map-2013], [@ILP-cost-only-scaling-2013], [@Chen:2014:seams],[@SEASS_2010_Michael_Maurer], [@11icde_smartsla_full], [@Chen:2013:iccs], [@Chen:2015:tsc-pending], [@Chen:2015:computer], [@gandhi2017model], [@ma2016auto], [@zhang2016container], [@awada2017improving]\ ### Control Theoretic Approach Advanced control theory is another widely investigated approach for autoscaling decision making in SSCAS because of its low latency and dynamic nature. Studies in this category could be either standard, i.e., they rely solely on the classic control theory; or extended where additional methods are considered. Among the others, standard controllers (e.g., Proportional-Derivative control [@2012-app-VM-mapping-2012] [@acdc09] [@EMA-CPU-memory-PD], Kalman control [@2014-kalman-pair-wised-coupling-on-tiers-2014] [@2014-kalman+rule-based-2014] and Fuzzy control [@2014-fuzzy-compare-to-JRao-2014] [@compare-to-Rao-fuzzy-2013] [@fuzzy-2-loop-control] ) are commonly designed as a sole approach to make autoscaling decisions in the cloud. Specifically, ARUVE [@2012-app-VM-mapping-2012] and CRAMP [@EMA-CPU-memory-PD] utilizes a Proportional-Derivative (PD) controller, where the proportional and derivative factors are not sensitive to a concrete QoS model while supporting proactive autoscaling of cloud services and applications in a shared hosting environment. Anglano et al. [@2014-fuzzy-compare-to-JRao-2014] and Albano et al. [@compare-to-Rao-fuzzy-2013] apply fuzzy control that is updated by fuzzy rules at runtime. The aim is to optimize both QoS, cost and energy by autoscaling hardware resources. Although the authors claim they can cope with any hardware resources, only CPU tuning is explored. They have also ignored the QoS interference. We have also found that control theoretic approaches can be sometime used with other algorithms to better facilitate the autoscaling decision making, forming extended controller: [@ICDCS2011], [@MIMO-fuzzy-hill-climbing-2013], [@2013-software-CP-only-2013], [@IWQOS11], [@TR-10-full-version], [@fuzzy-vm-interference]. Particularly, the gains in the controllers can be further tuned by optimization and/or machine learning algorithms, and this is especially useful for Model Predictive Control (MPC). Among others, Zhu and Agrawal [@TR-10-full-version] utilize a Proportional-Integral-Derivative (PID) and reinforcement learning controller for decision making with respect to adapting software control primitives. Such result is then tuned in conjunction with the hardware control primitives using exhaustive search. The QoS attributes and cost are formulated as weighted-sum relation. The autoscaling decision making process in APPLEware [@fuzzy-vm-interference] have relied on Model Predictive Control (MPC), with the aim to optimize a cost function that represents the local objectives and resource constraints at a point in time. The state of an application, together with the other autoscaling decisions from the neighboring VMs, are collectively considered in a quadratic programming solver. ### Search-Based Optimization A large amount of existing studies of SSCAS relies on search-based optimization, in which the decisions and trade-offs are extensively reasoned in a finite, but possibly large search space. Depending on the algorithms, search-based optimization for autoscaling decision making in the cloud can be either explicit or implicitthe former performs optimization as guided by explicit system models; while this process is not required for the later. - 4.7.3.1 Implicit search: The implicit and search-based optimization approaches for autoscaling decision making do not use QoS models. Similar to the control theoretic approaches, the implicit search is also limited in reasoning about the possible trade-offs. For example, the study from Xu et al. [@MASCOTS11] , [@MASCOTS11-bu-software-CP-full] applies a model-free Reinforcement Learning (RL) approach for adapting thread, CPU and memory for QoS and cost. The approach is however implicit, providing that there is neither explicit system models nor explicit optimization. The authors have considered QoS interference during autoscaling. Similarly, VScale [@parallel-RL-vertical-QoS-2013] utilizes RL for making autoscaling decisions, which are then achieved by vertical scaling. The RL is realized by using parallel learning, that is to say, the authors intend to speed up agent’s learning process of approximated model by learning in parallel, without visiting every state-action pair in a given environment. The approaches that rely on demand prediction (e.g., the Autoflex [@2013-scaling-select-one-predictor-and-error-correlation-2013], PRESS [@signal-resource-trend-prediction], [@ILP-cost-only-scaling-2013], [@2010-demand-pattern-match-2010], [@2013-self-organising-map-2013] and [@ensemble-prediction-VM-2014-full]) are also regarded as implicit search. This is because the autoscaling decision is directly predicted by the demand models, without the needs of reasoning and optimization. - 4.7.3.2 Explicit search: In search-based optimization category, the explicit approaches for autoscaling decision making rely on the explicit QoS models to evaluate and guide the search process. Depending on the different formulations of the decision making problem for autoscaling in the cloud, explicit search can reason about the effects of decisions and the possible trade-offs in details. According to our survey, we found three most commonly used formulations, these are single objective optimization, weighted-sum optimization, and pareto-based optimization. We discovered that it is not uncommon to optimize only a single objective (e.g., cost or profit) for SSCAS, providing that the requirements of the other objectives are satisfied (i.e., they are often regarded as constraints). For example, Kingfisher [@2011-cost-aware-v-h-scaling-2011] and Sedaghat et al. [@ILP-cost-only-scaling-2013] use Integer Linear Programming (ILP) to optimize the cost for scaling the CPU and memory for VMs of an application while regarding the demand for satisfying QoS as constraint. To apply explicit search-based optimization for SSCAS, the most widely solution for handling the multi-objectivity is to aggregate all related objectives into a weighted (usually weighted-sum) formulation, which converts the decision making process into a single objective optimization problem. The search-based algorithm include: exhaustive search [@3-stages-game-theory] [@tse1] [@kriging-controller] [@typical-navie-scaling-VM-number], auxiliary network flow model [@li-opt-clouds-4], force-directed search [@multitier-resalloc-Cloud11], binary search [@cache-static-ANN-bi-obj-2012]. For example, the FoSII project [@Compsac_2010_I_Brandic] [@SEASS_2010_Michael_Maurer] regards the autoscaling decision making as case based reasoning process, where the decision is made by looking for the similar historical cases using exhaustive search. The solution of the most similar case is reused o solve the current one. We noted that some studies have relied on more advanced and nonlinear search algorithms, ranging from relatively simple ones: dynamic programming [@2014-linear-centralized-decision-making-2014] and local-search strategy [@profit-cent-local-search], to more complex forms: grid search [@11icde_smartsla_full], decision tree search [@2014-profiling-decision-tree-scaling-2014] [@2014-decision-tree-software-CP-2014] and quadratic programming [@HPL-2008-123R1-mimo]. For example, CLOUDFARM [@2014-linear-centralized-decision-making-2014] addresses the decision making based on a weighted-sum utility function of all cloud-based application and services. The decision making process is formulated as a knapsack problem, which can be resolved by dynamic programming. We have also found that metaheuristic algorithms are popular for autoscaling decision making in SSCAS, because they can often efficiently address NP-hard problems with approximated results under no assumptions of the problem. The most common algorithms include: Tabu Search [@sla-provision], Genetic Algorithm (GA) [@software-RP-two-loops] [@2014-navie-ANN-GA-2014] [@BRGA-resource] , Particle Swarm optimization (PSO) [@software-RP-two-loops]. As an example, Zhu et al. [@sla-provision] formulate the autoscaling decision making as optimize a weighted-sum formulation of response time and cost. To optimize the objectives, the authors apply a hybrid Tabu Search, which relied on iterative gradient descent. Finally, pareto relation can explicitly handle multi-objectivity for autoscaling in the cloud without the need to specify weights on the objectives [@2014-eplison-GA-weigh-h-scaling-2014], [@wosp10sla], [@E3-R-extended], [@GA-full-simulation], [@Chen:2015:tsc-pending]. For example, in E$^3$-R [@E3-R-extended], the decision making problem is formulated using Pareto relation, where it is resolved by using Non-dominated Sorting Genetic Algorithm-II (NSGA-II). Further, the approach applies objective reduction technique with an aim to remove the objectives, which are not significantly conflicted with the others, from the decision making process. Chen and Bahsoon [@Chen:2015:tsc-pending] exploit Multi-Objective Ant Colony Optimization (MOACO) for trade-off decision making when autoscaling cloud-based services. The authors consider the trade-off between naturally conflicting objectives and between competing services (i.e., QoS interference). Further, a compromise-dominance mechanism is proposed to find well-compromised trade-off decision. Experimental Evaluation on SSCAS -------------------------------- Another important step in SSCAS research is to quantitatively evaluate the proposed solution, which is often achieved via experimental analysis. To this end, setting the experiments are of high importance to researchers in this field. Table \[tb:exp\] summarizes the common infrastructure, benchmarks and workload trace from the considered studies. As we can see that there are studies chose to use simulator for controlled experiments and ease of complexity. However, simulators may not fully capture the realistic environment. As a result, custom private cloud and public cloud are also exploited for both controlled and open experiments. Notably, custom private cloud can be much more flexible on choosing the underlying software, e.g., one may utilize the hypervisor or container directly or choose to use higher level software such as OpenStack [@open]. It is worth noting that custom private and public cloud can share similar underlying software and tools, but they may require expertise on different levels of abstraction. For example, one may choose to deploy SSCAS on a custom private cloud that makes use of Xen [@xen]the same hypervisor that underpins Amazon EC2 [@iaas] which contains additional high-level interfaces and restrictions. Benchmarks are not required for simulator, but it is crucial for both private and public cloud infrastructure. A wide range of benchmarks have been exploited to evaluate SSCAS, from simple web hosting to complex multi-tier software. Finally, the workload traces can be either synthetic in which a fixed pattern is generated by the workload generator (e.g., JMeter [@jmeter]); or real where recorded traces form different real domains are used to stress the benchmark and SSCAS. Implementation of Scaling ------------------------- The actual implementation of scaling depends on the underlying scenarios, e.g., the type of hypervisor/container, the cloud-based applications and the actual cloud control primitives among the others. Existing virtulization techniques have provided readily available commands and tools to support autoscaling at runtime. For example, if the underlying hypervisor was Xen [@xen], then resource such as CPU and memory of a VM, as well as create/destroy VMs can be scaled dynamically using the $xm$ command. Regarding the actual scaling methods, vertical scaling will have trivial effects on the states of the cloud-based applications, thus they can be directly applied using the command support by hypervisor/container. For horizontal scaling, making new replicas or removing old one needs consistency guarantee on stateful applications/services, which can be ensured by the underlying hypervisor through various readily available protocols. For example in Xen, horizontal scaling can be achieved via primary-backup replication or asynchronous checkpointing, etc. [\|p[2cm]{}\|p[2cm]{}\|X\|]{} &*Software and Tools\ &Simulator&CloudSim [@CloudSim], CDOSim [@full-simulation-model], CloudAnalyst [@CloudAnalysis], DCSim [@DCSim]\ &Private Cloud&Xen [@xen], VMWare ESXi [@vmware], KVM [@kvm], Docker [@docker], OpenStack [@open] , Eucalyptus [@euc]\ &Public Cloud&Amazon EC2 [@iaas], RackSpace [@rack], Azure [@azure], Google Compute Engine [@google]\ &RUBiS [@rubis], RUBBoS [@rubbos], TPC-W [@tpcw], WikiBench [@wikibench]\ &Synthetic trace, FIFA98 [@f98], Wikipedia [@wikipedia], ClarkNet [@clark]\ Reflections and Open Challenges for SSCAS Research {#sec:diss} ================================================== In this section, we reflect on the finding of our survey and taxonomy; state the open challenges as well as discuss industrial situation and pricing strategy for SSCAS. Discussion and Comparison on Existing SSCAS Research ---------------------------------------------------- We now discuss the most noticeable observations by reviewing the existing SSCAS research. We carefully position our discussions in light of the different logical aspects of SSCAS. ### The Levels of Self-Awareness and Self-Adaptivity in Cloud Autoscaling Systems {#the-levels-of-self-awareness-and-self-adaptivity-in-cloud-autoscaling-systems} Stimulus-awareness* has been considered in all the 109 studies as it is the most fundamental levels in self-awareness principles, because it is the basic requirement for a software system to be able to adapt. Time- (52%) and goal-awareness (52%) receives same attention due to the fact that the objective models often contain historical information and they can be used to reason about goals. In contrast, handling interaction- (13%) and meta-self-awareness (7%) are less popular as the former requires to handle QoS interference while the latter often come with extra complexity. While self-optimizing and self-configuring have been the major themes for SSCAS, we found very little studies that considered self-healing (5%) and only one work targets for self-protecting. This is obvious as the fundamental idea of autoscaling is not for security related purposes but for the performance related quality, which is often much more appealing for cloud consumers. Also, 67% of the studies has ignored the importance of specifying the required knowledge at the architecture level, entailing the risk of limited awareness [@epics]. ### Architectural Pattern {#architectural-pattern} The generic feedback (82%), particularly the single and close loop, has been the predominant architectural pattern for SSCAS. MAPE (15%) is ranked the second and OAD (3%) being the much less popular one, as OAD often assume the involvement of human decisions maker which is difficult in the case of SSCAS. The reason could be due to the fact that the feedback loops are flexible and simple to be realized, providing the basic components to achieve self-adaptivity. However, such design can limit the consideration of required knowledge for the autoscaling system to perform adaptations, or the consideration is rather simple and coarse-grained. In contrast, such an issues has been relaxed by OAD and MAPE as they are more stricted by predefining components to capture different aspects of a SSCAS. We see that MAPE is clearly more popular than OAD because the former can be good for separation of concepts (e.g., Analyze, Plan and Knowledge) and for expressing the sequential interactions between those concepts while the latter fails to capture runtime aspects of the SSCAS, as it is mainly designed for decoupling loops of different human activities. Nevertheless, these architectural patterns lack of fine-grained representation of the required knowledge. Thus, it is not immediately intuitive what level(s) of the knowledge is required by each logical aspect of a SSCAS. ### QoS Modeling {#qos-modeling} Both analytical model and machine learning model, included in 43% and 38% of the studies respectively, are widely exploited in QoS modeling for SSCAS while the simulation (10%) and hybrid (9%) approach are clearly less popular. This could be because analytical model is good for runtime efficiency, simplicity, interoperability, and they could be very effective if all of their assumptions are satisfied. However, analytical approaches generally require in-depth knowledge about the likely behaviors of the system being modeled, i.e., some knowledge about the system’s internal structure or environmental conditions. Such an issues is resolved by using machine learning model which are often assumptions free, and more importantly, they are able to continually evolve themselves at runtime in order to cope with dynamics and uncertainty. Nevertheless, depending on the learning algorithm, the resulting overhead can be high (e.g., the nonlinear ones) and the accuracy is sensitive to the situation (e.g., fluctuation of the data trend). In contrast, simulation exhibits static nature and it is restricted by a wide set of assumptions, including e.g., the distribution of workload and the effects of QoS interference, etc. needs complex human intervention and assumptions. However, it is believe that simulations model could be the most accurate way to model QoS when the all assumptions are satisfied [@full-simulation-model]. Hybrid model, as in 6 of the studies surveyed, could potentially combine the strengths from different models. We noted that only 13% of the studies intend to address QoS interferences when modeling the QoS in SSCAS. This might be because considering QoS interference will significantly increase the dimensionality in the model, which in turn, rendering the problem much more complex. Such a complexity makes human analysis very difficult, if not impossible. As a result, for those studies that do consider QoS interference, machine learning algorithms are often exploited. There are plenty of studies (61%) that consider dynamic (or semi-dynamic) structure of a QoS model (i.e., those that denoted as both dynamic and semi in Table \[tb:qos\]), however, the dynamic related to the input features have been rarely researched simultaneously, i.e, only 7% of the studies (for those that denoted as dynamic in Table \[tb:qos\]). Indeed, changing the inputs of a model could be useful only when the dimensionality of a model is high; that is to say, changing the input features might not cause significant difference if the considered total number of inputs is around, e.g., less than five. The majority of the dynamic (or semi-dynamic) modeling are machine learning based, leaving the static ones are largely analytical or simulation based. This is obvious, as the nature of those modeling approaches determine the extents to which they can be changed when they are built. A considerable amount of studies (65%) have claimed that their QoS models can work on any given inputs and/or output, as shown in Table \[tb:qos-inout\]. This happens mostly for machine learning and simulation approaches. However, during experimental analysis, the highest number of QoS that being modeled and the number of inputs were both four [@Chen:2014:ucc] [@Chen:2015:tse-pending]. The most commonly considered output is response time and the inputs are hardware resources, particularly CPU, memory and number of VM. This complies with the current trend in the cloud computing market. ### Granularity of Control {#granularity-of-control} In SSCAS research, the service/application level of granularity is the most popular one, which yields 45% of the studies. This is because focusing on the finest granularity of control can achieve the maximum level of scalability, which particularly fits the cloud. However, fine granularity of control is achieved in the expenses of the globally optimal quality of the cloud, since no interactions between service/application are considered. In contrast, focusing on cloud-level is another extreme, which trades scalability for global optimality. Considering multiple levels could be a solution to reach a better trade-off as discussed in a small amount of studies (8%), but how and when to select the levels to consider imposes additional challenges. Notably, most of the studies (72%) has relied on the control of each application in cloud regardless to the actual granularity of control. The reason being could be due to the fact that cloud-based application (or a collection of services) is the most crucial unit for consumers to experience the benefit of cloud, which is a common interest for both cloud consumers and providers. ### Decision Making Generally, rule based control is a highly intuitive approach for autoscaling decision making, and it also has negligible overhead. However, the static nature of the rules requires to assume all the possible conditions and the effects of those decisions that are mapped to the conditions, which is highly depending on the assumptions. To resolve such an issue, control theoretic approach appear to be an effective solutions as it is also efficient while require very little assumptions. However, the major drawback of control theoretic approaches is that they require to make many actuations on the physical system, in order to collect the ’errors’ for stabilizing itself. This means that amateur decisions are very likely to be made. In addition both approaches lack of handling multi-objectivity and the trade-off; they often fail to cope with the problem where there is a large number of autoscaling decisions, which is common for cloud. In contrast, search-based optimization, especially the explicit search, makes loose assumption about the number of autoscaling decisions and is able to find optimality (or near-optimality) under highly dynamic and uncertain environment. Therefore as we have shown, search-based optimization, either implicit or explicit, is the most popular approach for making decisions in SSCAS. We noted that there is a considerable amount studies (63%) that do not attempt to explicitly consider trade-off during decision making of SSCAS, as they assumed single objective or rely completely on human preferences. The reason might be attributed to the fact that such formalization is simple and straightforward, which can work well when there is a strong preference on an objective. The rest studies handle multi-objectivity via either weighted sum or pareto relation. QoS interference is again absent in many studies (84%) due to the fact that considering it will unavoidably increase the dimensionality (i.e., objectives) during the decisions making, leading to a more complex problem. This would make the problem unsolvable by many existing approaches. We found that most of the studies (78%) have claimed that their decision making approach could handle any given objectives, thus they have considered arbitrary QoS attributes and cost as the objectives in SSCAS since these are the most critical indicator for cloud-based services and applications. The highest number of objectives that were considered during the experiments are five [@Chen:2015:tsc-pending] though. CPU, memory and number of VM are the most popular control primitives in decision making because they are the most straightforward dimension to be scaled. However, only as little as 9% of the studies consider the interplay between software and hardware control primitives. To apply solution that can not handle a large search space of the decision making, one often reduce the search space by introducing fixed bundles, which is an assumption made by a considerable amount of studies, i.e., 34%. However, such reduction has the risk that some good solution can be ruled out during the process. Finally, while both vertical and horizontal scaling have been considered in the majority of the studies (66%), focusing solely on horizontal scaling is more popular than the vertical one as the former is more widely supported by major cloud vendors. Open Problems and Challenges for SSCAS research ----------------------------------------------- Drawing on the survey and taxonomy, in the following, we specify the open problems and challenges for future SSCAS research and make suggestions for potential research directions where appropriate. - Explicit Knowledge Representations are Required in SSCAS Architecture: As we can see from Section \[sec:sscas\], only 33% of the studies intend to discuss the required knowledge at the architecture level. This means that, in the remaining 67% work, it is more difficult to capture more complex and advanced levels of knowledge, as evident by the fact that most work does not go beyond the basic stimulus-awareness. Indeed, studies [@epics] [@2014_epics_handbook] [@7185305] [@epics_survey] [@Chen2016:book] have found that, for self-aware and self-adaptive software systems in general, the absence of explicit consideration for the fine-grained representation of the knowledge in the architecture can results in, e.g., improper inclusion of unnecessary knowledge and/or missing important knowledge that can improve adaptation quality when developing autoscaling systems. 67% studies which do not discuss knowledge at the architecture level implies that such an issue is often overlooked and it is remain unresolved in the SSCAS context, urging the need of further investigations. The challenge here lies in the fact of how can one systematically distinguish different levels of knowledge and how they can be architected into SSCAS in a principal way. We argue that the required levels of knowledge and their representations can be declared in light with the formal principle of self-awareness. In particular, a potential way is to follow the handbook [@2014_epics_handbook] for mapping different levels of knowledge into a concrete SSCAS architecture. - Multiple Loops can Create More Benefit for SSCAS Architecture: From Section \[sec:arch\] we noted that the majority of the studies has considered single loop, which could cause the problem of high coupling in the design of SSCAS. The multiple loops, on the other hand, helps to achieve better separation between different aspects of SSCAS, leading to fine-grained and localized adaptation. The challenge here is how many and at what levels of abstraction one should place the loops within the SSCAS Architecture. We suggest that designing multiple looped SSCAS architecture with respect to what levels of knowledge the system required could be a neat solution [@Chen:2015:computer]. - QoS Interference Should be Explicitly Handled in QoS Modeling and Decision Making Process: Our survey results (see Section \[sec:qos\] and \[sec:dm\]) indicate that only less than 16% of the studies took QoS interference into account. Missing QoS interference in the model and decision making could lead to incorrect or misleaded autoscaling decisions, as the cloud-based services would be unavoidably affected by the dynamic behaviors of its neighbors. However, incorporating QoS interference rises the challenge of dimensionality which causes the model and decision making much more complex. Therefore, this challenge calls for novel approach to reduce the dimensionality, or mitigate its negative effects, during the model and decision making in SSCAS [@Chen:2015:tse-pending] [@Chen:2015:tsc-pending]. - The Interplay Between Software Configuration and Hardware Resources is Important: Most existing studies of SSCAS focus on hardware resources as IaaS level only. However, as shown in [@2013-JRAO-most-closest-work-2013] [@software-RP-two-loops] [@2014-decision-tree-software-CP-2014] [@Chen:2015:tse-pending], various software configurations at PaaS level could interplay with each other and the hardware resources, which in turn, affect the QoS of cloud-based services. The challenge is how to create a holistic approach that combines both PaaS and IaaS level in SSCAS. - Dynamic Feature Selection is Required for QoS Modeling in SSCAS: Section \[sec:qos\] indicates that the majority of the existing studies have ignored primitive selection in the QoS modeling or have been relying on manual approach, because the assumed dimensions of inputs is rather limited. However, when both QoS interference and software configurations are considered, selecting the most significant features in the model becomes a crucial task since there could be an explosion of the primitives space [@Chen:2015:tse-pending] [@2013-single-learner-filter-wrapper-LR-2013]. Challenge here lies in how to evaluate the effectiveness of feature combination on the model accuracy while generating reasonable overhead. Given the arbitrary types of feature, which calls for generic and efficient feature selection design for SSCAS. - More Flexible Granularity of Control in SSCAS is Needed: Single, static and fixed granularity of control is predominately exploited in existing SSCAS research. To better handle dynamic and uncertainty, it could be more beneficial to introduce multiple granularity and/or dynamically adjust the granularity of control at runtime [@11icde_smartsla_full] [@Chen:2014:seams], as the granularity of control implies a trade-off between the global optimality of SSCAS and the imposed overhead. The challenge is how to explicitly capture the objective-dependency when designing granularity of control. - The Assumptions on the Bundles of Resources Needs to be Relaxed: From Section \[sec:dm\] we noted that while most of the studies have not constrained the possible autoscaling decisions with respect to the fixed bundles, there are still certain amount of studies that heavily rely on the fixed types of bundles, e.g., a search space of 57 VM instance types on Amazon EC2. However, renting bundles cannot and does not reflect the interests of consumers and the actual demand of their cloud-based services. We argue that future cloud autoscaling would inevitably needs to take arbitrary combination of software configurations and resources into account, as what has already been supported in Google Compute Engine [@google]. As a result, autoscaling decision making imposes a challenging problem that faces with an explosion of decision space (e.g, millions of alternatives), calling for novel and efficient approach to achieve optimal or near-optimal quality. - The Trade-off Between Conflicting Objectives Should be Explicitly Handled: As shown in Section \[sec:dm\], the approaches have mostly ignored trade-off. There is also certain amount of explicit search-based optimization studies has assumed only single objective. However, this can restrict the applicability of SSCAS as the decision making would fail in identifying good trade-off points or strongly bias to the single objective. Alternatively, there is also considerably large amount of studies exploit weighted sum objective aggregation, which embed the trade-off in a single representation. However, it is well-known that the relative weights are difficult to be tailored and a single aggregation could restrict the search, causing limitation when searching for good decisions spread over the search space. Further, achieving balanced trade-off have only being explored in very limited studies, e.g., [@Chen:2015:tsc-pending]. The challenge here is how to search for decisions that contain good convergence and diversity, and eventually selecting the one that has the most balanced trade-off for scaling. We advocate that stochastic optimization approach, particularly nature inspired algorithms, can be promising in addressing such a challenge. - More Real World Case Studies of SSCAS are Needed: We found that real world cases and scenarios of SSCAS, especially those with large scale and practical application, are absent in many studies. Indeed, those studies impose many challenges beyond the perspective of research, but they can be the only way to fully verify the potentials, effectiveness and impacts of SSCAS. Current Industrial Situation of SSCAS ------------------------------------- Industrial cloud providers (e.g., Amazon [@iaas] and RightScale [@rightscale]) have been relying on model-free, simple rule and policy based autoscaling approaches for decades. These approaches leave the difficult problem of how to specify rules to cloud consumers, which may work well in the beginning when the demand and complexity of cloud-based applications are simple and straightforward. However, recently the level of complexity (e.g., in terms of the number of cloud control primitives) of cloud is changing to a state that makes human analysis very difficult, especially under conflicting objectives and a large number of alternative autoscaling decisions [@2014-eplison-GA-weigh-h-scaling-2014], [@wosp10sla], [@E3-R-extended], [@Chen:2015:tsc-pending]. Specifically, as discussed in Section 5.1.5, those approaches suffer two significant pitfalls. (i) They requires understanding of the application and domain knowledge to determine the mapping between conditions and actions, which can significantly affect the quality of scaling [@al2013impact], [@software-RP-two-loops]. (ii) They cannot adapt to dynamically changing workload or state of the applications [@2014-navie-ANN-GA-2014], [@Chen:2015:tsc-pending]. As a result, engineering advanced SSCAS is an inevitable trend in this area; the reason why current big cloud providers have not yet widely implemented them could be due to the fact that SSCAS itself is not mature to the state which it can be reliably adopted. However, as our survey reports, researchers and practitioners have been working on overcoming these challenges for almost a decade. There has been some attempts to apply SSCAS commercially, for example, Microsoft Azure [@azure] has recently benefited from Aneka [@vecchiola2009aneka], a research effort supporting high level framework, which contains a more advanced and complicated SSCAS that relies on search-based optimization as part of its subsystems. Aneka’s work is an evidence of how pending industrial challenges had informed research; the results are now incorporated in Azure. Nevertheless, we envision more progress on enhanced, scalable and cost-effective effective solutions for both the cloud providers and consumers. Discussion on the Pricing for Cloud Autoscaling ----------------------------------------------- Indeed, more advanced autoscaling approaches in SSCAS (e.g., machine learning and search-based optimization) may impose additional computational resources. Furthermore, advances in autoscaling cannot be done in isolation of pricing (dynamic metering and pricing in particular), as both the cost and revenue are acknowledged among the drivers for the industrial need of more advanced solutions. However, upfront investment in additional recourse can be arguably paid off, in situations where scale and dynamic demand is effectively enabled. This can be observed through greater pay off through better utilization and SLA guarantee, which in turn, improve the overall reputation and thus attracting more consumers. As analyzed in numerous existing work [@lorido2014review][@qu2016auto][@GA-full-simulation][@Chen:2015:tse-pending][@Chen:2015:tsc-pending], the additional resources spent are actually marginal compared with the savings obtained through more accurate, effective autoscaling. In case the cloud provider wishes to charge the consumers for the services provided by SSCAS, there could be two ways to achieve this: (i) charging the computation utilized by the SSCAS through existing pricing schema, e.g., the reserved or spot instance from Amazon EC2. Here, the SSCAS is an optional service which would be priced as normal instance for the consumers’ application/services. (ii) The charge of SSCAS is combined with the normal price per time unit in existing pricing schema, e.g., instead of charging \$1/hour of an instance, it can be priced as \$1.3/hour where the extra \$0.3/hour is for the SSCAS. Here, the SSCAS is a default and mandatory service to the cloud consumers. Conclusion {#sec:con} ========== In this article, we survey the state-of-the-art research on SSCAS and provide a taxonomy based on our findings. Specifically, we review the literature with respect to the research questions presented in Section \[sec:rq\]. According to our survey, the key findings are: - Stimulus-, time- and goal-awareness are the most widely considered levels of knowledge in SSCAS. Self-configuring and Self-optimizing are the most popular self-adaptivity notions in SSCAS. - Feedback loop is the most commonly exploited architectural pattern for engineering SSCAS. - Analytical model and machine learning based model are prominent for QoS modeling in SSCAS. - Controlling at the level of service/application is the mostly applied granularity. - Search-based optimization is the most common approach for making autoscaling decisions. Apart from those observations, we also gain many insights on the open problems and challenges for future SSCAS research. The most noticeable ones are: - Explicit knowledge representations are required in SSCAS architecture. - Multiple loops can create non-trivial Benefits for SSCAS architecture. - QoS Interference should be explicitly handled in QoS modeling and decision making process. - The interplay between software configurations and hardware resources is non-trivial. - Dynamic Feature selection is required for QoS modeling in SSCAS. - More flexible granularity of control in SSCAS is needed: - The assumptions on the bundles of resources needs to be relaxed. - The trade-off between conflicting objectives should be explicitly handled. - More real world cases and scenarios of SSCAS are needed. We hope that our survey and taxonomy will motivate further research for more intelligent cloud autoscaling system and its interactions with the other problems in the cloud computing paradigm. [^1]: This work was supported by the Ministry of Science and Technology of China (Grant No. 2017YFC0804003), Science and Technology Innovation Committee Foundation of Shenzhen (Grant No. ZDSYS201703031748284), and EPSRC (Grant No. EP/J017515/01 and EP/K001523). [^2]: Service could refer to an entire application, or any conceptual part within an application. [^3]: Each autoscaling decision is a specific combination of configurations and/or resource provisions that achieves certain outcomes on the targeted objectives. [^4]: In queuing theory, $M$ denote Poisson distribution; $G$ denotes arbitrary distribution. A term $M/G/m$ refers to Poisson distribution of arrival rate, arbitrary distribution of service rate and there exists $m$ servers.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Binaural hearing aids communicate with each other through a wireless link for synchronization. A propagation model is needed to estimate the ear-to-ear link loss for such binaural hearing aids. The link loss is a critical parameter in a link budget to decide the sensitivity of the transceiver. In this paper, we have presented a model for the deterministic component of the ear-to-ear link loss. The model takes into account the dominant paths having most of the power of the creeping wave from the transceiver in one ear to the transceiver in other ear and the effect of the protruding part of the outer ear called pinna. Simulations are done to validate the model using in-the-ear (ITE) placement of antennas at 2.45 GHz on two heterogeneous phantoms of different age-group and body size. The model agrees with the simulations. The ear-to-ear link loss between the antennas for the binaural hearing aids in the homogeneous SAM phantom is compared with a heterogeneous phantom. It is found that the absence of the pinna and the lossless shell in the SAM phantom underestimate the link loss. This is verified by the measurements on a phantom where we have included the pinnas fabricated by 3D-printing.' author: - 'Rohit Chandra and Anders J Johansson [^1]' title: \| ©2013 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or distribution to servers or lists, or reuse of any copyrighted component of this work in other works. The complete paper can be found at: <http://ieeexplore.ieee.org/xpl/articleDetails.jsp?tp=&arnumber=6607217> and Digital Object Identifier :10.1109/TAP.2013.2283270\ A Link Loss Model for the On-body Propagation Channel for Binaural Hearing Aids --- Body Area Network (BAN), creeping waves, propagation Introduction ============ medical devices have revolutionized the field of medical sciences and have improved the quality of life of both diseased and healthy persons. They can be used for vital signal monitoring of patients as in Wireless Body Area Networks (WBAN) and can alert doctors about the critical condition of their patients. These devices can be used for rehabilitation of patients with disability, and also by healthy persons without any disease or disability e.g. by sportsmen and athletes. The main advantage of the wearable devices is that they are non-invasive and have limited risk of infection [@pf],[@dittmar]. The wearable medical devices placed at different positions on the body may communicate with each other for data exchange. In such scenarios, on-body propagation models become critical for proper estimation of the link budget. Various statistical and analytical on-body propagation models have been discussed in [@cotton]-[@conway]. In [@cotton], a time domain analysis and modeling of the on-body propagation characteristics was presented. The measurements were done in different scenarios like anechoic chamber, open office area, hallway, and outdoor environment and autocorrelation and cross-correlation statistics were presented at $2.45$ GHz. In [@fort], a simplified physical body area propagation model was derived using Maxwell’s equations. A channel model for wireless communication around the human body was developed in [@ryk]. The creeping wave is the phenomenon for the communication between the transceivers located on the opposite side of the body. An analytical propagation model of BAN channels based on the creeping wave theory was presented in [@alves] by re-using the theory of the wave propagation over the inhomogeneous earth’s surface [@wait]. Hearing aids are one such wearable medical device which is used for rehabilitation of hearing impaired persons. Binaural hearing aids are a system where there is a hearing aid in both the ears of the user. They communicate with each other for synchronization data [@chandra1] through the ear-to-ear on-body propagation channel. The ear-to-ear propagation channel has been discussed in [@zaso]-[@kvist3]. In [@zaso] it was shown that diffraction is the main propagation mechanism around the head for the link between the ears. This diffraction mechanism is the same as the creeping wave phenomenon. It was shown in our previous work [@chandra1] that in-the-ear (ITE) placement has less attenuation than in-the-canal (ITC) placement of the binaural hearing aids as the communication between the ITE antennas was through the creeping waves. In [@kvist1], the effect of the head size on the ear-to-ear radio propagation channel was examined. It was shown by the simulations that the variations in the head size may result in up to $10$ dB variation in the link loss. In [@bour], authors have shown by measurements with different antennas on the SAM head that small variations in the position of the antenna do not affect the ear-to-ear link loss whereas the operating frequencies have a larger impact on the link loss. Their result showed that the link loss increases with increase in the frequency. The importance of the correct orientation of the antenna to decrease the link loss was also shown. Homogeneous models were used for the simulations or for the measurements. The protruding part of the outer ear called pinna and the lossy skin are absent in the SAM head/homogeneous head. However, in [@chandra2] we have shown the significant effect of the pinna and the lossy skin on the ear-to-ear link loss. The propagation model for the communication between the ITE binaural hearing aids is required for proper estimation of the ear-to-ear link loss. The model for the propagation around the head presented in [@alves], can be used for this purpose. Since the model was developed for a simplified model of the head, treating it as a cylinder, there was a good agreement between the simulations and the measurements. However, this model does not take into account the losses due to the pinnas. In this paper, we have presented an analytical model which includes the losses of the pinnas. The cross-section of the head is modeled as a more realistic elliptical shape rather than a circular shape. It is shown that the two paths of the creeping waves are sufficient to describe the link loss. Simulations are done in SEMCAD-X [@semcad] which uses the FDTD method. A low-profile and compact ITE antenna [@chandra2] on the heterogeneous phantoms is used to verify the propagation model. Measurements are done to verify the effect of the pinna on the ear-to-ear link loss. Both the simulations and the measurements are done in $2.45$ GHz ISM band. The paper is organized as follows. In Section \[sec:num\_phantom\], the numerical phantoms used for the simulations are described. Section \[sec:model\] present the analytical model. Other factors apart from the pinna which may affect the ear-to-ear link loss are discussed in Section \[sec:factors\]. In Section \[sec:meas\], the measurement process and results are presented. Conclusions are presented in Section \[sec:conclusion\]. Numerical Phantoms {#sec:num_phantom} ================== Four numerical heterogeneous phantoms of different age group and gender and one homogeneous phantom are used for the simulations. The modified SAM head phantom with the ear canals [@chandra1], shown in Fig. \[SAM\_Phantom\_dim\], is used as a homogeneous phantom. The heterogeneous numerical phantoms used in this paper are provided by ITIS foundation [@itis] developed for the Virtual Family and Classroom project [@vfp]. They are whole body anatomical models consisting of more than $80$ tissues with different electrical properties. The models are that of a $34$ year old male called Duke, a $26$ year old female called Ella, a $15$ year old boy, Louis and a $11$ year old girl, Billie. They are shown in Fig. \[family\_heads\_dim\]. Since we are interested in the ear-to-ear link loss, only the truncated head with the neck of the phantoms are used for the simulations. This also has an advantage in terms of faster simulation. However, we have examined and presented the effect of the shoulders on the link loss. Table \[tissue\_parameter\] presents the electrical properties of the tissues present in the head and the neck of the phantoms at $2.45$ GHz. Tissue is the name of the tissue of the phantom of Virtual Family Project and Gabriel list gives the mapping of the phantom tissue to the Gabriel list obtained from [@ifac] for the electrical properties. The approximate head dimensions of the phantoms are shown in Table \[head\_dimension\] where the dimensions $l, h, m$ are shown in Fig. \[family\_heads\_total\]. ------------------------------------------------------------------------------------- -- -- -- [Tissue]{} & [Gabriel List]{} & [$\varepsilon_r$]{} & [$\sigma_e$ (S/m)]{}\ Air Internal & Air & 1 & 0\ Artery & Blood & 58.26 & 2.54\ Blood Vessel & Blood & 58.26 & 2.54\ Bone & Bone Cortical & 11.38 & 0.39\ Brain Grey Matter & Brain Grey Matter & 48.91 & 1.80\ Brain White Matter & Brain White Matter & 36.16 & 1.21\ Cartilage & Cartilage & 38.77 & 1.75\ Cerebellum & Cerebellum & 44.80 & 2.10\ Cerebrospinal fluid & Celebro Spinal Fluid & 66.24 & 3.45\ Commissura anterior & Brain White Matter & 36.16 & 1.21\ Commissura posterior & Brain White Matter & 36.16 & 1.21\ Connective tissue & Fat(mean) & 10.82 & 0.26\ Cornea & Cornea & 51.61 & 2.29\ Ear Cartilage & Cartilage & 38.77 & 1.75\ Ear Skin & Skin(Dry) & 38.00 & 1.46\ Eye Lens & Lens Cortex & 44.62 & 1.50\ Eye Sclera & Eye Tissue (Sclera) & 52.62 & 2.03\ Eye vitreous humor & Vitreous Humor & 68.20 & 2.47\ Fat & Fat & 5.28 & 0.10\ Hippocampus & Brain Grey Matter & 48.91 & 1.80\ Hypophysis & Gland & 57.20 & 1.96\ Hypothalamus & Gland & 57.20 & 1.96\ Intervertebral disc & Cartilage & 38.77 & 1.75\ Larynx & Cartilage & 38.77 & 1.75\ Mandible & Bone Cortical & 11.38 & 0.39\ Marrow Red & Bone Marrow(infiltrated) & 10.30 & 0.45\ Medulla Oblongata & Brain(average) & 42.53 & 1.51\ Midbrain & Brain(average) & 42.53 & 1.51\ Mucosa & Mucous Membrane & 42.85 & 1.59\ Muscle & Muscle(Transverse Fiber) & 38.00 & 1.46\ Nerve & Nerve & 30.14 & 1.08\ Pharynx & Air & 1 & 0\ SAT & Fat & 5.28 & 0.10\ Skin & Skin(Dry) & 38.00 & 1.46\ Skull & Bone Cortical & 11.38 & 0.39\ Spinal Cord & Spinal Chord & 30.14 & 1.08\ Teeth & Tooth & 11.38 & 0.39\ Tendon Ligament & Tendon & 43.12 & 1.68\ Thalamus & Brain Grey Matter & 48.91 & 1.80\ Thymus & Thymus & 57.20 & 1.96\ Thyroid gland & Gland & 57.20 & 1.96\ Tongue & Tongue & 52.62 & 1.80\ Vein & Blood & 58.26 & 2.54\ Vertebrae & Bone Cortical & 11.38 & 0.39\ ******** ------------------------------------------------------------------------------------- -- -- -- : Tissue Parameters[]{data-label="tissue_parameter"} \ Phantom Age(year) Gender l(mm) h(mm) s(mm) --------- ----------- -------- ------- ------- ------- SAM – Male 164 312 234 Duke 34 Male 155 251 238 Ella 26 Female 143 239 209 Louis 15 Male 152 241 214 Billie 11 Female 137 233 188 : Dimensions of the phantom’s head[]{data-label="head_dimension"} Link Loss Model {#sec:model} =============== It is well established that the creeping wave is the dominant phenomenon for the communication around the curved part of the body. In [@alves] and [@fort1], it was shown that the clockwise and the anti-clockwise creeping wave interfere with each other. The phantom in these investigations was symmetrical. However, the realistic phantoms are anatomically asymmetric and hence for the ear-to-ear channel there will be several paths such as over top of the head, over back of the head, over the front part of the head, etc. Investigations are done through simulations to determine the dominant paths i.e. the ones having most of the power of the creeping waves. Dominant Paths of the Creeping Waves ------------------------------------ The creeping waves from the transmitter in one ear can reach the receiver in the other ear by numerous paths. But it is not necessary that all these paths will have significant power as some paths will have high attenuation. Six different simulation cases are considered to determine the paths having significant power using the ITE antenna [@chandra1] on the Duke phantom. In these cases, one or more creeping paths are truncated by an uni-anisotropic perfectly matched layer (UPML) such that the creeping waves are blocked on these paths. The descriptions of the cases are presented in Table \[sim\_cases\]. The table lists the creeping paths and the blocked paths for the six cases. For example, for case (a), simulation domain includes the front and the top part of the head, allowing the waves to creep over these paths and truncating the back side of the head with an UMPL which blocks the creeping waves on the back side of the head. It should be noted here that the front, the back and the top are relative position w.r.t. the ear (antenna placement). These six cases are illustrated in Fig. \[PML\_head\] where the shadowed region represents the simulation domain. The link loss of all these cases is compared with the link loss when the complete head is used in the simulation domain. The simulation results are shown in Fig. \[link\_loss\_pml\]. Throughout the paper we present the plot of $S_{21}\|_{dB}$ which is related to the link loss ($LL$) in dB scale by: $LL\|_{dB} = -S_{21}\|_{dB}$. It should be noted that the frequency band of interest is $2.45$ GHz ISM band and hence all the comparison are done in this band where the antennas are matched to $50~\Omega$ source impedance. At other frequencies like around $2$ GHz and $2.9$ GHz, there is an extreme dip in the simulated $S_{21}$ for some scenarios which is either because of the high mismatch loss at these frequencies or strong destructive interference of waves coming from different paths. However, since these are not the frequencies of the interest, no further investigations are done for these extreme dips in the simulated $S_{21}$. Scenario Creeping Path Blocked Path ---------- ---------------- ---------------- Case (a) Front and Top Back Case (b) Top Front and Back Case (c) Top and Back Front Case (d) Front and Back Top Case (e) Front Top and Back Case (f) Back Front and Top : Simulation cases for determination of dominant path of creeping waves between the ears[]{data-label="sim_cases"} \ ![$S_{21}$ for different scenarios for determining the paths of the creeping waves carrying significant power. For the description of the cases in the legend refer Table \[sim\_cases\].[]{data-label="link_loss_pml"}](Duke_head_PML.pdf) ---------- --------------------- --------------------- --------------------- Scenario $\Delta S_{21}$(dB) $\Delta S_{21}$(dB) $\Delta S_{21}$(dB) \[2.4 GHz\] \[2.45 GHz\] \[2.5 GHz\] Case (a) 11.3 8.0 3.9 Case (b) 27.2 20.2 15.9 Case (c) -1.1 -2.5 -2.8 Case (d) -0.4 -1.6 -2.1 Case (e) 9.5 4.2 0.8 Case (f) -3.3 -5.0 -5.6 ---------- --------------------- --------------------- --------------------- : Difference between the link loss with and without UPML[]{data-label="diff_link_loss"} The difference between the link loss with these six simulation cases and the complete head is shown in Table \[diff\_link\_loss\] at $2.4$ GHz, $2.45$ GHz and $2.5$ GHz where $\Delta S_{21} = S_{21}\|_{complete\_head} - S_{21}\|_{UPML}$. A positive $\Delta S_{21}$ means increase in the link loss and a negative $\Delta S_{21}$ means decrease in the link loss when compared to the actual case when the complete head is included. It can be seen that whenever the back of the head is excluded from the simulation boundary (cf. case (a), (b), (e)) so that no wave creeps over the back side of the head, the link loss increased by $4.2$-$20.3$ dB at the central frequency of $2.45$ GHz. From this, it can be concluded that the creeping wave propagating over the back of the head is the strongest. This can also be confirmed by case (f) as the link loss is lowest when only back path is included. However, when the top of the head is excluded from the simulation boundary as in case (d), the link loss decreased by $1.6$ dB at $2.45$ GHz. Hence, the top creeping wave does not carry any significant power. This observation is also supported by the increase in the link loss by $20.3$ dB at $2.45$ GHz in case (b) where the waves creeps only over the head top. In case (d), where the front and the back of the head is included, the difference is small. Hence, the front and the back path carry most of the significant power. With these observations we can conclude that the two paths, one going around the back and the other going around the front part of the head should give a good estimation of the ear-to-ear link loss. Effect of the outer lossy skin and the Pinna -------------------------------------------- The presence of the outer lossy skin and the pinnas on the heterogeneous phantoms introduces extra loss when compared with the SAM head. In [@chandra2], it was shown that for low profile antennas which are very close to the head surface, the loss because of the pinnas could be up to $13$ dB. The presence of the pinnas may introduce an additional loss as the strongest creeping wave which creeps over back part of the head passes through them and hence gets attenuated. The loss because of the pinnas could be theoretically calculated by the sum of the propagation loss in a lossy medium (pinna) and the reflection loss as described in [@chandra2]. To verify the effect of the pinnas on the ear-to-ear link loss, an approximate pinna is modeled for the SAM phantom with the lossy shell having the electrical properties same as that of the SAM liquid which is approximately same for the ear-cartilage of the Duke phantom. The electrical properties of the pinna is assigned the same values (permittivity, $\varepsilon_r$ = $39.2$ and conductivity, $\sigma_e$ = $1.8$ S/m) as the SAM liquid to maintain the homogeneity. The height of the pinna is taken as $18$ mm which is same as that of the pinna in the Duke phantom. A semicircular structure of the pinna resulted in $7$ dB further increment in the link loss whereas a more realistic structure of the pinna resulted in $11$ dB increment in the link loss [@chandra2]. A reverse methodology is implemented to verify the effect of the presence of the pinna and the lossy skin on the Duke phantom. The tissues of the ear in the Duke phantom, namely, the ear cartilage and the ear skin (see Table \[tissue\_parameter\]) is assigned the electrical properties of air. This is done to remove the effect of the pinna. It resulted in decrease in the ear-to-ear link loss by $8$ dB at $2.45$ GHz. Next, the skin of the phantom (which is the outer most layer) is assigned the electrical properties of lossless SAM shell (permittivity, $\varepsilon_r$ = $3.7$ and conductivity, $\sigma_e$ = $0$). This further decreased the link loss by $13$ dB at $2.45$ GHz. The effects are illustrated in Fig. \[no\_pinna\_duke\]. The link loss for the SAM phantom with a realistic pinna and a lossy shell approaches to that of the Duke phantom in the ISM band as seen in Fig. \[link\_loss\_SAM\_Duke\]. Moreover, the reverse methodology shows the significant effect of the pinna and the lossy skin on the ear-to-ear link loss. Hence, it can be concluded that any numerical or real phantom with an outer lossless shell should be used with a caution while estimating the link loss for the on-body propagation and that the homogeneous phantoms should have a lossy outer shell. The phantoms should have the pinnas to get more accurate estimate of the ear-to-ear link loss. The shape of the pinna is also critical as more realistic pinna model estimates more accurate results. ![$S_{21}$ for Homogeneous vs. Heterogeneous Phantom using the ITE Antenna. The value in the square braces in the legend is the link loss at 2.45GHz.[]{data-label="link_loss_SAM_Duke"}](SAM_vs_Duke_ITE_Pinna.pdf) ![Effect of Pinna and Lossy Skin on Ear-to-Ear link loss for Duke[]{data-label="no_pinna_duke"}](No_Pinna_Air_tissues.pdf) ![Elliptical fit for Duke in the transverse plane at the level of the antennas. $a = 115.8$ mm is the semi-major axis of the ellipse and $b = 75.1$ is the semi-minor axis of the ellipse. $d$ is the length of the path of the creeping wave going over the back side of the head and $p$ is the perimeter of the ellipse. $x$ is the offset of the ears from the center of the ellipse[]{data-label="duke_ellipse"}](Duke_ellipse_paper.pdf) With the observations that the two paths of the creeping waves, namely, the front path and the back path are sufficient to describe the ear-to-ear link loss and the significant effect of the pinna and the lossy skin on the link loss, the analytical model for the ear-to-ear link loss is developed in the next section. Complete Link Loss Model ------------------------ In this section the ear-to-ear link loss analytical model based on the creeping wave is developed. The model is for low profile antennas such that the height of the antenna above the body surface is negligible. The head is approximated by a more realistic elliptical cross-section rather than a circular cross-section. The elliptical approximation is illustrated in Fig. \[duke\_ellipse\] for Duke head. In [@alomainy], it was shown that the human body can be approximated by a metallic cylinder. Thus, the head is assumed to be metallic. The loss because of the waves creeping over the metallic surface is considered as an approximation for the waves creeping over the lossy medium. The lowest mode for the creeping wave is considered. A simpler version of the model for an elliptical approximation of the human torso was presented in our previous work [@chandraEucap2012]. Here, the model with the additional effect of the pinna is presented. The creeping wave field on the elliptical path can be written as: $$\label{elecfield_ellipse} {\bf E} = {\bf E}_0e^{-L}$$ where the reference field ${\bf E}_0$ over the conducting surface for the vertical polarization at a distance $s$ is given by [@alves]: $$\label{eo} {\bf E}_0 = 2\sqrt{\frac{\eta_0}{2\pi}}{\frac{\sqrt{P_{TX}G_{TX}}}{s}}e^{-jks}$$ where $P_{TX}$ is the feeding power, $k=\frac{2\pi}{\lambda_0}$ is the wave number in the free space and $G_{TX}$ is the gain of the transmitting antenna. $L$ is the complex attenuation factor representing the loss on the surface. It is product of creeping distance and creeping attenuation per unit length. For the elliptical path over a metallic surface, it is given by [@balanis]: $$\begin{aligned} \label{l} L = \frac{(k)^{1/3}}{2}\left(\frac{3\pi ab}{4} \right)^{2/3}e^{\frac{j\pi}{6}} &&\nonumber \\ .\int_{\varphi_1}^{\varphi_2}\frac{ab}{\sqrt{[a^4\cos^2\varphi + b^4\sin^2\varphi][a^2\cos^2\varphi + b^2\sin^2\varphi]}}\mathrm{d}\varphi\end{aligned}$$ where $a$ is the semi-major axis and $b$ is the semi-minor axis of the ellipse. $\varphi_1$ and $\varphi_2$ are the exit point angle at the transmitter and the trapping point angle at the receiver respectively as shown in Fig. \[duke\_ellipse\]. The total field at the receiver can be written as: $$\label{E_total} {\bf E} = {\bf E}_{f} + {\bf E}_{b}$$ where the subscript $f$ is for the wave creeping over the front part of the head and $b$ is for the back of the head. ${\bf E}_{f}$ is given by: $$\label{elecfield} {\bf E}_{f} = {\bf E}_{0f}e^{-L_f}$$ where the reference field ${\bf E}_{0f}$ for the front part of the head is given by (\[eo\]) with $s = p - d$, where $p$ is the perimeter of the ellipse and $d$ is the elliptical arc length between the ears at back side of the head (as shown in Fig. \[duke\_ellipse\]). The arc length, $t$, of the creeping wave between the the angle $\varphi_1$ and $\varphi_2$ on the elliptical surface can be expressed as [@balanis]: $$\label{arc_length} t = ab\int_{\varphi_1}^{\varphi_2}\frac{(a^4\cos^2\varphi+ b^4\sin^2\varphi)^\frac{1}{2}}{(a^2\cos^2\varphi+ b^2\sin^2\varphi)^\frac{3}{2}}\mathrm{d}\varphi$$ $L_f$ can be calculated by (\[l\]) with the integration from $-\varphi_1$ to $\varphi_2 = \pi+\varphi_1$, where $\varphi_1$ is given by: $$\label{varphi1} \varphi_1 = \tan^{-1}\left[\frac{a}{b}\frac{x}{\sqrt{a^2-x^2}}\right]$$ where $x$ is the offset of the ear from the center of the ellipse as shown in Fig. \[duke\_ellipse\]. $d$ can be calculated by substituting the proper values of $\varphi_1$ and $\varphi_2$ and $p$ can be calculated by substituting $\varphi_1 = 0$ and $\varphi_2 = 2\pi$ in (\[arc\_length\]). The wave going over the back of the head has to pass through the pinnas and hence there will be reflection and absorption loss apart from the surface attenuation. $E_{b}$ is given by: $$\label{elecfield_back} {\bf E}_{b} = {\bf E}_{0b}e^{-L_b}T^2_{pinna}e^{-2\alpha R}$$ The reference field, ${\bf E}_{0b}$ is given by (\[eo\]) with $s = d$. For $L_b$ in (\[l\]), $\varphi_1$ is same as that in (\[varphi1\]) and $\varphi_2 = \pi-\varphi_1$. $T_{pinna}$ is the equivalent transmission coefficient for the air-pinna-air interface given by [@molisch] as: $$\label{T} T_{pinna} = \frac{T_1 T_2e^{-j\alpha_d}}{1+\rho_1 \rho_2 e^{-2j\alpha_d}}$$ where $T_1$ and $T_2$ are the transmission coefficient of the air-to-pinna and pinna-to-air interface respectively. $\rho_1$ and $\rho_2$ are the reflection coefficient of the air-to-pinna and the pinna-to-air interface respectively. $\alpha_d$ is the electrical length of the pinna with permittivity $\epsilon_r$ for the normally incident waves and is given by $\alpha_d = \frac{2\pi\sqrt{\epsilon_r}R}{\lambda}$. Since the vertical polarization is dominant for the used antenna (described in the next section), TM waves are assumed. $e^{-\alpha R}$ in (\[elecfield\_back\]) represents the absorption loss in the pinna of average thickness $R$ with $\alpha = \|Im[k_{pinna}]\|$ where $k_{pinna}$ is complex wavenumber of the lossy medium (pinna). It should be noted that the terms $T_{pinna}$ and $e^{-\alpha R}$ are squared to represent the effect of the both pinnas. The received power is: $$\label{E_total_e} P_{RX} = \frac{\|{\bf E}\|^2 A_{RX}}{2\eta_0}$$ Substituting $A_{RX} = G_{RX}\frac{\lambda^2}{4\pi}$ and ${\bf E} = {\bf E}_{f} + {\bf E}_{b} $, the link loss $\left(\frac{P_{TX}}{P_{RX}} \right)$ between the perfectly matched antennas in the dB scale can be written as: $$\begin{aligned} \label{LinkLossdB} LL\|_{dB} &=& -10\mathrm{log}\left[\frac{G_{RX}G_{TX}\lambda^2}{4\pi^2}\left(\vline \frac{e^{-L_f}}{p-d}e^{-jk(p-d)}+\right.\right. \nonumber \\ && \left.\left. \frac{e^{-L_b}T^2_{pinna}e^{-2\alpha R}}{d}e^{-jkd}\vline\right)^2\right]\end{aligned}$$ Verification of the Analytical Model ------------------------------------ The analytical model is verified for the phantoms Duke and Billie, having the largest and the smallest head respectively. The parameters of Duke’s and Billie’s head after the elliptical fit are described in Table \[analytical\]. The values of $L_f$, $L_b$, creeping loss (losses excluding the gain of the antennas) and the the total ear-to-ear link loss obtained from the analytical model are also shown. The gain of the antenna is calculated from the simulations as described in [@alves]. The gain of the antenna in the azimuthal plane for the vertical polarized E-field (solid line) and horizontal polarized (dot line) is shown in Fig. \[Duke\_Gain\]. It can be seen from the radiation pattern that the vertical polarization (normal to the head surface) is dominant. Also, the pattern is asymmetrical i.e. the two directions in which creeping waves exit or reach the antenna have different gain as one side has the pinna. Moreover, the two antennas on different ears have different gain due to the mismatch. Hence, the highest gain of the two antenna at $\varphi = 0^\circ$ is taken as the model assumes perfect match. The comparison between the simulated link loss and the link loss calculated with the analytical model in the ISM band is shown in Fig. \[sim\_ana\]. The difference between the analytical and the simulated link loss is within $4$ dB for both the phantoms showing a good agreement. [\|c\|\|c\|\|c\|\|c\|\|c\|\|c\|]{} Freq.(GHz) & Gain(dBi) & $L_f$ & $L_b$ & CL$^{}$(dB) & LL[$^\dagger$]{}(dB)\ \ \ 2.4 & -10.0 & 4.30+2.48i & 3.69+2.13i & 61.34 & 81.34\ 2.45 & -9.95 & 4.33+2.50i & 3.72+2.15i & 61.80 & 81.70\ 2.5 & -10.2 & 4.36+2.52i & 3.74+2.16i & 62.27 & 82.67\ \ \ 2.4 & -7.0 & 4.05+2.34i & 3.37+1.95i & 56.40 & 70.40\ 2.45 & -6.7 & 4.08+2.35i & 3.39+1.96i & 56.85 & 70.25\ 2.5 & -6.6 & 4.10+2.37i & 3.41+1.97i & 57.29 & 70.49\ $^$CL: Creeping Loss (losses excluding the antenna gain)\ $^\dagger$LL: Total ear-to-ear link loss (including the antenna gain) ![Link Loss: Simulated vs. Analytical[]{data-label="sim_ana"}](S21_sim_analytical.pdf) Limitations and Discussions about the Analytical Model ------------------------------------------------------ The analytical model gives the deterministic component of the ear-to-ear link loss. In the indoor scenarios where multi-path components (MPCs) are present, the link loss could be higher or lower depending upon the interference of the MPCs. The model is valid for the antennas which are very close to the human body so that the dominant propagation mechanism is through the creeping waves. The elliptical fit of the head is done when the models are looking parallel to the transversal plane. Duke was looking parallel to the transversal plane at the at the ear level (xy plane in the simulations) but for Billie, the model is rotated by $5^{\circ}$ to make her vision parallel to the xy plane. The gain of the antenna is also critical as it is difficult to measure with the body [@alves], but could result in significant differences in the link loss. The difference in the link loss from the analytical model between Duke and Billie is about $11$ dB at $2.45$ GHz. $5$ dB out of this $11$ dB is the difference in the creeping loss attributed to the difference in the head size and the ear thickness. The rest $6$ dB is due to the difference in the gain of the antennas. Hence, a more appropriate method of calculating the gain of the antennas with the body may result in more accurate results. Moreover, as discussed in [@chandraEucap2012], there is no well defined measure (to best of our knowledge) such as the gain for the coupling from the antenna to the creeping wave over the human body which involves near-field effects. Hence, using the standard gain which is calculated in far-field is an effective approximation. Other Factors Affecting the Ear-to-Ear Link Loss {#sec:factors} ================================================ Effect of Different Head Sizes ------------------------------ The effect of the head size on the ear-to-ear propagation channel has been discussed in [@kvist], [@kvist1] and [@chandra2]. In our previous work [@chandra2], we have shown by the simulations on the heterogeneous phantoms that different head size will have different link loss as shown in Fig. \[link\_loss\_phantoms\]. This can be attributed to the fact that the different head size will result in different path lengths of the creeping waves going behind the head and front of the head and also because of the gain variations of the antenna resulting from anatomical variations in different phantoms. ![$S_{21}$ for different phantoms. From left to right: Duka, Ella, Louis, Billie[]{data-label="link_loss_phantoms"}](S21_Diff_Phantom.pdf) Effect of the Shoulders ----------------------- The shoulders may also affect the ear-to-ear link loss as the reflections from the shoulders may result in multipath components interfering either constructively or destructively with the creeping waves resulting in different power levels. The effect of the shoulders on Duke phantom can be observed in Fig. \[shoulder1\]. It can be seen that the difference between the link loss with and without the shoulders is within $2$ dB in the ISM band. Thus, there is a possibility to exclude the shoulders in order to decrease the simulation time [@chandra2]. ![$S_{21}$ with shoulders and without shoulders for Duke phantom[]{data-label="shoulder1"}](withShoulder_withoutShoulderITE.pdf) Measurements {#sec:meas} ============ Measurements are carried out to verify the effect of the pinna. The measurement setup and the measured scenarios are described below. Measurement Setup ----------------- The phantom shown in Fig. \[phantom\] is used for the measurements. The phantom is hollow and is filled with the tissue stimulating liquid mimicking the average electrical properties of the human head tissues at $2.45$ GHz (permittivity, $\varepsilon_r$ = $39.2$ and conductivity, $\sigma_e$ = $1.80$ S/m). The phantom does not have the pinnas. External pinnas are manufactured and attached to the phantom. The measured scenarios are (a) ear-to-ear link loss without the pinnas (b) ear-to-ear link loss with the pinna. Return loss of the antennas are also measured in all these scenarios. The measurement is done in an anechoic chamber of the EIT department, Lund University. S-parameters, $S_{11}$ and $S_{21}$, are measured at $1601$ frequency points in $1.95$ GHz to $2.95$ GHz frequency band using HP-8720C Vector Network Analyzer (VNA). The output power is set to $-10$ dBm. TOSM calibration is done to calibrate the VNA using a fabricated calibration kit. The calibration kit is fabricated in order to remove the effect of the semi-rigid cable which is attached to the antenna. ![Phantom used for the measurement. The placement of the antennas and the pinna is can also be seen.[]{data-label="phantom"}](meas_phantom_pinna.pdf) Fabricated Pinna ---------------- The pinnas are commercially fabricated by the 3D-printing technology by Shapeways [@shapeways]. Plastic is used as the material. They are hollow from inside and holes are made on them to facilitate the filling of the hollow pinnas with the tissue stimulating liquid. After filling the pinnas with the liquid, the holes are closed with a glue gun. The pinnas are glued on the phantom with the help of a thin double sided tape. The fabricated pinnas are shown in Fig \[pinna\_fab\]. The phantom with the pinna and the antenna placement is shown in Fig \[phantom\]. ![Pinnas fabricated by the 3D-printing[]{data-label="pinna_fab"}](pinna_fab.pdf) Antennas -------- A monopole antenna is designed and optimized on the SAM phantom with the pinnas to work in $2.45$ GHz band. The antenna is loaded with a disc and then the disc is shorted to the ground plane to miniaturize the size and make the antenna low profile. The fabricated antenna with the dimensions is shown in Fig. \[fab\_antenna\]. The central conductor of the semi-rigid coaxial cable, UT-85-H-M17 provided by Rosenberger [@rosenberger] is used as the central conductor in the monopole antenna. The outer conductor is then soldered to the ground plane of the antenna (also shown in Fig. \[fab\_antenna\]). The length of the cable is $1$ m with the SMA connector at the other end of the cable. Hence, the calibration kit is fabricated to remove the affect of this semi-rigid cable. The semi-rigid cable has outer diameter of $2.2$ mm. It is used in order to place the antenna close to the head because directly attaching the SMA connector to the antenna could have resulted in much more air gap between the antenna and the head. The ground plane and the loading disc are etched out from a Teflon laminate of $0.8$ mm thickness having $0.35$ $\mu$m thick copper layer. ![Fabricated Antenna[]{data-label="fab_antenna"}](fab_antenna.pdf) Measured Results and Discussions -------------------------------- The return loss of the fabricated antenna is measured on the phantom, with and without the pinnas. It is observed that the antenna is mismatched when placed on the phantom without the pinna but the presence of the pinna improves the matching. The mismatch loss in the absence of the pinnas are removed from the measured $S_{21}$ for proper comparison by treating the return loss in the presence of the pinnas as the reference. This is done as: $$\label{s21_comp} \|S_{21}\|_{no\_pinna}\|^{2} = \|S_{21}\|_{no\_pinna}\|_{meas}^{2} \times F$$ where $F$ is the correction factor for removal of the mismatch loss without the pinna with the antenna with the pinna as the reference, given by: $$\label{mismatch_norm} F = \frac{(1-\|S_{11}\|_{pinna}^{2})\times (1-\|S_{22}\|_{pinna}^{2})}{(1-\|S_{11}\|_{no\_pinna}^{2}) \times (1-\|S_{22}\|_{no\_pinna}^{2})}$$ where $\|.\|$ represents the absolute value of the S-parameters in linear scale. The measured results are shown in Fig. \[meas\_s21\]. It can be seen that the link loss without the pinna is $4$ dB less than the link loss with the pinna at $2.45$ GHz. The difference is not as significant as discussed in Section III(B). The possible reason for this could be the fact that the antennas used for the measurements are not low profile as the antennas used in simulations. Moreover, the phantom used for the measurement does not have the ear canals. Hence, the antennas are not shadowed to the extent as that in simulations. Nevertheless, the increase in the link loss because of the pinnas can still be observed. ![Measured $S_{21}$ with and without the pinna in the $2.45$ GHz ISM Band[]{data-label="meas_s21"}](withEar_vs_WithoutEarMeasISM.pdf) Conclusions {#sec:conclusion} =========== An analytical model based on creeping wave for the ear-to-ear link loss was presented. It is useful in estimating the deterministic component of the link loss between the transceivers for binaural hearing aids. The model takes into account the losses on the surface of the head, the loss due to the pinna and models the cross-section of the head with an elliptical shape. Through the simulations it was observed that the two paths of the creeping wave, one going over the back of the head and the other in front of the head at the level of the ears were the dominant paths of the creeping waves carrying most of the power. Hence, a two path analytical model was developed. The verification of the model was done through the simulations on two heterogeneous phantoms of different age (34 year and 11 year) and head size. The difference between the model and the simulations was found to be within $4$ dB in the $2.45$ GHz ISM band showing a good agreement. The main benefit of the model lies in fast estimation of the link loss when compared with time and memory consuming numerical simulations. With the estimated value of the link loss, the sensitivity of the hearing aids could be decided. A comparison between the SAM phantom and a heterogeneous phantom was done which showed a significant effect of the pinna and the lossy skin. It was observed that the presence of the pinnas increased the link loss. This effect of the pinnas was verified through the measurements on a phantom where external pinnas manufactured by 3D-printing were attached. Acknowledgment {#acknowledgment .unnumbered} ============== This research work has been done as a part of the Ultra Portable Devices (UPD) project. The authors would like to thank Swedish Foundation for Strategic Research (SSF) for funding the Ultra Portable Devices (UPD) project at Lund University [1]{} P.F. Binkley, “Predicting the Potential of Wearable Technology", IEEE Eng. Med. Biol. Mag., vol. 22,, no. 3, pp. 23-27, June 2003 A. Dittmar; R. Meffre; F. De Oliveira; C. Gehin and G. Delhomme, “Wearable Medical Devices Using Textile and Flexible Technologies for Ambulatory Monitoring", in Proc. 27th Annu. Int. Conf. of the Eng. in Med. and Biology Soc., IEEE-EMBS 2005, pp.7161-7164, 2005 S.L. Cotton, G.A. Conway and W.G. Scanlon, “A Time-Domain Approach to the Analysis and Modeling of On-Body Propagation Characteristics Using Synchronized Measurements at 2.45 GHz", IEEE Trans. Antennas Propag., vol.57, no.4, pp.943-955, April 2009 A. Fort, F. Keshmiri, G.R. Crusats, C. Craeye, and C. Oestges, “A Body Area Propagation Model Derived From Fundamental Principles: Analytical Analysis and Comparison With Measurements", IEEE Trans. Antennas Propag., vol.58, no.2, pp.503-514, Feb. 2010 G. Roqueta, A. Fort; C. Craeye and C. Oestges, “Analytical Propagation Models for Body Area Networks", in Proc. IET Seminar on Antennas and Propag. for Body-Centric Wireless Commun., pp. 90-96, April 2007 J. Ryckaert, P. De Doncker, R. Meys; A. de Le Hoye and S. Donnay, “Channel model for wireless communication around human body", Electronics Letters, vol.40, no.9, pp. 543-544, April 2004 T. Alves, B. Poussot, J.-M. Laheurte, “Analytical Propagation Modeling of BAN Channels Based on the Creeping-Wave Theory", IEEE Trans. Antennas Propag., vol.59, no.4, pp.1269-1274, April 2011 G.A. Conway, W.G. Scanlon, S.L. Cotton and M.J.Bentum, “An analytical path-loss model for on-body radio propagation", in Proc. URSI Int. Symp. on Electromagnetic Theory, pp. 332-335, August 2010 J.R. Wait, “The ancient and modern history of EM ground-wave propagation", IEEE Antennas Propagat. Mag., vol.40, no.5, pp.7-24, Oct 1998 R. Chandra and A. J Johansson, “Miniaturized antennas for link between binaural hearing aids", in Proc. 32nd Annu. Int. Conf. of the IEEE Engg. in Med. and Biol. Soc., (EMBC’10), pp.688-691, Aug.2010 T. Zasowski, G. Meyer, F. Althaus and A. Wittneben, “UWB signal propagation at the human head", IEEE Trans. Microw. Theory Tech., vol.54, no.4, pp. 1836-1845, June 2006 S.H. Kvist, J. Thaysen and K.B. Jakobsen, “Investigation of the ear-to-ear radio propagation channel", in Proc. 5th European Conf. on Antennas and Propagation (EUCAP’11), pp. 3640-3644, April 2011 S.H. Kvist, J. Thaysen and K.B. Jakobsen, “The effect of the head size on the ear-to-ear radio-propagation channel for body-centric wireless networks", in Proc. 2010 Loughborough Antennas and Propagation Conf. (LAPC’10), pp.345-348, Nov. 2010 B. Nour and O. Breinbjerg, “Measurement and characterization of the path loss for ear-to-ear wireless communication", in Proc. 5th European Conf. on Antennas and Propagation (EUCAP’11), pp.1621-1625, April 2011 S.H. Kvist, J. Thaysen, K.B. Jakobsen, “Ear-to-ear on-body channel fading in the ISM-band for tangentially-polarized antennas", in Proc. Loughborough Antennas and Propagation Conf. (LAPC’11), Nov. 2011 S.H. Kvist, J. Thaysen, K.B. Jakobsen, “Polarization of unbalanced antennas for ear-to-ear on-body communications at 2.45 GHz", in Proc. Loughborough Antennas and Propagation Conf. (LAPC’11), Nov. 2011 R. Chandra and A.J Johansson, “Influence on the ear-to-ear link loss from heterogeneous head phantom variations", in Proc. 5th European Conf. on Antennas and Propagation (EUCAP’11), pp.1612-1615, April 2011 \[Online\]. Available: http://www.speag.com/products/semcad/solutions/ \[Online\]. Available: http://www.itis.ethz.ch A. Christ et al., “The Virtual Family–Development of anatomical CAD models of two adults and two children for dosimetric simulations", Physics in Medicine and Biology,vol.55, no.2, pp. N23-N38, Jan. 2010 \[Online\]. Available: http://www.niremf.ifac.cnr.it/tissprop/ A. Fort, J. Ryckaert, C. Desset, P. De Doncker, P. Wambacq and L. Van Biesen, “Ultra-wideband channel model for communication around the human body", IEEE J. Sel. Areas Commun, vol.24, no.4, pp. 927-933, April 2006 Yan Zhao, Yang Hao, A. Alomainy and C. Parini, “UWB on-body radio channel modeling using ray theory and subband FDTD method, IEEE Trans. Microw. Theory Tech, vol.54, no.4, pp. 1827-1835, June 2006 R. Chandra and A.J Johansson, “An Elliptical Link Loss model for Wireless Propagation around Human Torso", presented 6th European Conf. on Antennas and Propagation (EUCAP’12), April 2011 C. Balanis and L. Peters, Jr., “Aperture radiation from an axially slotted elliptical conducting cylinder using geometrical theory of diffraction, IEEE Trans. Antennas Propag., vol.17, no.4, pp.507-513, Jul 1969 Andreas F. Molisch, Wireless Communications, IEEE Press/John Wiley $\&$ Sons, Ltd., 2009 \[Online\]. Available: http://www.shapeways.com/ \[Online\]. Available: http://www.rosenberger.de/ [^1]: Rohit Chandra and Anders J Johansson are with the Department of Electrical and Information Technology, Lund University, Lund, Sweden, e-mail: [email protected]	{ "pile_set_name": "ArXiv" }
--- abstract: 'We calculate the conductivity matrix of a weakly coupled quark-gluon plasma at the leading-log order. By setting all quark chemical potentials to be identical, the diagonal conductivities become degenerate and positive, while the off-diagonal ones become degenerate but negative (or zero when the chemical potential vanishes). This means a potential gradient of a certain fermion flavor can drive backward currents of other flavors. A simple explanation is provided for this seemingly counter intuitive phenomenon. It is speculated that this phenomenon is generic and most easily measured in cold atom experiments.' author: - 'Jiunn-Wei Chen' - 'Yen-Fu Liu' - Shi Pu - 'Yu-Kun Song' - Qun Wang title: 'Negative Off-Diagonal Conductivities in a Weakly Coupled Quark Gluon Plasma' --- Introduction ============ Hydrodynamics describes the evolution of a fluid perturbed away from thermal equilibrium by long wave length fluctuations. The long wave length physics (long compared with the mean field path of particle collisions) can be systematically described by an expansion of space-time derivatives on classical fields with prefactors called transport coefficients. These transport coefficients encode the physics of short (compared with the mean free path) distance and are inputs to hydrodynamics. But they can be computed, in principle, once the microscopic theory of the system is known. We are interested in computing the transport coefficients in Quantum Chromodynamics (QCD) with $N_{f}$ flavors of massless quarks at finite temperature ($T$) and chemical potentials ($\mu _{a}$, $a=1,2,\cdots ,N_{f}$). The leading transport coefficients at the first derivative order include the shear viscosity ($\eta $), bulk viscosity ($\zeta $), and the conductivity matrix ($\lambda $). The shear viscosity of QCD has attracted a lot of attention recently. Its ratio with the entropy density ($s$) extracted from the hot and dense matter created at Relativistic Heavy Ion Collider (RHIC) [Arsene:2004fa,Adcox:2004mh,Back:2004je,Adams:2005dq]{} just above the phase transition temperature ($T_{c}$) yields $1/(4\pi )\leq \eta /s\leq 2.5/(4\pi )$ at $T_{c}\leq T\leq 2T_{c}$ [@Song:2010mg], which is close to a conjectured universal lower bound of $1/(4\pi )$ [@Kovtun:2004de] inspired by the gauge/gravity duality [Maldacena:1997re,Gubser:1998bc,Witten:1998qj]{}. This value of $\eta /s$ cannot be explained by extrapolating perturbative QCD result [Arnold:2000dr,Arnold:2003zc,Chen:2010xk,Chen:2011km]{}. The smallest $\eta /s$ is likely to exist near $T_{c}$ [@Csernai:2006zz; @Chen:2006iga] (see, e.g., Ref. [@Chen:2011km] for a compilation and more references). Alsofinite $\mu $ results suggests that $\eta /s$ is smaller at smaller $\mu $. This is based on results of perturbative QCD at $T\gg $ $T_{c}$ [Chen:2012jc]{} and of a hadronic gas at $T\ll $ $T_{c}$ and small $\mu $ [Chen:2007xe]{}. It is speculated that the same pattern will persist at $T_{c}$ such that the smallest $\eta /s$ might exist near $T_{c}$ with $\mu =0$ [Chen:2012jc]{}. For the bulk viscosity, the sum rule study [Kharzeev:2007wb,Karsch:2007jc]{} shows that $\zeta $ increases rapidly near $T_{c}$ when $T$ approaches $T_{c}$ from above. This is consistent with the lattice gluon plasma result near $T_{c}$ [@Meyer:2010ii] and perturbative QCD result [@Arnold:2006fz] at much higher $T$. This, when combined with pion gas results below $T_{c}$ [Chen:2007kx,FernandezFraile:2008vu,Lu:2011df,Dobado:2011qu,Chakraborty:2010fr]{}, suggests that $\zeta /s$ has a local maximum near $T_{c}$ (see, e.g., [Chen:2011km]{} for a compilation). Unlike $\eta /s$, perturbative QCD result shows very small $\mu $ dependence in $\zeta /s$ [@Chen:2012jc]. Note that at high $\mu $, there are also bulk viscosities governed by the weak interaction such as the Urca processes which have consequences in neutron star physics [Dong:2007mb,Alford:2006gy,Alford:2008pb,Sa’d:2006qv,Sa’d:2007ud,Wang:2010ydb]{}. These are quite different from the transport coefficients from the strong interaction mentioned above. The perturbative QCD calculations of $\eta $ and $\zeta $ with finite $\mu $ were performed at the leading-log (LL) order of the strong coupling constant ($g$) expansion in Ref. [@Chen:2012jc]. Either $T$ or $\mu $ in the calculation is much larger than $\Lambda _{QCD}$ which is the scale where QCD becomes non-perturbative. But the calculation is not applicable to the color superconducting phase at $\mu /T\rightarrow \infty $, since the vacuum in the calculation has no symmetry breaking. In this work, we apply the same perturbative QCD approach to compute the conductivity matrix $\lambda $ at the LL order. The conductivity is an important transport coefficient which plays an essential role in the evolution of electromagnetic fields in heavy ion collisions [Huang:2013iia,McLerran:2013hla]{}. The conductivity in strongly coupled quark gluon plasma was calculated with lattice QCD [@Ding:2010ga; @Amato:2013naa] and Dyson-Schwinger equation [@Qin:2013aaa]. We first review the constraints from the second law of thermal dynamics (i.e. the entropy production should be non-negative) which show that the particle diffusion, heat conductivities, and electric conductivity are all unified into one single conductivity in this system. When $N_{f}>1$, the conductivity becomes a $N_{f}\times N_{f}$ matrix. We then show through the Boltzmann equation that the conductivity matrix $\lambda $ at the LL order is symmetric and positive definite ($\sum_{a,b}\lambda _{ab}X^{a}X^{b}>0$ for any real, non-vanishing vector $X$). The former is a manifestation of the Onsager relation while the latter is a manifestation of the second law of thermal dynamics. For simplicity, we show the numerical results of $\lambda $ with all fermion chemical potential to be identical. In this limit, there are only two independent entries in $\lambda $. All the diagonal matrix elements are degenerate and positive since $\lambda $ is positive definite. However, the off-diagonal matrix elements are degenerate but negative at finite $\mu $. This means a gradient $\bm{\nabla }\mu _{a}$ can drive a current of flavor $a $ alone the gradient direction, but it will also drive currents of different flavors in the opposite direction. This backward current phenomenon might seem counter intuitive, but we find that it is generic and it has a simple explanation. We speculate that this phenomenon might be most easily measured in cold atom experiments. Entropy principle in hydrodynamics \[sub:Entropy-principle copy(1)\] ==================================================================== Single flavor case ------------------ Let us start from the hydrodynamical system with only one flavor of quark of electric charge $Q$. The energy-momentum conservation and current conservation yield $$\begin{aligned} \partial _{\mu }T^{\mu \nu }& =QF^{\nu \lambda }j_{\lambda }, \notag \\ \partial _{\mu }j^{\mu }& =0, \label{eq:conservation_01}\end{aligned}$$where $T^{\mu \nu }$ is the energy-momentum tensor, $j^{\mu }$ is the quark current and $F^{\nu \lambda }$ is the electromagnetic field strength tensor. The long wave length physics can be systematically described by the expansion of space-time derivatives$$\begin{aligned} T^{\mu \nu }& =T_{(0)}^{\mu \nu }+\varepsilon T_{(1)}^{\mu \nu }+\varepsilon ^{2}T_{(2)}^{\mu \nu }+\cdots , \notag \\ j^{\mu }& =j_{(0)}^{\mu }+\varepsilon j_{(1)}^{\mu }+\varepsilon ^{2}j_{(2)}^{\mu }+\cdots ,\end{aligned}$$where we have used the parameter $\varepsilon $ to keep track of the expansion and we will set $\varepsilon =1$ at the end. $F^{\nu \lambda }$ is counted as $\mathcal{O}(\varepsilon )$. We will then assume the system is isotropic and homogeneous in thermal equilibrium so there is no special directions or intrinsic length scales macroscopically. We also assume the underlying microscopic theory satisfies parity, charge conjugation and time reversal symmetries such that the antisymmetric tensor $\varepsilon ^{\mu \nu \alpha \beta }$ does not contribute to $T^{\mu \nu }$ and $j^{\mu }$. Also, we assume the system is fluid-like, describable by one (and only one) velocity field (the conserved charged is assumed to be not broken spontaneously, otherwise the superfluid velocity needs to be introduced as well). Also, at $\mathcal{O}(\varepsilon ^{0})$, the system is in local thermal equilibrium, i.e. the system is in equilibrium in the comoving frame where the fluid velocity is zero. With these assumptions, we can parametrize $$\begin{aligned} T_{(0)}^{\mu \nu }& =\left( \epsilon +P\right) u^{\mu }u^{\nu }-Pg^{\mu \nu }, \notag \\ j_{(0)}^{\mu }& =nu^{\mu }, \notag \\ T_{(1)}^{\mu \nu }& =\Pi \left( u^{\mu }u^{\nu }-g^{\mu \nu }\right) +\pi ^{\mu \nu }+h^{\mu }u^{\nu }+h^{\nu }u^{\mu }, \notag \\ j_{(1)}^{\mu }& =\nu ^{\mu }, \label{T}\end{aligned}$$where $g^{\mu \nu }=$diag($+,-,-,-$ ) and $\epsilon $, $P$ and $n$ are the energy density, pressure and number density, respectively. The fluid velocity $u^{\mu }=(u^{0},\boldsymbol{u})=(1,\boldsymbol{v})/\sqrt{1-\boldsymbol{v}^{2}}$ and $u^{\mu }u_{\mu }=1.$ $\Pi $, $\pi ^{\mu \nu }$, $h^{\mu }$ and $\nu ^{\mu }$ are the bulk viscous pressure, shear viscous tensor, heat flow vector and diffusion current. They satisfy the orthogonal relations, $\pi ^{\mu \nu }u_{\nu }=\nu ^{\mu }u_{\mu }=h^{\mu }u_{\mu }=0$. The covariant entropy flow is given by [@Israel:1979wp; @Pu:2011vr] $$S^{\mu }=\beta Pu^{\mu }+\beta T^{\mu \nu }u_{\nu }-\beta \mu j^{\mu }=su^{\mu }+\beta h^{\mu }-\beta \mu \nu ^{\mu }, \label{eq:entropy_flow_01}$$where $\beta =1/T$ and $s=\beta (\epsilon +P-\mu n)$ is the entropy density. Taking the space-time derivative of $S^{\mu }$, then using the Gibbs-Duhem relation $d\epsilon =Tds+\mu dn$ and the conservation equations ([eq:conservation\_01]{}), we obtain the equation for entropy production: $$\begin{aligned} \partial _{\mu }S^{\mu }& =-\nu ^{\mu }\left[ \partial _{\mu }(\beta \mu )+\beta QE_{\mu }\right] +h^{\mu }\left( \partial _{\mu }\beta +\beta u^{\nu }\partial _{\nu }u_{\mu }\right) \notag \\ & +\beta \pi ^{\mu \nu }\partial _{\langle\mu }u_{\nu \rangle}-\beta \Pi \partial \cdot u, \label{div-s-1}\end{aligned}$$where the symmetric traceless tensor $\partial _{\langle \mu }u_{\nu \rangle} $ is defined by, $$\partial _{\langle\mu }u_{\nu \rangle}=\frac{1}{2}\left[ \Delta _{\mu \alpha }\Delta _{\nu \beta }+\Delta _{\nu \alpha }\Delta _{\mu \beta }-\frac{2}{3}\Delta _{\mu \nu }\Delta _{\alpha \beta }\right] \partial ^{\alpha }u^{\beta },$$and where$\;\Delta ^{\mu \nu }=g^{\mu \nu }-u^{\mu }u^{\nu }$ and $E^{\mu }\equiv F^{\mu \nu }u_{\nu }$ is the electric field in the comoving frame. At $\mathcal{O}(\varepsilon )$, $\partial _{\mu }T^{\mu \nu }\simeq \partial _{\mu }T_{(0)}^{\mu \nu }=0$. This equation yields $$\partial _{\mu }\beta +\beta u^{\nu }\partial _{\nu }u_{\mu }=\frac{n}{\epsilon +P}[\partial _{\mu }(\beta \mu )+\beta QE_{\mu }], \label{eq:relation_01}$$where we have used the thermodynamic equation $dP=\beta (\epsilon +P)dT+nTd(\beta \mu )$. This identity simplifies Eq. (\[div-s-1\]) to $$\begin{aligned} \partial _{\mu }S^{\mu }& =-\left( \nu ^{\mu }-\frac{n}{\epsilon +P}h^{\mu }\right) [\partial _{\mu }(\beta \mu )+\beta QE_{\mu }] \notag \\ & +\beta \pi ^{\mu \nu }\partial _{\left\langle \mu \right. }u_{\left. \nu \right\rangle }-\beta \Pi \partial \cdot u. \label{eq:h_nu_02-1}\end{aligned}$$ The second law of thermodynamics requires $\partial _{\mu }S^{\mu }\geq 0$. It can be satisfied if, up to terms orthogonal to $\partial \cdot u$, $\partial _{\left\langle \mu \right. }u_{\left. \nu \right\rangle }$ and $[\partial _{\mu }(\beta \mu )+\beta QE_{\mu }]$, $\Pi $, $\pi ^{\mu \nu }$, $h^{\mu }$ and $\nu ^{\mu }$ have the following forms at $\mathcal{O}(\varepsilon )$: $$\begin{aligned} \pi ^{\mu \nu }& =2\eta \partial ^{\left\langle \mu \right. }u^{\left. \nu \right\rangle }, \notag \\ \Pi & =-\zeta \partial \cdot u, \notag \\ \nu ^{\mu }-\frac{n}{\epsilon +P}h^{\mu }& =\lambda \Delta ^{\mu \nu }\left[ \partial _{\nu }(\beta \mu )+\beta QE_{\nu }\right] , \label{eq:definition_01}\end{aligned}$$where $\Delta ^{\mu \nu }$ is inserted because $\nu ^{\mu }u_{\mu }=h^{\mu }u_{\mu }=0$. The coefficients $\eta $, $\zeta $ and $\lambda $ are transport coefficients with names of shear viscosity, bulk viscosity and conductivity, respectively. The second law of thermodynamics requires these transport coefficients to be non-negative. On the right hand side of Eq. (\[eq:definition\_01\]), the three vectors $\partial _{\nu }\mu $, $\partial _{\nu }\beta $ and $E_{\nu }$ form a unique combination and share the same transport coefficient $\lambda $ [Israel:1979wp]{}. It is obtained by assuming $\partial _{\mu }T_{(0)}^{\mu \nu }=0$ and $T_{(0)}^{\mu \nu }$ has the ideal fluid form described in Eq.(\[T\]). In general, we do not expect this to be true in all systems (e.g. a solid might not have the ideal fluid description) and hence there could be more transport coefficients. Conventionally, the transport coefficients corresponding to $\partial _{\nu }\mu $, $\partial _{\nu }\beta $ and $E_{\nu }$ are called particle diffusion, heat conductivity, and electric conductivity, respectively. In hydrodynamics, the choice of the velocity field is not unique. One could choose $\mathbf{u}$ to align with the momentum density $T^{0i}\mathbf{\hat{i}}$ or the current $\mathbf{j}$, or their combinations. However, the system should be invariant under the transformation $u_{\mu }\rightarrow u_{\mu }^{^{\prime }}=u_{\mu }+\varepsilon \delta u_{\mu }$ as long as $u_{\mu }^{^{\prime }2}=1$ is maintained (or $u^{\mu }\delta u_{\mu }=0$ at $\mathcal{O}(\varepsilon )$). Under this transformation, $h_{\mu }\rightarrow $ $h_{\mu }^{\prime }=h_{\mu }+\left( \epsilon +P\right) \delta u_{\mu }$ and $\nu _{\mu }\rightarrow $ $\nu _{\mu }^{\prime }=\nu _{\mu }+n\delta u_{\mu }$. However, the entropy production equation (\[eq:h\_nu\_02-1\]) remains invariant under this transformation. In this paper, we will be working at the Landau frame with $\mathbf{u}$ proportional to the momentum density $T^{0i}\mathbf{\hat{i}}$ such that $T^{0i}=0$ in the comoving frame. Then $$\mathbf{h}=0,\quad \bm{\nu }=\lambda \lbrack -\bm{\nabla }(\beta \mu )+\beta Q\mathbf{E}] \label{aa}$$ from Eq. (\[eq:definition\_01\]). $\lambda $ is positive, the sign makes sense for particle diffusion and electric conduction because the diffusion is from high to low density and positively charged particles move along the $\mathbf{E}$ direction. However, heat conduction induces a flow from low to high temperature! This result is counter intuitive. This is because $\bm{\nabla }T$ induces a momentum flow $\mathbf{h}$. If we choose to boost the system to the Landau frame where $\mathbf{h}=0$, then the physics is less transparent. For particle diffusion and electric conduction this is not a problem, because one could have particles and antiparticles moving in opposite directions and still keep the net momentum flow zero. The physics of heat conduction becomes clear in the Eckart frame where $\mathbf{u}$ is proportional to the current $\mathbf{j}$ and we have $$\bm{\nu }=0,\quad \mathbf{h}=-\frac{\epsilon +P}{n}\lambda \lbrack -\bm{\nabla }(\beta \mu )+\beta Q\mathbf{E}].$$ In this frame, the direction of heat conduction is correct (while the physics of particle diffusion and electric conduction become less transparent). As expected, $\mathbf{h}$ stays finite when $\mu =Q=0$ but $\bm{\nabla }\beta \neq 0$. Multi-flavor case ----------------- When the flavor of massless quarks is increased to $N_{f}$, then there are $N_{f}$ conserved currents $j_{a}^{\mu }$ (the conserved electric current is just a combination of them). The hydrodynamical equations becomes $$\begin{aligned} \partial _{\mu }T^{\mu \nu }& =\sum_{a=1}^{N_{f}}Q_{a}F^{\nu \lambda }j_{a,\lambda }, \notag \\ \partial _{\mu }j_{a}^{\mu }& =0,\ a=1,2,\cdots ,N_{f}. \label{eq:conservation_01-1}\end{aligned}$$Then the entropy production yields $$\begin{aligned} \partial _{\mu }S^{\mu } &=&-\sum_{a=1}^{N_{f}}\left( \nu _{a}^{\mu }-\frac{n_{a}}{\epsilon +P}h^{\mu }\right) \left[ \partial _{\mu }(\beta \mu _{a})+\beta Q_{a}E_{\mu }\right] +\beta \pi ^{\mu \nu }\partial _{\left\langle \mu \right. }u_{\left. \nu \right\rangle }-\beta \Pi \partial \cdot u \notag \\ &\geq &0\end{aligned}$$Working in the Landau frame, we have$$\bm{\nu }_{a}=\sum_{b=1}^{N_{f}}\lambda _{ab}[-\bm{\nabla }(\beta \mu _{b})+\beta Q_{b}\mathbf{E}]. \label{d1}$$Our task is to compute the $\lambda $ matrix which can be achieved by setting $\bm{\nabla }(\beta \mu _{b})\neq 0$ but $\mathbf{E}=0$. The second law of thermodynamics dictates $\lambda $ being a positive definite matrix. Effective kinetic theory \[sec:Effective-kinetic-theory\] ========================================================= We will use the Boltzmann equation to compute our LL result of $\lambda $. It has been shown that Boltzmann equation gives the same leading order result as the Kubo formula in the coupling constant expansion in a weakly coupled $\phi ^{4}$ theory [@Jeon:1994if; @Hidaka:2010gh] and in hot QED [@Gagnon:2007qt], provided the leading $T$ and $\mu $ dependence in particle masses and scattering amplitudes are included. This conclusion is expected to hold in perturbative QCD as well [@Arnold:2002zm]. The Boltzmann equation of a quark gluon plasma describes the evolution of the color and spin averaged distribution function $\tilde{f}_{p}^{i}(x)$ for particle $i$ ($i=g,q_{a},\bar{q}_{a}$ with $a=1.2...N_{f}$ for gluon, $N_{f}$ quarks and $N_{f}$ anti-quarks): $$\frac{d\tilde{f}_{p}^{i}(x)}{dt}=\mathcal{\tilde{C}}_{i}, \label{eq:BE_01}$$where $\tilde{f}_{p}^{i}(x)$ is a function of space-time $x^{\mu }=(t,\mathbf{x})$ and momentum $p^{\mu }=(E_{p},\mathbf{p})$. For the LL calculation, we only need to consider two-particle scattering processes denoted as $c_{1}c_{2}\rightarrow c_{3}c_{4}$. The collision term has the form $$C_{c_{1}c_{2}\rightarrow c_{3}c_{4}}\equiv \int_{k_{1}k_{2}k_{3}}d\Gamma _{c_{1}c_{2}\rightarrow c_{3}c_{4}}\left[ \tilde{f}_{k_{1}}^{c_{1}}\tilde{f}_{k_{2}}^{c_{2}}\tilde{F}_{p}^{c_{3}}\tilde{F}_{k_{3}}^{c_{4}}-\tilde{F}_{k_{1}}^{c_{1}}\tilde{F}_{k_{2}}^{c_{2}}\tilde{f}_{p}^{c_{3}}\tilde{f}_{k_{3}}^{c_{4}}\right] . \label{definition of C ab-cd}$$where $\tilde{F}^{g}=1+\tilde{f}^{g}$ and $\tilde{F}^{q(\bar{q})}=1-\tilde{f}^{q(\bar{q})}$ and $$d\Gamma _{c_{1}c_{2}\rightarrow c_{3}c_{4}}=\frac{1}{2E_{p}}\|M_{c_{1}c_{2}\rightarrow c_{3}c_{4}}\|^{2}\prod\limits_{i=1}^{3}\frac{d^{3}k_{i}}{(2\pi )^{3}2E_{k_{i}}}(2\pi )^{4}\delta ^{(4)}(k_{1}+k_{2}-k_{3}-p), \label{gamma ab-cd}$$where $\|M_{c_{1}c_{2}\rightarrow c_{3}c_{4}}\|^{2}$ is the matrix element squared with all colors and helicities of the initial and final states summed over. The scattering amplitudes can be regularized by hard thermal loop propagators and in this paper we use the same scattering amplitudes as in Ref. [@Arnold:2003zc] (see also Table I of Ref. [@Chen:2012jc]). Then the collision term for a quark of flavor $a$ is $$\begin{aligned} N_{q}\mathcal{\tilde{C}}_{q_{a}}& =\frac{1}{2}C_{q_{a}q_{a}\leftrightarrow q_{a}q_{a}}+C_{q_{a}\bar{q}_{a}\leftrightarrow q_{a}\bar{q}_{a}}+\frac{1}{2}C_{gg\leftrightarrow q_{a}\bar{q}_{a}}+C_{q_{a}g\leftrightarrow q_{a}g} \notag \\ & +\sum\limits_{b,b\neq a}^{N_{f}-1}(C_{q_{a}q_{b}\leftrightarrow q_{a}q_{b}}+C_{q_{a}\bar{q}_{b}\leftrightarrow q_{a}\bar{q}_{b}}+C_{q_{b}\bar{q}_{b}\leftrightarrow q_{a}\bar{q}_{a}}),\end{aligned}$$where $N_{q}=2\times 3=6$ is the quark helicity and color degeneracy factor and the factor $1/2$ is included when the initial state is formed by two identical particles. Similarly, $$N_{g}\mathcal{\tilde{C}}_{g}=\frac{1}{2}C_{gg\leftrightarrow gg}+\sum\limits_{a=1}^{N_{f}}(C_{gq_{a}\leftrightarrow gq_{a}}+C_{g\bar{q}_{a}\leftrightarrow g\bar{q}_{a}}+C_{q_{a}\bar{q}_{a}\leftrightarrow gg}), \label{C22g}$$where $N_{g}=2\times 8=16$ is the gluon helicity and color degeneracy factor. In equilibrium, the distributions are denoted as $f^{q_{a}(\bar{q}_{a})}$ and $f^{g}$, with $$\begin{aligned} f_{p}^{g}& =\frac{1}{e^{u\cdot p/T}-1}, \label{distribution g} \\ f_{p}^{q_{a}(\bar{q}_{a})}& =\frac{1}{e^{(u\cdot p\mp \mu _{a})/T}+1}, \label{distribution q}\end{aligned}$$where $T$ is the temperature, $u$ is the fluid four velocity and $\mu _{a}$ is the chemical potential for the quark of flavor $a$. They are all space time dependent. The thermal masses of gluon and quark/anti-quark for external states (the asymptotic masses) can be computed via [Arnold:2002zm,Mrowczynski:2000ed]{} $$\begin{aligned} m_{g}^{2}& =\sum_{i}N_{i}C_{i}\frac{2g^{2}}{d_{A}}\int \frac{d^{3}p}{(2\pi )^{3}2E_{p}}f_{p}^{i}, \label{mg} \\ m_{q}^{2}& =m_{\bar{q}}^{2}=2C_{F}g^{2}\int \frac{d^{3}p}{(2\pi )^{3}2E_{p}}(2f_{p}^{g}+f_{p}^{q}+f_{p}^{\bar{q}}), \label{thermal mass g}\end{aligned}$$where $d_{A}=8$, $C_{g}=C_{A}=3$, and $C_{q(\bar{q})}=C_{F}=4/3$. This yields $$\begin{aligned} m_{g}^{2}& =\frac{C_{A}}{6}g^{2}T^{2}+\sum_{a=1}^{N_{f}}\frac{C_{F}}{16}g^{2}(T^{2}+\frac{3}{\pi ^{2}}\mu _{a}^{2}), \\ m_{q_{a}}^{2}& =\frac{1}{4}C_{F}g^{2}\left( T^{2}+\frac{\mu _{a}^{2}}{\pi ^{2}}\right) , \label{eqi thermal mass g}\end{aligned}$$where we have set $E_{p}=\|\mathbf{p}\|$ in the integrals on the right hand sides of Eqs. (\[mg\]) and (\[thermal mass g\]). The difference from non-vanishing masses is of higher order. In this work, we only need the fact that the thermal masses are proportional to $g^{2}$ for the LL results. Linearized Boltzmann equation ----------------------------- Matching to the derivative expansion in hydrodynamics, we expand the distribution function of particle $i$ as a local equilibrium distribution plus a correction $$\tilde{f}_{p}^{i}(x)=f_{p}^{i}-\varepsilon f_{p}^{i}(1\mp f_{p}^{i})\chi ^{i}, \label{eq:expansion_01}$$where the upper/lower sign corresponds to the femion/boson distribution. Inserting Eq. (\[eq:expansion\_01\]) into Eq. (\[eq:BE\_01\]), we can solve the linearized Boltzmann equation by keeping linear terms in space-time derivatives. Here we neglect the viscous terms related to $\partial _{\mu }u_{\nu }$ in $\chi ^{i}$ and consider only the $\bm{\nabla }(\beta \mu _{a})$ terms. At the zeroth order, $\mathcal{O}(\varepsilon ^{0})$, the system is in local thermal equilibrium and the Boltzmann equation (\[eq:BE\_01\]) is satisfied, $\mathcal{\tilde{C}}[f_{p}^{i}]=0$. At $\mathcal{O}(\varepsilon )$, the left hand side of the Boltzmann equation yields $$\frac{df_{p}^{g}}{dt}=-\beta f_{p}^{g}F_{p}^{g}\sum_{a=1}^{N_{f}}\left[ \frac{n_{a}T}{\epsilon +P}\mathbf{p}\cdot \bm{\nabla }(\beta \mu _{a})\right] , \label{b1}$$and $$\frac{df_{p}^{q_{a}(\bar{q}_{a})}}{dt}=-\beta f_{p}^{q_{a}(\bar{q}_{a})}F_{p}^{q_{a}(\bar{q}_{a})}\sum_{b=1}^{N_{f}}\left( \frac{n_{b}T}{\epsilon +P}\mp \frac{T}{E_{p}^{q_{a}(\bar{q}_{a})}}\delta _{ab}\right) \mathbf{p}\cdot \bm{\nabla }(\beta \mu _{b}). \label{b2}$$To derive this result, we have used $\partial _{\mu }u^{0}=0$ in the local fluid rest frame where $u^{\mu }=(1,0,0,0)$ and $\partial _{\mu }T_{(0)}^{\mu \nu }=0$ and $\partial _{\mu }j_{a(0)}^{\mu }=0$ which yields$$\begin{aligned} \frac{\partial \epsilon }{\partial t}& =-\left( \epsilon +p\right) \bm{\nabla }\cdot \mathbf{u} \label{energy density time derivation} \\ \frac{\partial \mathbf{u}}{\partial t}& =-\frac{\bm{\nabla }P}{\epsilon +p} \notag\end{aligned}$$and$$\frac{\partial n_{a}}{\partial t}=-n_{a}\bm{\nabla }\cdot \mathbf{u}. \label{quark number density time derivation}$$And then by applying thermodynamic relations, we can replace the time derivatives of $T$ , $\mu $ and $\mathbf{u}$ with spatial derivatives: $$\begin{aligned} \frac{\partial T}{\partial t}& =-T\left( \frac{\partial P}{\partial \epsilon }\right) _{n}\bm{\nabla }\mathbf{\cdot u}, \notag \\ \frac{\partial \mu }{\partial t}& =-\left[ \mu \left( \frac{\partial P}{\partial \epsilon }\right) _{n}+\left( \frac{\partial P}{\partial n}\right) _{\epsilon }\right] \bm{\nabla }\mathbf{\cdot u}, \label{dT/dt} \\ \frac{\partial \mathbf{u}}{\partial t}& =-\beta \bm{\nabla }T-\sum_{a=1}^{N_{f}}\frac{n_{a}T}{\epsilon +p}\bm{\nabla }\left( \frac{\mu _{a}}{T}\right) . \notag\end{aligned}$$Those relations lead to Eqs.(\[b1\],\[b2\]). To get the right hand side of the Boltzmann equation at $\mathcal{O}(\varepsilon )$, we parametrize $\chi ^{i}$ of Eq. (\[eq:expansion\_01\]) as $$\begin{aligned} \chi ^{i}& =\beta \sum_{a=1}^{N_{f}}\mathbf{A}^{ia}(p)\cdot \bm{\nabla }(\beta \mu _{a}), \notag \\ \mathbf{A}^{ia}(p)& =A^{ia}(\|\mathbf{p}\|)\hat{\mathbf{p}}. \label{c1}\end{aligned}$$The matrix $A^{ia}$ is $\left( 2N_{f}+1\right) \times N_{f}$. We will see there are $\left( 2N_{f}+1\right) \times N_{f}$ equations to constrain them. For each Boltzmann equation, we have a linear combination of $N_{f}$ terms of $\bm{\nabla }(\beta \mu _{a})$. Since each $\bm{\nabla }(\beta \mu _{a})$ is linearly independent to each other, thus there are $N_{f}$ equations for each Boltzmann equation. Totally we have $2N_{f}+1$ Boltzmann equations, thus we have $\left( 2N_{f}+1\right) \times N_{f}$ equations to solve for $A^{ia}$. These equations are $$\frac{n_{a}\mathbf{p}}{\epsilon +P}=\beta \frac{1}{f_{p}^{g}F_{p}^{g}}\frac{1}{N_{g}}\left[ \frac{1}{2}\mathbf{C}_{gg\text{ }\rightarrow gg}^{a}+\sum\limits_{c=1}^{N_{f}}\left( \mathbf{C}_{gq_{c}\text{ }\rightarrow gq_{c}}^{a}+\mathbf{C}_{g\bar{q}_{c}\rightarrow g\bar{q}_{c}}^{a}+\mathbf{C}_{q_{c}\bar{q}_{c}\rightarrow gg}^{a}\right) \right] , \label{eq:constr-g}$$$$\begin{aligned} \left( \frac{n_{a}}{\epsilon +P}-\frac{1}{E_{p}^{q_{b}}}\delta _{ab}\right) \mathbf{p}& =\frac{\beta }{f_{p}^{q_{b}}F_{p}^{q_{b}}}\frac{1}{N_{q}}\left[ \frac{1}{2}\mathbf{C}_{q_{b}q_{b}\rightarrow q_{b}q_{b}}^{a}+\mathbf{C}_{q_{b}\bar{q}_{b}\rightarrow q_{b}\bar{q}_{b}}^{a}+\frac{1}{2}\mathbf{C}_{g\text{ }g\text{ }\rightarrow q_{b}\bar{q}_{b}}^{a}+\mathbf{C}_{q_{b}g\rightarrow q_{b}g}^{a}\right. \notag \\ & +\left. \sum\limits_{c,c\neq b}^{N_{f}-1}\left( \mathbf{C}_{q_{b}q_{c}\rightarrow q_{b}q_{c}}^{a}+\mathbf{C}_{q_{b}\bar{q}_{c}\rightarrow q_{b}\bar{q}_{c}}^{a}+\mathbf{C}_{q_{c}\bar{q}_{c}\rightarrow q_{b}\bar{q}_{b}}^{a}\right) \right] , \label{eq:constr-q}\end{aligned}$$and $$\begin{aligned} \left( \frac{n_{a}}{\epsilon +P}+\frac{1}{E_{p}^{\bar{q}_{b}}}\delta _{ab}\right) \mathbf{p}& =\frac{\beta }{f_{p}^{\bar{q}_{b}}F_{p}^{\bar{q}_{b}}}\frac{1}{N_{q}}\left[ \frac{1}{2}\mathbf{C}_{\bar{q}_{b}\bar{q}_{b}\rightarrow \bar{q}_{b}\bar{q}_{b}}^{a}+\mathbf{C}_{\bar{q}_{b}q_{b}\rightarrow \bar{q}_{b}q_{b}}^{a}+\frac{1}{2}\mathbf{C}_{gg\text{ }\rightarrow \bar{q}_{b}q_{b}}^{a}+\mathbf{C}_{\bar{q}_{b}g\rightarrow \bar{q}_{b}g}^{a}\right. \notag \\ & +\left. \sum\limits_{c,c\neq b}^{N_{f}-1}\left( \mathbf{C}_{\bar{q}_{b}\bar{q}_{c}\rightarrow \bar{q}_{b}\bar{q}_{c}}^{a}+\mathbf{C}_{\bar{q}_{b}q_{c}\rightarrow \bar{q}_{b}q_{c}}^{a}+\mathbf{C}_{\bar{q}_{c}q_{c}\rightarrow \bar{q}_{b}q_{b}}^{a}\right) \right] , \label{eq:constr-qb}\end{aligned}$$where $$\begin{aligned} \mathbf{C}_{c_{1}c_{2}\rightarrow c_{3}c_{4}}^{a}(p)& \equiv \int_{k_{1}k_{2}k_{3}}d\Gamma _{c_{1}c_{2}\rightarrow c_{3}c_{4}}f^{c_{1}}f^{c_{2}}F^{c_{3}}F^{c_{4}} \notag \\ & \times \left[ \mathbf{A}^{c_{1}a}(k_{1})+\mathbf{A}^{c_{2}a}(k_{2})-\mathbf{A}^{c_{3}a}(k_{3})-\mathbf{A}^{c_{4}a}(p)\right] .\end{aligned}$$Formally we can rewrite these linearized Boltzmann equations in a compact form, $$\left\vert \mathbf{S}^{a}\right\rangle =\mathcal{C}_{\lambda }\left\vert \mathbf{A}^{a}\right\rangle , \label{linearized equation}$$where $\left\vert \mathbf{S}^{a}\right\rangle $ and $\left\vert \mathbf{A}^{a}\right\rangle $ are both vectors of $\left( 2N_{f}+1\right) $ components and $\mathcal{C}_{\lambda }$ is a $\left( 2N_{f}+1\right) \times \left( 2N_{f}+1\right) $ matrix. Conductivity matrix ------------------- In the kinetic theory, the quark current of flavor $a$ is $$j_{a}^{\mu }=N_{q}\int \frac{d^{3}p}{\left( 2\pi \right) ^{3}}\frac{p^{\mu }}{E^{q_{a}}}\left( \tilde{f}_{p}^{q_{a}}-\tilde{f}_{p}^{\bar{q}_{a}}\right) .$$Expanding this expression to $\mathcal{O}(\varepsilon )$ and matching it to Eq.(\[d1\]), we have$$\lambda _{ab}=\frac{N_{q}\beta }{3}\int \frac{d^{3}p}{(2\pi )^{3}}\frac{1}{E^{q_{a}}}\left( f^{q_{a}}F^{q_{a}}\mathbf{p\cdot }\mathbf{A}^{q_{a}b}-f^{\bar{q}_{a}}F^{\bar{q}_{a}}\mathbf{p\cdot }\mathbf{A}^{\bar{q}_{a}b}\right) . \label{e1}$$Since we are working in the Landau frame, we should impose the Landau-Lifshitz condition $$0=T^{0j}=-\sum_{i}N_{i}\int \frac{d^{3}p}{\left( 2\pi \right) ^{3}}f^{i}F^{i}\chi ^{i}p^{j}.$$This implies $$\sum_{i}N_{i}\int \frac{d^{3}p}{\left( 2\pi \right) ^{3}}f^{i}F^{i}\mathbf{p\cdot A}^{ia}=0.$$We can use these constraints to rewrite Eq.(\[e1\]) as$$\begin{aligned} \lambda _{ab}& =-\frac{\beta }{3}\left\{ N_{g}\int \frac{d^{3}p}{(2\pi )^{3}}f^{g}F^{g}\frac{n_{a}}{\epsilon +P}\mathbf{p\cdot A}^{gb}\right. \notag \\ & +\sum_{c=1}^{N_{f}}N_{q}\int \frac{d^{3}p}{(2\pi )^{3}}f^{q_{c}}F^{q_{c}}\left( \frac{n_{a}}{\epsilon +P}-\frac{1}{E^{q_{a}}}\delta _{ca}\right) \mathbf{p}\cdot \mathbf{A}^{q_{c}b} \notag \\ & \left. +\sum_{c=1}^{N_{f}}N_{q}\int \frac{d^{3}p}{(2\pi )^{3}}f^{\bar{q}_{c}}F^{\bar{q}_{c}}\left( \frac{n_{a}}{\epsilon +P}+\frac{1}{E^{q_{a}}}\delta _{ca}\right) \mathbf{p}\cdot \mathbf{A}{}^{\bar{q}_{c}b}\right\} . \label{01}\end{aligned}$$This form can be schematically written as $$\lambda _{ab}=\langle \mathbf{A}^{b}\|\mathbf{S}^{a}\rangle =\langle \mathbf{A}^{b}\|\mathcal{C}_{\lambda }\left\vert \mathbf{A}^{a}\right\rangle , \label{f3}$$where we have used Eq.(\[linearized equation\]) for the second equality. More explicitly, $$\begin{aligned} \lambda _{ab}& =\frac{\beta ^{2}}{24}\left( D_{gg\rightarrow gg}^{ab}+\sum_{c=1}^{N_{f}}D_{q_{c}q_{c}\rightarrow q_{c}q_{c}}^{ab}+\sum_{c=1}^{N_{f}}D_{\bar{q}_{c}\bar{q}_{c}\rightarrow \bar{q}_{c}\bar{q}_{c}}^{ab}\right) \notag \\ & +\frac{\beta ^{2}}{6}\sum_{c=1}^{N_{f}}\left( D_{q_{c}\bar{q}_{c}\rightarrow gg}^{ab}+D_{gq_{c}\rightarrow gq_{c}}^{ab}+D_{g\bar{q}_{c}\rightarrow g\bar{q}_{c}}^{ab}+D_{q_{c}\bar{q}_{c}\rightarrow q_{c}\bar{q}_{c}}^{ab}\right) \notag \\ & +\frac{\ \beta ^{2}}{12}\sum_{\substack{ c,d=1 \\ c\neq d}}^{N_{f}}\left( D_{q_{c}q_{d}\rightarrow q_{c}q_{d}}^{ab}+D_{\bar{q}_{c}\bar{q}_{d}\rightarrow \bar{q}_{c}\bar{q}_{d}}^{ab}+2D_{q_{c}\bar{q}_{d}\rightarrow q_{c}\bar{q}_{d}}^{ab}+2D_{q_{c}\bar{q}_{c}\rightarrow q_{d}\bar{q}_{d}}^{ab}\right) , \label{f1}\end{aligned}$$where $$\begin{aligned} D_{c_{1}c_{2}\rightarrow c_{3}c_{4}}^{ab}\equiv & \int \prod_{i=1}^{4}\frac{d^{3}p_{i}}{(2\pi )^{3}2E_{i}}(2\pi )^{4}\delta ^{4}(k_{1}+k_{2}-k_{3}-k_{4}) \notag \\ & \times \|M_{c_{1}c_{2}\rightarrow c_{3}c_{4}}\|^{2}f_{k_{1}}^{c_{1}}f_{k_{2}}^{c_{2}}F_{k_{3}}^{c_{3}}F_{k_{4}}^{c_{4}} \notag \\ & \times \left[ \mathbf{A}^{c_{1}a}(k_{1})+\mathbf{A}^{c_{2}a}(k_{2})-\mathbf{A}^{c_{3}a}(k_{3})-\mathbf{A}^{c_{4}a}(k_{4})\right] \notag \\ & \cdot \left[ \mathbf{A}^{c_{1}b}(k_{1})+\mathbf{A}^{c_{2}b}(k_{2})-\mathbf{A}^{c_{3}b}(k_{3})-\mathbf{A}^{c_{4}b}(k_{4})\right] . \label{f2}\end{aligned}$$ From Eq.(\[linearized equation\]), it is clear that if $$\mathbf{A}_{0}^{ia}(p)=\mathbf{p}, \label{0}$$then from momentum conservation this implies $$\mathcal{C}_{\lambda }\left\vert \mathbf{A}_{0}^{a}\right\rangle =0.$$Those modes are called zero modes (denoted by the subscribe $0$ in Eq.([0]{})). They would have been a problem for Eq.(\[f3\]) unless $\langle \mathbf{S}^{a}\|\mathbf{A}_{0}^{a}\rangle =0$, but this is guaranteed from the total momentum conservation at $\mathcal{O}(\varepsilon )$,$$\frac{d}{dt}\sum_{i}\int dp^{3}\mathbf{p}f_{p}^{i}(x)=0,$$and Eqs.(\[b1\],\[b2\]). Thus, we can just solve for $\left\vert \mathbf{A}^{a}\right\rangle $ in Eq.(\[f3\]) by discarding the zero modes. From Eqs.(\[f1\]) and (\[f2\]), we can see easily that $\lambda _{ab}=\lambda _{ba}$. This is a manifestation of the Onsager relation which appears when particle scattering is symmetric under the time-reversal transformation. We can also see that $\lambda $ is positive definite. The Leading-Log Results with Identical Chemical Potentials[sec:Numerical-results]{} =================================================================================== Now we are ready to solve the conductivity matrix $\lambda $. Our strategy to solve for $\lambda _{ab}$ is to make use of Eq.(\[f3\]) to solve for $\|\mathbf{A}^{a}\rangle $ from $\lambda _{aa}$ (no summation over $a$). Once all the $\|\mathbf{A}^{a}\rangle $ are obtained, $\lambda _{ab}$ can be computed. Also, in solving for $\lambda _{aa}$, one can use the standard algorithm to systematically approach the answer from below [Chen:2011km]{}. The dependence on the strong coupling constant is similar to that in shear viscosity—it is inversely proportional to the scattering rate which scales as $g^{4}\ln g^{-1}$ with the $\ln g^{-1}$ dependence coming from regularizing the collinear infrared singularity by the thermal masses of quarks or gluons. $\lambda $ is of mass dimension two, thus we will present our result in the normalized conductivity $$\tilde{\lambda}\equiv \lambda \beta ^{2}g^{4}\ln g^{-1}$$such that $\tilde{\lambda}$ is dimensionless and coupling constant independent. For simplicity, we will concentrate on the linear response of a thermal equilibrium system with all fermion chemical potentials to be identical, i.e. $\mu _{a}=\mu $ for all $a$’s but each $\bm{\nabla }\left( \beta \mu _{a}\right) $ could be varied independently. This symmetry makes all the diagonal matrix elements (denoted as $\lambda _{qq}$) identical and all the off-diagonal ones (denoted as $\lambda _{qq^{\prime }}$) identical. $\lambda _{qq}$ and $\lambda _{qq^{\prime }}$ are even in $\mu $ (and so are $\tilde{\lambda}_{qq}$ and $\tilde{\lambda}_{qq^{\prime }}$) because our microscopic interaction (in vacuum) is invariant under charge conjugation, thus $\lambda $ should be invariant under $\mu _{a}\rightarrow -\mu _{a}$. It is easy to diagonalize $\lambda $. One eigenvalue is$$\lambda _{+}/N_{f}\equiv \lambda _{qq}+(N_{f}-1)\lambda _{qq^{\prime }}, \label{eq:lambda_+}$$corresponding to the conductivity of the flavor singlet total quark current ($\lambda _{+}$ is the total quark current conductivity)$$\bm{\nu }=\sum_{a=1}^{N_{f}}\bm{\nu }_{a}=-\lambda _{+}\sum_{a=1}^{N_{f}}\frac{\bm{\nabla }(\beta \mu _{a})}{N_{f}}.$$The other $(N_{f}-1)$ eigenvalues are degenerate with the value $$\lambda _{-}\equiv \lambda _{qq}-\lambda _{qq^{\prime }}. \label{eq:lambda_-}$$They are the conductivities of the flavor non-singlet currents $$\Delta \bm{\nu }_{a}\equiv \bm{\nu }_{a}-\bm{\nu }_{1}=-\lambda _{-}\bm{\nabla }[\beta (\mu _{a}-\mu _{1})],$$ with $a=2,...,N_{f}$. ![(color online). The normalized diagonal conductivity $\tilde{\protect\lambda}_{qq}$ (upper panel) and off-diagonal conductivity $\tilde{\protect\lambda}_{qq^{\prime }}$ (lower panel) as functions of $(\protect\mu /T)^{2}$ for different $N_{f}$. []{data-label="fig. 1"}](conductivity-LL-equal-mu-lambda-aa.eps) ![(color online). The normalized diagonal conductivity $\tilde{\protect\lambda}_{qq}$ (upper panel) and off-diagonal conductivity $\tilde{\protect\lambda}_{qq^{\prime }}$ (lower panel) as functions of $(\protect\mu /T)^{2}$ for different $N_{f}$. []{data-label="fig. 1"}](conductivity-LL-equal-mu-lambda-ab.eps) ![(color online). The normalized conductivities $\tilde{\protect\lambda}_{+}$ (upper panel) and $\tilde{\protect\lambda}_{-}$ (lower panel) as functions of $(\protect\mu /T)^{2}$ for different $N_{f}$ . []{data-label="fig. 2"}](conductivity-LL-equal-mu-lambda-plus.eps) ![(color online). The normalized conductivities $\tilde{\protect\lambda}_{+}$ (upper panel) and $\tilde{\protect\lambda}_{-}$ (lower panel) as functions of $(\protect\mu /T)^{2}$ for different $N_{f}$ . []{data-label="fig. 2"}](conductivity-LL-equal-mu-lambda-minus.eps) $\tilde{\lambda}_{qq}$ and $\tilde{\lambda}_{qq^{\prime }}$ are shown as functions of $(\mu /T)^{2}$ in Fig. \[fig. 1\] for various $N_{f}$ with $N_{f}\leq 16$ such that the system is asymptotically free, while $\tilde{\lambda}_{+}$ and $\tilde{\lambda}_{-}$ are shown in Fig. \[fig. 2\] (note that there is no $\tilde{\lambda}_{qq^{\prime }}$ or $\tilde{\lambda}_{-}$ for $N_{f}=1$). The fact that the matrix $\lambda $ is positive definite makes $\tilde{\lambda}_{qq}$, $\tilde{\lambda}_{+}$ and $\tilde{\lambda}_{-}$ positive, but it imposes no constraint on the sign of $\tilde{\lambda}_{qq^{\prime }}$. When $(\mu /T)^{2}\rightarrow 0$, we can expand $\tilde{\lambda}_{qq}=a_{0}+a_{1}(\mu /T)^{2}+...$, and $\tilde{\lambda}_{qq^{\prime }}=a_{0}^{\prime }+a_{1}^{\prime }(\mu /T)^{2}+...$. We find $a_{0}^{\prime }=0$ for all $N_{f}$ while the values of $a_{0},a_{1}$ and $a_{1}^{\prime }$ for different $N_{f}$ are tabulated in Table. \[power expansion coefficients lambdaqq\]. Our result for $a_{0}$ agrees within $0.1\%$ to that of Arnold, Moore and Yaffe (AMY) calculated up to $N_{f}=6$ listed in Table III of Ref. [@Arnold:2000dr]. The $a_{0}^{\prime }=0$ property is due to a bigger symmetry enjoyed by the LL results: if we just change all the quarks of flavor $a$ into anti-quarks while the rest of the system stays the same, then as far as collision is concerned, the other quarks and the gluons will not feel any difference. This is because the LL result only depends on two-particle scattering, and although this action could change the sign of certain amplitudes, it does not change the collision rate. For example, the amplitudes of $q_{a}q_{b}\rightarrow $ $q_{a}q_{b}$ and $\bar{q}_{a}q_{b}\rightarrow $ $\bar{q}_{a}q_{b}$ ($a\neq b$) have different signs because one of the couplings changes sign when we change the color into its anti-color, but the amplitude squared is of the same. This makes the diagonal terms even in all the chemical potentials $$\lambda _{aa}\left( \pm \mu _{1},\pm \mu _{2},\cdots ,\pm \mu _{N_{f}}\right) =\lambda _{aa}\left( \mu _{1},\mu _{2},\cdots ,\mu _{N_{f}}\right) ,$$while the off-diagonal term $\lambda _{ab}$ is odd in $\mu _{a}$ and $\mu _{b}$ but even in other chemical potentials $$\begin{aligned} & \lambda _{ab}\left( \pm \mu _{1},\cdots ,\pm \mu _{a},\cdots ,\pm \mu _{b},\cdots ,\pm \mu _{N_{f}}\right) \notag \\ & =\text{sign}\left( \mu _{a}\mu _{b}\right) \lambda _{ab}\left( \mu _{1},\cdots ,\mp \mu _{a},\cdots ,\mp \mu _{b},\cdots ,\mu _{N_{f}}\right) . \label{lab}\end{aligned}$$Thus, at the LL order, $\lambda $ becomes diagonal when all the chemical potentials vanish. To understand the other features of $\tilde{\lambda}_{qq}$ and $\tilde{\lambda}_{qq^{\prime }}$, we first turn to $\tilde{\lambda}_{+}$ and $\tilde{\lambda}_{-}$ in the $(T/\mu )^{2}\rightarrow 0$ limit. In this large chemical potential limit, the quark contribution dominates over those of anti-quark and gluon. The Fermi-Dirac distribution function $f^{q_{a}}$ of quark $q_{a}$ multiplied by its Pauli blocking factor $F^{q_{a}}$ can be well approximated by a $\delta -$function, $f_{{}}^{q_{a}}F_{{}}^{q_{a}}\simeq T\delta (E_{p}-\mu )$. We then first set $\bm{\nabla }(\beta \mu _{a})=\bm{\nabla }(\beta \mu )$ for all $a$ so all the currents $\bm{\nu }_{a}$ becomes identical. $\lambda _{+}$ can be rewritten as $\sum_{a,b}\lambda _{ab}$, and Eq. (\[01\]) yields$$\lambda _{+}\simeq -\frac{\beta }{3}N_{q}\int \frac{d^{3}p}{(2\pi )^{3}}T\delta (E_{p}-\mu )\mathbf{p}\cdot \mathbf{A}\sum_{a,b,c=1}^{N_{f}}\left( \frac{n_{b}}{\epsilon +P}-\frac{1}{\mu }\delta _{ca}\right) .$$The summation gives $N_{f}^{2}\left( \frac{n\mu }{\epsilon +P}-1\right) \propto N_{f}^{2}Ts/(\epsilon +P)\propto N_{f}^{2}T^{2}/\mu ^{2}$ and $\lambda _{+}\propto N_{f}^{2}T^{2}A$. On the other hand, Eq.(\[f1\]) gives $\lambda _{+}\propto N_{f}^{4}\mu ^{2}A^{2}$ where $N_{f}^{4}$ comes from summing the $a$, $b$, $c$, $d$ indices of $D_{q_{c}q_{d}\rightarrow q_{c}q_{d}}^{ab}$ and we have used $f_{k_{1}}^{c_{1}}f_{k_{2}}^{c_{2}}F_{k_{3}}^{c_{3}}F_{k_{4}}^{c_{4}}\propto T^{2}$ in Eq.(\[f2\]). These two conditions yield $\lambda _{+}\propto N_{f}^{0}T^{4}/\mu ^{2}$. This is indeed what happens in Fig. \[fig. 2\] at large $\mu $ (although the $1/\mu ^{2}$ dependence is not so obvious in this plot but we have checked this at much larger $\mu ^{2}/T^{2}$). We can perform the similar counting to the scaling of $\lambda _{-}$. From Eq. (\[01\]), $\lambda _{-}\propto N_{f}\mu ^{2}A$ and from Eq.(\[f1\]) $\lambda _{-}\propto N_{f}^{3}\mu ^{2}A^{2}$. Thus, $\lambda _{-}\propto \mu ^{2}/N_{f}$ which is also observed in Fig. \[fig. 2\]. The main difference in $\lambda _{+}/N_{f}$ and $\lambda _{-}$ is the $T/\mu $ dependence—$\lambda _{-}$ has no cancellation factor of $\left( \frac{n\mu }{\epsilon +P}-1\right) \propto T^{2}$ in large $\mu $. The different $\mu $ scaling between $\lambda _{+}$ and $\lambda _{-}$ at large $\mu $ is due to collisions, which change the direction of the current and reduce the conductivity. While both flavor singlet and non-singlet fermions can collide among themselves, they do not collide with each other (the scattering amplitude vanishes). Thus, when $\mu $, the flavor singlet chemical potential, is increased, the flavor singlet current experiences more collisions. Therefore the flavor singlet conductivity $\lambda _{+}$ is reduced. For the flavor non-singlet current, the increase of $\mu $ does not affect the collision. However, it will increase the averaged fermi momentum such that the induced current and the flavor non-singlet conductivity $\lambda _{-}$ will be increased. Given the large $\mu $ behavior of $\tilde{\lambda}_{+}$ and $\tilde{\lambda}_{-}$, the large $\mu $ behavior of $\tilde{\lambda}_{qq}$ and $\tilde{\lambda}_{qq^{\prime }}$ is now easily reconstructed: $\tilde{\lambda}_{qq}\simeq (N_{f}-1)\tilde{\lambda}_{-}/N_{f}\propto (N_{f}-1)\mu ^{2}/T^{2}N_{f}^{2}$ ($N_{f}\geq 2$) and $\tilde{\lambda}_{qq^{\prime }}\simeq -\tilde{\lambda}_{-}/N_{f}\propto -\mu ^{2}/T^{2}N_{f}^{2}$. The sign of $\tilde{\lambda}_{qq^{\prime }}$ can be best understood from the flavor non-singlet current effect such that a gradient of $\mu _{a}$ induces anti-$b$ currents ($b\neq a$) and yields $\tilde{\lambda}_{qq^{\prime }}<0$. We can then interpolate $\tilde{\lambda}_{qq^{\prime }}$ to $\tilde{\lambda}_{qq^{\prime }}=0$ at zero $\mu $. There is no non-trivial structure at intermediate $\mu $. For $\tilde{\lambda}_{qq}$, the $N_{f}=1$ curve seems to be at odd with other $N_{f}$ curves, but this anomaly disappears when viewed in the $\tilde{\lambda}_{+}$ plot. The fact that $\lambda _{qq}>0$ while $\lambda _{qq^{\prime }}<0$ at finite $\mu $ is intriguing. It means a gradient $\bm{\nabla }\mu _{a}$ can drive a current $\bm{\nu }_{a}$ along the $-\bm{\nabla }\mu _{a}$ direction, but it will also drive currents of different flavors in the opposite direction. This backward current phenomenon seems counter intuitive at the first sight. But the physics behind is just that the flavor singlet current experiences more collisions in a flavor singlet medium than the flavor non-singlet ones. If the medium is flavor non-singlet, e.g. $\mu _{1}=-\mu _{2}\neq 0$ while the other chemical potentials all vanish, then the flavor non-singlet current $\bm{\nu }_{2}-\bm{\nu }_{1}$ will experience more collisions than the flavor singlet current. Therefore, we will have $\lambda _{12}>0$. This is consistent with Eq.(\[lab\]) derived from the symmetry of the LL order along. Thus the simple explanation based on collisions that we presented above seems quite generic. It might happen in other systems such as cold atoms as well. In that case, cold atom experiments might be the most promising ones to observe this backward current phenomenon. $\begin{tabular}{\|c\|r\|r\|c\|\|c\|c\|c\|c\|} \hline $N\_[f]{}$ & $a\_[0]{}$\;\;\ & $a\_[1]{}$\;\quad\ & $a\_[1]{}\^$ & $N\_[f]{}$ & $a\_[0]{}$ & $a\_[1]{}$ & $a\_[1]{}\^$ \\ \hline 1 & 14.3676 & -0.3077 & - & 9 & 7.8019 & 2.1076 & -0.7572 \\ \hline 2 & 12.9989 & 1.7347 & -5.0372 & 10 & 7.3806 & 1.9880 & -0.6404 \\ \hline 3 & 11.8688 & 2.3969 & -3.3569 & 11 & 7.0025 & 1.8766 & -0.5487 \\ \hline 4 & 10.9197 & 2.5757 & -2.3922 & 12 & 6.6612 & 1.7731 & -0.4754 \\ \hline 5 & 10.1113 & 2.5680 & -1.7906 & 13 & 6.3517 & 1.6791 & -0.4159 \\ \hline 6 & 9.4145 & 2.4791 & -1.3909 & 14 & 6.0697 & 1.5917 & -0.3668 \\ \hline 7 & 8.8076 & 2.3600 & -1.1117 & 15 & 5.8117 & 1.5121 & -0.3260 \\ \hline 8 & 8.2743 & 2.2319 & -0.9090 & 16 & 5.5747 & 1.4384 & -0.2916 \\ \hline \end{tabular} \ $ Summary \[sec:Conclusion-and-discussion\] ========================================= We have calculated the conductivity matrix of a weakly coupled quark-gluon plasma at the leading-log order. By setting all quark chemical potentials to be identical, the diagonal conductivities become degenerate and positive, while the off-diagonal ones become degenerate but negative (or zero when the chemical potential vanishes). This means a potential gradient of a certain fermion flavor can drive backward currents of other flavors. A simple explanation is provided for this seemingly counter intuitive phenomenon. It is speculated that this phenomenon is generic and most easily measured in cold atom experiments. Acknowledgement: SP thanks Tomoi Koide and Xu-guang Huang for helpful discussions on the Onsager relation. JWC thanks Jan M. Pawlowski for useful discussions and the U. of Heidelberg for hospitality. JWC, YFL and SP are supported by the CTS and CASTS of NTU and the NSC (102-2112-M-002-013-MY3) of ROC. YKS is supported in part by the CCNU-QLPL Innovation Fund under grant No. QLPL2011P01. This work is also supported by the National Natural Science Foundation of China under grant No. 11125524 and 11205150, and in part by the China Postdoctoral Science Foundation under the grant No. 2011M501046. [99]{} I. Arsene et al. \[BRAHMS Collaboration\], Nucl. Phys. A 757, 1 (2005) \[nucl-ex/0410020\]. K. Adcox et al. \[PHENIX Collaboration\], Nucl. Phys. A 757, 184 (2005) \[nucl-ex/0410003\]. B. B. Back, M. D. Baker, M. Ballintijn, D. S. Barton, B. Becker, R. R. Betts, A. A. Bickley and R. Bindel et al., Nucl. Phys. A 757, 28 (2005) \[nucl-ex/0410022\]. J. Adams et al. \[STAR Collaboration\], Nucl. Phys. A 757, 102 (2005) \[nucl-ex/0501009\]. H. Song, S. A. Bass, U. Heinz, T. Hirano and C. Shen, Phys. Rev. Lett. 106, 192301 (2011) \[Erratum-ibid. 109, 139904 (2012)\]. P. Kovtun, D. T. Son and A. O. Starinets, Phys.Rev. Lett. 94, 111601 (2005). J. M. Maldacena, Adv. Theor. Math. Phys. 2, 231 (1998) \[Int. J. Theor. Phys. 38, 1113 (1999)\]. S. S. Gubser, I. R. Klebanov and A. M. Polyakov, Phys. Lett. B 428, 105 (1998) E. Witten, Adv. Theor. Math. Phys. 2, 253 (1998) P. B. Arnold, G. D. Moore and L. G. Yaffe, JHEP 0011, 001 (2000) \[hep-ph/0010177\]. P. B. Arnold, G. D. Moore and L. G. Yaffe, JHEP 0305, 051 (2003) \[hep-ph/0302165\]. J.-W. Chen, J. Deng, H. Dong and Q. Wang, Phys. Rev. D 83, 034031 (2011) \[Erratum-ibid. D 84, 039902 (2011)\] \[arXiv:1011.4123 \[hep-ph\]\]. J. -W. Chen, J. Deng, H. Dong and Q. Wang, Phys. Rev. C 87, 024910 (2013) \[arXiv:1107.0522 \[hep-ph\]\]. L. P. Csernai, J. I. Kapusta and L. D. McLerran, Phys. Rev. Lett. 97, 152303 (2006). J.-W. Chen and E. Nakano, Phys. Lett. B 647, 371 (2007) \[hep-ph/0604138\]. J.-W. Chen, Y. -F. Liu, Y.-K. Song and Q. Wang, Phys. Rev. D 87, 036002 (2013) \[arXiv:1212.5308 \[hep-ph\]\]. J.-W. Chen, Y. -H. Li, Y. -F. Liu and E. Nakano, Phys. Rev. D 76, 114011 (2007) \[hep-ph/0703230\]. D. Kharzeev, K. Tuchin, JHEP 0809, 093 (2008). F. Karsch, D. Kharzeev, K. Tuchin, Phys. Lett. B663, 217-221 H. B. Meyer, JHEP 1004, 099 (2010). P. B. Arnold, C. Dogan and G. D. Moore, Phys. Rev. D 74, 085021 (2006) \[hep-ph/0608012\]. J.-W. Chen and J. Wang, Phys. Rev. C 79, 044913 (2009) \[arXiv:0711.4824 \[hep-ph\]\]. D. Fernandez-Fraile and A. Gomez Nicola, Phys. Rev. Lett. 102, 121601 (2009). E. Lu, G. D. Moore, Phys. Rev. C83, 044901 (2011). A. Dobado, F. J. Llanes-Estrada and J. M. Torres-Rincon, Phys. Lett. B 702, 43 (2011). P. Chakraborty and J. I. Kapusta, Phys. Rev. C 83, 014906 (2011). H. Dong, N. Su and Q. Wang, Phys. Rev. D 75, 074016 (2007) \[astro-ph/0702104\]. M. G. Alford and A. Schmitt, J. Phys. G 34, 67 (2007) \[nucl-th/0608019\]. M. G. Alford, M. Braby and A. Schmitt, J. Phys. G 35, 115007 (2008) \[arXiv:0806.0285 \[nucl-th\]\]. B. A. Sa’d, I. A. Shovkovy and D. H. Rischke, Phys. Rev. D 75, 065016 (2007) \[astro-ph/0607643\]. B. A. Sa’d, I. A. Shovkovy and D. H. Rischke, Phys. Rev. D 75, 125004 (2007) \[astro-ph/0703016\]. X. Wang and I. A. Shovkovy, Phys. Rev. D 82, 085007 (2010) \[arXiv:1006.1293 \[hep-ph\]\]. X. -G. Huang and J. Liao, Phys. Rev. Lett. 110, 232302 (2013) \[arXiv:1303.7192 \[nucl-th\]\]. L. McLerran and V. Skokov, arXiv:1305.0774 \[hep-ph\]. H. -T. Ding, A. Francis, O. Kaczmarek, F. Karsch, E. Laermann and W. Soeldner, Phys. Rev. D 83, 034504 (2011) \[arXiv:1012.4963 \[hep-lat\]\]. A. Amato, G. Aarts, C. Allton, P. Giudice, S. Hands and J. -I. Skullerud, arXiv:1307.6763 \[hep-lat\]. S. -x. Qin, arXiv:1307.4587 \[nucl-th\]. W. Israel and J. M. Stewart, Annals Phys. 118, 341 (1979). S. Pu, arXiv:1108.5828 \[hep-ph\]. S. Jeon, Phys. Rev. D 52, 3591 (1995). Y. Hidaka and T. Kunihiro, Phys. Rev. D 83, 076004 (2011). J. S. Gagnon and S. Jeon, Phys. Rev. D 76, 105019 (2007). P. B. Arnold, G. D. Moore and L. G. Yaffe, JHEP 0301, 030 (2003) \[hep-ph/0209353\]. S. Mrowczynski and M. H. Thoma, Phys. Rev. D 62, 036011 (2000) \[hep-ph/0001164\].	{ "pile_set_name": "ArXiv" }
--- abstract: 'The chiral magnetic effect (CME) is an exact statement that connects via the axial anomaly the electric current in a system consisting of interacting fermions and gauge field with chirality imbalance that is put into a strong external magnetic field. Experimental search of the magnetically induced current in QCD in heavy ion collisions above a pseudocritical temperature hints, though not yet conclusive, that the induced current is either small or vanishing. This would imply that the chirality imbalance in QCD above $T_c$ that could be generated via topological fluctuations is at most very small. Here we present the most general reason for absence (smallness) of the chirality imbalance in QCD above $T_c$. It was recently found on the lattice that QCD above $T_c$ is approximately chiral spin (CS) symmetric with the symmetry breaking at the level of a few percent. The CS transformations mix the right- and left-handed components of quarks. Then an exact CS symmetry would require absence of any chirality imbalance. Consequently an approximate CS symmetry admits at most a very small chirality imbalance in QCD above $T_c$. Hence the absence or smallness of an magnetically induced current observed in heavy ion collisions could be considered as experimental evidence for emergence of the CS symmetry above $T_c$.' address: 'Institute of Physics, University of Graz, A-8010 Graz, Austria' author: - 'L. Ya. Glozman' title: 'The chiral magnetic effect and the chiral spin symmetry in QCD above $T_c$' --- QCD; chiral spin symmetry; high temperatures; chiral magnetic effect Introduction ============ According to the Atiah-Singer theorem local topological fluctuations of the gluonic field in Euclidean space with nonzero topological charge induce creation of chiral quarks such that the number of the right-handed quarks exceeds the number of the left-handed quarks by the topological charge. This process produces a chirality imbalance. If such a system is put into a strong external magnetic field, then this process should give rise to the magnetically induced electric current. This phenomenon is called the chiral magnetic effect (CME) [@K1; @K2] and the exact relation between the external magnetic field and magnetically induced current based on axial anomaly is $$\label{cme} {\mbox{\boldmath$J$\unboldmath}} = \sigma_5 {\mbox{\boldmath$B$\unboldmath}},$$ where the chiral magnetic conductivety $\sigma_5$ is expressed in terms of the chiral chemical potential as $$\label{pot} \sigma_5 = N_c \sum _f \frac {Q^2_f e^2}{2\pi^2} \mu_5$$ with $Q_f e$ being the electric charge of the quark with the flavor $f$ and $$\label{mu} \mu_5 = \frac{\mu_R - \mu_L}{2}$$ being the axial chemical potential, i.e a difference of the chemical potentials for the right- and left-handed quarks. The axial chemical potential parametrises the chirality imbalance induced by local topological fluctuations. Experimental search of CME at RHIC and LHC suggests that the magnetically induced current is at most very small, though there are many uncertanties and direct extraction of the current from data is impossible [@W1; @W2; @W3]. If the magnetic field that is formed during collision of two nuclei is sufficiently large and at the same time a magnetically induced electric current vanishes or small, as hinted by experimental data, then one infers from (\[cme\]) that chirality imbalance in QCD is either absent or very small. This obviously requires a convincing explanation. It was recently suggested [@G1] and then observed on the lattice [@R1; @R2; @R3] that QCD at RHIC and LHC temperatures is approximately chiral spin (CS) symmetric [@G2; @G3]. This symmetry, that includes chiral symmetry $U(1)_A$ as a subgroup, is not a symmetry of the Dirac Lagrangian but is a symmetry of the Lorentz-invariant fermion charge and consequently is a symmetry of the chromoelectric interactions in QCD in a given reference frame. The chromomagnetic interaction as well as the quark kinetic term break this symmetry. Observation of this symmetry in $T_c - 3 T_c$ interval implies that a dominant physics at this temperatures is due to the chromoelectric interaction between chirally symmetric quarks, that are bound into the color-singlet objects (“strings”), and a contribution of the chromomagnetic interaction as well as of the quark kinetic energy is at least much smaller. The CS symmetry is not a symmetry of the Dirac Lagrangian and consequently implies that there are no free deconfined quarks. A salient feature of the CS-transformations is that they mix the right- and left-handed quarks. Then exact CS symmetry in QCD above $T_c$ would require that the chirality imbalance should be absent since this imbalance is proportional the difference of number of the right- and left-handed quarks. Hence an exact CS-symmetry would require vanishing of the magnetically induced electric current in heavy ion collisions. In reality this symmetry is not exact and is broken at a few percent level. Then this approximate CS symmetry admits only a very small magnetically induced current. Stated otherwise a nonobservation or observation of only small magnetically induced current provided that the external magnetic field is sufficiently large could be considered as an experimental evidence of the CS symmetry observed on the lattice. This short paper is structured in the following way. In sections 2 and 3 we overview already known results about CS symmetry and its observation on the lattice at high temperatures. This will allow to avoid reading the preceeding papers. Then in section 4 we present the key argument of this paper and will conclude in section 5. Chiral spin symmetry ==================== The chiral spin $SU(2)_{CS}$ transformation was defined in ref. [@G2] as a transformation that rotates in the space of the right- and left-handed Weyl spinors $$\label{W} \left(\begin{array}{c} R\\ L \end{array}\right) \rightarrow \exp \left(i \frac{\varepsilon^n \sigma^n}{2}\right) \left(\begin{array}{c} R\\ L \end{array}\right)\; .$$ In terms of the Dirac spinors $\psi$ this transformation can be written via $\gamma$-matrices [@G3] $$\label{V-defsp} \psi \rightarrow \psi^\prime = \exp \left(i \frac{\varepsilon^n \Sigma^n}{2}\right) \psi = \exp \left(i \frac{\varepsilon^n \sigma^n}{2}\right) \left(\begin{array}{c} R\\ L \end{array}\right)\; ,$$ where the generators $\Sigma^n$ of the four-dimensional reducible representation are $$\Sigma^n = \{\gamma_0,-i \gamma_5\gamma_0,\gamma_5\}. \label{SIGCS}$$ The $su(2)$ algebra is automatically satisfied for these three generators, $$[\Sigma^a,\Sigma^b]=2i\epsilon^{abc}\Sigma^c. \label{algebra}$$ The $U(1)_A$ group is a subgroup of $SU(2)_{CS}$. In Euclidean space with the $O(4)$ symmetry all four directions are equivalent and one can use any Euclidean hermitian $\gamma$-matrix $\gamma_k$, $k=1,2,3,4$ instead of Minkowskian $\gamma_0$: $$\Sigma^n = \{\gamma_k,-i \gamma_5\gamma_k,\gamma_5\}, \label{SIGCS}$$ $$\gamma_i\gamma_j + \gamma_j \gamma_i = 2\delta^{ij}; \qquad \gamma_5 = \gamma_1\gamma_2\gamma_3\gamma_4. \label{gamma}$$ The $su(2)$ algebra is satisfied with any $k=1,2,3,4$. The direct product of the $SU(2)_{CS}$ group with the flavor group $SU(2)_{CS} \times SU(N_F)$ can be extended to a $SU(2N_F)$ group. This group includes the chiral symmetry $SU(N_F)_L \times SU(N_F)_R \times U(1)_A$ as a subgroup. The $SU(2N_F)$ transformations are given by $$\psi \rightarrow \psi^\prime = \exp\left(i \frac{\epsilon^m T^m}{2}\right) \psi,$$ with $m=1,2,...,(2N_F)^2-1$ and the set of $(2N_F)^2-1$ generators being $$\begin{aligned} T^m=\{ (\tau^a \otimes {1}_D), ({1}_F \otimes \Sigma^n), (\tau^a \otimes \Sigma^n) \}\end{aligned}$$ where $\tau$ are the flavor generators (with the flavor index $a$) and $n=1,2,3$ is the $SU(2)_{CS}$ index. The fundamental vector of $SU(2N_F)$ at $N_F=2$ is $$\Psi =\begin{pmatrix} u_{\textsc{R}} \\ u_{\textsc{L}} \\ d_{\textsc{R}} \\ d_{\textsc{L}} \end{pmatrix}.$$ The $SU(2)_{CS}$ and $SU(2N_F)$ groups are not symmetries of the Dirac Lagrangian. At the same time they are symmetries of the Lorentz-invariant fermion charge $$Q = \int d^3x \bar \psi(x) \gamma_0 \psi(x) = \int d^3x \psi^\dagger(x) \psi(x). \label{Q}$$ This salient feature allows us to use the $SU(2)_{CS}$ and $SU(2N_F)$ symmetries to distinguish the chromoelectric and chromomagnetic interactions in a given reference frame because the chromoelectric interaction is influenced only by the color charge while the chromomagnetic interaction is dictated by the spatial current. The latter current is not $SU(2)_{CS}$ and $SU(2N_F)$ symmetric. More specifically the (chromo)electric and (chromo)magnetic fields in Minkowski space in a given reference frame are different fields. Interaction of fermions with the gauge field in Minkowski space-time can be split in a given reference frame into temporal and spatial parts: $$\overline{\psi} \gamma^{\mu} D_{\mu} \psi = \overline{\psi} \gamma^0 D_0 \psi + \overline{\psi} \gamma^i D_i \psi . \label{cl}$$ The covariant derivative $D_{\mu}$ includes interaction of the matter field $\psi$ with the gauge field ${\boldsymbol{A}}_\mu$, $$D_{\mu}\psi =( \partial_\mu - ig \frac{{\boldsymbol{t}}\cdot {\boldsymbol{A}}_\mu}{2})\psi.$$ The temporal term contains interaction of the color-octet charge density $$\bar \psi (x) \gamma^0 \frac{{\boldsymbol{t}}}{2} \psi(x) = \psi (x)^\dagger \frac{{\boldsymbol{t}}}{2} \psi(x) \label{den}$$ with the chromoelectric part of the gluonic field. It is invariant under $SU(2)_{CS}$ and $SU(2N_F)$ since it is invariant under any unitary transformation that acts in Dirac and/or flavor spaces. The $SU(2)_{CS}$ transformations defined via the Euclidean Dirac matrices can be identically applied to Minkowski Dirac spinors without any modification of the generators. The spatial part contains the quark kinetic term and the interaction of the spatial current $\bar \psi (x) \gamma^i \frac{{\boldsymbol{t}}}{2} \psi(x)$ with the chromomagnetic field. It breaks $SU(2)_{CS}$ and $SU(2N_F)$. We conclude that interaction of the chromoelectric and chromomagnetic components of the gauge field with quarks in a given reference frame can be distinguished by symmetry. Of course, in order to discuss the chromoelectric and chromomagnetic components of the gluonic field one needs to fix a reference frame. The hadron invariant mass is the rest frame energy. Consequently, to address physics of hadron mass one should discuss energy in the hadron rest frame. At high temperatures the Lorentz invariance is broken and a natural frame is the medium rest frame. Emergence of chiral spin and $SU(2N_F)$ symmetries in QCD above $T_c$ ===================================================================== Above the chiral restoration pseudocritical temperature $T_c$ one apriori expects in observables chiral $SU(2)_L \times SU(2)_R$ symmetry because the quark condensate vanishes during a very smooth crossover at temperatures between 100 and 200 MeV. A pseudocritical temperature $T_c$ determined from the chiral susceptibility is around 155 MeV in $N_F=2+1$ QCD [@F]. This symmetry above the crossover is evidenced by degeneracy of correlators connected by the chiral transformation. While the axial anomaly is a pertinent property of QCD its effect is determined by the topological charge density. There are strong indications from the lattice that the $U(1)_A$ symmetry is also effectively restored above $T_c$ [@JLQCD1; @JLQCD3] which suggests that the local topological fluctuations in Euclidean space in QCD are at least very strongly suppressed above $T_c$. It is a matter of the present debates whether the $U(1)_A$ restoration happens at the same temperature as of $SU(2)_L \times SU(2)_R$ or at a slightly higher temperature [@hotQCD]. The $U(1)_A$ restoration is evidenced by degeneracy of correlators connected by the $U(1)_A$ transformation. $SU(2)_L \times SU(2)_R$ and $U(1)_A$ transformation properties of the $J=1$ operators age given in the left panel of Fig. \[F1\]. ![Transformations between $J=1$ operators, $i=1,2,3$. The left columns indicate the $SU(2)_L \times SU(2)_R$ representation for every operator. Red and blue arrows connect operators which transform into each other under $SU(2)_L \times SU(2)_R$ and $U(1)_A$, respectively. Green arrows connect operators that belong to $SU(2)_{CS}$, $k=4$ triplets. Purple arrow shows the $SU(4)$ 15-plet. The $f_1$ operator is is a singlet of $SU(4)$. The Fig. is from Ref. [@G3].[]{data-label="F1"}](multiplets_a.pdf "fig:"){width="0.45\linewidth"} ![Transformations between $J=1$ operators, $i=1,2,3$. The left columns indicate the $SU(2)_L \times SU(2)_R$ representation for every operator. Red and blue arrows connect operators which transform into each other under $SU(2)_L \times SU(2)_R$ and $U(1)_A$, respectively. Green arrows connect operators that belong to $SU(2)_{CS}$, $k=4$ triplets. Purple arrow shows the $SU(4)$ 15-plet. The $f_1$ operator is is a singlet of $SU(4)$. The Fig. is from Ref. [@G3].[]{data-label="F1"}](multiplets_b.pdf "fig:"){width="0.45\linewidth"} In the right panel of the same Fig. we present transformation properties of the same operators with respect to $SU(2)_{CS}$ and $SU(4)$ [@G3]. If one observes on the lattice degeneracy of correlators that are connected by $SU(2)_{CS}$ and $SU(4)$ that would be a signal for emergence of these symmetries. On the r.h.s. of Fig. \[tcorr\] we show temporal correlators $$C_\Gamma(t) = \sum\limits_{x, y, z} \braket{\mathcal{O}_\Gamma(x,y,z,t) \mathcal{O}_\Gamma(\mathbf{0},0)^\dagger}, \label{eq:momentumprojection}$$ at a temperature $T = 1.2 T_c$ calculated in $N_F=2$ QCD with a chirally symmetric Dirac operator [@R3]. Here $\mathcal{O}_\Gamma(x,y,z,t)$ is an operator that creates a quark-antiquark pair with fixed quantum numbers. Summation over $x,y,z$ projects out the rest frame. ![ Temporal correlation functions for $12 \times 48^3$ lattices. The l.h.s. shows correlators calculated with free noninteracting quarks with manifest $U(1)_A$ and $SU(2)_L \times SU(2)_R$ symmetries. The r.h.s. presents full QCD results at a temperature $1.2 T_c$, which shows multiplets of all $U(1)_A$, $SU(2)_L \times SU(2)_R$, $SU(2)_{CS}$ and $SU(4)$ groups. The Fig. is from Ref. [@R3]. []{data-label="tcorr"}](fig2a_48x12.pdf "fig:") ![ Temporal correlation functions for $12 \times 48^3$ lattices. The l.h.s. shows correlators calculated with free noninteracting quarks with manifest $U(1)_A$ and $SU(2)_L \times SU(2)_R$ symmetries. The r.h.s. presents full QCD results at a temperature $1.2 T_c$, which shows multiplets of all $U(1)_A$, $SU(2)_L \times SU(2)_R$, $SU(2)_{CS}$ and $SU(4)$ groups. The Fig. is from Ref. [@R3]. []{data-label="tcorr"}](fig2b_48x12.pdf "fig:") Correlators of the isovector scalar (S) and isovector pseudoscalar (PS) operators are connected by the $U(1)_A$ transformation and their degeneracy indicates restoration of this symmetry. If there is a tiny splitting of the S and PS correlators then it should be so small so that it cannot be seen in the present lattice data. This strongly suggests that the topological transitions are at least severely suppressed above $T_c$. An approximate degeneracy of the $a_1$, $b_1$, $\rho_{(1,0)+(0,1)}$ and $\rho_{(1/2,1/2)_b}$ correlators indicates emergent $SU(2)_{CS}$ and $SU(4)$ symmetries. Their breaking is estimated at the level of less than $ 5\%$. A similar multiplet structure is seen in spatial correlators in the temperature range $T_c - 3 T_c$ [@R1; @R2]. On the l.h.s of Fig. \[tcorr\] we present correlators calculated with noninteracting quarks on the same lattice. They represent a QGP at a very high temperature where due to asymptotic freedom the quark-gluon interaction can be neglected. Dynamics of free quarks are governed by the Dirac equation and only $U(1)_A$ and $SU(2)_L \times SU(2)_R$ chiral symmetries exist. A qualitative difference between the pattern on the l.h.s. and the pattern on the r.h.s of Fig. \[tcorr\] is remarkable. The temporal correlators are directly connected to measurable spectral density. $SU(2)_{CS}$ and $SU(4)$ symmetries of the t-correlators imply the same symmetries of spectral densities. Chiral spin transformation properties of the axial chemical potential ===================================================================== Now we arrive at the key point of this note: What are implications of the emerging chiral spin symmetry above $T_c$ on axial chemical potential term in effective action? The quark chemical potential $ \mu \psi (x)^\dagger \psi(x)$ and the axial chemical potential $ \mu_5 \psi (x)^\dagger \gamma_5 \psi(x)$ terms can be present in the QCD action, $$S = \int_{0}^{\beta} dt \int d^3x \overline{\psi} [ \gamma_{\mu} D_{\mu} + \mu \gamma_4 + \mu_5 \gamma_4 \gamma_5 + m] \psi,$$ in the $SU(2)_{CS}$ and $SU(2N_F)$ symmetric regime only if they are invariant with respect to the chiral spin transformations, i.e. are the chiral spin singlets. The chemical potential term is indeed invariant, i.e. transforms into itself upon the $SU(2)_{CS}$ transformation (\[V-defsp\]) $$\psi (x)^\dagger \psi(x) \longrightarrow \psi (x)^\dagger \psi(x).$$ The axial chemical potential $ \psi (x)^\dagger \gamma_5 \psi(x)$ term is not invariant under (\[V-defsp\]) and transforms into a superposition of three terms: $$\psi (x)^\dagger \gamma_5 \psi(x) \longrightarrow \alpha \psi (x)^\dagger \gamma_4 \psi(x) + \beta \psi (x)^\dagger \gamma_4 \gamma_5 \psi(x) + \gamma \psi (x)^\dagger \gamma_5 \psi(x),$$ i.e. it transforms under the triplet representation of $SU(2)_{CS}$ with coefficients $\alpha,\beta,\gamma$ being determined by three rotation angles. Then such a term [is not allowed]{} in a $SU(2)_{CS}$-symmetric theory. In other words, the $SU(2)_{CS}$ symmetry prohibiths existence of a finite axial chemical potential. The same statement can be understood with a less formal language. The axial chemical potential parametrizes an excess of the right-handed quarks over the left-handed quarks. The $SU(2)_{CS}$ transformation mixes the right- and left-handed quarks, see (\[W\]). Consequently a given finite excess of the right- over the left-handed quarks cannot be $SU(2)_{CS}$-symmetric. Only a vanishing axial chemical potential is consistent with chiral spin symmetry. We conclude that if the emerged chiral spin symmetry were exact, then the chiral magnetic conductivety $\sigma_5$ in (\[cme\]) must vanish. [Exact $SU(2)_{CS}$ symmetry requires vanishing of the magnetically induced electrical current even if the external magnetic field is large.]{} Of course, in reality the chiral spin symmetry above $T_c$ is not exact: it is broken at a few percent level. Then the magnetically induced electric current can be only very small as it would originate only from the CS symmetry breaking contributions. The topological fluctuations with a nonzero topological charge do indeed break the CS symmetry. However their role in the QCD dynamics above $T_c$ can be only very small because of a very good CS symmetry. This result explains a vanishing or very small magnetically induced electric current above $T_c$ as it is suggested by the present experimental data. Stated otherwise, if future experiments confirm a smallness or absence of the magnetically induced current provided that the transient external magnetic field is sufficiently strong, it would be an experimental evidence of the chiral spin symmetry above $T_c$. Conclusions =========== A formation of multiplets in correlators described by the chiral spin $SU(2)_{CS}$ and $SU(4)$ groups [@G2; @G3] in the range $T_c - 3 T_c$ was observed on the lattice [@R1; @R2; @R3]. These symmetries include the chiral $U(1)_A$ and $SU(2)_L \times SU(2)_R$ as subgroups. These are not symmetries of the free Dirac action and they are not consistent with free deconfined quarks. In the medium rest frame the chromoelectric interaction is invariant under both $SU(2)_{CS}$ and $SU(4)$ transformations, while the chromomagnetic interaction as well as the quark kinetic term break them. The emergence of these symmetries in the $T_c$ - $3 T_c$ window suggests that the chromomagnetic field disappears or is strongly suppressed, while the confining chromoelectric field is still active. This implies that the physical degrees of freedom are chirally symmetric quarks bound by the chromoelectric interaction into color-singlet objects without chromomagnetic effects. This regime of QCD was named as a “stringy fluid”. These symmetries are broken at a few percent level. If the $SU(2)_{CS}$ and $SU(4)$ symmetries were exact it would require a vanishing of the chirality imbalance. Consequently the magnetically induced current would exactly vanish even if the external magnetic field be very strong. Then a tiny magnetically induced current can be only related to the CS symmetry breaking dynamics which is however much less important than a confining chromoelectric interaction that binds the chirally symmetric quarks into color singlet objects. This conclusion is drawn from the smallness of the CS symmetry breaking. Confirmation in experiments of smallness of the magnetically induced current or its absence provided that the external magnetic field is sufficiently strong could be considered as an experimental verification of the chiral spin and $SU(2N_F)$ symmetries in QCD above $T_c$. The author is thankful to S. Voloshin for discussion of the present experimental situation as well as to M. Chernodub, T. Cohen and C. Lang for their reading of the ms and discussions. [0]{} D. E. Kharzeev, L. D. McLerran and H. J. Warringa, Nucl. Phys. A [803]{} (2008) 227. K. Fukushima, D. E. Kharzeev and H. J. Warringa, Phys. Rev. D [78]{} (2008) 074033. D. E. Kharzeev, J. Liao, S. A. Voloshin and G. Wang, Prog. Part. Nucl. Phys. [88]{} (2016) 1. S. A. Voloshin, Phys. Rev. C [98]{} (2018) 054911. see review plenary talks at Quark Matter 2018 and Quark Matter 2019 conferences L. Ya. Glozman, Proceedings of the “Critical point and onset of deconfinement” conference, Wroclav, Poland, May 30 - June 4, 2016,[Acta Physica Polonica B, Proceedings Supplement]{}, [10]{} (2017) 583 \[arXiv:1610.00275\]. C. Rohrhofer, Y. Aoki, G. Cossu, H. Fukaya, L. Y. Glozman, S. Hashimoto, C. B. Lang and S. Prelovsek, Phys. Rev. D [96]{} (2017) 094501 Erratum: Phys. Rev. D [99]{} (2019) 039901. C. Rohrhofer, Y. Aoki, G. Cossu, H. Fukaya, C. Gattringer, L. Y. Glozman, S. Hashimoto, C. B. Lang and S. Prelovsek, Phys. Rev. D [100]{} (2019) 014502. C. Rohrhofer, Y. Aoki, L. Y. Glozman and S. Hashimoto, Phys. Lett. B [802]{} (2020) 135245. L. Y. Glozman, Eur. Phys. J. A [51]{} (2015) 27. L. Y. Glozman and M. Pak, Phys. Rev. D [92]{} (2015) 016001. Y. Aoki, S. Borsanyi, S. Durr, Z. Fodor, S. D. Katz, S. Krieg and K. K. Szabo, JHEP [0906]{} (2009) 088. G. Cossu, S. Aoki, H. Fukaya, S. Hashimoto, T. Kaneko, H. Matsufuru and J. I. Noaki, Phys. Rev. D [87]{} (2013) 114514. Erratum: \[ Phys. Rev. D [88]{} (2013) 019901\]. A. Tomiya, G. Cossu, S. Aoki, H. Fukaya, S. Hashimoto, T. Kaneko and J. Noaki, Phys. Rev. D [96]{} (2017) 034509. Addendum: \[ Phys. Rev. D [96]{} (2017) 079902\]. A. Bazavov [et al.]{}, Phys. Rev. D [100]{} (2019) 094510.	{ "pile_set_name": "ArXiv" }
--- abstract: \| Estimation of Distribution Algorithms (EDAs) are one branch of Evolutionary Algorithms (EAs) in the broad sense that they evolve a probabilistic model instead of a population. Many existing algorithms fall into this category. Analogous to genetic drift in EAs, EDAs also encounter the phenomenon that updates of the probabilistic model not justified by the fitness move the sampling frequencies to the boundary values. This can result in a considerable performance loss. This paper proves the first sharp estimates of the boundary hitting time of the sampling frequency of a neutral bit for several univariate EDAs. For the UMDA that selects $\mu$ best individuals from $\lambda$ offspring each generation, we prove that the expected first iteration when the frequency of the neutral bit leaves the middle range $[\tfrac 14, \tfrac 34]$ and the expected first time it is absorbed in 0 or 1 are both $\Theta(\mu)$. The corresponding hitting times are $\Theta(K^2)$ for the cGA with hypothetical population size $K$. This paper further proves that for PBIL with parameters $\mu$, $\lambda$, and $\rho$, in an expected number of $\Theta(\mu/\rho^2)$ iterations the sampling frequency of a neutral bit leaves the interval $[\Theta(\rho/\mu),1-\Theta(\rho/\mu)]$ and then always the same value is sampled for this bit, that is, the frequency approaches the corresponding boundary value with maximum speed. For the lower bounds implicit in these statements, we also show exponential tail bounds. If a bit is not neutral, but neutral or has a preference for ones, then the lower bounds on the times to reach a low frequency value still hold. An analogous statement holds for bits that are neutral or prefer the value zero. author: - \| Benjamin Doerr\ Laboratoire d’Informatique (LIX)\ CNRS\ École Polytechnique\ Institute Polytechnique de Paris\ Palaiseau, France - \| Weijie Zheng\ Shenzhen Key Laboratory of Computational Intelligence\ University Key Laboratory of Evolving Intelligent Systems of Guangdong Province\ Department of Computer Science and Engineering\ Southern University of Science and Technology\ Shenzhen, China title: 'Sharp Bounds for Genetic Drift in EDAs[^1]' --- Introduction ============ Estimation of Distribution Algorithms (EDAs) are evolutionary algorithms (EAs) that evolve a probabilistic model instead of a population. An iteration of an EDA usually consists of three steps. (i) Based on the current probabilistic model, a population of individuals is sampled. (ii) The fitness of this population is determined. (iii) Update of the probabilistic model: Based on the fitness of this population and the probabilistic model, a new probabilistic model is computed. Different probabilistic models and update strategies form different specific algorithms in this branch. In multivariate EDAs, the probabilistic model contains dependencies among the variables. Examples for multivariate EDAs include Mutual-Information-Maximization Input Clustering [@BonetIV96], Bivariate Marginal Distribution Algorithm [@PelikanM99], the Factorized Distribution Algorithm [@MuhlenbeinM99], the Extended Compact Genetic Algorithm [@HarikLS06], and many other. For univariate EDAs, the bit positions of the probabilistic model are mutually independent. Univariate EDAs include Population-Based Incremental Learning (PBIL) [@Baluja94; @BalujaC95] with special cases Univariate Marginal Distribution Algorithm (UMDA) [@MuhlenbeinP96] and Max-Min Ant System with iteration-best update (MMAS$_{ib}$) [@NeumannSW10], and the Compact Genetic Algorithm (cGA) [@HarikLG98]. Since the dependencies in multivariate EDAs bear significant difficulties for a mathematical analysis, almost all theoretical results for EDAs regard univariate models [@KrejcaW18survey]. This paper also deals exclusively with univariate EDAs. In evolutionary algorithms, it is known that the frequencies of bit values in the population are not only influenced by the contribution of the bit to the fitness, but also by random fluctuation stemming from other bits having a stronger influence on the fitness. These random fluctuations can even lead to certain bits converging to a single value different from the one in the optimal solution. This effect is called genetic drift [@Kimura64; @AsohM94]. Genetic drift also happens in EDAs. Gonz[á]{}lez, Lozano, and Larra[ñ]{}aga [@GonzalezLL01] showed that for the $2$-dimensional [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}function, the sampling frequency of PBIL can converge to any search point in the search space with probability near to $1$ if the initial sampling frequency goes to that search point and the learning rate goes to 1. Droste [@Droste06] noticed the possibility of the cGA getting stuck, but he only analyzed the runtime conditional on being finite, no analysis of genetic drift or stagnation times was given. Costa, Jones, and Kroese [@CostaJK07] proved that a constant smoothing parameter for the Cross Entropy (CE) algorithm (which is equivalent to a constant learning rate $\rho$ for PBIL) results in that the probability mass function converges to a unit mass at some random candidate, but no convergence speed analysis was given. In summary, as Krejca and Witt said in [@KrejcaW18survey], the genetic drift in EDAs is a general problem of martingales, that is, that a random process with zero expected change will eventually stop at the absorbing boundaries of the range. Witt [@Witt19] and Lengler, Sudholt, and Witt [@LenglerSW18] recently showed that genetic drift can result in a considerable performance loss on the [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}function. In this work, we shall quantify this effect asymptotically precise for several EDAs and this via proven results. The few previous works in this direction have obtained the following results. Friedrich, Kötzing, and Krejca [@FriedrichKK16] showed that for the cGA, the expected frequency of a neutral bit is arbitrary close to the borders $0$ or $1$ after $\omega(K^2)$ generations. Though not stated in [@FriedrichKK16], from Corollary 9 in [@FriedrichKK16], we can derive an upper bound of $O(K^2)$ for the expected time of leaving the interval $[\frac 14, \frac 34]$, and $O(K^2 \log K)$ for the expected hitting time of a boundary value. For the UMDA, the situation is similar [@FriedrichKK16]. After $\omega(\mu)$ iterations, the frequencies are arbitrary close to the boundaries and the expected hitting time can be shown to be $O(\mu \log \mu)$ via similar arguments as above. Sudholt and Witt [@SudholtW16] mentioned that the boundary hitting time of the cGA is $\Theta(K^2)$, but without a complete proof (in particular, because they did not discuss what happens once the frequency leaves the interval $[\frac 16, \frac 56]$). Although Krejca and Witt [@KrejcaWitt17] focused on the lower bound of the runtime of the UMDA on [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}, we can derive from it that the hitting time of the boundary $0$ is at least $\Omega(\mu)$. This follows from the drift of $\phi$ in Lemma 9 in [@KrejcaWitt17] together with the additive drift theorem [@HeY01]. Our results: While the results above give some indication on the degree of stability of PBIL and the cGA, a sharp proven result is still missing. This paper overcomes this shortage and gives precise asymptotical hitting times for PBIL (including the UMDA and the MMAS$_{ib}$) and the cGA. With a simultaneous analysis of UMDA and cGA, we prove that for the UMDA selecting $\mu$ best individuals from $\lambda$ offspring on some $D$-dimensional problem, the expected number of iterations until the frequency of the neutral bit is absorbed in 0 or 1 for the UMDA without margins or when the frequency hits the margins $\{1/D,1-1/D\}$ for the UMDA with such margins is $\Theta(\mu)$, and the corresponding hitting time is $\Theta(K^2)$ for the cGA with hypothetical population size $K$. This paper also gives a precise asymptotical analysis for PBIL: In expectation in $\Theta(\mu/\rho^2)$ generations the sampling frequency of a neutral bit leaves the interval $[\Theta(\rho/\mu),1-\Theta(\rho/\mu)]$ and then always the same value is sampled for this bit. For the lower bounds implicit in these estimates we prove an exponential tail bound in Corollary \[cor:tail\]. We also extend the lower bound results to bits that are neutral or have a preference for some bit value (Section \[sec:lowerP\]). For example, we prove that for PBIL it takes an expected number of $\Omega(\mu/\rho^2)$ iterations until the sampling frequency of a bit that is neutral or prefers a one (neutral or prefers a zero) reaches the interval $[0,\tfrac 14]$ ($[\tfrac 34, 1]$). The corresponding reaching time is $\Omega(K^2)$ for the cGA. The remainder of this paper is organized as follows. Section \[sec:IntroEDA\] briefly introduces PBIL and the cGA under the umbrella of the $n$-Bernoulli-$\lambda$-EDA framework proposed in [@FriedrichKK16]. Our notation for our results is fixed in Section \[sec:Notation\]. Section \[sec:lower\] and Section \[sec:upper\] discuss how fast the frequency of a neutral bit approaches the boundaries. Section \[sec:lowerP\] extends the lower bound results of Section \[sec:lower\] to bits that are neutral or have some preference. Finally, in Section \[sec:disc\] we argue how our results allow to interpret existing research results and how they give hints on how to choose the parameters of these EDAs. The $n$-Bernoulli-$\lambda$-EDA Framework {#sec:IntroEDA} ========================================= Since the $n$-Bernoulli-$\lambda$-EDA framework proposed in [@FriedrichKK16] covers many well-known EDAs including PBIL and cGA, we use it to make precise these two EDAs. We note that often margins like $1/D$ and $1-1/D$ are used, that is, the frequencies are restricted to stay in the interval $[1/D,1-1/D]$. This prevents the frequencies from reaching the absorbing states $0$ and $1$. To ease the presentation, we regard the EDAs without such margins. We note that, trivially, the time to reach an absorbing state is not smaller than the time to reach a margin value. Hence an upper bound on the hitting time of the absorbing states is also an upper bound for the time to reach or exceed the margin values. Our main result on lower bounds, Theorem \[thm:lower\], shows a lower bound for the time to reach a frequency value in $[0,\frac14] \cup [\frac 34, 1]$. This again is a lower bound for the time to reach (or exceed) the margin values or the absorbing states. The $n$-Bernoulli-$\lambda$-EDA framework for maximizing a function $f: \{0,1\}^D \rightarrow {\ensuremath{\mathbb{R}}}$ is shown in Algorithm \[alg:EDA\]. By suitably specifying the update scheme $\phi$, we derive PBIL and the cGA. The general idea of population-based incremental learning (PBIL) is to sample $\lambda$ individuals from the current distribution, select $\mu$ best of them, and use these (with a learning rate of $\rho$) and the current distribution to define the new distribution. Formally, the update scheme is $$p_j^t=\varphi(p^{t-1}, (X_i, f(X_i))_{i=1,\dots,\lambda})_j=(1-\rho)p_j^{t-1}+\frac{\rho}{\mu}\sum\limits_{i=1}^{\mu}\tilde{X}_{i,j}^t, \label{eq:pbilupdate}$$ where $\rho$ is the learning rate and $\tilde{X}_1^t,...,\tilde{X}_\mu^t$ are the selected $\mu$ best individuals from the $\lambda$ offspring. The cross entropy algorithm (CE) has various definitions according to the problems to be solved. The basic CE algorithm for discrete optimization [@CostaJK07] samples $N$ individuals from the current distribution, selects $N_b$ best of them, and uses these (with a time-dependent smoothing rate of $\alpha_t$) and the current distribution to define the new distribution. The formal update scheme is (\[eq:pbilupdate\]) with $\mu$, $\lambda$ and $\rho$ respectively replaced by $N_b$, $N$ and $\alpha_t$. The basic CE is equal to PBIL except that the learning rate is fixed for PBIL, whereas CE utilizes time-dependent learning rates. When referring to the CE algorithm in this paper, we mean this version from [@CostaJK07], but we denote its parameters by $\mu$, $\lambda$ and $\rho_t$ instead of $N_b$, $N$ and $\alpha_t$ to reflect the similarity with PBIL. Two special cases of PBIL have been regarded in the literature. The univariate marginal distribution algorithm (UMDA) only uses the samples of this current iteration to define the next probabilistic model, hence it is equivalent to PBIL with a learning rate of $\rho=1$. The $\lambda$-max-min ant system ($\lambda$-MMAS) only selects the best sampled individual and the current model to construct the new model, hence it is the special case with $\mu=1$. $X_{i,j}^t \leftarrow 1$ with probability $p_{i}^{t-1}$, $X_{i,j}^t \leftarrow 0$ with probability $1-p_{i}^{t-1}$; $p^t\leftarrow\varphi(p^{t-1}, (X_i, f(X_i))_{i=1,\dots,\lambda})$; \[alg:EDA\] The compact genetic algorithm (cGA) with hypothetical population size $K$, not necessarily an integer, samples two individuals and then changes the frequency of each bit by an absolute value of $1/K$ towards the bit value of the better individual (unless the two sampled individuals have identical values in this bit). Formally, we have $\lambda=2$ in the $n$-Bernoulli-$\lambda$-EDA framework and the update scheme is $$\begin{split} p_j^t=\varphi(p^{t-1}, (X_i, f(X_i))_{i=1,\dots,\lambda})_j= \begin{cases} p_j^{t-1}+\tfrac{1}{K}, & \text{if $X_{(1),j}^t>X_{(2),j}^t$}\\ p_j^{t-1}-\tfrac{1}{K}, & \text{if $X_{(1),j}^t<X_{(2),j}^t$}\\ p_j^{t-1}, & \text{if $X_{(1),j}^t=X_{(2),j}^t$},\\ \end{cases} \end{split} \label{eq:cgaupdate}$$ where $\{X_{(1)}^t,X_{(2)}^t\} = \{X_1^t,X_2^t\}$ such that $f(X_{(1)}^t) \ge f(X_{(2)}^t)$. We shall always make the following well-behaved frequency assumption (first called so in [@Doerr19gecco], but made in many earlier works already): We assume that any two frequencies the cGA can reach differ by a multiple of $1/K$. In the case of no margins, this means that the cGA can only use frequencies in $\{0, 1/K, 2/K, \dots, 1\}$. Note that $K$ needs to be even so that the initial frequency $1/2$ is also a multiple of $1/K$. When using the margins $1/D$ and $1- 1/D$, the set of reachable frequency boundaries is $\{1/D, 1/D+1/K, 1/D+2/K, \dots, 1 - 1/D\}$. To have $1/2$ in this set, $1 - 2/D$ needs to be an even multiple of $1/K$. Notation Used in Our Analyses {#sec:Notation} ============================= Genetic drift is usually studied via the behavior of a neutral bit. Let $f: \{0,1\}^D \rightarrow {\ensuremath{\mathbb{R}}}$ be an arbitrary fitness function with a neutral bit. Without loss of generality, let the first bit of the fitness function $f$ be neutral, that is, we have $f(0,X_2,\dots,X_D)=f(1,X_2,\dots,X_D)$ for all $X_2,\dots,X_D\in\{0,1\}$. Then we can simply assume that $\tilde{X}_{i,1}^t=X_{i,1}^t,i=1,\dots,\mu$ in (\[eq:pbilupdate\]), and $X_{(1),1}^t=X_{1,1}^t, X_{(2),1}^t=X_{2,1}^t$ in (\[eq:cgaupdate\]). Let $p_t=p_1^t$ be the frequency of the neutral bit after generation $t$. For PBIL, we have $$\begin{split} p_t= \begin{cases} \frac{1}{2}, &t=0,\\ (1-\rho)p_{t-1}+\frac{\rho}{\mu}\sum\limits_{i=1}^{\mu}X_{i,1}^t, &t\geq 1, \end{cases} \end{split} \label{eq:pbil}$$ where the $X_{i,1}^t$ are independent $0,1$ random variables with $\Pr[X_{i,1}^t=1]=p_{t-1}$. For the cGA, we have $$\begin{split} p_t= \begin{cases} \frac{1}{2}, &t=0,\\ \begin{cases} p_{t-1}+\frac{1}{K}, & \text{if $X_{1,1}^t>X_{2,1}^t$}\\ p_{t-1}-\frac{1}{K}, & \text{if $X_{1,1}^t<X_{2,1}^t$}\\ p_{t-1}, & \text{if $X_{1,1}^t=X_{2,1}^t$}\\ \end{cases} , &t\geq 1, \end{cases} \end{split}$$ where $X_{1,1}^t$ and $X_{2,1}^t$ are independent $0,1$ random variables with $ \Pr[X_{1,1}^t=1]= \Pr[X_{2,1}^t=1]=p_{t-1}$. We observe that this random process $(p_t)$ is independent of $f,D$, and, in the case of PBIL, $\lambda$. We also have $$E[p_t\mid p_{t-1}]=p_{t-1},$$ that is, both PBIL and the cGA are balanced in the sense of [@FriedrichKK16]. Finally, let $T=\min\{t\mid p_t \in \{0,1\}\}$ be the hitting time of the absorbing states 0 and 1. We are now ready to prove our matching upper and lower bounds for the hitting time $T$. We start with the lower bounds in Section \[sec:lower\] as these are easier to prove and thus a good warm-up for the upper bound proofs in Section \[sec:upper\]. Lower Bounds on the Boundary Hitting Time {#sec:lower} ========================================= In this section, we prove the following lower bounds for the hitting times of the absorbing states. The expectations of hitting times are asymptotically equal to (and necessarily not less than) the expected times of leaving the frequency range $(\frac 14, \frac 34)$, so we now determine these, which are also of independent interest. Consider using an $n$-Bernoulli-$\lambda$-EDA to optimize some function $f$ with a neutral bit. Let $T_0$ denote the first time the frequency of the neutral bit is in $[0,\frac 14] \cup [\frac 34, 1]$. For PBIL without margins, we have $E[T_0]=\Omega(\tfrac{\mu}{\rho^2})$, in particular, $E [T_0]=\Omega({\mu})$ for the UMDA and $E [T_0 ]=\Omega(\tfrac{1}{\rho^2})$ for the $\lambda$-MMAS. For the cGA, we have $E[T_0]=\Omega(K^2)$. \[thm:lower\] For PBIL, building on the notation introduced in Section \[sec:Notation\], we consider the random process $$Z_{t\mu+a}=(1-\rho)p_t\mu+\rho p_{t}(\mu-a)+\rho\sum\limits_{i=1}^aX_{i,1}^{t+1},$$ where $t=0,1,\dots$, and $a=0,1,\dots,\mu-1$. For $a=0$, we obviously have $Z_{t\mu}/{\mu}=p_{t}$, that is, the $Z$-process contains the process $(p_t)$ we are interested in. Noting that $Z_{(t+1)\mu}$ can also be written as $Z_{t\mu+\mu}=(1-\rho)p_t\mu+\rho p_{t}(\mu-\mu)+\rho\sum\limits_{i=1}^{\mu}X_{i,1}^{t+1}$, it is also not difficult to see that for all $k=0,1,\dots$, we have $$\begin{split} \Pr[Z_{k+1}={}&Z_k+\rho-\rho p_t\mid Z_{1},\dots,Z_{k}]=p_t,\\ \Pr[Z_{k+1}={}&Z_k+0-\rho p_t\mid Z_1,\dots,Z_k]=1-p_t, \end{split} \label{eq:nprob}$$ where $t=0,1,\dots$. Consequently, $$E[Z_{k+1}\mid Z_1,\dots,Z_k]=Z_k$$ and the sequence $Z_0,Z_1,Z_2,\dots$ is a martingale. For $k=1,2,\dots$, let $R_k=Z_k-Z_{k-1}$ define the martingale difference sequence. By (\[eq:nprob\]), $$\|R_k\|\le \max\{\rho(1-p_t),\rho p_t\} \le \rho.$$ By the Hoeffding-Azuma inequality for maxima and minima (Theorem 3.10 and (41) in [@McDiarmid98], note that in (41) the absolute value should be inside the maximum, that is, $\max_k \|\sum_{i=1}^k Y_i\|$, as can be seen from the proof), we have $$\Pr \left[ \max\limits_{k=1,\dots,t\mu}\left\|\sum\limits_{i=1}^k{R_i} \right\|\ge M \right]\le 2\exp \left(-\frac{M^2}{2t\mu\rho^2} \right). \label{eq:azumaMM}$$ Recalling $Z_0 = \frac {\mu}2$ and $p_t = Z_{t\mu}/\mu$, we have $$\Pr \left[ \max\limits_{k=1,\dots,t} \left\| p_k - \tfrac 12 \right\| \ge M/\mu \right] \le \Pr \left[ \max\limits_{k=1,\dots,t\mu} \left\|\sum\limits_{i=1}^k{R_i} \right\|\ge M \right]. \label{eq:pZ}$$ Combining and with $M=\mu/4$, we obtain $$\Pr \left[ \max\limits_{k=1,\dots,t} \left\| p_k - \tfrac 12 \right\| \ge \tfrac 14 \right] \le 2 \exp\left(-\frac{\mu}{32t\rho^2}\right).$$ Consequently, since $T_0=\min \{t \mid \|p_t-\tfrac{1}{2} \| \ge \tfrac{1}{4} \}$, we have $$E [T_0 ] \ge (1-2\exp (-\mu/(32t\rho^2) ) )(t+1)$$ for all $t\in {\ensuremath{\mathbb{N}}}$. Taking, e.g., $t=\mu/(32\rho^2)$, gives the desired result $E [T_0]=\Omega({\mu}/{\rho^2})$. For the cGA, we may simply regard the process $Z_k=p_k$. Since for all $k=0,1,\dots$, $$\begin{aligned} \Pr[Z_{k+1}={}&Z_k+\tfrac{1}{K}\mid Z_{1},\dots,Z_{k}]=p_k(1-p_k),\\ \Pr[Z_{k+1}={}&Z_k-\tfrac{1}{K}\mid Z_1,\dots,Z_k]=p_k(1-p_k),\\ \Pr[Z_{k+1}={}&Z_k\mid Z_1,\dots,Z_k]=1-2p_k(1-p_k),\end{aligned}$$ we have $E [Z_{k+1}\mid Z_1,\dots,Z_k ]=Z_k$. The martingale difference sequence $R_k:=Z_k-Z_{k-1}$ satisfies $\|R_k\|\le\tfrac{1}{K}$. By the Hoeffding-Azuma inequality, we have $$\Pr \left[ \max\limits_{k=1,\dots,t} \left\|p_k - \tfrac 12\right \| \ge M \right] = \Pr \left[ \max\limits_{k=1,\dots,t} \left\|\sum\limits_{i=1}^k{R_i} \right \| \ge M \right] \le 2 \exp \left(-\frac{M^2K^2}{2t} \right).$$ With $M = \frac 14$ and $t=K^2/32$, we have $E[T_0] =\Omega(K^2)$. We note that the lower bound proof for PBIL can be extended to CE, either by simply replacing $\rho$ by the supremum $\rho_{\sup} = \sup\{\rho_t \mid t \in {\ensuremath{\mathbb{N}}}\}$ and obtaining a lower bound of $\Omega(\mu/\rho_{\sup}^2)$, or by replacing $t\rho^2$ in by $\sum_{s=1}^t \rho_t^2$. With a suitable choice of $t$, this gives a bound taking into account the particular values of $(\rho_t)$. We omit the details. Since it might be useful to not only know a bound on the expected hitting time, but also a tail bound, e.g., to combine this with a union bound over all frequencies, we separately formulate the following statements, which were all shown in the proof of Theorem \[thm:lower\]. \[cor:tail\] Consider using an $n$-Bernoulli-$\lambda$-EDA to optimize some function $f$ with a neutral bit. Let $p_t, t = 0, 1, 2, \dots$ denote the frequency of the neutral bit after iteration $t$. 1. If the EDA is PBIL with learning rate $\rho$ and selection size $\mu$, then for all $\gamma > 0$ and $T \in {\ensuremath{\mathbb{N}}}$ we have $$\Pr[\forall t \in [0..T] : \|p_t - \tfrac 12\| < \gamma] \ge 1 - 2 \exp\left(-\frac{\gamma^2 \mu}{2\rho^2 T}\right).$$ 2. If the EDA is the cGA with hypothetical population size $K$, then for all $\gamma > 0$ and $T \in {\ensuremath{\mathbb{N}}}$ we have $$\Pr[\forall t \in [0..T] : \|p_t - \tfrac 12\| < \gamma] \ge 1 - 2 \exp\left(-\frac{\gamma^2 K^2}{2 T}\right).$$ Upper Bounds on the Boundary Hitting Time {#sec:upper} ========================================= We now prove that, roughly speaking, the lower bounds shown in the previous section are asymptotically tight. To prove our upper bounds, we use the following two auxiliary lemmas. For all $z\ge 0$ and $z_0>0$, we have $$\sqrt z\leq \sqrt z_0+\tfrac{1}{2}z_0^{-1/2}(z-z_0)-\tfrac{1}{8}z_0^{-3/2}(z-z_0)^2+\tfrac{1}{16}z_0^{-5/2}(z-z_0)^3.$$ \[lem:z\] For the convenience of the proof, let $x=\sqrt z$ and $a=\sqrt {z_0}$. We consider the function $$\begin{split} g(x)={}&x-a-\tfrac{1}{2}a^{-1}(x^2-a^2)+\tfrac{1}{8}a^{-3}(x^2-a^2)^2-\tfrac{1}{16}a^{-5}(x^2-a^2)^3\\ ={}&-\tfrac{1}{16}a^{-5}x^6+\tfrac{5}{16}a^{-3}x^4-\tfrac{15}{16}a^{-1}x^2+x-\tfrac{5}{16}a \end{split}$$ and show that $g(x) \le 0$. Since $$g'(x)=-\tfrac{3}{8}a^{-5}x^5+\tfrac{5}{4}a^{-3}x^3-\tfrac{15}{8}a^{-1}x+1$$ and $$\begin{split} g''(x)={}&-\tfrac{15}{8}a^{-5}x^4+\tfrac{15}{4}a^{-3}x^2-\tfrac{15}{8}a^{-1}\\ ={}&-\tfrac{15}{8}a^{-5}(x^4-2a^2x^2+a^4)=-\tfrac{15}{8}a^{-5}(x^2-a^2)^2\le 0, \end{split}$$ we know that $g'(x)$ is monotonically decreasing. Since $g'(0)=1$ and $g'(a)=0$, we observe that $g(x)$ increases in $[0,a)$ and decreases in $[a,\infty)$. Therefore, $g(x)\le g(a)=0$. An easy calculation gives the following second-order and third-order central moments of the frequency of the neutral bit in PBIL and the cGA. For PBIL, we have $$\begin{split} {\mathrm{Var}\xspace}[p_t\mid p_{t-1}]={}&\frac{\rho^2}{\mu}p_{t-1}(1-p_{t-1}),\\ E[(p_t-E[p_t\mid p_{t-1}])^3\mid p_{t-1}]={}&\frac{\rho^3}{\mu^2}p_{t-1}(1-p_{t-1})(1-2p_{t-1}). \end{split}$$ For the cGA, we have $$\begin{split} {\mathrm{Var}\xspace}[p_t\mid p_{t-1}]={}&\frac{2}{K^2}p_{t-1}(1-p_{t-1}),\\ E[(p_t-E[p_t\mid p_{t-1}])^3\mid p_{t-1}]={}&0. \end{split}$$ \[lem:moments\] For PBIL, note that $\sum_{i=1}^{\mu}X_{i,1}^t \sim \operatorname{Bin}(\mu, p_{t-1})$. Thus we have $$\begin{aligned} E\left[\left(\sum_{i=1}^{\mu}X_{i,1}^t\right)^2 \,\middle\|\, p_{t-1}\right] ={}& \mu(p_{t-1} - p_{t-1}^2+\mu p_{t-1}^2),\\ E\left[\left(\sum_{i=1}^{\mu}X_{i,1}^t\right)^3 \,\middle\|\, p_{t-1}\right] ={}& \mu(p_{t-1}+3(\mu-1)p_{t-1}^2+(\mu-1)(\mu-2)p_{t-1}^3).\end{aligned}$$ Hence, with (\[eq:pbil\]), we compute $$\begin{aligned} {\mathrm{Var}\xspace}[p_t\mid {}&p_{t-1}]=E[(p_t-E[p_t\mid p_{t-1}])^2 \mid p_{t-1}] = E\left[\left(-\rho p_{t-1}+\frac{\rho}{\mu}\sum_{i=1}^{\mu}X_{i,1}^t\right)^2 \,\middle\|\, p_{t-1}\right]\\ ={}& E\left[\rho^2 p_{t-1}^2 -2\rho p_{t-1}\frac{\rho}{\mu}\sum_{i=1}^{\mu}X_{i,1}^t+\frac{\rho^2}{\mu^2}\left(\sum_{i=1}^{\mu}X_{i,1}^t\right)^2 \,\middle\|\, p_{t-1}\right]\\ ={}& \rho^2 p_{t-1}^2 -2\rho p_{t-1}\frac{\rho}{\mu}\mu p_{t-1}+\frac{\rho^2}{\mu^2} \mu(p_{t-1} - p_{t-1}^2+\mu p_{t-1}^2)\\ ={}& \rho^2 p_{t-1}^2 -2\rho^2 p_{t-1}^2 +\frac{\rho^2}{\mu}(p_{t-1} - p_{t-1}^2+\mu p_{t-1}^2)\\ ={}& \frac{\rho^2}{\mu}p_{t-1}(1-p_{t-1})\end{aligned}$$ and $$\begin{aligned} E[(p_t-{}&E[p_t\mid p_{t-1}])^3\mid p_{t-1}] = E\left[\left(-\rho p_{t-1}+\frac{\rho}{\mu}\sum_{i=1}^{\mu}X_{i,1}^t\right)^3 \,\middle\|\, p_{t-1}\right]\\ ={}&E\left[-\rho^3 p_{t-1}^3+3\rho^2 p_{t-1}^2 \frac{\rho}{\mu}\sum_{i=1}^{\mu}X_{i,1}^t -3\rho p_{t-1}\frac{\rho^2}{\mu^2}\left(\sum_{i=1}^{\mu}X_{i,1}^t\right)^2+\frac{\rho^3}{\mu^3}\left(\sum_{i=1}^{\mu}X_{i,1}^t\right)^3\,\middle\|\, p_{t-1}\right]\\ ={}&-\rho^3 p_{t-1}^3+3\frac{\rho^3}{\mu} p_{t-1}^2 \mu p_{t-1}-3 \frac{\rho^3}{\mu^2} p_{t-1} \mu(p_{t-1} - p_{t-1}^2+\mu p_{t-1}^2) \\ {}&+ \frac{\rho^3}{\mu^3} \mu(p_{t-1}+3(\mu-1)p_{t-1}^2+(\mu-1)(\mu-2)p_{t-1}^3)\\ ={}& \frac{\rho^3}{\mu^2} (-\mu^2 p_{t-1}^3+ 3\mu^2 p_{t-1}^3-3\mu(p_{t-1}^2 - p_{t-1}^3+\mu p_{t-1}^3)+p_{t-1}\\ {}&+3(\mu-1)p_{t-1}^2+(\mu-1)(\mu-2)p_{t-1}^3)\\ ={}&\frac{\rho^3}{\mu^2} (p_{t-1}+3(\mu-1-\mu)p_{t-1}^2+(-\mu^2 +3\mu^2+3\mu-3\mu^2+(\mu-1)(\mu-2))p_{t-1}^3)\\ ={}&\frac{\rho^3}{\mu^2} (p_{t-1}-3p_{t-1}^2+2p_{t-1}^3)=\frac{\rho^3}{\mu^2}p_{t-1}(1-p_{t-1})(1-2p_{t-1}).\end{aligned}$$ Similarly, for the cGA, we compute $$\begin{aligned} {\mathrm{Var}\xspace}[p_t\mid {}&p_{t-1}]=E[(p_t-E[p_t\mid p_{t-1}])^2 \mid p_{t-1}]\\ ={}&p_{t-1}(1-p_{t-1}) \left(\frac{1}{K}\right)^2+ p_{t-1}(1-p_{t-1})\left(-\frac{1}{K}\right)^2=\frac{2}{K^2}p_{t-1}(1-p_{t-1})\end{aligned}$$ and $$\begin{aligned} E[(p_t-{}&E[p_t\mid p_{t-1}])^3\mid p_{t-1}] =p_{t-1}(1-p_{t-1}) \left(\frac{1}{K}\right)^3+ p_{t-1}(1-p_{t-1})\left(-\frac{1}{K}\right)^3=0. \qedhere\end{aligned}$$ We are now ready to prove the following upper bounds for the hitting time of the absorbing states of the frequency of a neutral bit. We consider EDAs without margins here, but it is clear that the upper bounds on the hitting times of absorbing states also hold for the hitting times of margins when they are present. Consider using an $n$-Bernoulli-$\lambda$-EDA to optimize some function $f$ with a neutral bit. - If the EDA is PBIL with $\rho < 1$, including the case of the $\lambda$-MMAS, then the following holds. Let $c\in (\tfrac{1}{2},\tfrac{1}{\sqrt 2})$. We say that the frequency $p_t$ of the neutral bit runs away from time $t$ on if 1. $p_t \le c \frac{\rho}{\mu}$ and in all iterations $t' > t$ all samples have a zero in the neutral bit, or 2. $p_t \ge 1 - c \frac{\rho}{\mu}$ and in all iterations $t' > t$ all samples have a one in the neutral bit. For $\tilde{T}$ denoting the first $t$ such that $p_t$ runs away from time $t$ on, we have $E[\tilde{T}] = O(\tfrac{\mu}{\rho^2})$. - If the EDA is the UMDA, that is, PBIL with $\rho = 1$, then the first hitting time $T$ of the absorbing states $\{0, 1\}$ satisfies $E[T] = O(\mu)$. - For the cGA, the expected first time to reach an absorbing state satisfies $E[T]=O(K^2)$. \[thm:upper\] Let $q_t=\min\{ p_t,1-p_t\}$ and $Y_t=\sqrt{q_t}$. Then $T=\min\{t\mid q_t=0\}$ and $\tilde{T}=\min\{t\mid q_t \le c\tfrac{\rho}{\mu}\}$. Due to the symmetry, we just discuss the case when $q_{t-1}=p_{t-1}$. Obviously, $p_{t-1}\le \tfrac{1}{2}$ in this case. Let us assume that $p_{t-1} > c\tfrac{\rho}{\mu}$. Using Lemma \[lem:z\] with $z=p_t$ and $z_0=p_{t-1}$, we have $$\begin{split} E[\sqrt{p_t}\mid p_{t-1}] \le{}& E[Y_{t-1}\mid p_{t-1}]+\tfrac{1}{2}p_{t-1}^{-1/2}E[p_t-p_{t-1}\mid p_{t-1}]\\ {}&-\tfrac{1}{8}p_{t-1}^{-3/2}E [(p_t-p_{t-1})^2\mid p_{t-1}] + \tfrac{1}{16}p_{t-1}^{-5/2}E [(p_t-p_{t-1})^3\mid p_{t-1} ] \end{split}$$ and thus $$\begin{split} E[Y_{t-1}{}&-\sqrt{p_t}\mid Y_{t-1}] \ge-\tfrac{1}{2}p_{t-1}^{-1/2}E[p_t-p_{t-1}\mid p_{t-1}]\\ +{}&\tfrac{1}{8}p_{t-1}^{-3/2}E [(p_t-p_{t-1})^2\mid p_{t-1} ] -\tfrac{1}{16}p_{t-1}^{-5/2}E [(p_t-p_{t-1})^3\mid p_{t-1} ]. \end{split} \label{eq:drift}$$ We analyze PBIL first. We start by showing that, regardless of $p_0$, the expected time to reach $p_t \in P \coloneqq [0,c\rho/{\mu}] \cup [1-c{\rho}/{\mu},1]$ is $O({\mu}/{\rho^2})$. Via Lemma \[lem:moments\], we have $$\begin{split} E[Y_{t-1}{}&-\sqrt{p_t}\mid Y_{t-1}]\\ \ge{}&\frac{1}{8}p_{t-1}^{-3/2}\frac{\rho^2}{\mu}p_{t-1}(1-p_{t-1})- \frac{1}{16}p_{t-1}^{-5/2}\frac{\rho^3}{\mu^2}p_{t-1}(1-p_{t-1})(1-2p_{t-1})\\ ={}&\frac{\rho^2}{16\mu}p_{t-1}^{-1/2}(1-p_{t-1}) \left(2-\frac{\rho}{\mu p_{t-1}}(1-2p_{t-1}) \right)\\ \ge{}&\frac{\rho^2}{16\mu}p_{t-1}^{-1/2}(1-p_{t-1})\left(2-\frac{1}{c}\right), \end{split}$$ where the last estimate follows from $p_{t-1} \ge c\rho/\mu$ and from the fact that $0<p_{t-1}\le \tfrac 12$ implies $0 \le 1-2p_{t-1}<1$. Since $p_{t-1}\le \tfrac{1}{2}$, we have ${p_{t-1}^{-1/2}(1-p_{t-1})\ge \tfrac{\sqrt 2}{2}}$. Hence $E[Y_{t-1}-\sqrt{p_t}\mid Y_{t-1}]\ge c_1\rho^2/\mu$, where $c_1=\tfrac{\sqrt2}{32}(2-\tfrac{1}{c})$. Using $q_t=\min\{p_t,1-p_t\}$, we have $$E[Y_{t-1}-Y_t\mid Y_{t-1}]\ge E[Y_{t-1}-\sqrt{p_t}\mid Y_{t-1}]\ge c_1\rho^2/\mu.\label{eq:driftA}$$ By artificially modifying the process $(Y_t)$ once it goes below $c\rho/\mu$, e.g., by defining $(\tilde{Y}_t)$ via $\tilde{Y}_t=Y_t$ if $Y_t\ge c\rho/\mu$ and $\tilde{Y}_t=0$ otherwise, we can ensure that we have a drift of $E[\tilde Y_{t-1} - Y_t \mid Y_t-1 > 0] \ge c_1 \rho^2 / \mu$ until we reach zero. Such an artificial extension of a process beyond the region of interest, to the best of our knowledge, was in the theory of evolutionary algorithms first used in [@DoerrHK11]. With this artificial extension we can now use the Additive Drift Theorem [@HeY01] with target $\tilde{Y}_t = 0$ and $\tilde{Y}_0=\sqrt{\tfrac{1}{2}}$ and obtain that the expected time for the $\tilde{Y}$-process to reach or go below $\sqrt{c \rho / \mu}$, equivalently to the $p_t$ process reaching $P$, is at most $\frac{\tilde{Y}_0}{c_1 \rho^2/\mu}=\frac{16}{2-1/c} \mu/\rho^2=O(\mu/\rho^2)$. Now we discuss the neutral frequency’s behavior once it has reached $P$. W.l.o.g. let $p_t \le c\rho/\mu$. Then the probability that all of the next $\mu\lceil 1/\rho \rceil$ samplings have a zero in the neutral bit is at least $$\begin{aligned} (1-p_t){}&^{\mu\lceil 1/\rho \rceil} \ge \left(1-\frac{c\rho}{\mu}\right)^{\mu\lceil 1/\rho \rceil} \ge \left(1-\frac{c\rho}{\mu}\right)^{\mu\frac{2}{\rho}} = \left(1-\frac{c\rho}{\mu}\right)^{2c\left(\frac{\mu}{c\rho}-1\right)}\left(1-\frac{c\rho}{\mu}\right)^{2c} \\ \ge{}& \exp(-2c)\left(1-2c\frac{c\rho}{\mu}\right) \ge \exp(-2c)(1-2c^2) > 0,\end{aligned}$$ where the second inequality uses $\lceil 1/\rho \rceil \le 2/\rho$ since $\rho \le 1$, the antepenultimate inequality uses the Bernoulli’s inequality, the penultimate inequality uses $\mu \ge 1$ and $\rho \le 1$, and the last inequality uses $c < 1/\sqrt 2$. In this case, the frequency after these $\lceil 1/\rho \rceil$ iterations is $$\begin{aligned} p_{t+\lceil 1/\rho \rceil}=(1-\rho)^{\lceil 1/\rho \rceil}p_t \le (1-\rho)^{1/\rho}p_t\le \frac{p_t}{e} \le \frac{c}{e}\frac{\rho}{\mu}. \end{aligned}$$ Therefore, with a similar calculation, it is easy to see that the probability that all of the next $\mu\lceil 1/\rho \rceil$ samplings have a zero in the neutral bit (from the $(t+\lceil 1/\rho \rceil+1)$-th iteration to the $(t+2\lceil 1/\rho \rceil)$-th iteration) is at least $(\exp(-2c)(1-2c^2))^{1/e}$, and ${p_{t+2/\rho}\le (c/e^2)(\rho/\mu)}$. A simple induction gives that the probability that all samplings have a zero in the neutral bit from the $(t+{(n-1)\lceil 1/\rho \rceil}+1)$-th iteration to the $(t+{n\lceil 1/\rho \rceil})$-th iteration is at least $(\exp(-2c)(1-2c^2))^{1/e^{n-1}}$. Therefore, the probability that only zeros are sampled in the neutral bit is at least $$\begin{aligned} \prod_{i=0}^{\infty}(\exp(-2c){}&(1-2c^2))^{1/e^{i}} = (\exp(-2c)(1-2c^2))^{\sum_{i=0}^{\infty} 1/e^{i}} \\ ={}& (\exp(-2c)(1-2c^2))^{1/(1-e^{-1})}>0,\end{aligned}$$ where the last inequality uses $\exp(-2c)(1-2c^2) > 0$. Let us divide the run of the EDA into phases. The first phase starts with the first iteration, each subsequent phase starts with the iteration following the end of the previous phase. A phase ends when for the first time after reaching in this phase a $p_t$-value in $P$ an unexpected value is sampled in the neutral bit. That is, when a one is sampled if the first $p_t$-value in $P$ is in $[0,c \frac{\rho}{\mu}]$ or when a zero is sampled when the first $p_t$-value is at least $1 - c \frac{\rho}{\mu}$. By the above, we know the following about these phases. Each phase, regardless of the past, has a positive (constant) probability of not ending. We call this a successful phase. Consequently, there is an expected constant number of phases, one of which successful (namely the last). In each phase, successful or not, it takes an expected time of $O(\mu/\rho^2)$ until the frequency of the neutral bit reaches a value in $P$. In the successful phase, the frequency then runs away. For the unsuccessful phases, we now show that the phase ends after an expected number of additional $O(1/\rho)$ iterations after reaching a frequency value in $P$. Note that this means analyzing a run of the algorithm starting (in iteration $t+1$) with the neutral frequency $p_t$ in $P$, say w.l.o.g. in $[0,c \frac{\rho}{\mu}]$, conditional on the event that at some future time a one is sampled in this bit. Let $U$ be the event that the phase under investigation is unsuccessful. Let $X \in \{1,2, \dots\}$ be minimal such that in iteration $t+X$ a one is sampled in the neutral bit of a selected individual. Conditional on $U$, the random variable $X$ is well-defined (that is, finite). For $X=s$ to hold, in particular no one can be sampled in the iterations $t+1, \dots, t+(s-1)$, and this implies not sampling a one in iteration $t+(s-1)$ when the current value of the frequency is $p_t (1-\rho)^{s-1}$. Consequently, the expected length (number of iterations) of an unsuccessful phase is $$\begin{aligned} E[X \mid U] &= \sum_{s = 1}^\infty s \Pr[X = s \mid U] = \frac{1}{\Pr[U]} \sum_{s=1}^\infty s \Pr[X = s] \nonumber\\ &\le \frac{1}{\Pr[U]} \sum_{s=1}^\infty s \mu p_t (1-\rho)^{(s-1)}\label{eq:sum}\end{aligned}$$ using a union bound over the $\mu$ samples in iteration $t+(s-1)$. To estimate this expectation, we first compute $\Pr[U]$. For any $k \in {\ensuremath{\mathbb{N}}}$, we have $$\begin{aligned} \Pr[U] & \ge \Pr[X \le k] = 1 - \Pr[X > k]\\ & = 1 - \prod_{i=1}^{k} \Pr[X > i \mid X > i-1]\\ & = 1 - \prod_{i=0}^{k-1} \left(1 - p_t(1 - \rho)^i\right)^\mu \\ & \ge 1 - \exp\left(-\mu p_t \sum_{i=0}^{k-1}(1-\rho)^i\right)\\ & = 1 - \exp\left(-\mu p_t \frac{1 - (1-\rho)^{k}}{1 - (1-\rho)}\right)\\ & \ge 1 - \left(1 - \frac 12 \mu p_t \frac{1 - (1-\rho)^{k}}{\rho}\right) = \mu p_t \frac{1 - (1-\rho)^{k}}{2 \rho}\end{aligned}$$ using the well-known estimates $1+x \le \exp(x)$ valid for all $x \in {\ensuremath{\mathbb{R}}}$ and $\exp(-x) \le 1 - \frac x2$ valid for all $0 \le x \le 1$. Taking the supremum over all $k \in {\ensuremath{\mathbb{N}}}$, we obtain $\Pr[U] \ge \frac{\mu p_t}{2\rho}$. To estimate the infinite sum in , we first recall the elementary formula $\sum_{s=1}^\infty s x^s = \frac{x}{(1-x)^2}$ for $0 < x < 1$, which follows from computing $A := \sum_{s=1}^\infty s x^s = x \sum_{s=1}^\infty (s-1) x^{s-1} + \sum_{s=1}^\infty x^s = xA + \frac{x}{1-x}$ and solving for $A$. From this, we obtain $$\begin{aligned} \sum_{s=1}^\infty s \mu p_t (1-\rho)^{(s-1)} = \mu p_t \frac{1}{\rho^2}\end{aligned}$$ and finally $$\begin{aligned} E[X \mid U] & \le \frac{\mu p_t \frac{1}{\rho^2}}{\frac{\mu p_t}{2\rho}} = \frac{2}{\rho}.\end{aligned}$$ Consequently, an unsuccessful phase in total takes an expected number of $O(\mu / \rho^2) + O(1 / \rho) = O(\mu / \rho^2)$ iterations. By Wald’s equation, recalling that we have an expected constant number of unsuccessful iterations, we see that the total time until the frequency of the neutral bit runs away is $O(\mu / \rho^2)$ iterations. For the cGA, in a similar manner as in the first part of the analysis for PBIL, by Lemma \[lem:moments\], equation (\[eq:drift\]) becomes $$\begin{split} E[Y_{t-1}-{}&\sqrt{p_t}\mid Y_{t-1}] \ge\frac{1}{8}p_{t-1}^{-3/2}\frac{2}{K^2}p_{t-1}(1-p_{t-1})\\ ={}&\frac{1}{4}p_{t-1}^{-1/2}\frac{1-p_{t-1}}{K^2} \ge \frac{1}{4}\frac{\sqrt2}{2}\frac{1}{K^2}=\frac{\sqrt 2}{8}\frac{1}{K^2}. \end{split}$$ Hence, $$E[Y_{t-1}-Y_t\mid Y_{t-1}]\ge E[Y_{t-1}-\sqrt{p_t}\mid Y_{t-1}]\ge \tfrac{\sqrt 2}{8}/K^2.$$ Via the Additive Drift Theorem [@HeY01] and $Y_0=\sqrt{\tfrac{1}{2}}$, we know that the expected time for the $Y$-process to reach zero is at most $Y_0 / \tfrac{\sqrt 2}{8K^2} = 4K^2$. We now briefly show that the upper bound proof can, under suitable assumptions, also be applied to CE with small modifications. Assume that the learning rate sequence $(\rho_t)$ has both supremum and infimum, and let ${\rho_{\sup}=\sup\{\rho_t \mid t\in {\ensuremath{\mathbb{N}}}\}}$ and ${\rho_{\inf}=\inf\{\rho_t \mid t\in {\ensuremath{\mathbb{N}}}\}}$. Consider the first generation when the frequency reaches ${\tilde{P} \coloneqq [0,c\rho_{\sup}/{\mu}] \cup [1-c{\rho_{\sup}}/{\mu},1]}$. Following similar arguments as above, we can obtain that the corresponding value in the right side of (\[eq:driftA\]) becomes $c_1\rho_{\inf}^2/\mu$, and hence the expected reaching time is $O(\mu/\rho_{\inf}^2)$. For the neural frequency’s behavior once it has reached $P$, we discuss the case when there exists a positive constant $c'<2$ so that $\rho_{\sup}/\rho_{\inf} \le c'$. In this case, we refine $c\in (1/2, \sqrt{1/(2c')})$. Then we can obtain that the probability that all samplings have a zero in the neutral bit from the $(t+i\lceil 1/\rho_{\inf} \rceil+1)$-th iteration to the $(t+(i+1)\lceil 1/\rho_{\inf} \rceil)$-th iteration is at least $$\begin{aligned} \left(\exp\left(-\frac{2c\rho_{\sup}}{\rho_{\inf}}\right)\left(1-\frac{2c^2\rho_{\sup}^2}{{\rho_{\inf}}}\right)\right)^{1/e^i}\ge \left(\exp(-2cc')(1-2c^2c')\right)^{1/e^i}>0\end{aligned}$$ for $i=0,1,\dots$, and the frequency after these $\lceil 1/\rho_{\inf} \rceil$ iterations is at most $c\rho_{\sup}/{(e^{i+1}\mu)}$. Hence, the probability that only zeros are sampled in the neutral bit is at least $$\begin{aligned} \left(\exp(-2cc')(1-2c^2c')\right)^{1/(1-e^{-1})}>0.\end{aligned}$$ Similarly, we could calculate that an unsuccessful phase ends after an expected number of additional $O(\rho_{\sup} / \rho_{\inf}^2)$ iterations after reaching a frequency value in $\tilde{P}$. Hence, for CE, the total time until the frequency of the neutral bit runs away is $O(\mu / \rho_{\inf}^2)$ iterations. We note that Theorem \[thm:lower\] and Theorem \[thm:upper\] give sharp bounds for several hitting times. For the UMDA without margins, the expected first time when the frequency of the neutral bit is absorbed in 0 or 1 is $\Theta(\mu)$, and the corresponding hitting time is $\Theta(K^2)$ for the cGA. For PBIL without margins and any $c\in (1/2, 1/\sqrt 2)$, the expected first time that the frequency of the neutral bit hits $c\rho/\mu$ or $1-c\rho/\mu$ is $\Theta({\mu}/{\rho^2})$. As discussed in the second paragraph in Section \[sec:IntroEDA\], these results also hold for the hitting time of the margins $\{1/D, 1-1/D\}$ when running EDAs with such margins. Extending the Lower Bounds to Bits with Preference: Domination Results {#sec:lowerP} ====================================================================== In the previous Sections \[sec:lower\] and \[sec:upper\], we discussed how fast neutral bits approach the boundaries of the frequency range. In many situations, e.g., for the benchmark functions [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}or [[<span style="font-variant:small-caps;">LeadingOnes</span>]{}]{}, bits are not neutral, but are neutral or have a preference of one bit-value (here the value one). Precisely, we say some bit, w.o.l.g., the first bit, of the fitness function $f$ is neutral or prefers a one (we also say weakly prefers a one) if and only if $$f(0,X_2,\dots,X_D) \le f(1,X_2,\dots,X_D)$$ for all $X_2,\dots,X_D\in\{0,1\}$. We say that the bit weakly prefers a zero if $f(0,X_2,\dots,X_D) \ge f(1,X_2,\dots,X_D)$ for all $X_2,\dots,X_D\in\{0,1\}$. If seems natural that for a bit that weakly prefers a one, the time for its frequency to reach or go below a certain value satisfies the same lower bounds as proven for neutral bits, and an analogous statement should be true for bits that weakly prefer a zero. This is what we show in this section. To prove this result, we first establish the following dominance result, which we expect to be useful also beyond this work. It in particular shows that when comparing two runs of an EDA, the first one starting with a higher frequency in a neutral bit than the second, then in the next generation the frequency of the neutral bit in the first run stochastically dominates the one in the second run. This statement remains true if the bit in the first run is not neutral, but weakly prefers ones. A simple induction extends this statement to all generations. While not important for our work, we add that we believe that the lemma below does not remain true when both functions can be such that the first bit weakly prefers a one. Also, simple examples show that our claim is false for the cGA without well-behaved frequencies. Consider using an $n$-Bernoulli-$\lambda$-EDA to optimize (i) some function $f$ such that the first bit weakly prefers a one and (ii) some function $g$ with the first bit being neutral. Assume that the first process is started with a frequency vector $u^0$ and the second with a frequency vector $v^0$ such that $u^0_i = v^0_i$ for $i=2, \dots, D$, and $u^0_1 \ge v^0_1$. Assume that in the case of the cGA, the well-behaved frequency assumption holds. Let $u^t$ and $v^t$ be the corresponding frequency vectors generated in the $t$-th generation. Then $u^t_1 \succeq v^t_1$ for all $t \in {\ensuremath{\mathbb{N}}}$. Analogously, if $f$ is such that the first bit weakly prefers a zero and we start with $u^0_1 \le v^0_1$, then $u^t_1 \preceq v^t_1$ for all $t \in {\ensuremath{\mathbb{N}}}$. \[lem:domi\] We only show the result for weak preference of a one as the other statement can be shown in an analogous fashion or by regarding $(-f, -g, 1-u, 1-v)$ instead of $(f,g,u,v)$. We first show the claim for the first iteration and later argue that an easy induction shows the claim for any time $t$. For PBIL (or CE), we recall from Section \[sec:Notation\] that in the second process in an iteration $t$ started with a frequency $v^{t-1}_1$ of the neutral first bit of $g$, the next frequency of this neutral bit is distributed as $$(1-\rho) v^{t-1}_1 + \rho \frac 1 \mu Y,\label{eq:pbilupdateneutral}$$ where $Y \sim \operatorname{Bin}(\mu,v^{t-1}_1)$. In the first process, a closer inspection of the update rule shows the frequency of the bit weakly preferring a one changes from $u^{t-1}_1$ to $$u_1^{t} \sim (1-\rho) u^{t-1}_1 + \rho \frac 1 \mu X,\label{eq:pbilupdateprefone}$$ where $X \succeq \operatorname{Bin}(\mu,v^{t-1}_1)$. If $u^0_1 \ge v^0_1$, then $\operatorname{Bin}(\mu,u^0_1)$ stochastically dominates $\operatorname{Bin}(\mu,v^0_1)$, and hence $u^1_1 \succeq v^1_1$ by and . For the cGA with the well-behaved frequency assumption, we note that $u^0_1 \ge v^0_1$ implies $u^0_1 = v^0_1$ or $u^0_1 \ge v^0_1 + 1/K$. We only regard the latter, more interesting case. We show $u^1_1 \succeq v^1_1$ using the definition of domination, that is, that ${\Pr[u^1_1 \le \lambda] \le \Pr[v^1_1 \le \lambda]}$ holds for all $\lambda \in {\ensuremath{\mathbb{R}}}$. We discuss differently the following three cases. - Assume $\lambda <v^0_1$. Since $u^0_1 - 1/K \ge v^0_1 > \lambda$ from our assumption, we have ${\Pr[u^1_1 \le \lambda] = 0 \le \Pr[v^1_1 \le \lambda]}$. - Assume $v^0_1 \le \lambda < u^0_1$. In this case, $\Pr[u^1_1 \le \lambda] \le u^0_1 (1-u^0_1) \le \tfrac 14$ and $\Pr[v^1_1 \le \lambda] = 1 - v^0_1 (1-v^0_1) \ge 1 - \tfrac 14$, which gives the claim. - Assume $\lambda \ge u^0_1$. Since $v^0_1+1/K \le u^0_1 \le \lambda$ from our assumption, we have $\Pr[v^1_1 \le \lambda] = 1 \ge \Pr[u^1_1 \le \lambda]$. Hence, we have $u^1_1 \succeq v^1_1$. To extend our proof to arbitrary generation $t$, we note that if we have $u_1^{t-1} \succeq v_1^{t-1}$, then (see, e.g., [@Doerr19tcs Theorem 12]) we can find a coupling of the two probability spaces describing the states of the two algorithms at the start of iteration $t$ in such a way that for any point $\omega$ in the coupling probability space we have $u_1^{t-1} \ge v_1^{t-1}$. Conditional on this $\omega$, we can use the above argument for one iteration and obtain $u_1^{t} \succeq v_1^{t}$. This implies that we also have $u_1^{t} \succeq v_1^{t}$ without conditioning on an $\omega$. From Lemma \[lem:domi\], now easily derive that our lower bounds shown in Section \[sec:lower\], suitable adjusted, also hold for bits that weakly prefer one value. \[thm:lowerP1\] Consider using an $n$-Bernoulli-$\lambda$-EDA to optimize some function $f$ with a bit weakly preferring a one. Let $p_t, t = 0, 1, 2, \dots$ denote the frequency of this bit after iteration $t$. Let $T_0 = \min\{t \mid p_t \le \tfrac 14\}$ denote the first time this frequency is in $[0,\frac 14]$. 1. Let the EDA be PBIL with learning rate $\rho$ and selection size $\mu$. Then $E[T_0]=\Omega(\tfrac{\mu}{\rho^2})$, in particular, $E [T_0]=\Omega({\mu})$ for the UMDA and $E [T_0 ]=\Omega(\tfrac{1}{\rho^2})$ for the $\lambda$-MMAS. Again for PBIL, for all $\gamma > 0$ and $T \in {\ensuremath{\mathbb{N}}}$ we have $$\Pr[\forall t \in [0..T] : p_t > \tfrac 12 - \gamma] \ge 1 - 2 \exp\left(-\frac{\gamma^2 \mu}{2\rho^2 T}\right).$$ 2. Let the EDA be the cGA with hypothetical population size $K$. Then $E[T_0]=\Omega(K^2)$ and for all $\gamma > 0$ and $T \in {\ensuremath{\mathbb{N}}}$ we have $$\Pr[\forall t \in [0..T] : p_t > \tfrac 12 - \gamma] \ge 1 - 2 \exp\left(-\frac{\gamma^2 K^2}{2 T}\right).$$ Let $g$ be some function with first bit truly neutral, let $\tilde p_t, t = 0, 1, 2, \dots$ denote the frequency of this bit after iteration $t$, and let $\tilde T_0 = \min\{t \mid \tilde p_t \le \tfrac 14\}$ denote the first time this frequency is in $[0,\frac 14]$. Noting that $\tilde p_0 = p_0 = \tfrac 12$, we apply Lemma \[lem:domi\] and observe that $p_t \succeq \tilde p_t$ for all $t$. This together with Corollary \[cor:tail\] shows the tail bounds. From $p_t \succeq \tilde p_t$ for all $t$, we also deduce $T_0 \succeq \tilde T_0 \succeq \min\{t \mid p_t \in [0,\tfrac 14] \cup [\tfrac 34, 1]\} =: T_0'$ and thus $E[T_0] \ge E[T_0']$. By Theorem \[thm:lower\], $T_0'$ satisfies the lower bounds we claim for the expectation of $T_0$, and so does $T_0$ itself. In an analogous fashion, we obtain the corresponding result for bits weakly preferring a zero. \[thm:lowerP0\] Consider using an $n$-Bernoulli-$\lambda$-EDA to optimize some function $f$ with a bit weakly preferring a zero. Let $p_t, t = 0, 1, 2, \dots$ denote the frequency of this bit after iteration $t$. Let $T_0 = \min\{t \mid p_t \ge \tfrac 34\}$ denote the first time this frequency is in $[\frac 34, 1]$. 1. Let the EDA be PBIL with learning rate $\rho$ and selection size $\mu$. Then $E[T_0]=\Omega(\tfrac{\mu}{\rho^2})$, in particular, $E [T_0]=\Omega({\mu})$ for the UMDA and $E [T_0 ]=\Omega(\tfrac{1}{\rho^2})$ for the $\lambda$-MMAS. Again for PBIL, for all $\gamma > 0$ and $T \in {\ensuremath{\mathbb{N}}}$ we have $$\Pr[\forall t \in [0..T] : p_t < \tfrac 12 + \gamma] \ge 1 - 2 \exp\left(-\frac{\gamma^2 \mu}{2\rho^2 T}\right).$$ 2. Let the EDA be the cGA with hypothetical population size $K$. Then $E[T_0]=\Omega(K^2)$ and for all $\gamma > 0$ and $T \in {\ensuremath{\mathbb{N}}}$ we have $$\Pr[\forall t \in [0..T] : p_t < \tfrac 12 + \gamma] \ge 1 - 2 \exp\left(-\frac{\gamma^2 K^2}{2 T}\right).$$ We have just extended our previous lower bounds to the case of bits preferring a particular value. One may ask whether similar results can be obtained for upper bounds as well. Let us comment on this question. Let us, as in Theorem \[thm:lowerP1\] and its proof, denote by $p_t$ the frequencies of a bit preferring a one and by $T_0$ the first time this frequency has reached or exceeded a particular value (e.g., $\tfrac 34$ or the upper boundary of the frequency range). Let us denote by $\tilde p_t$ and $\tilde T_0$ the corresponding random variables for a neutral bit. Then again $p_t \succeq \tilde p_t$ implies $T_0 \preceq \tilde T_0$, so (informally speaking or made precise via a coupling argument) $p_t$ reaches the target not later than $\tilde p_t$. However, we do not have any good upper bounds on $\tilde T_0$, neither on its expectation nor in the domination sense. On the technical side, the reason is that we regarded the symmetric process $q_t = \min\{p_t, 1-p_t\}$ in Section \[sec:upper\]. The true reason is that also the process itself (when regarding a neutral bit) is symmetric: With probability $\tfrac 12$ each, the first visit to a boundary is to $\tfrac 1D$ and to $1 - \frac 1D$. However, if the first visit is to $\tfrac 1D$, then it takes quite some time to reach $1 - \tfrac 1D$. Consequently, the distribution of the first hitting time of $1 - \tfrac 1D$ is not well concentrated, and consequently, its expectation might be significantly larger than the first hitting time of $\{\tfrac 1D, 1 - \tfrac 1D\}$. For this reason, we currently do not see how our domination arguments allow to deduce from our results on neutral bits reasonable upper bounds on hitting times of frequencies of bits with weak preferences. However, we expect that in most situations where bits with weak preferences occur, one would rather try to exploit the preference to show stronger upper bounds than in the neutral case. For this reason, trying to retrieve information from the neutral case might not be too interesting anyway. Discussion {#sec:disc} ========== Just like classic evolutionary algorithms, EDAs are subject to genetic drift and this can, even when using margins for the frequency range, lead to a suboptimal performance. For several classical EDAs, this paper proved the first sharp estimates of the expected time the sampling frequency of a neutral bit takes to leave the middle range $[\frac 13, \frac 34]$ or to reach the boundaries. These times, roughly speaking, are $\Theta(K^2)$ iterations for the cGA and $\Theta(\mu/\rho^2)$ iterations for PBIL (and consequently $\Theta(\mu)$ for its special case UMDA). These results are useful both to interpret existing performance results and to set the parameters right in future applications of EDAs. As an example of the former, we note that the recent work [@LehreN19foga] shows that the UMDA with $c \log D \le \mu = o(D)$, $c$ a sufficiently large constant, with $\lambda \le 71 \mu$, and with the margins $1/D$ and $1 - 1/D$, has a weak performance of $\exp(\Omega(\mu))$ on the $D$-dimensional <span style="font-variant:small-caps;">DeceptiveLeadingBlocks</span> benchmark function. This runtime is at least some unspecified, but most likely large polynomial in $D$; it is super-polynomial as soon as $\mu$ is chosen super-logarithmic. For our work, we know that the expected time for the frequency of a neutral bit to reach the boundaries is only $O(\mu)$ iterations. Since the <span style="font-variant:small-caps;">DeceptiveLeadingBlocks</span> function, similar to the classic [[<span style="font-variant:small-caps;">LeadingOnes</span>]{}]{}function, has many bits that for a long time behave like neutral, a value of $\mu = o(D)$ results in that a constant fraction of these currently neutral bits will have reached the boundaries at least once within the first $D$ iterations. Hence also without looking at the proof of the result in [@LehreN19foga], which indeed exploits the fact that frequencies reach the margins to show the weak performance, our results already indicate that the weak performance might be caused by the use of parameter values leading to strong genetic drift. For a practical use of EDAs, our tail bounds of Corollary \[cor:tail\] can be helpful. As a quick example, assume one wants to optimize some function via the cGA and one is willing to spend a computational budget of $F$ fitness evaluations. Since the cGA performs two fitness evaluations per iteration, this is equivalent to saying that we have a budget of $T = F/2$ iterations. From Corollary \[cor:tail\](b), with $\gamma = 1/4$, and a simple union bound over the $D$ bits, we see that the probability that one of the (temporarily) neutral bits leaves the middle range $[\tfrac 14, \frac 34]$ is at most ${D \cdot 2 \exp(-\frac{\gamma^2 K^2}{2T})}$. Consequently, by using a parameter value of $K \ge \frac 1\gamma \sqrt{F \ln(20D)}$, we obtain that with probability at least 90% no neutral bit leaves the middle range (and, with the results of Section \[sec:lowerP\], no bit that weakly prefers one bit value leaves the middle range into the opposite direction). Phrased differently, this means that within this time frame, only those bits approach the boundaries for which there is a sufficiently strong signal from the objective function. While this consideration cannot determine optimal parameters for each EDA and each objective function, it can at least prevent the user from taking parameters that are likely to give an inferior performance due to genetic drift. Since genetic drift has been shown to lead to a poor performance in the past, we strongly recommend to choose the parameters $K$ and $\mu$ large enough so that estimates based on Corollary \[cor:tail\] guarantee that bits without a fitness signal stay in the middle range. Acknowledgements {#acknowledgements .unnumbered} ---------------- This work was supported by a public grant as part of the Investissement d’avenir project, reference ANR-11-LABX-0056-LMH, LabEx LMH, in a joint call with Gaspard Monge Program for optimization, operations research and their interactions with data sciences. This work was also supported by National Key R&D Program of China (Grant No. 2017YFC0804003); the Program for Guangdong Introducing Innovative and Enterpreneurial Teams (Grant No. 2017ZT07X386); Shenzhen Peacock Plan (Grant No. KQTD2016112514355531); the Science and Technology Innovation Committee Foundation of Shenzhen (Grant No. ZDSYS201703031748284) and the Program for University Key Laboratory of Guangdong Province (Grant No. 2017KSYS008). [DBIJV96]{} Hideki Asoh and Heinz M[ü]{}hlenbein. On the mean convergence time of evolutionary algorithms without selection and mutation. In [International Conference on Parallel Problem Solving from Nature, PPSN 1994]{}, pages 88–97. Springer, 1994. Shummet Baluja. Population-based incremental learning: A method for integrating genetic search based function optimization and competitive learning. Technical report, Carnegie Mellon University, Pittsburgh, PA, USA, 1994. Shumeet Baluja and Rich Caruana. Removing the genetics from the standard genetic algorithm. In [International Conference on Machine Learning, ICML 1995]{}, pages 38–46. Elsevier, 1995. Andre Costa, Owen Dafydd Jones, and Dirk Kroese. Convergence properties of the cross-entropy method for discrete optimization. , 35(5):573–580, 2007. Jeremy S. De Bonet, Charles Lee Isbell Jr., and Paul A. Viola. Mimic: Finding optima by estimating probability densities. In [Advances in Neural Information Processing Systems, NIPS 1996]{}, pages 424–430. [MIT]{} Press, 1996. Benjamin Doerr, Edda Happ, and Christian Klein. Tight analysis of the (1+1)-[EA]{} for the single source shortest path problem. , 19:673–691, 2011. Benjamin Doerr. Analyzing randomized search heuristics via stochastic domination. , 773:115–137, 2019. Benjamin Doerr. A tight runtime analysis for the [cGA]{} on jump functions: [EDA]{}s can cross fitness valleys at no extra cost. In [Genetic and Evolutionary Computation Conference, GECCO 2019]{}, pages 1488–1496. [ACM]{}, 2019. Stefan Droste. A rigorous analysis of the compact genetic algorithm for linear functions. , 5:257–283, 2006. Tobias Friedrich, Timo K[ö]{}tzing, and Martin S. Krejca. cannot be balanced and stable. In [Genetic and Evolutionary Computation Conference, GECCO 2016]{}, pages 1139–1146. ACM, 2016. Cristina Gonz[á]{}lez, Jose A. Lozano, and Pedro Larra[ñ]{}aga. The convergence behavior of the [PBIL]{} algorithm: a preliminary approach. In [International Conference on Artificial Neural Nets and Genetic Algorithms, ICANNGA 2001]{}, pages 228–231. Springer, 2001. Georges R. Harik, Fernando G. Lobo, and David E. Goldberg. The compact genetic algorithm. In [International Conference on Evolutionary Computation, ICEC 1998]{}, pages 523–528. IEEE, 1998. Georges R. Harik, Fernando G. Lobo, and Kumara Sastry. Linkage learning via probabilistic modeling in the extended compact genetic algorithm ([ECGA]{}). In [Scalable Optimization via Probabilistic Modeling]{}, pages 39–61. Springer, 2006. Jun He and Xin Yao. Drift analysis and average time complexity of evolutionary algorithms. , 127:51–81, 2001. Martin S. Krejca and Carsten Witt. Lower bounds on the run time of the univariate marginal distribution algorithm on [OneMax]{}. In [Foundations of Genetic Algorithms, FOGA 2017]{}, pages 65–79. ACM, 2017. Martin S. Krejca and Carsten Witt. Theory of estimation-of-distribution algorithms. , abs/1806.05392, 2018. Per Kristian Lehre and Phan Trung Hai Nguyen. On the limitations of the univariate marginal distribution algorithm to deception and where bivariate [EDA]{}s might help. In [Foundations of Genetic Algorithms, FOGA 2019]{}, pages 154–168, 2019. Johannes Lengler, Dirk Sudholt, and Carsten Witt. Medium step sizes are harmful for the compact genetic algorithm. In [Genetic and Evolutionary Computation Conference, [GECCO]{} 2018]{}, pages 1499–1506. [ACM]{}, 2018. Colin McDiarmid. Concentration. In [Probabilistic Methods for Algorithmic Discrete Mathematics]{}, pages 195–248. Springer, Berlin, 1998. Heinz Mühlenbein and Thilo Mahnig. - a scalable evolutionary algorithm for the optimization of additively decomposed functions. , 7(4):353–376, 1999. Kimura Motoo. Diffusion models in population genetics. , 1(2):177–232, 1964. Heinz M[ü]{}hlenbein and Gerhard Paass. From recombination of genes to the estimation of distributions [I]{}. [B]{}inary parameters. In [International Conference on Parallel Problem Solving from Nature, PPSN 1996]{}, pages 178–187. Springer, 1996. Frank Neumann, Dirk Sudholt, and Carsten Witt. A few ants are enough: [ACO]{} with iteration-best update. In [Genetic and Evolutionary Computation Conference, [GECCO]{} 2010]{}, pages 63–70. [ACM]{}, 2010. Martin Pelikan and Heinz M[ü]{}hlenbein. The bivariate marginal distribution algorithm. In [Advances in Soft Computing]{}, pages 521–535. Springer, 1999. Dirk Sudholt and Carsten Witt. Update strength in [EDAs]{} and [ACO]{}: How to avoid genetic drift. In [Genetic and Evolutionary Computation Conference, GECCO 2016]{}, pages 61–68. ACM, 2016. Carsten Witt. Upper bounds on the running time of the univariate marginal distribution algorithm on [OneMax]{}. , 81:632–667, 2019. Weijie Zheng, Guangwen Yang, and Benjamin Doerr. Working principles of binary differential evolution. In [Genetic and Evolutionary Computation Conference, GECCO 2018]{}, pages 1103–1110. ACM, 2018. [^1]: A small subset of the results presented in this work were already stated, without proof or proof idea, in the conference paper [@ZhengYD18 Theorem 4.5], namely that the expected time the frequency of a neutral bit takes to hit the absorbing states $0$ or $1$ is $\Theta(K^2)$ for cGA and $\Theta(\mu)$ for UMDA. Our work now extends the UMDA result to the PBIL, strenghthens all lower bounds by regarding the event of leaving the middle range $[\frac 14,\frac 34]$ of the frequency range, adds a tail bound for the lower bounds, and adds domination arguments allowing to extend the lower bounds to bits that are neutral or prefer a particular value. Also, complete proofs are given for all results. Both authors contributed equally to this work and both act as corresponding authors.	{ "pile_set_name": "ArXiv" }
--- abstract: 'A model of six-species food web is studied in the viewpoint of spatial interaction structures. Each species has two predators and two preys, and it was previously known that the defensive alliances of three cyclically predating species self-organize in two-dimensions. The alliance-breaking transition occurs as either the mutation rate is increased or interaction topology is randomized in the scheme of the Watts-Strogatz model. In the former case of temporal disorder, via the finite-size scaling analysis the transition is clearly shown to belong to the two-dimensional Ising universality class. In contrast, the geometric or spatial randomness for the latter case yields a discontinuous phase transition. The mean-field limit of the model is analytically solved and then compared with numerical results. The dynamic universality and the temporally periodic behaviors are also discussed.' author: - Beom Jun Kim - Jianbin Liu - Jaegon Um - 'Sung-Ik Lee' title: 'Instability of defensive alliances in the predator-prey model on complex networks' --- The rock-scissors-paper (RSP) game [@ecoli] gives a typical three-strategy model of cyclical predator-prey food chain if rock, scissors, and paper are replaced by different competitive species in biology. In this simple game, a rock beats a pair of scissors, a pair of scissors beat a sheet of paper, and a sheet of paper beats a rock. As generalizations of RSP game, the food webs composed of different number of species show a variety of interesting behaviors [@pre54-6186]. Especially, the models proposed by Szabó et al. in Refs. [@szabo1] and [@szabo2] have been shown to self-organize to form the defensive alliances, which becomes unstable as the mutation rate is increased. ![(a) The interaction topology used in this work (Model A in Ref. [@szabo1]). Arrow e.g., from 0 to 1 indicates that the species 0 is a predator to the species 1 (and thus the species 1 is the prey to the species 0). As time goes on the initial random configuration in (b) evolves to (c), where only three species \[(0,2,4) or (1,3,5)\] exist to form the defensive alliance. Different colors indicate different species in (b) and (c). (Color online.) []{data-label="figa"}](fig1a.eps "fig:"){width="15.00000%"} ![(a) The interaction topology used in this work (Model A in Ref. [@szabo1]). Arrow e.g., from 0 to 1 indicates that the species 0 is a predator to the species 1 (and thus the species 1 is the prey to the species 0). As time goes on the initial random configuration in (b) evolves to (c), where only three species \[(0,2,4) or (1,3,5)\] exist to form the defensive alliance. Different colors indicate different species in (b) and (c). (Color online.) []{data-label="figa"}](fig1b.eps "fig:"){width="15.00000%"}![(a) The interaction topology used in this work (Model A in Ref. [@szabo1]). Arrow e.g., from 0 to 1 indicates that the species 0 is a predator to the species 1 (and thus the species 1 is the prey to the species 0). As time goes on the initial random configuration in (b) evolves to (c), where only three species \[(0,2,4) or (1,3,5)\] exist to form the defensive alliance. Different colors indicate different species in (b) and (c). (Color online.) []{data-label="figa"}](fig1c.eps "fig:"){width="15.00000%"} In the present study, we use the model A in Ref. [@szabo1] described by the interaction topology in Fig. \[figa\](a), where the six species weave a food web that is more complicated than the RSP game: Each species has two predators, two preys and one neutral interacting partner. There indeed exists ecological systems showing the cyclic dominance: Examples include the three-morph mating system in the side-blotched lizard in which each morph dominates another morph when rare [@lizard]. Another example is the system of three different populations of [Escherichia coli]{} [@ecoli] composed of toxin-producing (T), toxin-resistant (R), and toxin-sensitive (S) strains. If the growth rate of each group satisfies S $>$ R $>$ T, the RSP game captures the cyclic dominance in the system: T dominates S (S is killed by T), S dominates R due to the higher growth rate, and R dominates T from the same reason. If bacteria produce two different toxins, the food web is constructed from nine different species as has been well studied in Ref. [@szabo2]. The main interest here is to study the effect of the spatial interaction structure and for that purpose we play the predator-prey game on the complex network structure of the Watts-Strogatz (WS) network [@WS] constructed on two-dimensions (2D) as follows: (1) We first build the 2D $L\times L$ ($N \equiv L^2)$ regular square lattice. (2) Every bond is visited once and with the rewiring probability $\alpha$ is rewired to the randomly chosen other site. The above procedure then yields a network structure which possesses characteristics such as the short characteristic path length [@WS; @networkreview]. Once the network is built as described above, the time evolution of the system obeys the following rules [@szabo1]: (1) Six species are scattered randomly on a square lattice as in Fig. \[figa\](b), then (2) the species on each randomly chosen site is mutated to one of its predating species with the mutation rate $P$. (3) If no mutation occurs (with the probability $1-P$), one of the nearest neighbors is chosen to interact, and the dominant one survives and invades the subordinated one. For example, if the pair of species 0 and 1 are chosen, the species 1 is replaced by the species 0 \[see Fig. \[figa\](a)\]. If two neutral partners has been chosen, i.e., no arrow connects the two species in Fig. \[figa\](a), nothing happens. For the 2D regular square lattice corresponding to the rewiring probability $\alpha=0$, it has been found that the defensive alliances composed of three species, (0,2,4) or (1,3,5), are spontaneously formed at a small mutation rate and the other species that does not belong to the alliance dies out [@szabo1] \[see Fig. \[figa\](c)\]. As the mutation rate is increased, it has been shown [@szabo1] that the defensive alliance becomes unstable and there occurs well-defined phase transition of the universality class of the 2D Ising model [@textbook], which is expected because of the existence of the $Z_2$ symmetry: Interchange of two alliances (0,2,4) $\leftrightarrow$ (1,3,5) do not change the game rules, which we call here [inter-alliance symmetry]{}. The spontaneous breaking of the defensive alliance originates either from the high mutation rate $P$ or from the high degree of structural randomness controlled by the rewiring probability $\alpha$. In other words, the instability of the defensive alliance is induced either by the temporal randomness or by the structural randomness. In reality, these two different types of randomness may coexist. However, from the practical computational difficulty we in this paper only investigate the effect of each randomness separately. We first investigate the phase transition for the 2D regular square lattice corresponding to the WS network with the rewiring probability $\alpha=0$. Although this was previously studied [@szabo1], we in this work use the extensive finite-size scaling analysis to confirm not only the 2D Ising static universality class but also to identify the dynamic universality class detected by the dynamic critical exponent $z$. In order to describe the alliance breaking transition we define the order parameter which functions as a magnetization in the Ising model: $$\label{eq:m} m=(c_0+c_2+c_4)-(c_1+c_3+c_5),$$ where $c_s \equiv N_s/N$ $(s=0,\cdots ,5)$ is the density of species $s$ with $N_s$ being the number of sites occupied by $s$. When $P$ is sufficiently small, the defensive alliance is formed with $c_0 \approx c_2 \approx c_4 \approx 1/3$ and $c_1 \approx c_3 \approx c_5\approx 0$ or vice versa, leading to $m \approx \pm 1$ (ordered phase). When the system recovers its full symmetry at a high mutation rate, all species have the same density, yielding $m \approx 0$. Only for convenience we use $$\mu \equiv \ln(1/P)$$ as a control parameter. High mutation rates correspond to the small values of the mutation parameter $\mu$, and thus qualitatively speaking, the physical meaning of $\mu$ resembles that of the inverse temperature in the standard statistical mechanics. Numerical simulations are performed on systems of sizes varying from $L=32$ to 192 with the periodic boundary conditions employed. After equilibration for $10^4$ steps per site, which is sufficiently long enough, the thermal average is computed for later $10^4$ steps at least. The measured quantities are $\langle \|m\| \rangle$, the Binder’s cumulant [@binder] $$U_L \equiv 1-\frac{\langle m^4 \rangle}{3\langle m^2 \rangle^2},$$ and the susceptibility $$\chi \equiv N \bigl(\langle m^2\rangle-\langle \|m\|\rangle^2 \bigr),$$ where $\langle...\rangle$ denotes the thermal average. In order to study dynamic critical behavior, we also measure the autocorrelation function as a function of time $t$ defined by $$\label{eq:Ct1} C(t) \equiv \langle \|m(t)m(0)\|\rangle -\langle \|m\|\rangle^2 .$$ The finite-size scaling forms of measured quantities are written as [@FSS] $$\begin{aligned} & & \langle \|m\| \rangle = L^{-\beta/\nu}\widetilde{m}\bigl((\mu-\mu_c) L^{1/\nu}\bigr), \label{eq:mscale} \\ & & U_L = \widetilde{U}\bigl((\mu-\mu_c)L^{1/\nu}\bigr), \\ & & \chi = L^{\gamma/\nu}\widetilde{\chi}\bigl((\mu-\mu_c)L^{1/\nu}\bigr), \\ & & C(t)/C(0) = \widetilde{C}(t L^{-z}), \label{eq:Ct2}\end{aligned}$$ where $\widetilde{m},\widetilde{U},\widetilde{\chi}$, and $\widetilde{C}$ are suitable scaling functions, $\mu_c$ is the critical value of $\mu$, and $\beta$, $\gamma$, $\nu$ are standard critical exponents [@textbook] while $z$ is the dynamic critical exponent defined at $\mu_c$ from the divergence of the relaxation time scale ($\tau \sim L^z$). ![ Phase transition in the 2D regular lattice in terms of the mutation parameter $\mu$. (a) The order parameter $\langle \|m\| \rangle$ versus $\mu$ clearly shows the existence of the alliance breaking transition. (b) A unique crossing at $\mu_c=6.50(4)$ with $\beta/\nu\approx1/8$ is obtained from the finite-size scaling form of $\langle \|m\| \rangle$. (c) All data points for $\langle \|m\| \rangle$ in (b) collapse into to a smooth curve by using scaled variables. (d) Binder’s cumulant at different sizes cross at $\mu_c=6.50(4)$ in agreement with (b). The inset in (d) shows that $\Delta U_L \sim L^{1/\nu}$ to yield $\nu \approx 1$, where $\Delta U_L \equiv U_L(\mu_1) - U_L(\mu_2)$ with $\mu_{1,2}$ are two adjacent points near $\mu_c$. (e) The susceptibility $\chi$. The inset shows a log-log plot of $\chi$ versus $L$ at $\mu_c$ leading to $\gamma \approx 7/4$. (f) The autocorrelation function $C(t)/C(0)$ at $\mu_c$ versus $tL^{-1.9}$ with $z=1.9$. All curves for different sizes collapse well to a single curve, indicating that the dynamic critical exponent $z=1.9(2)$. (Color online)[]{data-label="fig2d"}](fig2a.eps "fig:"){width="25.00000%"}![ Phase transition in the 2D regular lattice in terms of the mutation parameter $\mu$. (a) The order parameter $\langle \|m\| \rangle$ versus $\mu$ clearly shows the existence of the alliance breaking transition. (b) A unique crossing at $\mu_c=6.50(4)$ with $\beta/\nu\approx1/8$ is obtained from the finite-size scaling form of $\langle \|m\| \rangle$. (c) All data points for $\langle \|m\| \rangle$ in (b) collapse into to a smooth curve by using scaled variables. (d) Binder’s cumulant at different sizes cross at $\mu_c=6.50(4)$ in agreement with (b). The inset in (d) shows that $\Delta U_L \sim L^{1/\nu}$ to yield $\nu \approx 1$, where $\Delta U_L \equiv U_L(\mu_1) - U_L(\mu_2)$ with $\mu_{1,2}$ are two adjacent points near $\mu_c$. (e) The susceptibility $\chi$. The inset shows a log-log plot of $\chi$ versus $L$ at $\mu_c$ leading to $\gamma \approx 7/4$. (f) The autocorrelation function $C(t)/C(0)$ at $\mu_c$ versus $tL^{-1.9}$ with $z=1.9$. All curves for different sizes collapse well to a single curve, indicating that the dynamic critical exponent $z=1.9(2)$. (Color online)[]{data-label="fig2d"}](fig2b.eps "fig:"){width="25.00000%"} ![ Phase transition in the 2D regular lattice in terms of the mutation parameter $\mu$. (a) The order parameter $\langle \|m\| \rangle$ versus $\mu$ clearly shows the existence of the alliance breaking transition. (b) A unique crossing at $\mu_c=6.50(4)$ with $\beta/\nu\approx1/8$ is obtained from the finite-size scaling form of $\langle \|m\| \rangle$. (c) All data points for $\langle \|m\| \rangle$ in (b) collapse into to a smooth curve by using scaled variables. (d) Binder’s cumulant at different sizes cross at $\mu_c=6.50(4)$ in agreement with (b). The inset in (d) shows that $\Delta U_L \sim L^{1/\nu}$ to yield $\nu \approx 1$, where $\Delta U_L \equiv U_L(\mu_1) - U_L(\mu_2)$ with $\mu_{1,2}$ are two adjacent points near $\mu_c$. (e) The susceptibility $\chi$. The inset shows a log-log plot of $\chi$ versus $L$ at $\mu_c$ leading to $\gamma \approx 7/4$. (f) The autocorrelation function $C(t)/C(0)$ at $\mu_c$ versus $tL^{-1.9}$ with $z=1.9$. All curves for different sizes collapse well to a single curve, indicating that the dynamic critical exponent $z=1.9(2)$. (Color online)[]{data-label="fig2d"}](fig2c.eps "fig:"){width="25.00000%"}![ Phase transition in the 2D regular lattice in terms of the mutation parameter $\mu$. (a) The order parameter $\langle \|m\| \rangle$ versus $\mu$ clearly shows the existence of the alliance breaking transition. (b) A unique crossing at $\mu_c=6.50(4)$ with $\beta/\nu\approx1/8$ is obtained from the finite-size scaling form of $\langle \|m\| \rangle$. (c) All data points for $\langle \|m\| \rangle$ in (b) collapse into to a smooth curve by using scaled variables. (d) Binder’s cumulant at different sizes cross at $\mu_c=6.50(4)$ in agreement with (b). The inset in (d) shows that $\Delta U_L \sim L^{1/\nu}$ to yield $\nu \approx 1$, where $\Delta U_L \equiv U_L(\mu_1) - U_L(\mu_2)$ with $\mu_{1,2}$ are two adjacent points near $\mu_c$. (e) The susceptibility $\chi$. The inset shows a log-log plot of $\chi$ versus $L$ at $\mu_c$ leading to $\gamma \approx 7/4$. (f) The autocorrelation function $C(t)/C(0)$ at $\mu_c$ versus $tL^{-1.9}$ with $z=1.9$. All curves for different sizes collapse well to a single curve, indicating that the dynamic critical exponent $z=1.9(2)$. (Color online)[]{data-label="fig2d"}](fig2d.eps "fig:"){width="25.00000%"} ![ Phase transition in the 2D regular lattice in terms of the mutation parameter $\mu$. (a) The order parameter $\langle \|m\| \rangle$ versus $\mu$ clearly shows the existence of the alliance breaking transition. (b) A unique crossing at $\mu_c=6.50(4)$ with $\beta/\nu\approx1/8$ is obtained from the finite-size scaling form of $\langle \|m\| \rangle$. (c) All data points for $\langle \|m\| \rangle$ in (b) collapse into to a smooth curve by using scaled variables. (d) Binder’s cumulant at different sizes cross at $\mu_c=6.50(4)$ in agreement with (b). The inset in (d) shows that $\Delta U_L \sim L^{1/\nu}$ to yield $\nu \approx 1$, where $\Delta U_L \equiv U_L(\mu_1) - U_L(\mu_2)$ with $\mu_{1,2}$ are two adjacent points near $\mu_c$. (e) The susceptibility $\chi$. The inset shows a log-log plot of $\chi$ versus $L$ at $\mu_c$ leading to $\gamma \approx 7/4$. (f) The autocorrelation function $C(t)/C(0)$ at $\mu_c$ versus $tL^{-1.9}$ with $z=1.9$. All curves for different sizes collapse well to a single curve, indicating that the dynamic critical exponent $z=1.9(2)$. (Color online)[]{data-label="fig2d"}](fig2e.eps "fig:"){width="25.00000%"}![ Phase transition in the 2D regular lattice in terms of the mutation parameter $\mu$. (a) The order parameter $\langle \|m\| \rangle$ versus $\mu$ clearly shows the existence of the alliance breaking transition. (b) A unique crossing at $\mu_c=6.50(4)$ with $\beta/\nu\approx1/8$ is obtained from the finite-size scaling form of $\langle \|m\| \rangle$. (c) All data points for $\langle \|m\| \rangle$ in (b) collapse into to a smooth curve by using scaled variables. (d) Binder’s cumulant at different sizes cross at $\mu_c=6.50(4)$ in agreement with (b). The inset in (d) shows that $\Delta U_L \sim L^{1/\nu}$ to yield $\nu \approx 1$, where $\Delta U_L \equiv U_L(\mu_1) - U_L(\mu_2)$ with $\mu_{1,2}$ are two adjacent points near $\mu_c$. (e) The susceptibility $\chi$. The inset shows a log-log plot of $\chi$ versus $L$ at $\mu_c$ leading to $\gamma \approx 7/4$. (f) The autocorrelation function $C(t)/C(0)$ at $\mu_c$ versus $tL^{-1.9}$ with $z=1.9$. All curves for different sizes collapse well to a single curve, indicating that the dynamic critical exponent $z=1.9(2)$. (Color online)[]{data-label="fig2d"}](fig2f.eps "fig:"){width="25.00000%"} Figure \[fig2d\] summarizes the numerical results for the phase transition in the 2D regular lattice. The order parameter $\langle \|m\| \rangle$ shown in Fig. \[fig2d\](a) exhibits the existence of the transition, which is analyzed in detail in Fig. \[fig2d\](b) and (c) by using the finite-size scaling form in Eq. (\[eq:mscale\]). The critical point $\mu_c = 6.50(4)$ as well as critical exponents $\beta \approx 1/8$ and $\nu \approx 1$ are obtained, which are confirmed again from the finite-size scaling of the Binder’s cumulant shown in Fig. \[fig2d\](d). The divergence of the susceptibility in Fig. \[fig2d\](e) is analyzed to get $\gamma \approx 7/4$. At $\mu = \mu_c$, we compute the autocorrelation function (\[eq:Ct1\]) and plot it in Fig. \[fig2d\](f) in accord with the scaling form in Eq. (\[eq:Ct2\]): All curves at different sizes collapse well to a single curve with the dynamic critical exponent $z \approx 2$. All these findings clearly confirm that the alliance breaking transition induced by the mutation belongs to the 2D Ising universality class as was known in Ref. [@szabo1]. We next study the phase transition induced by the spatial randomness introduced via the long-range shortcut in the WS network model. One can motivate the study along this direction since in real systems, the spatial interaction topology among species can be much more complicated than the nearest-neighbor interaction on a regular square lattice. Distinguished from the regular network, the small-world network [@WS; @networkreview], which captures characteristics of many real networks very well, has a remarkably small average path length similar to the globally coupled network (or the mean-field case). The three-strategy RSP game has been studied on some small-world networks and the periodic flourishes of three strategies were found to happen during the time evolution [@szabo3]. ![Phase transition in the WS network in terms of the rewiring probability $\alpha$. The mutation parameter is set to $\mu = 7.0$. (a) The order parameter $\langle \|m\| \rangle$ as a function of $\alpha$ shows a sudden drop at $\alpha_c \approx 0.28$. The change of $\langle \|m\| \rangle$ becomes steeper as $L$ is increased, indicating that the transition is discontinuous one. (b) The normalized histogram $H(m)$ of $m$ for $L=1024$ is displayed at $\alpha = 0.27, 0.28$, and 0.29. The abrupt change in the form of the histogram between $\alpha = 0.28$ and 0.29 clearly confirms again discontinuous transition. (Color online) []{data-label="figWS"}](fig3a.eps "fig:"){width="25.00000%"}![Phase transition in the WS network in terms of the rewiring probability $\alpha$. The mutation parameter is set to $\mu = 7.0$. (a) The order parameter $\langle \|m\| \rangle$ as a function of $\alpha$ shows a sudden drop at $\alpha_c \approx 0.28$. The change of $\langle \|m\| \rangle$ becomes steeper as $L$ is increased, indicating that the transition is discontinuous one. (b) The normalized histogram $H(m)$ of $m$ for $L=1024$ is displayed at $\alpha = 0.27, 0.28$, and 0.29. The abrupt change in the form of the histogram between $\alpha = 0.28$ and 0.29 clearly confirms again discontinuous transition. (Color online) []{data-label="figWS"}](fig3b.eps "fig:"){width="25.00000%"} We fix the mutation parameter to $\mu = 7.0$, which is well inside the ordered phase with the defensive alliance formed for the 2D regular square lattice (see Fig. \[fig2d\]). Use of a very large value of $\mu$, corresponding to very small mutation rate, was found to make the system approaches a local dynamic fixed point and then the system stays there forever. At $\mu = 7.0$, which is small enough to ensure the equilibration and large enough to make the system ordered at small $\alpha$, the system is found to undergo a phase transition at $\alpha = \alpha_c \approx 0.28(1)$ as displayed in Fig. \[figWS\]. The abrupt drop down of the order parameter $\langle \|m\| \rangle$ at the transition as displayed in Fig. \[figWS\](a), together with the change of the histogram $H(m)$, normalized to satisfy $\sum_m H(m) = 1$, in Fig. \[figWS\](b), clearly indicates that the transition is of the discontinuous nature in a sharp contrast to the finding for 2D regular lattice at $\alpha = 0$ (see Fig. \[fig2d\]). In more details, the sudden change of the peak position of $H(m)$ between $\alpha=0.28$ and 0.29 in Fig. \[figWS\](b) is interpreted as a strong evidence of a discontinuous transition. Continuous transition in general exhibits the continuous shift of the peak position toward $m=0$ as the critical point is approached from the ordered phase. ![Time evolution of density of species $c_s(t)$ for the WS network for $L=1024$ at $\mu = 7.0$. (a) At $\alpha = 0.27$, the inter-alliance symmetry is broken, suggesting $m \neq 0$, and each species in the alliance dominates all the other species in a time-periodic fashion. It is interesting to note that the non-dominant alliance is also formed and even the member of the non-dominant alliance prevails periodically. (b) At $\alpha = 0.29$, the system recovers its full symmetry and each species is equivalent to others. However, the time-evolution of $c_s(t)$ is still periodic, in contrast to the disordered phase in 2D regular lattice. The time $t$ is measured after stationarity is achieved. (Color online)[]{data-label="fig:csWS"}](fig4a.eps "fig:"){width="25.00000%"}![Time evolution of density of species $c_s(t)$ for the WS network for $L=1024$ at $\mu = 7.0$. (a) At $\alpha = 0.27$, the inter-alliance symmetry is broken, suggesting $m \neq 0$, and each species in the alliance dominates all the other species in a time-periodic fashion. It is interesting to note that the non-dominant alliance is also formed and even the member of the non-dominant alliance prevails periodically. (b) At $\alpha = 0.29$, the system recovers its full symmetry and each species is equivalent to others. However, the time-evolution of $c_s(t)$ is still periodic, in contrast to the disordered phase in 2D regular lattice. The time $t$ is measured after stationarity is achieved. (Color online)[]{data-label="fig:csWS"}](fig4b.eps "fig:"){width="25.00000%"} In general, one can study the phase diagram of the model in the 2D parameter space of $(\mu, \alpha)$. Due to the practical difficulty to cover the whole parameter space, we in this work only explore the phase transitions along the two straights line in the $(\mu, \alpha)$ plane: One on the axis $(\mu, \alpha =0)$, and the other on the line ($\mu = 7.0, \alpha)$. Both parameters $\mu$ and $\alpha$ control the amounts of randomness (temporal and spatial ones, respectively), and accordingly the ordered phase with the defensive alliance becomes unstable as either $\mu$ is decreased (i.e., the mutation rate $P$ is increased) or $\alpha$ is increased. Not only the phase transitions have different natures (continuous one belonging to the 2D Ising universality class for the former, and discontinuous one for the latter), but also the ordered and disordered phases in each case are very much different in terms of the time evolution of densities of species. In the ordered and disordered phase on the axis of $\alpha=0$, the density of each species does not fluctuate much but stays at almost the same level: For ordered phase at $\mu > \mu_c$, $c_s \approx 1/3$ for $s \in \{ \mbox{alliance} \}$ and $c_s \approx 0$ otherwise, while for the disordered phase at $\mu < \mu_c$, $c_s \approx 1/6$ for all species. In contrast, the time evolution $c_s(t)$ for the case of the WS network at $(\mu = 7.0, \alpha)$ is strikingly different. In Fig. \[fig:csWS\](a), the time evolutions of densities of species are shown for $\mu = 7.0$ and $\alpha = 0.27 (< \alpha_c)$. It is clearly shown that the inter-alliance $Z_2$ symmetry is broken, indicating that the system has a nonzero value of the order parameter. Another very important observation one can make from Fig. \[fig:csWS\](a) is that each species within the alliance ($s=0,2,4$) cyclically dominates all the others in a very regular way, and very interestingly, the species that does not belong to the dominant alliance also prevails in a time-periodic manner. Even in the disordered phase at $\alpha > \alpha_c$, the periodic dominance persists while the inter-alliance symmetry is fully recovered \[see Fig. \[fig:csWS\](b)\]. ![Numerical results for the mean-field case. (a) The order parameter $\langle \|m\| \rangle$ versus $\mu$ for $N=32^2, 64^2, \cdots, 512^2$. As $N$ is increased the region for the ordered phase shifts towards higher values of $\mu$, indicating that the defensive alliance vanishes in the thermodynamic limit at any nonzero value of the mutation rate. (b) Time evolution of the order parameter $m(t)$ is not periodic at all, different from the WS network shown in Fig. \[fig:csWS\]. (Color online)[]{data-label="fig:MF"}](fig5a.eps "fig:"){width="25.00000%"}![Numerical results for the mean-field case. (a) The order parameter $\langle \|m\| \rangle$ versus $\mu$ for $N=32^2, 64^2, \cdots, 512^2$. As $N$ is increased the region for the ordered phase shifts towards higher values of $\mu$, indicating that the defensive alliance vanishes in the thermodynamic limit at any nonzero value of the mutation rate. (b) Time evolution of the order parameter $m(t)$ is not periodic at all, different from the WS network shown in Fig. \[fig:csWS\]. (Color online)[]{data-label="fig:MF"}](fig5b.eps "fig:"){width="25.00000%"} We finally investigate the mean-field limit of the game [@rspmf], where all individual species interact with all the other species in the system. The master equation for the number $N_s$ of species $s$ is given by $$\begin{aligned} \Delta N_s & = & P\left(-c_s + \frac{c_{i_1}}{2} + \frac{c_{i_2}}{2}\right) \nonumber \\ &+ &2(1-P)c_s( c_{i_1} + c_{i_2} - c_{j_1} - c_{j_2}) ,\end{aligned}$$ where $\Delta N_s \equiv N_s(t+1) - N_s(t)$, $i_1$ and $i_2$ ($j_1$ and $j_2$) are two preys (two predators) of $s$. For example, the species $s=0$ has $i_1 = 1, i_2 = 2, j_1 = 4, j_2 = 5$ \[see Fig. \[figa\](a)\]. The term proportional to $P$ describes the decrease of $N_s$ by the mutation from $s$ to its predators and increase of $N_s$ by the mutation from its preys to $s$. The other term originates from the interaction of $s$ to other species: When $s$ meets its preys (predators) $N_s$ is increased (decreased). The numerical factor two in front of $(1-P)$ is due to the interaction of other species with $s$. If we start from the situation when $c_0 = c_2 = c_4 = (1 + m)/6$ and $c_1 = c_3 = c_5 = (1-m)/6$, with the order parameter $m$ in Eq. (\[eq:m\]), the above master equation is reduced to a very simple form $$\frac{dm}{dt} = m(t+1) - m(t) = -NPm,$$ yielding the solution $$m(t) = m(0) \exp(-t/\tau)$$ with $\tau = PN$. The mean-field solution indicates that there is no ordered phase at any nonzero value of the mutation rate, i.e., [the defensive alliance cannot be formed in the mean-field limit]{}. It is also noteworthy that the time-periodic behavior observed for the WS network in Fig. \[fig:csWS\] ceases to exist in the mean-field case. The simulation results displayed in Fig. \[fig:MF\] for the mean-field model are in perfect agreements with the above analytic findings: The ordered phase that appears to exist for small system sizes drifts away toward the region of higher value of $\mu$ as $N$ is increased, suggesting the disappearance of the ordered phase in the thermodynamic limit \[see Fig. \[fig:MF\](a)\]. There is no time-periodic behavior of $m$ in equilibrium as shown in Fig. \[fig:MF\](b), where $t$ is measured after equilibration. In summery, we have investigated the instability of the defensive alliances for the simple food web of six species in three different spatial interaction structures: the 2D local regular square lattice, the WS network, and the globally coupled mean-field network. The temporal randomness imposed by the mutations as well as the spatial randomness in the interaction structure tuned by the rewiring probability in the WS network has been shown to make the defensive alliance unstable. When the mutation rate is increased for the 2D square lattice the alliance breaking transition has been clearly identified to belong to the 2D Ising universality class due to the common $Z_2$ symmetry. On the other hand, when the rewiring probability is increased, the transition becomes discontinuous, and around the transition the natures of the ordered and disordered phases are very different from the 2D square lattice. The mean-field model has also been studied analytically and numerically with the results that the defensive alliance cannot be formed at any value of the mutation rate and that the time-periodic behavior observed in the WS network is not seen any longer. B.J.K. was supported by grant No. R01-2005-000-10199-0 from the Basic Research Program of the Korea Science and Engineering Foundation. J.U. and S.-I.L. acknowledge the support by the Ministry of Science and Technology of Korea through the Creative Research Initiative Program. The numerical calculations have been performed in the computer cluster Iceberg at Ajou University. B. Kerr, M.A. Riley, M.W. Feldman, and B.J.M. Bohannan, Nature (London) [418]{}, 171 (2002). L. Frachebourg, P.L. Krapivsky, and E. Ben-Naim, Phys. Rev. E [54]{}, 6186 (1996). G. Szabó and T. Czárán, Phys. Rev. E [64]{}, 042902 (2001). G. Szabó and T. Czárán, Phys. Rev. E [63]{}, 061904 (2001). B. Sinervo and C.M. Lively, Nature (London) [380]{}, 240 (1996). D.J. Watts and S.H. Strogatz, Nature (London) [393]{}, 440 (1998). M.E.J. Newman, SIAM Review [45]{}, 167 (2003); R. Albert and A-L. Barabási, Rev. Mod. Phys. [74]{}, 47 (2002). R.J. Baxter, Exactly Solved Models in Statistical Mechanics (Academic Press, London, 1982); C.J. Thompson, Classical Equilibrium Statistical Mechanics (Clarendon Press, Oxford, 1988). K. Binder, Phys. Rev. Lett. [47]{}, 693 (1981); Rep. Prog. Phys. [60]{}, 487 (1997). P. Peczak, A.M. Ferrenberg, and D.P. Landau, Phys. Rev. B [43]{}, 6087 (1991); G.A. Baker Jr. and N. Kawashima, J. Phys. A: Math. Gen. [29]{}, 7183 (1996). H. Hong, B.J. Kim, and M.Y. Choi, Phys. Rev. E [66]{}, 018101 (2002); B.J. Kim, H. Hong, P. Holme, G.S. Jeon, P. Minnhagen, and M. Y. Choi, [ibid]{}. [64]{}, 056135 (2001); B.J. Kim, P. Minnhagen, S. K. Oh, and J. S. Chung, Phys. Rev. B [64]{}, 024406 (2001). G. Szabó, A. Szolnoki, and R. Izsák, J. Phys. A: Math Gen. [37]{}, 2599 (2004); A. Szolnoki and G. Szabó, Phys. Rev. E [70]{}, 037102 (2004). The mean-field version of the RSP game was studied in M. Frean and E.R. Abraham, Proc. R. Soc. Biol. Sci. B [268]{}, 1323 (2001).	{ "pile_set_name": "ArXiv" }
--- author: - 'D. Donato' - 'M. Gliozzi' - 'R. M. Sambruna' - 'J. E. Pesce' date: 'Received: ; accepted: ' title: '[Chandra]{} Observations of the X–ray Environment of BL Lacs' --- Introduction ============ Orientation-based unification models for Active Galactic Nuclei have been successful in explaining the rich variety of observed properties in the various classes of AGN (Urry & Padovani 1995; Antonucci 1993). According to these schemes, the various AGNs are the same intrinsic object, powered by accretion of the host galaxy gas onto a super-massive black hole, seen at different orientation angles with respect to a preferred axis. In the case of radio-loud AGN, the different subclasses are due to the different orientation of the relativistic jet, with blazars (BL Lacertae objects and Flat Spectrum Radio Quasars) corresponding to the more aligned sources, and radio galaxies (Fanaroff-Riley I and II) being their parent populations (e.g., Urry & Padovani 1995 and references therein). Previous studies at radio, IR, and optical wavelengths show that for BL Lacs (the relatively local, low-luminosity version of blazars) the parent population is most likely represented by FR I radio galaxies, although some sources may be occasionally hosted by FR IIs (e.g., Kollgaard et al. 1992). Indeed, it is possible to predict the beamed luminosity function of BL Lacs from the luminosity function of FRIs invoking a Lorentz factor $\gamma \sim3-20$, depending on the wavelength (Padovani & Urry 1991; Urry et al. 1991). A model-independent probe of these ideas is represented by studies of the near environment. If FRIs are misaligned versions of BL Lacs, the larger-scale environment (host galaxy and cluster of galaxies) of the two classes should be the same, as they are not affected by beaming. Studies at optical-IR wavelengths showed that BL Lacs and FRIs reside in giant elliptical galaxies and, on average, in poor clusters of Abell richness class 0 or less (Falomo et al. 1999, Wurtz et al. 1997). Similarly, the X-ray environment of BL Lacs and FRIs should be similar. [ROSAT]{}, and more recently [Chandra]{}, observations of FRIs established that the X-ray cores of these sources are usually embedded in diffuse soft X-ray emission which is interpreted as the thermal halo of the host galaxy on kpc scales (Worrall et al. 2001, Canosa et al. 1999). However, because of the limited sensitivity and spatial resolution of [ROSAT]{}, it was not possible to perform similar observations for most BL Lacs, where the bright core dominates the X-ray emission. The only exception is PKS 0521–365, where an unusually large and bright X-ray halo was detected with the HRI (Hardcastle at al. 1999). Separating the diffuse X-ray emission from the core X-rays requires high angular resolution and improved sensitivity, such as afforded by the [Chandra]{} X-ray Observatory (e.g., Birkinshaw et al. 2002). Motivated by these considerations, we acquired [Chandra]{} imaging observations of a sample of nearby ($z<0.1$) BL Lacs during cycle 1, to study their kpc-scale X-ray environment and to compare to those of FRIs. As our sample contains sources claimed to reside in optical clusters of galaxies (Falomo et al. 1999), [Chandra]{} observations can be used to confirm independently the presence of the cluster and quantify physical properties of the intracluster gas. Based on their spectral energy distributions, BL Lacs can be classified as Low-energy peaked BL Lacs (LBLs), when the radio-to-X-ray spectral index ($\alpha_{\rm rx}$) is larger than 0.75; and as High-energy peaked BL Lacs (HBLs), when $\alpha_{\rm rx}$0.75 (Padovani & Giommi 1995). There is much debate on the origin of the LBL-HBL division, which appears to be continuous with luminosity (Fossati et al. 1997; Sambruna et al. 1996). One of the goals of our GO1 proposal was to investigate whether LBLs and HBLs exist in different X-ray environments, to ascertain whether the ambient gas can affect their different observed jet properties. The outline of the paper is as follows. In Sect. 2 we discuss the sample selection criteria and in Sect. 3 the observations and data analysis. Results of the spatial and spectral are given in Sects. 4 and 5, respectively. The conclusions are presented in Sect. 6. The analysis of serendipitous X-ray sources is given in Appendix A. Throughout this paper, $H_0=75$ km s$^{-1}$ Mpc$^{-1}$ and $q_0=0.5$ are adopted. With this choice, 1 corresponds to 71 kpc for PKS 0548–322 and for BL Lac, 54 kpc for 3C 371, 73 kpc for PKS 2005–489, 62 kpc for 1ES 2321+419, and 47 kpc for 1ES 2344+514. Sample selection ================ The targets were selected from two complete, flux-limited samples: the 1 Jy sample of radio-selected BL Lacs, and the [Einstein]{} Slew Survey (1ES) sample of X-ray selected BL Lacs. Sources with redshifts larger than $z$=0.2 were excluded, to optimize the [Chandra]{} resolution and for comparison with available samples of FRIs (Worrall et al. 2001). Sources with $z<0.03$ were also excluded, because the physical size of the field of view is too small to study the intracluster gas ($\leq0.5$ Mpc), should they reside in a cluster of galaxies. In this range we found 4 LBLs in the 1 Jy sample. For each one we chose 2 HBLs matched in redshift (HBLs are more common at low redshifts), in order to have similar redshift distributions for the two types of BL Lacs. The final list contained 12 targets. Only 6 targets were awarded [Chandra]{} time, including 4 HBLs and 2 LBLs. The targets are listed in Table 1, together with their basic properties and classification. Imaging data at longer wavelengths are available for all 6 sources, for comparison with the ACIS images. Two of the sources (PKS 0548–322 and PKS 2005–489) are known to reside in a relatively rich optical cluster of galaxies (Falomo et al. 1995; Pesce et al. 1995). The [Chandra]{} observation of 3C 371, where an X-ray counterpart to the optical jet was found, was previously discussed in Pesce et al. (2001), with main emphasis on the jet properties. Here, we will neglect the X-ray jet for this source and focus on the extended X-ray environment. [lccccccc]{} Object Name & R.A. & Dec. & z & $N_{\rm H,Gal}$ & Type & $L_{2-10 {\rm keV}}$&Ref.\ (1) & (2) & (3) &(4)& (5) & (6) & (7) &(8)\ PKS 0548–322 & 05 50 40.80 & -32 16 17.80 & 0.069 & 2.52 & HBL & 31.9 & KU98\ 3C 371 & 18 06 50.60 & +69 49 28.10 & 0.051 & 4.73 & LBL & 1.5 & DO01\ PKS 2005–489 & 20 09 25.40 & -48 49 54.00 & 0.071 & 5.08 & HBL & 62.2 & PA01\ BL Lac & 22 02 43.30 & +42 16 39.80 & 0.069 & 21.6 & LBL & 6.1 & RA02\ 1ES 2321+419 & 23 23 52.10 & +42 10 59.00 & 0.059 & 10.9 & HBL & 1.5 & PE96\ 1ES 2344+514 & 23 47 04.80 & +51 42 17.40 & 0.044 & 16.8 & HBL & 3.9 & DO01\ [Columns]{}: 1=Source name; 2=Right Ascension at J2000; 3=Declination at J2000; 4=Redshift; 5=Galactic column density in units of $10^{20} {\rm cm^{-2}}$. The Galactic values was derived from the nh program at HEASARC (based on Dickey & Lockman 1990); 6=Source type: HBL=High–Energy Peaked BL Lacertae object, LBL=Low–Energy Peaked BL Lacertae object; 7=Intrinsic X–ray luminosity (in units of $10^{43}$ erg s$^{-1}$) between 2 and 10 keV from data in literature; 8=Reference for the X–ray luminosity: KU98=Kubo et al. 1998; DO01=Donato et al. 2001; PA01=Padovani et al. 2001; RA02=Ravasio et al. 2002; PE96=Perlman et al. 1996. Observations and Data Reduction =============================== The [Chandra]{} observations of the 6 objects were carried out in early 2000. A log of the observations is reported in Table 2, together with information about the source intensities. All sources were observed with ACIS-I at the aimpoint of the I3 chip, except 3C 371 which was observed with ACIS-S, at the aimpoint of S3. The choice of S3 for 3C 371 is due to the fact that it was known to have a synchrotron optical jet. As a science goal for this target was to find the X-ray counterpart of the synchrotron jet, the softer response of a back-illuminated CCD such as S3 was needed to maximize the detection of the X-ray jet. The short exposures (2–6 ks) were designed to search for the X-ray diffuse emission around the cores, but are clearly insufficient to study in detail their physical properties. Again an exception is 3C 371, where an exposure of 10 ks was requested, in order to detect the X-ray jet (see, e.g., Sambruna et al. 2002). Since the main purpose of this work is to observe the large-scale environments of blazars, the observations were made with the full CCD operational and the standard ACIS frame time (3.24 sec), even if it was expected that the central sources would have produced a high X-ray count rate ($\geq1$ c/s). With this set-up it is likely that more than one photon will be detected in a single CCD pixel within a single integration time. This effect, called pile-up, leads to a distortion (hardening) of the spectrum of the core and a reduction in the measured count rate relative to the true incoming rate. Pile-up becomes important for $\geq0.1$ c/frame. The pile-up effect also produces an image of the core that is different from the original source image: it appears somewhat extended and flat-topped, and in the cases of most extreme pile-up with a central hole, where the pixels have no counts. In this case the pile-up is strong enough that the total amplitude of the event is larger than the on-board threshold ($\sim$15 keV) and is rejected. This rejection is visible in the images of the central core region of PKS 0548–322 and PKS 2005–489. For the image analysis we used the `CIAO` 2.2.1 software package and followed the standard reduction criteria, using the latest files provided by the [Chandra]{} X–ray Center. We inspected the light curves of the background in all cases to search for possible background fluctuations. Only for PKS 0548–322 we found a short variation of the background intensity. The corresponding time interval was removed from the analysis. We restricted our analysis to the 0.3–8.0 keV energy range, where the instrument is better calibrated and the background is negligible. The exposure for BL Lac was divided into two segments taken at different times for operational reasons (Table 2). The 2 images of BL Lac were combined using the `CIAO` tool [mergeall]{} in order to increase the signal-to-noise ratio, for a total exposure of 3.1 ks. [lccccccc]{} Object Name & Seq. Num.& Exposure & Start Date & Net Rate & Bkg Rate & c/f\ (1) & (2) & (3) &(4) & (5) & (6) & (7)\ PKS 0548–322 & 700145 & 4559 &1999/12/31 & 0.235$\pm$0.007 & 4.71 & 0.7516\ 3C 371 & 700146 & 10120 &2000/03/21 & 0.263$\pm$0.005 & 4.56 & 0.8334\ PKS 2005–489 & 700147 & 5858 &2000/10/07 & 0.491$\pm$0.009 & 2.61 & 1.5719\ BL Lac & 700148 & 2117 &2000/01/07 & 0.171$\pm$0.009 & 0.14 & 0.5464\ & 700196 & 1039 &2000/02/16 & 0.172$\pm$0.013 & 0.08 & —\ 1ES 2321+419 & 700149 & 4628 &2000/02/08 & 0.253$\pm$0.007 & 0.84 & 0.8114\ 1ES 2344+514 & 700150 & 2727 &2000/02/08 & 0.185$\pm$0.008 & 1.17 & 0.5912\ [Columns]{}: 1=Source name; 2=Sequence number of the observation; 3=Net Chandra exposure in seconds after data screening; 4=Observation start date (Year-Month-Day); 5=Net photon count rate (c/s) of the source in the energy range 0.3–8 keV; 6=Background photon count rate (in units of $10^{-3}$ s$^{-1}$) obtained rescaling to the same extraction region used for the net source count rate; 7=Count/frame evaluated with PIMMs and using spectral information from the literature (see previous Table for the references). The input flux in PIMMs is such that the output flux corresponds to the measured one. Spatial Analysis ================ The main aim of the spatial analysis is to investigate whether diffuse emission is present around the BL Lac objects or whether their X-ray radial profile is entirely consistent with a point source. In the latter case, no contribution to the local X–ray emission from a diffuse component is present. The first step of the spatial analysis was to produce a smooth image of the field of view. For example, Fig. 1 shows the adaptively smoothed Chandra images of 1ES 2344+514 and PKS 0548-322, in the 0.3–8 keV energy range. Before extracting the radial profiles of the sources, we removed from the images instrumental features (i.e., spikes produced by out-of-time events) and field sources. To find the field sources we used the detect tool [wavdetect]{} with the default value ($10^{-9}$) for the threshold for identifying a pixel as belonging to a source (see Sect. 7). For 3C 371, we also removed the X-ray jet (Pesce et al. 2001) by using an elliptical region (1.6$\times$ 1.1) centered on the jet. The same region was substituted by a region of identical shape and size located symmetrically with respect to the core, in order to restore a possible halo contribution. Using the tools [dmextract]{} and [dmtcalc]{}, we extracted the radial profiles on a scale typical of clusters (i.e., approximately 250 kpc, see for example De Grandi et al. 1999) that correspond to a radius of about 200 on our images, depending on the redshifts. In the case of the ACIS-I observations, this radius encompasses the boundary between the various CCDs. The estimated average influence of the gap between the single CCDs on the surface brightness is of the order of 4%, therefore the surface brightness should be considered as a lower limit. A series of annular regions, with increments of the radius of 2, was used to extract the radial profile. For 3C 371, the only source observed with ACIS-S, the number of the annuli was limited by the distance between the source and the edge of S3. While the responses of ACIS-I CCDs are not dramatically different, the response of S3 and S2 are different. The background regions were extracted from annular, circular, or box regions external to the annular regions centered on the sources. These regions are free of obvious sources. In order to evaluate the instrumental response to a point source, we used the tool [mkpsf]{} to create an image of the Point Spread Function (PSF) of an on-axis point source, normalized to the source flux. The PSF changes with source position and photon energy, and is created by interpolation of the medium resolution library of pre-launch calibration files (the PSF hypercube library). The [mkpsf]{} tool is able to create only monochromatic PSFs. However, a monochromatic PSF is too simple a model to describe adequately the spatial distribution of our sources. We therefore improved the model to find a better representation of the source PSF, following a method similar to that used by Worrall et al. (2001). Our improvement consists on merging 8 different monochromatic PSFs chosen and weighted on the basis of the source energy spectra between 0.3 and 8 keV. This method can be summarized as follows: 1\) We first extracted the energy distributions of the photons from a circular region centered on the source. The radii of these regions vary from 5 to 10, depending on the source brightness. 2\) In each case, we sampled the entire photon distribution by choosing 8 discrete energy values. The number of counts at any of these energy values corresponds to the energy weight (see Fig. 2). The energies and weights are shown in Table 3 and are indicated with E and W, respectively. 3\) Using [mkpsf]{} we created 8 monochromatic PSFs at the position of the AGN on the detector for each of the 8 sampled energies, and coadded them. Each PSF was weighted by its relative normalization (the weights). [lcccccccc]{} Object Name & $E_1, W_1$& $E_2, W_2$& $E_3, W_3$& $E_4, W_4$& $E_5, W_5$& $E_6, W_6$& $E_7, W_7$& $E_8, W_8$\ (1) & (2) & (3) &(4) & (5) & (6) & (7) &(8) & (9)\ PKS 0548–322 & 0.82, 34 & 1.22, 38 & 1.37, 30 & 1.62, 21 & 1.87, 19 & 2.07, 11 & 2.77, 8 & 3.47, 7\ 3C 371 & 0.47, 58 & 0.87, 58 & 1.17, 37 & 1.47, 37 & 1.87, 26 & 2.17, 23 & 2.72, 17 & 4.02, 10\ PKS 2005–489 & 0.57, 77 & 0.87, 117 & 1.17, 102 & 1.27, 91 & 1.42, 70 & 1.52, 63 & 1.77, 38 & 1.97, 36\ BL Lac & 0.87, 7 & 1.22, 14 & 1.37, 10 & 1.62, 15 & 2.47, 8 & 3.07, 9 & 3.62, 6 & 4.17, 8\ 1ES 2321+419 & 0.77, 21 & 1.12, 28 & 1.32, 32 & 1.62, 28 & 1.92, 17 & 2.47, 16 & 2.92, 13 & 3.67, 11\ 1ES 2344+514 & 0.52, 5 & 1.02, 15 & 1.27, 15 & 1.52, 12 & 1.72, 7 & 2.17, 6 & 2.57, 6 & 4.82, 7\ [Columns]{}: 1=Source name; 2-9=Energy (in keV) and weight of each monochromatic PSF that has been coadded in order to have a more realistic representation of the source PSF (see Sect. 4). [lcccccccc]{} Object Name & $A_0$& $A_1$ & $A_2$ & $A_3$ & $A_4$ & $A_5$& $A_6$ &$A_7$\ & & $\times 10^{-2}$& & $\times 10^{-5}$&$\times 10^{-5}$& &$\times 10^{-4}$&\ (1) & (2) & (3) &(4) & (5) & (6) & (7) & (8) & (9)\ PKS 0548–322 & 0.17 & 5.5 & 5.2 & 0.5 & 0.7 & 0.34 & 1.4 & 12.5\ 3C 371 & 0.17 & 3.0 & 9.0 & 1.0 & 1.1 & 0.36 & 1.0 & 16.6\ PKS 2005–489 & 0.29 & 6.5 & 10.0 & 3.7 & 4.8 & 0.09 & 2.0 & 12.0\ BL Lac & 0.30 & 2.0 & 7.1 & 14.0 & 0.2 & 0.02 & 1.8 & 70.0\ 1ES 2321+419 & 0.12 & 2.8 & 6.5 & 14.0 & 0.5 & 0.30 & 1.7 & 85.0\ 1ES 2344+514 & 0.32 & 3.0 & 9.0 & 4.5 & 1.0 & 1.64 & 0.8 & 15.0\ [Columns]{}: 1=Source name; 2-9=Values of the parameters of the analytical function used to fit the composed PSF (see Sect. 4 and Eq. (1)). Once we obtained the composite PSF, we searched for a fair analytical representation that describes it. This empirical representation is obtained using an 8-parameter function: $$PSF(\rm{x})= A_0[A_1\frac{e^{-\frac{\rm{x}}{A_2}}}{\rm{x}^2}-A_3+A_4e^{-\frac{(\rm{x}-15)^2}{A_5}}+A_6e^{-\frac{(\rm{x}-13)^2}{A_7}}]$$ For each source, we fitted the composite PSF with Eq. (1) and found the best-fit values of the 8 parameters (Table 4). The instrumental radial profile is not expected to be a good representation for the central region of strongly piled-up sources. Therefore, since no `CIAO` task exists yet to correct the PSF for pile-up, we normalized the instrumental radial profile at $r$=2: at larger distances the pile-up effect on the PSF is negligible, according to the [Chandra]{} Proposers’ Observatory Guide. For each source we superimposed the composite PSF (weighted sum of the 8 monochromatic PSFs) on the surface brightness versus the radial distance. Due to the pile-up effect on the spectral distribution of the photons (spectral hardening), this method is likely to overestimate the contribution of the hard photons. The result is a slightly broader PSF, which will not cause any false detection of extended emission but rather underestimate its contribution. After submitting this paper, a script that automatically performs the same task (`ChaRT`) was released by the X–ray Center (`CXC`). We performed the analysis of the radial profiles using `ChaRT` and found consistent results with our method above. In Fig. 3 we plot the radial profiles of the 6 sources. The dashed line indicates the profile of the composite PSF obtained using Eq. (1) and the parameters in Table 4. The horizontal dotted line represents the background. In the case of 3C 371, the radial profile is consistent with the PSF and no excess emission is detected in our exposure. In the case of BL Lac, 1ES 2321+419, and 1ES 2344+514, there is some evidence for weak excess emission between 18 and 30. Excess emission over the PSF is clearly present for PKS 0548–322 and PKS 2005–489 up to $\sim$ 150 (180 kpc) and 130 (160 kpc), respectively (Fig. 3). A $\beta$-model (King profile) was used to model the excess X-ray emission in all cases. We find that the $\beta$ model provides a significant improvement only in the case of BL Lac, PKS 0548–322, and PKS 2005–489. The free parameters of the $\beta$-model are the value of $\beta$ and the core radius, $r_{\rm c}$. For PKS 0548–322, the reduced $\chi^2$ of the fit is $\chi^2_{\rm r}=1.23$ with 142 degrees of freedom (d.o.f.), with $\beta=0.35 \pm$0.20 and $r_{\rm c}$=23.2$\pm$2.5. For PKS 2005–489, $\chi^2_{\rm r}=1.0$ (138 d.o.f.), with $\beta=0.56 \pm$0.08 and $r_{\rm c}$=12.1$\pm$1.2. For BL Lac, the model parameters are poorly constrained; we thus fixed $\beta$ to 0.6 (Worrall & Birkinshaw 1994) and left $r_{\rm c}$ as the only free parameter of the fit. We obtain $\chi^2_{\rm r}=0.94$ for 24 d.o.f. and $r_{\rm c}$=14.7$\pm$12.6. All the errors are 1$\sigma$. The core radius implies a physical radius of the diffuse emission of 28 kpc in PKS 0548–322, 15 kpc in PKS 2005–489, and 18 kpc for BL Lac. These values are typical of the X-ray halo of the host galaxy (Birkinshaw et al. 2002, Fabbiano et al. 1992), which is a giant elliptical in all cases (Falomo & Kotilainen 1999). Fig. 4 shows the PSF plus $\beta$ model fit for PKS 0548–322 and PKS 2005–489, where the parameters are best constrained. As is apparent from the observed radial profiles of PKS 0548–322 and PKS 2005–489, diffuse emission in these two sources extends well beyond the fitted value of the core radii, up to $\sim$ 130-150, or several hundreds of kiloparsecs. This scale is typical of clusters, and indeed both sources are known to reside in optical clusters (Pesce et al. 1995). However, the radial profiles are suggestive of two components, one (the galaxy halo) up to 40, plus a large-scale tail related to the cluster. Thus, we attempted a fit to the radial profiles with the PSF plus two $\beta$ models. In the first model, the core radius was fixed to the halo radius of 23 and 12 for PKS 0548–322 and PKS 2005–489, respectively. The parameters of the second $\beta$-model were left free to vary. No improvement was obtained, and the core radius of the second $\beta$ model is totally unconstrained. This is because the core radius of the cluster extends on a scale larger than our field of view, making its determination difficult. Thus, while the radial profiles show clear evidence for the presence of X-ray emission on cluster scales, it is not possible to constrain its spatial properties with the present data. However, given its isotropic distribution, the cluster X-ray emission dominates the X-ray spectrum, even when the latter is extracted on the scale of the galaxy halo (see below). Indeed, the cluster contribution is still present along the line of sight when extracting the spectrum on small spatial scales. To evaluate the luminosities of the diffuse emission, we extracted the count rate from annular regions around the sources. For BL Lac, the annular region has inner and outer radii of 2035, respectively. The measured 0.3–8 keV count rate is 0.011$\pm$0.002 c/s. The count rate was converted into flux assuming a thermal Raymond-Smith spectrum with Galactic column density, temperature $kT=1$ keV, and fixing abundances to 0.2 solar. The derived observed flux is $F_{0.4-5\rm keV}=6.7\times 10^{-14}$ erg cm$^{-2}$ s$^{-1}$, corresponding to an absorption-corrected luminosity $L_{0.4-5\rm keV}=1 \times 10^{42}$ erg s$^{-1}$, typical of FRIs halos (e.g., Worrall 2002). For PKS 0548–322 and PKS 2005–489, the integrated count rate in 20–40 includes the contribution of the cluster. To evaluate the halo luminosity, we subtracted the cluster contribution from the total count rate in 20-40, rescaling the cluster surface brightness to the halo area and assuming a uniform cluster emission profile. We derive 0.019$\pm$0.002 c/s and 0.038$\pm$0.003 c/s in 0.3–8 keV for PKS 0548–322 and PKS 2005–489, respectively. Assuming a thermal Raymond-Smith spectrum with Galactic column density, temperature $kT=1$ keV, and abundances fixed to 0.2 solar, these count rates correspond to an observed flux of $F_{0.4-5\rm keV}=1.4 \times 10^{-13}$ erg cm$^{-2}$ s$^{-1}$ and intrinsic luminosity $L_{0.4-5\rm keV}=1.3\times 10^{42}$ erg s$^{-1}$ for PKS 0548–322, and $F_{0.4-5\rm keV}=3.1\times 10^{-13}$ erg cm$^{-2}$ s$^{-1}$ and $L_{0.4-5\rm keV}=3.1\times 10^{42}$ erg s$^{-1}$ for PKS 2005–489. An alternative way to estimate the luminosity of the halo is based on the integration of the $\beta$-model over the extended component region. Integrating between 20 and 40 we obtained results fully consistent with those derived above. With this method, we can also evaluate the integrated luminosity over the entire 0–40 range. We found that the 0.4–5 keV unabsorbed fluxes for BL Lac, PKS 0548–322, and PKS 2005–489 are $F_{0.4-5\rm keV}=2.8\times 10^{-13}$ erg cm$^{-2}$ s$^{-1}$, $F_{0.4-5\rm keV}=3.3\times 10^{-13}$ erg cm$^{-2}$ s$^{-1}$, $F_{0.4-5\rm keV}=7.6\times 10^{-13}$ erg cm$^{-2}$ s$^{-1}$, respectively. The corresponding intrinsic luminosities are $L_{0.4-5\rm keV}=2.9\times 10^{42}$ erg s$^{-1}$ for BL Lac $L_{0.4-5\rm keV}=3.3\times 10^{42}$ erg s$^{-1}$ for PKS 0548–322 $L_{0.4-5\rm keV}=7.8\times 10^{42}$ erg s$^{-1}$ for PKS 2005–489. The assumption of uniform cluster emission is unrealistic, as the cluster radial profile may increase toward the center of the cluster, where the gas is denser/hotter. Thus, the above count rates represent an upper limit to the X-ray emission of the galaxy’s halo. For the cluster emission, the count rates extracted from 40–150/130 are 0.021 $\pm$ 0.002 c/s for PKS 0548–322, and 0.013 $\pm$ 0.001 c/s for PKS 2005–489. Assuming the parameters from the spectral analysis (see below), the corresponding observed fluxes and intrinsic luminosities are $F_{0.4-5\rm keV}=2.9\times 10^{-13}$ and $L_{0.4-5\rm keV}=3.\times 10^{42}$ for PKS 0548–322, and $F_{0.4-5\rm keV}=1.3\times 10^{-13}$ and $L_{0.4-5\rm keV}=1.3\times 10^{42}$ for PKS 2005–489, typical of cluster with Abell richness 1 (Mulchaey & Zabludoff 1998; Mushotzky 1998; Mahdavi et al. 1997). For 3C 371, 1ES 2321+419, and 1ES 2344+514, we evaluated upper limits to the extended halo. Count rates were extracted in an annulus of inner and outer radii 20 and 40, respectively, and converted into flux assuming a Raymond-Smith thermal model with $kT=1$ keV, abundances fixed to 0.2 solar, and Galactic column densities. We derive $L_{0.4-5\rm keV}=1.2\times 10^{41}$ erg s$^{-1}$ for 3C 371, $L_{0.4-5\rm keV}=5.2\times 10^{40}$ erg s$^{-1}$ for 1ES 2321+419, and $L_{0.4-5\rm keV}=4 \times 10^{40}$ erg s$^{-1}$ for 1ES 2344+514. In conclusion, diffuse X-ray emission is clearly present in 3 out of 6 sources of the sample (PKS 0548–322, PKS 2005–489, and BL Lac). In two other sources, 1ES 2321+419 and 1ES 2344+514, there is some weak evidence for faint diffuse emission. The diffuse emission in BL Lac, and possibly in the two 1ES sources, is on the scale of the galaxy’s halo and has a luminosity typical of FRIs. In the two PKS sources, there is evidence for diffuse emission on both the galaxy halo and the cluster scale. However, it is not possible to model the cluster spatial properties. Spectral Analysis ================= The goal of this section is to investigate the X-ray properties of the diffuse X-ray emission in PKS 0548–322 and PKS 2005–489. To this end, we extracted and analyzed the X-ray spectra of this component. The ACIS-I spectra of the diffuse emission was extracted using an annulus centered on the core. We used a large extraction region (inner radius 20, outer radius 150/130) to increase the signal-to-noise ratio. We chose this high value for the inner radius in order to reduce contamination of the PSF wings from the AGN. An issue with such a large extraction region, however, is that it encompasses 4 CCDs that have different spectral responses. To circumvent this difficulty, we divided the extraction region in four sectors, each one covering a single CCD. The resulting four spectra were fitted jointly within XSPEC v.11.0.1. To account for the recently observed quantum efficiency decay of ACIS, possibly caused by molecular contamination of the ACIS filters, we have applied a time-dependent correction to the ACIS quantum efficiency based on the presently available information from the CXC. The model used to fit the spectra is the thermal plasma model APEC, with column density fixed to the Galactic value (Table 1), solar abundances fixed at 0.2, and temperature $kT$ ranging between 0 and 15 keV. We found $kT=6.3^{+8.7}_{-2.8}$ keV, $\chi^2_{\rm r}=0.42/47$, for PKS 0548-322, and $kT=4.6^{+2.0}_{-1.7}$ keV, $\chi^2_{\rm red}=0.97/54$ for PKS 2005–489. No improvements are obtained adding a second component (a power law or a second thermal model). The fitted temperatures are typical of the X-ray emission from the hot intracluster gas in rich cluster (De Grandi & Molendi 2002). In an attempt to separate the contributions of the galaxy halo and of the cluster, we extracted two X-ray spectra, the first in the range 20–40 (to maximize the halo) and the second in 40–130/150, for both sources. We found that the best-fit is obtained using the APEC model and abundances fixed to 0.2 solar. Spectral analysis of the first spectrum shows that the fitted temperature (with range between 0 and 15 keV) is still high ($kT=9.5^{+5.5}_{-4.6}$ keV with $\chi^2_{\rm r}=0.62/6$ for PKS 0548–322 and $kT=3.7^{+1.8}_{-1.3}$ keV with $\chi^2_{\rm r}=0.90/18$ for PKS 2005–489), indicating that the cluster emission is still dominant. For the second spectrum, the parameters are: $kT=6.0^{+9.0}_{-2.9}$ keV, $\chi^2_{\rm r}=0.39/40$, for PKS 0548-322, and $kT=12.1^{+2.9}_{-8.1}$ keV, $\chi^2_{\rm r}=0.98/34$ for PKS 2005–489. Summary and Conclusions ======================= We presented ACIS-I/S observations of 6 BL Lacertae objects (four HBLs and two LBLs), aimed at detecting the diffuse circumnuclear X-ray emission predicted by the unification models. The short exposures, 2–6 ks (10 ks for 3C 371), were optimized to detect diffuse emission, but are insufficient to study their properties in detail. Diffuse X-ray emission was convincingly detected in 3/6 cases (PKS 0548–322, PKS 2005–489, and BL Lac), with marginal evidence in an additional 2 sources (1ES 2321+419 and 1ES 2344+514). The extended X-ray emission is on scales of several kiloparsec and has a luminosity similar to FRI galaxies studied with [Chandra]{} (Worrall et al. 2002). This qualitatively supports the unification models for radio-loud AGN, which states that BL Lacs and FRI galaxies are the same intrinsic objects, seen at different orientations of the relativistic jets, with BL Lacs being the sources more closely aligned with the line of sight. The presence of an extended X-ray emission around the X-ray cores, interpreted as a thermal bremsstrahlung from a galactic atmosphere, can trace the medium whose pressure may confine the jets that are pointing toward us. For PKS 0548–322 and PKS 2005–489 we evaluated the density and the external pressure of the X-ray component. The gas density profile is derived by deprojection of the $\beta$-model used for the decomposition of the surface brightness profile (e.g., Ettori 2000). We adopted the cooling function value (e.g., Sarazin 1988) for the hot gas temperature and abundance derived from the spectral analysis. The resulting central gas density is $1.4\times 10^{-3} ~\rm{atoms~cm}^{-3}$ for PKS 0548–322 and $7.4\times 10^{-3} ~\rm{atoms~cm}^{-3}$ for PKS 2005–489. The external pressures evaluated at the characteristic radii (28 kpc for PKS 0548–322 and 15 kpc for PKS 2005–489) are $P_{\rm ext} = 5.2\times 10^{-14} ~\rm{N~m}^{-2}$ and $P_{\rm ext} = 2.2\times 10^{-13} ~\rm{N~m}^{-2}$, respectively. Comparing the external gas pressure to the internal jet pressure can provide useful information on the propagation of the jet. As the jets of BL Lacertae objects are seen at small angles with respect to the line of sight, the internal jet pressure can only be derived via modeling of the SED of the blazar. Ghisellini et al. (2002) estimated the total jet pressure for a sample of blazars, including FSRQs, LBLs, and HBLs. In the case of our sources (considered extreme HBLs), the inferred pressure ranges from $\sim 2\times10^{-3}~\rm{N~m}^{-2}$ to $\sim 6\times10^{-2}~\rm{N~m}^{-2}$ (Fig. 2 of Ghisellini et al. 2002). These are several orders of magnitude larger than the external gas pressure, implying the jets cannot be confined by the external pressure of the galactic atmosphere, at least on parsec scales. This is not unexpected, since it is commonly accepted that the jets start out supersonic near the central black hole. One of the goals of the [Chandra]{} observations was to determine whether HBLs and LBLs are characterized by different environments. The two classes exhibit different properties in terms of their Spectral Energy Distributions (e.g., Donato et al. 2001 and references therein), and the decreasing luminosity from LBLs to HBLs can be interpreted as a reduction in the jet power from the central energy source. One can thus ask the question of whether the difference in jet power in HBLs and LBLs is due to different gaseous environments. For example, one could hypothesize that the less powerful HBL jets live in a denser medium that is more efficient in decelerating them. Unfortunately, the present sample is insufficient to draw firm conclusions: of the 5 sources exhibiting diffuse X-ray emission, 4 are HBLs and only 1 is an LBL. More fundamentally, the short ACIS exposure prevents us from a detailed spectral analysis of the circumnuclear gas. Deeper follow-up observations are needed to this aim, as well as a larger, statistical sample of both HBLs and LBLs. On the optical side, we note that recent observations reveal no differences in the host galaxy properties of the two BL Lac subclasses (Urry et al. 2000, Scarpa et al. 2000). Diffuse X-ray emission on scales of 100 kpc or more, typical of a cluster of galaxies, is present in PKS 0548–322 and PKS 2005–489. This finding independently confirms previous claims, based on optical imaging, that these two BL Lacs reside in clusters of moderate-to-poor richness. However, the spatial parameters of the X-ray emission of the cluster are unconstrained in our ACIS images, while the X-ray emission of the cluster gas dominates the spectrum integrated along the line of sight. The cluster temperature and luminosity appears consistent (within 2$\sigma$) with the values found for normal clusters, i.e., clusters with no AGN activity (e.g., Mahdavi et al. 1997, Mulchaey & Zabludoff 1998; Mushotzky 1998; Markevitch 1998), suggesting that the presence of the AGN does not affect the global properties of the gas. On the other hand, it is also interesting that the two brightest X-ray BL Lacs of the sample reside in an extended environment, suggesting that the gas can affect nuclear activity. For example, Ellingson et al. (1991) suggested that nuclear activity could be triggered by galaxy-galaxy interactions and merging in the cluster core. Such a possibility is supported in the case of PKS 0548–322 by the presence of tidal interactions between the host galaxy of the BL Lac and nearby companion galaxies as seen in the optical image (Falomo et al. 1995). In conclusion, we detected diffuse X-ray emission around the cores of 3, and possibly 5, BL Lacertae objects within our short exposures. The core radii and luminosities of the diffuse X-ray emission are similar to those observed in FRI radio galaxies. This finding supports current unification models for radio-loud AGN, which attempt at unifying BL Lacs and FRI galaxies through orientation. Future work will include deeper and additional X-ray observations of a larger, statistical sample of BL Lacs, in order to put this conclusion on a firmer basis and to determine in greater detail the physical properties of the diffuse gas. Antonucci, R. 1993, ARA&A, 31, 473 Barthel, P. D. 1989, ApJ, 336, 606 Birkinshaw, M., Worrall, D. M., & Hardcastle M.J. 2002, MNRAS, 335, 142 Canosa, C. M., Worrall, D. M., Hardcastle, M. J., & Birkinshaw, M. 1999, MNRAS, 310, 30 De Grandi, S., Böhringer, H., Guzzo, L., et al. 1999, ApJ, 514, 148 De Grandi, S., & Molendi, S. 2002, ApJ, 567, 163 Dickey J. M., & Lockman, F. J. 1990, ARA&A, 28, 215 Donato, D., Ghisellini, G., Tagliaferri, G., & Fossati, G. 2001, A&A, 375, 739 Ellingson, E., Yee, H. K. C., & Green, R. F. 1991, ApJ, 371, 49 Ettori, S. 2000, MNRAS, 311, 313 Fabbiano, G., Kim, D. -W., & Trinchieri, G. 1992, ApJS, 80, 531 Falomo, R., & Kotilainen, J. K. 1999, A&A, 352, 85 Falomo, R., Pesce, J. E., & Treves, A. 1995, ApJ, 438, L9 Fossati, G., Maraschi, L., Celotti, A., Comastri, A., & Ghisellini, G. 1997, MNRAS, 289, 136 Ghisellini, G., Celotti, A., & Costamante, L. 2002, A&A, 386, 833 Giommi, P., Barr, P., Pollock, A. M. T., Garilli, B., & Maccagni, D. 1990, ApJ, 356, 432 Hardcastle, M. J., Worrall, D. M., & Birkinshaw, M. 1999, MNRAS, 305, 246 Kollgaard, R. I., Wardle, J. F. C., Roberts, D. H., & Gabuzda, D. C. 1992, AJ, 104, 1687 Kubo, H., Takahashi, T., Madejski, G., et al. 1998, ApJ, 504, 693 Mahdavi, A., Böehringer, H., Geller, M. J., & Ramella, M. 1997, ApJ, 483, 68 Markevitch, M. 1998, ApJ, 504, 27 Mulchaey, J. S., & Zabludoff, A. I. 1998, ApJ, 496, 73 Mushotzky, R. 1998, PNAS, 95, 72 Padovani, P., & Giommi, P. 1995, ApJ, 444, 567 Padovani, P., & Urry, C. M. 1991, ApJ, 368, 373 Padovani, P., Costamante, L., Giommi, P., et al. 2001, MNRAS, 328, 931 Perlman, E. S., Stocke, J. T., Schachter, J.F., et al. 1996, ApJS, 104, 251 Pesce, J. E., Falomo, R., & Treves, A. 1995, AJ, 110, 1554 Pesce, J. E., Sambruna, R. M., Urry, C. M., & Scarpa, R. 2000, HEAD, 32.0302 Pesce, J. E., Sambruna, R. M., Tavecchio, F., et al. 2001, ApJ, 556, L79 Ravasio, M., Tagliaferri, G., Ghisellini, G., et al. 2002, A&A, 383, 763 Sambruna, R. M., Maraschi, L., & Urry, C. M. 1996, ApJ, 463, 444 Sambruna, R. M., Maraschi, L., Tavecchio, F., et al. 2002, ApJ, 571, 206 Scarpa, R., Urry, C. M., Falomo, R., Pesce, J. E., & Treves, A. 2000, ApJ, 532, 740 Sarazin, C. L. 1988, [X-ray emission from clusters of galaxies]{}, Cambridge University Press Urry, C. M., & Padovani, P. 1995, PASP, 107, 803 Urry, C. M., Padovani, P., & Stickel, M. 1991, ApJ, 382, 501 Urry, C. M., Scarpa, R., O’Dowd, et al. 2000, ApJ, 532, 816 Veron-Cetty, M. -P., & Veron, P. 2001, A&A, 374, 92 Worrall, D. M. 2002, NewAR, 46, 121 Worrall, D. M., & Birkinshaw, M. 1994, ApJ, 427, 134 Worrall, D. M., & Birkinshaw, M. 2000, ApJ, 530, 178 Worrall, D. M., Birkinshaw, M., & Hardcastle, M. J. 2001, MNRAS, 326, L7 Wurtz, R., Stocke, J. T., Ellingson, E., & Yee, H. K. C. 1997, ApJ, 480, 547 APPENDIX A Serendipitous sources in the fields A glance at the images reveals the presence of several point sources in the fields of the targets. We used the `CIAO` tool [wavdetect]{} to search for serendipitous X-ray sources in the f.o.v. In the algorithm, the parameter [scale]{} (a list of radii, in image pixels, of Mexican Hat wavelet functions) was left free to range between 1 and 16, and the parameter [threshold]{} (the number of detected spurious sources in a pixel map) held fixed at [$10^{-9}$]{}. The algorithm returns a list of elliptical regions that define the positions and the shapes of the detected sources. We next used the coordinates from [wavdetect]{} and its associated error regions to search for their optical counterparts on ESO archival plates. We used circular search regions with radii of 5because it is known that most reprocessed ACIS-I observations have an offset of up to 1.5. In Table 5 we list the coordinates (J2000), the net counts, and the detection significance in $\sigma$, where $\sigma$ is the ratio of the net source counts to the “Gehrels error” of the background counts (for more details, see [wavdetect]{} manual at the following URL http://asc.harvard.edu/toolkit/pimms.jsp). We report the hardness ratios, defined as the difference between the count rates in the 2–8 keV (hard band) and those in the 0.3–2 keV (soft band), divided by the 0.3–8 keV count rate. A positive value of the hardness ratio suggests that the source has a flat spectrum, while a positive value suggests a steep spectrum. Also listed in Table 5 are the ESO identification codes, the angular distance of the optical source from the X-ray source, and the apparent magnitudes in the red and blue bands. The NED and Simbad on-line catalogs were searched to find additional information on the optical counterparts. We now comment on individual fields. ----- -------------- ------------- ------- ------- ------------ ---------------- ------- ------- ------- ------ R.A. Dec. $c/s$ HR ($\sigma$) ESO Code Dist. $M_R$ $M_B$ Note (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) PKS 0548–322 A 05 50 52.17 -32 10 52.5 91 -0.65 36.85 U0525-02418996 0.560 16.3 16.8 B 05 50 36.63 -32 17 33.3 193 -0.59 74.42 U0525-02415715 0.406 14.6 14.8 (a) C 05 50 31.50 -32 17 39.2 28 -0.86 11.49 U0525-02414666 0.506 17.8 20.1 D 05 50 27.25 -32 13 49.0 17 -0.65 7.62 U0525-02413729 0.563 18.1 20.7 E 05 50 21.63 -32 16 49.4 18 -0.22 7.97 U0525-02412549 0.727 17.9 18.9 (b) F 05 51 14.78 -32 13 03.4 12 -0.83 5.12 U0525-02423572 1.284 11.9 12.0 G 05 51 13.26 -32 25 47.5 32 0.06 11.11 - - - - 3C 371 A 18 07 37.87 69 51 12.1 8 0.00 3.61 - - - - B 18 07 26.57 69 46 26.6 38 -0.79 11.26 U1575-03911444 0.521 19.2 19.2 C 18 07 04.30 69 49 32.8 7 0.14 2.88 - - - - D 18 06 50.17 69 46 23.1 7 -1.00 2.59 - - - - E 18 07 27.11 69 45 32.1 28 -0.79 7.87 - - - - F 18 06 58.78 69 43 59.0 36 -0.17 10.59 - - - - G 18 06 03.36 69 51 17.4 19 -0.79 7.34 U1575-03907608 1.382 16.9 19.3 PKS 2005–489 A 20 09 41.06 -48 47 22.9 8 -0.25 3.76 - - - - B 20 09 13.88 -48 46 38.0 8 -0.25 3.74 - - - - C 20 09 07.14 -48 55 35.2 33 -1.00 14.55 U0375-38531345 0.385 9.7 11.4 (c) D 20 09 52.53 -48 56 08.9 30 -0.20 12.43 U0375-38543383 0.536 18.2 19.0 E 20 09 36.39 -48 42 02.9 11 -0.27 4.59 U0375-38539109 1.104 18.2 19.4 F 20 09 30.34 -48 56 19.8 7 -0.71 3.53 U0375-38537485 1.560 18.2 19.9 G 20 08 30.15 -48 50 33.2 17 -0.53 6.06 - - - - BL LAC A 22 02 53.68 42 17 48.5 32 -0.94 16.11 U1275-16943177 2.131 13.0 14.5 B 22 02 51.03 42 13 02.2 15 -0.07 7.28 U1275-16942082 0.588 18.0 18.2 C 22 02 43.81 42 13 45.3 9 0.56 4.90 - - - - D 22 02 14.64 42 13 53.4 38 -0.58 18.69 - - - - E 22 03 11.04 42 14 36.9 10 -0.20 4.83 - - - - F 22 03 09.86 42 13 37.0 6 -0.33 3.40 U1275-16950337 0.420 13.9 15.0 G 22 03 03.50 42 14 01.2 8 -0.50 4.22 - - - - H 22 02 54.06 42 11 03.6 19 -0.26 8.94 U1275-16943366 2.604 17.4 18.1 I 22 02 30.81 42 11 29.2 7 0.14 3.32 - - - - L 22 02 18.88 42 22 28.9 41 0.07 1.88 - - - - 1ES 2321+514 A 23 24 06.48 42 11 52.9 16 0.12 8.27 - - - - B 23 23 40.64 42 05 29.0 13 0.38 6.34 U1275-18369935 0.463 17.9 18.6 C 23 23 35.36 42 05 51.8 18 -0.78 8.70 - - - - D 23 23 33.08 42 07 42.1 9 -0.33 4.18 - - - - E 23 24 16.16 42 14 28.7 12 -0.33 5.94 - - - - F 23 24 02.15 42 05 21.2 8 0.50 4.25 U1275-18374011 1.440 16.4 17.1 G 23 23 27.33 42 08 45.9 8 0.25 3.57 - - - - H 23 24 01.26 42 00 46.5 18 -0.56 6.13 - - - - I 23 23 43.62 42 01 44.4 13 -0.54 5.22 U1275-18370478 3.179 18.4 18.6 L 23 23 26.60 42 05 53.7 13 -0.54 5.51 - - - - M 23 23 11.51 42 06 36.6 17 -0.53 5.69 - - - - 1ES 2344+514 A 23 47 27.14 51 40 18.4 9 -0.56 5.56 U1350-18688025 1.522 18.7 19.1 B 23 47 10.40 51 40 32.3 8 -0.25 3.90 U1350-18682932 0.944 16.5 17.9 C 23 46 36.64 51 42 30.5 16 -0.50 7.86 - - - - D 23 47 34.55 51 43 52.2 7 -1.00 3.36 U1350-18690283 0.840 11.2 11.8 (d) E 23 47 23.94 51 41 36.8 7 -0.43 3.47 - - - - F 23 48 06.02 51 40 16.9 9 -0.56 4.18 - - - - G 23 47 32.77 51 34 42.0 4 -1.00 1.81 - - - - ----- -------------- ------------- ------- ------- ------------ ---------------- ------- ------- ------- ------ [Columns]{}: 1=Detected source; 2=Right Ascension at J2000; 3=Declination at J2000; 4=Net count rate in 0.3–8 keV; 5=Hardness ratio, defined as $(h-s)/(h+s)$ where $h$ is the count rate in 2–8 keV and $s$ in 0.3–2 keV; 6=Detection significance (in $\sigma$, see Sect. 7); 7=ESO identification code; 8=Distance (in ) between the Chandra coordinate and the optical one; 9-10=Optical magnitude in the red and blue band; 11=(a) Galaxy 2MASXi J0550366-321733; (b) ROSAT X–ray source 1WGA J0550.3-3216 XrayS; (c) Star CCDM J20091-4856AB or HD 190857; (d) Star GSC 03650-00158. [PKS 0548–322]{}: In the [Chandra]{} field we found 7 serendipitous point sources (Table 5). Optical counterparts are found for all sources, except source G (see Fig. 5, left, for the overlay of the X-ray isocontours on the optical image). The X–ray sources detected with [wavdetect]{} are marked with capital letters. Source E coincides with 1WGA J0550.3-3216, an X–ray source previously identified by the satellite. Source B is a galaxy from the 2MASS catalog (2MASXi J0550366-321733), studied in previous optical works (Pesce et al. 1995, Falomo et al. 1995). This spiral galaxy coincides with galaxy G4 in the PKS 0548–322 cluster (Falomo et al. 1995), and has a redshift of $z$=0.072. Falomo et al. (1995) found strong emission lines in the optical spectrum and classified the source as an extreme Fe II emitting AGN. We measure $\sim$200 X-ray counts from the nucleus of this galaxy in the energy range 0.3–8 keV, sufficient for a crude spectral analysis. The X-ray spectrum is well fitted ($\chi^2_{\rm r}=1.37/16$) by a single power law with absorption fixed to Galactic, and photon index $\Gamma=2.21\pm 0.33$. This is consistent with the canonical slope of $\Gamma$=1.8 measured for Seyferts and other lower-luminosity AGN. The observed 2–10 keV flux is $F_{2-10 \rm keV}=1.4\times 10^{-13}$ ergs cm$^{-2}$ s$^{-1}$, corresponding to an intrinsic luminosity of $L_{2-10 \rm keV}=1.4\times 10^{42}$ erg s$^{-1}$, assuming a redshift of 0.072. [3C 371]{}: The Western elongation of the soft X-ray emission in Fig. 1 is due to the X–ray jet (Pesce et al. 2001). Seven X–ray sources are detected in the [Chandra]{} field but only two sources, B and G, have an optical counterpart. [PKS 2005–489]{}: In the X–ray image we detected 7 point sources. Sources D, E, and F have very weak optical counterparts, and source C corresponds to a F8V double or multiple star (CCDM J20091-4856AB also called HD 190857). For the positions of the detected X–ray sources on the optical image, see Fig. 5, right. [BL Lac]{}: Ten serendipitous sources are detected in the [Chandra]{} field. Only four of them (A, B, F, and H) have optical counterparts. [1ES 2321+419]{}: Eleven field sources are detected. Only three of them (B, F, and I) have optical counterparts. [1ES 2344+514]{}: Seven X-ray sources are found in this field. The weak X–ray source D coincides with a bright star (GSC 03650-00158).	{ "pile_set_name": "ArXiv" }
--- abstract: \| Decentralized coded caching is studied for a content server with $N$ files, each of size $F$ bits, serving $K$ active users, each equipped with a cache of distinct capacity. It is assumed that the users’ caches are filled in advance during the off-peak traffic period without the knowledge of the number of active users, their identities, or the particular demands. User demands are revealed during the peak traffic period, and are served simultaneously through an error-free shared link. A new decentralized coded caching scheme is proposed for this scenario, and it is shown to improve upon the state-of-the-art in terms of the required delivery rate over the shared link, when there are more users in the system than the number of files. Numerical results indicate that the improvement becomes more significant as the cache capacities of the users become more skewed.\ author: - bibliography: - 'Report\_Asilomar.bib' title: \| Decentralized Coded Caching with Distinct\ Cache Capacities --- Coded caching, decentralized caching, network coding, proactive caching. Introduction {#Intro} ============ The ever-increasing mobile data traffic has imposed a great challenge on the current network architectures. The growing demand is typically addressed by increasing the achievable data rates; however, moving content to the network edge has recently emerged as a promising alternative solution as it reduces both the bandwidth requirements and the delay. The use of edge caching is further motivated by the continuous drop in the cost of memory. In this paper, we consider an extreme form of edge caching, in which contents are stored directly at user terminals in a proactive manner. Proactive caching of popular content during off-peak traffic periods also helps flattening the high temporal variability of traffic. [@DowdyCaching; @AlmerothCacing]. Proactive caching is performed in two phases: The placement phase takes place during off-peak traffic hours, when the resources are abundant, and the users’ caches are filled by the server without knowing the future user demands. When the user demands are revealed, the delivery phase is performed, in which a common message is transmitted from the server to all the users over the shared communication channel. Each user decodes its requested file by combining the bits received in the delivery phase with the contents of its local cache. The goal is to minimize the delivery rate, which guarantees that all the user demands are satisfied, independent of the demand combination of the users. Research on caching over the past decade has mainly focused on the placement phase in order to identify the most popular contents to be cached locally at user terminals [@baev2008approximation; @BorstCaching]. Recently, coded caching scheme was introduced in [@MaddahAliCentralized] for proactive caching, and it is shown that by storing and transmitting coded contents, and designing the placement and delivery phases jointly, it is possible to significantly reduce the delivery rate compared to uncoded caching. A centralized caching scenario is studied in [@MaddahAliCentralized], in which the number and the identities of the users are known in advance by the server. This allows coordination of the cache contents across the users during both the placement and delivery phases; such that, by carefully placing pieces of contents in user caches, a maximum number of multicasting opportunities are created for tranmission during the delivery phase. Several other recent work has considered centralized coded caching, and the required delivery rate has been further reduced [@ZhiChenXOR; @MohammadDenizTCom; @KaiWanUncodedCaching]. In practice, however, the number or the identity of active users that will participate in the delivery phase might not be known in advance during the placement phase. In such a scenario, called decentralized coded caching, coordination across users is not possible during the placement phase. However, Maddah-Ali and Niesen proposed a scheme that randomly caches parts of each content at each user, and can still exploit multicasting opportunities in the delivery phase, albeit limited compared to the centralized setting [@MaddahAliDecentralized]. Decentralized coded caching has been studied in various other settings, e.g., files with different popularities [@NiesenNonuniform; @JiArXivNonuniform], and distinct lengths [@ZhangDistinctFileSizes], online caching [@PedarsaniOnlineCaching], etc. Most of the existing literature on coded caching assume identical cache sizes across users. Recently, in [@WangHeterogenous] decentralized caching to users with heterogeneous cache sizes is studied, and by extending the scheme proposed in [@MaddahAliDecentralized] to this scenarios, authors have shown that significant gains can still be obtained compared to uncoded caching. In this paper, we propose a novel decentralized caching algorithm for users with distinct cache capacities. We show that the proposed scheme requires a smaller delivery rate than the one achieved in [@WangHeterogenous]. The simulation results illustrate that the improvement in the delivery rate is more significant when the distribution of the cache capacities across users is more skewed. The rest of this paper is organized as follows. The system model is introduced in Section \[SystemModel\]. In Section \[s:Results\], we introduce the proposed coded caching scheme, analyze its performance in terms of the delivery rate. The performance of the proposed caching scheme is compared with the state-of-the-art result, and some numerical results are presented in Section \[s:Comparison\]. We conclude the paper in Section \[Conc\]. Notations: The set of integers $\left\{ 1, ..., K \right\}$ is denoted by $\left[ 1:K \right]$. Notation $\oplus$ illustrates the bitwise XOR operation. For two sets $Q$ and $P$, $Q \backslash P$ is a set including the members of $Q$ and excluding the members of $P$. Notation $\left\| {.} \right\|$ represents cardinality of a set or length of a file. We use the notation $\bar \oplus$ to represent the bitwise XOR operation between binary sequences with different lengths. The arguments of $\bar \oplus$ are first zero-padded to have the same length as the longest argument, and then they are bitwise XOR-ed. System Model {#SystemModel} ============ A server with $N$ independent $F$-bit files, $W_1, ..., W_N$, is considered, where each file is assumed to be uniformly distributed over $\left[ 1:2^F \right]$. There are $K$ active users, $U_1, ..., U_K$, where user $U_k$ is equipped with a cache of capacity $M_{k}F$ bits, with $M_k \le N$, $\forall k$. We denote the cache capacities by vector $\mu \triangleq (M_1,\ldots,M_K)$. Let $Z_k$ denote the contents of $U_k$’s cache at the end of the placement phase. Unlike in centralized coded caching [@MaddahAliCentralized], cache contents are independent of the number of users, their identities, or the user requests. User requests are revealed after the placement phase, where $d_k \in \left[ 1:N \right]$ denotes the file requested by user $U_k$, $k=1, \ldots, K$. User demands are served simultaneously through an error-free shared link. Let $X$ denote the $RF$-bit message transmitted over the shared link by the server to enable each user $U_k$ to decode its requested file $W_{d_k}$, together with its local cache content. Our goal is to characterize the minimum rate $R(\mu)$; such that, each user can decode its desired file with arbitrarily small probability of error, independent of the particular demand combination. Decentralized Coded Caching {#s:Results} =========================== We first illustrate our decentralized coded caching scheme on the following example. \[DecDistCacheSizes\] Consider a caching system with $N = 2$ files $W_1$ and $W_2$, and $K = 4$ users. It is assumed that the cache capacity of user $U_k$ is ${M_k} = {\left( {1/2} \right)^{4 - k}}M$, $\forall k \in \left[ 1:4 \right]$. In the placement phase, user $U_k$ caches a random $M_{k}F/2$ bits of each file independently. Since there are $N=2$ files in the database, a total of $M_{k}F$ bits are cached by user $U_k$. When $N<K$, it can be shown that the worst-case user demands happens when $N$ users with the smallest cache capacities have different requests. For this particular example, we have $M_1 \le \cdots \le M_4$, and the worst-case happens when users $U_1$ and $U_2$ request distinct files. Hence, we can assume the worst-case demand combination of $d_k = 1$, if $k=1,3$, and $d_k = 2$, otherwise. The contents served in the delivery phase are divided into three distinct parts, where $X_i$ is delivered in part $i$, for $i=1, 2, 3$. Thus, the common message is $X=\left( X_1, X_2, X_3 \right)$. We further divide the message $X_2$ into three pieces $X_2^1$, $X_2^2$, and $X_2^3$. Below, we explain the purpose of each part in detail. 1. In the first part of the delivery phase, the bits of each requested file which have not been cached by any user are directly delivered by the server. The following contents are delivered in this part. $X_1=\left( W_{1,\left\{ \emptyset \right\}}, W_{2,\left\{ \emptyset \right\}} \right)$. 2. The bits of the file requested by a user having been cached by another user are transmitted in the second part of the delivery phase. The server first delivers each user the bits of its requested file which are in the cache of one user with the same request. Then, each user receives the bits of its requested file which are in the cache of a single user with different request. By delivering the following contents, user $U_k$ can obtain the bits of file $W_{d_k}$ having been cached in user $U_l$, for $k,l \in \left[1:4\right]$, such that $l \ne k$. $X_2^1$ $=$ $\left( W_{1,\left\{ 3 \right\}} \right.$ $\bar \oplus$ $W_{1,\left\{ 1 \right\}}$, $W_{2,\left\{ 4 \right\}}$ $\bar \oplus$ $\left. W_{2,\left\{ 2 \right\}} \right)$, $X_2^2$ $=$ $\left( W_{1,\left\{ 4 \right\}} \right.$ $\bar \oplus$ $W_{1,\left\{ 2 \right\}}$, $W_{2,\left\{ 3 \right\}}$ $\bar \oplus$ $\left. W_{2,\left\{ 1 \right\}} \right)$, $X_2^3$ $=$ $\left( W_{1,\left\{ 2 \right\}} \right.$ $\bar \oplus$ $\left. W_{2,\left\{ 1 \right\}} \right)$. 3. In the last part, the server delivers the users the bits of their requested files which have been cached by more than one another user. Accordingly, each user $U_k$, $\forall k \in \left[ 1:4 \right]$, can obtain all the bits of file $W_{d_k}$ which are in the cache of users in any set $S \subset \left[ {1:4} \right]\backslash \left\{ k \right\}$, where $\left\| S \right\| \ge 2$, after receiving the following contents. $X_3$ $=$ $\left( W_{1,\left\{ 2,3 \right\}} \right.$ $\bar \oplus$ $W_{2,\left\{ 1,3 \right\}}$ $\bar \oplus$ $W_{1,\left\{ 1,2 \right\}}$, $W_{1,\left\{ 2,4 \right\}}$ $\bar \oplus$ $W_{2,\left\{ 1,4 \right\}}$ $\bar \oplus$ $W_{2,\left\{ 1,2 \right\}}$, $W_{1,\left\{ 3,4 \right\}}$ $\bar \oplus$ $W_{1,\left\{ 1,4 \right\}}$ $\bar \oplus$ $W_{2,\left\{ 1,3 \right\}}$, $W_{2,\left\{ 3,4 \right\}}$ $\bar \oplus$ $W_{1,\left\{ 2,4 \right\}}$ $\bar \oplus$ $W_{2,\left\{ 2,3 \right\}}$, $W_{1,\left\{ 2,3,4 \right\}}$ $\bar \oplus$ $W_{2,\left\{ 1,3,4 \right\}}$ $\bar \oplus$ $W_{1,\left\{ 1,2,4 \right\}}$ $\bar \oplus$ $\left. W_{2,\left\{ 1,2,3 \right\}} \right)$. After these parts, each user can decode all the missing bits of its desired file. To find the delivery rate, we first note that, by the law of large number, the length of the subfile $W_{k,V}$, for any set $V \subset \left[ {1:K} \right]$, is approximately given by $$\label{SizeSubfileExample} \left\| {{W_{k,V}}} \right\| \approx \prod\limits_{i \in V} {\left( {\frac{{{M_i}}}{2}} \right)} \prod\limits_{j \in \left[ {1:4} \right]\backslash V} {\left( {1 - \frac{{{M_j}}}{2}} \right)} F, \quad \forall k \in \left[ {1:K} \right].$$ For the example under consideration, when $M = 1$, i.e., $\mu = \left\{ {1/8,1/4,1/2,1} \right\}$, the delivery rate is $1.758$, while the delivery rate of the scheme proposed in [@WangHeterogenous] for this setting is $2.681$. Hence, the proposed scheme provides $34.43\%$ reduction in the delivery rate compared to the state-of-the-art result for this example. Placement Phase --------------- Since the active users are not known in advance in the decentralized setting, cache contents cannot be coordinated among the users. Similarly to the placement phases of the the decentralized coded schemes in the literature [@MaddahAliDecentralized; @WangHeterogenous], user $U_k$ caches a random $M_{k}F/N$ bits of each file independently, for $k=1, ..., K$. Since $N$ files are hosted in the database, a total of $M_{k}F$ bits are cached by each user $U_k$, and hence, the corresponding cache-capacity constraint is satisfied. For any set $V \subset \left[ {1:K} \right]$, let $W_{i,V}$ represent the bits of file $W_i$ that have been exclusively cached by the users in set $V$ at the end of the placement phase, i.e., $W_{i,V} \subset Z_k$, $\forall k \in V$, and $W_{i,V} \cap Z_k = \emptyset$, $\forall k \in \left[ 1:K \right] \backslash V$. Delivery Phase -------------- User demands are revealed at the beginning of the delivery phase. Without loss of generality, we re-label the users such that the first $K_1$ users, referred to as group $G_1$, have the same request $W_1$, the next $K_2$ users, group $G_2$, request file $W_2$, and so on so forth. For notational convenience, we define ${S_i} \buildrel \Delta \over = \sum\limits_{l = 1}^i {{K_l}}$. Therefore, the user demands are as follows: $$\label{DemandsGeneralCase} d_k = i,\quad \mbox{for $i=1, ..., N,$ and $k=S_{i-1} + 1, ..., S_i$},$$ where we set $S_0 = 0$. We further order the users within a group according to their cache sizes, and assume, without loss of generality, that ${M_{{S_{i - 1}} + 1}} \le {M_{{S_{i - 1}} + 2}} \le \cdots \le {M_{{S_i}}}$, for $i= 1, \ldots ,N$. The proposed delivery phase is presented in Algorithm \[DeliveryHeterogenous\]. For any general demand combination described above, the delivery phase presented in Algorithm \[DeliveryHeterogenous\] contains two procedures CODED DELIVERY and RANDOM DELIVERY, and in each case the server chooses the one with the smaller delivery rate. Below, we explain these two procedures in detail. The CODED DELIVERY procedure includes three distinct parts, where the content delivered in part $i$ is denoted by $X_i$, $i=1,2,3$, and the common message $X=\left( X_1, X_2, X_3 \right)$ is sent to all the users during the delivery phase. The message transmitted in part 2, $X_2$, is further divided into three pieces $X_2^1$, $X_2^2$, and $X_2^3$, i.e., $X_2=\left( X_2^1, X_2^2, X_2^3 \right)$. Based on the aforementioned placement phase, the main motivation of the CODED DELIVERY procedure is to enable each user to recover the missing bits of its requested file which have been cached by $i$ other users, $\forall i \in \left\{ 0, ..., K-1 \right\}$. In Part 1 of the this procedure, each user receives the bits of its requested file which have not been cached by any user. The purpose of Part 2 is to enable each user to obtain all the missing bits of its request that have been cached by another single user. First, consider the message $X_2^1$. For $i=1, ..., N$, each user $k \in \left[ S_{i-1}+1:S_{i} \right]$ (i.e., $U_k \in G_i$), has access to bits $W_{i,\{k\}}$ locally in its cache, and with $X_2^1$ it can decode all the pieces $W_{i,\left\{ l \right\}}$, $\forall l \in \left[ S_{i-1}+1:S_{i} \right]$, i.e., the bits of its demand $W_i$, which are in the cache of another user in the same group, and no other user. Delivering the messages $X_2^2$ and $X_2^3$ together helps the users to decode the bits of their requested files having been cached by a single user in other groups; that is, after receiving $\bigcup\limits_{k = {S_{j - 1}} + 1}^{{S_j} - 1} \left( W_{i,\left\{ k \right\}} \bar \oplus W_{i,\left\{ {k + 1} \right\}}\right)$, $\bigcup\limits_{k = {S_{i - 1}} + 1}^{{S_i} - 1} \left( W_{j,\left\{ k \right\}} \bar \oplus W_{j,\left\{ {k + 1} \right\}} \right)$, and $W_{i,\left\{S_{j-1}+1\right\}}$ $\bar \oplus$ $W_{j,\left\{S_{i-1}+1\right\}}$, the users in both groups $G_i$ and $G_j$ can obtain the missing bits of their requested files that have been cached by a user in another group, for $i=1, ..., N-1$ and $j=i+1, ..., N$ (and no other user). Note that, having received $X_2^2$, the third message $W_{i,\left\{S_{j-1}+1\right\}}$ $\bar \oplus$ $W_{j,\left\{S_{i-1}+1\right\}}$ is the smallest number of bits (based on the assumption ${M_{{S_{l - 1}} + 1}} \le {M_{{S_{l - 1}} + 2}} \le \cdots \le {M_{{S_l}}}$, $\forall l \in \left[ 1:N \right]$) that enable all the users in both groups $G_i$ and $G_j$ to obtain the missing bits of their desired files that are in the cache of users in the other group, for $i=1, ..., N-1$ and $j=i+1, ..., N$. Part 3 of our algorithm is the same as the delivery phase proposed in [@WangHeterogenous Algorithm 2], and it is performed to send the users the missing bits of their requests that have been cached by more than one user. Finally, in the RANDOM DELIVERY procedure, as in the DELIVERY procedure of [@MaddahAliDecentralized], the server transmits enough random linear combinations of the bits of file $W_i$ to the users in group $G_i$ to make sure they all can decode the file, for $i=1, \ldots, N$. Delivery Rate Analysis ---------------------- In the following, we evaluate the delivery rate of the proposed caching scheme for the worst-case user demands. Consider first the case $N \ge K$. It can be argued in this case that the worst-case user demands happens if each file is requested by at most one user. Hence, by re-ordering the users, for the worst-case user demands, we have $K_i = 1,$ for $1 \le i \le N$, and $K_i = 0,$ otherwise. In this case, it can be shown that the CODED DELIVERY procedure requires a lower delivery rate than the RANDOM DELIVERY procedure; hence, the server uses the former. In this case, it is possible to simplify the CODED DELIVERY procedure such that, only message $X_2^3$ is transmitted in Part 2, when $N \ge K$, i.e., $X_2 = X_3^2$. The corresponding common message, $X=\left( X_1, X_2^3, X_3 \right)$, transmitted over the CODED DELIVERY procedure, reduced to the delivery phase of [@WangHeterogenous Algorithm 2]. Thus, the proposed scheme achieves the same delivery rate as [@WangHeterogenous Algorithm 2] when $N \ge K$. Next, we consider the case $N<K$. It is possible to show that the worst-case user demands in this case happens when $N$ users with the smallest cache capacities all request different files, i.e., they end up in different groups. The delivery rate of the proposed delivery phase when $N<K$ is presented in the following theorem. The proof of the worst-case demand distribution as well as Theorem \[TheDelRateDecDistCacheSizes\] are skipped due to space limitations; however, they can be found in the longer version of the paper in [@Asilomar16]. \[TheDelRateDecDistCacheSizes\] In a decentralized caching system with $N$ files in the database, each of size $F$ bits, and $K$ users with cache capacities $\mu = \left\{ {{M_1},...,{M_K}} \right\}$, such that $M_1 \le M_2 \le \cdots \le M_K$, the following delivery rate-cache capacity trade-off is achievable when $N<K$: $$\begin{aligned} \label{OurDeliveryRateHeterogenous} R_c\left( \mu \right) = & \min \left\{ \sum\limits_{i = 1}^K {\left[ {\prod\limits_{j = 1}^i {\left( {1 - \frac{{{M_j}}}{N}} \right)} } \right]} \right. \nonumber\\ & \quad \;\; \left.- \Delta {R_1}\left( \mu \right) - \Delta {R_2}\left( \mu \right),\sum\limits_{i = 1}^N \left( {1 - \frac{{{M_i}}}{N}} \right) \right\},\end{aligned}$$ where \[DeltaR\] $$\begin{aligned} \label{DeltaRone} \Delta {R_1}\left( \mu \right) =& \left( {K - N} \right)\prod\limits_{l = 1}^K {\left( {1 - \frac{{{M_l}}}{N}} \right)},\\ \Delta {R_2}\left( \mu \right) =& \left[ {\sum\limits_{k = 1}^{K - N} {\left( {\frac{ (k-1) M_{k + N} }{{N - {M_{k + N}}}}} \right)} } \right]\prod\limits_{l = 1}^K {\left( {1 - \frac{{{M_l}}}{N}} \right)}. \label{DeltaRtwo}\end{aligned}$$ ![Illustration of cache capacity distribution normalized by $\sum\limits_{k = 1}^K {{M_k}}$ for different $\alpha$ values, and $K=50$. The $x$-axis corresponds to the user index $k$.[]{data-label="CacheCapacitiesDistribution"}](Cache_Capacities_Distribution.pdf) Comparison with the State-of-the-Art and Numerical Results {#s:Comparison} ========================================================== In this section, the proposed caching scheme is compared with the scheme proposed in [@WangHeterogenous] both analytically and numerically. We note that, although the scheme presented in [@WangHeterogenous] is for $N \ge K$, it can also be applied to the case $N < K$, and the same delivery rate as [@WangHeterogenous Theorem 2], denoted here by $R_b(\mu)$, can be achieved. Hence, in the following, when we refer to the scheme stated in [@WangHeterogenous Algorithm 2] for $N < K$, we consider its generalization to this scenario. When $N<K$, according to [@WangHeterogenous Theorem 2] and , we have $$\label{DeliveryRateHeterogenousComparison} {R_b}\left( \mu \right) - {R_c}\left( \mu \right) \ge \Delta {R_1}\left( \mu \right) + \Delta {R_2}\left( \mu \right) {\mathrel{\overset{\makebox[0pt]{\mbox{\normalfont\tiny\sffamily (a)}}}{>}}}0.$$ The inequality (a) holds as long as $N < K$. Therefore, when the number of files in the database is smaller than the number of active users in the delivery phase, the proposed coded caching scheme requires a smaller delivery rate than the one presented in [@WangHeterogenous]. For the numerical results, we consider an exponential cache distribution among users, such that the cache capacity of user $U_k$ is given by $$\begin{aligned} M_k = {\alpha ^{K - k}}M,\end{aligned}$$ where $0 \le \alpha \le 1$, for $k=1, \ldots, K$, and $M$ denote the maximum cache capacity in the system. Thus, we have $\mu = \left\{ {{\alpha ^{K - 1}}M,{\alpha ^{K - 2}}M, \ldots, M} \right\}$, such that $M_1 \le M_2 \le \cdots \le M_K$. The distribution of cache capacities normalized by $\sum\limits_{k = 1}^K {{M_k}}$, i.e., $M_k/\sum\limits_{k = 1}^K {{M_k}}$ denoted by ${\bar M}_K$, $\forall k \in \left[1:K \right]$, is demonstrated in Fig. \[CacheCapacitiesDistribution\] for different values of $\alpha$, when $K=50$. Observe that, the smaller the value of $\alpha$, the more skewed the cache capacity distribution across users become. In the special case of $\alpha=1$, we obtain the homogeneous cache capacity model studied in [@MaddahAliDecentralized]. In Fig. \[N50K70\], the delivery rate of the proposed scheme, $R_c(\mu)$, is compared with that of the coded scheme proposed in [@WangHeterogenous], i.e., $R_b(\mu)$, when $N=50$, $K=70$, and $\alpha = 0.97$. The delivery rate is plotted in this figure versus the largest cache capacity in the system, $M$. As expected the performance improves, i.e., the delivery rate reduces as $M$ increases. We also clearly observe that the proposed scheme outperforms the scheme presented in [@WangHeterogenous]. The improvement is particularly significant for lower values of $M$. The cut-set lower bound for this setting is also included in the figure. Although the delivery rate of the proposed scheme approaches the lower bound for relatively small values of $M$, there is still a gap for large values of $M$, which may as well be due to the looseness of the lower bound. In order to see the effect of skewness of the cache capacities on the delivery rate, in Fig. \[N\_75\_K\_90\_AlphaVary\], the delivery rate of different schemes are plotted as a function of $\alpha \in \left[ 0.9, 1 \right]$, for $N=30$, $K=45$, and the largest cache capacity of $M=2$. The delivery rate of the proposed decentralized coded caching scheme is lower than the one presented in [@WangHeterogenous] for the whole range of $\alpha$ values under consideration, while the gain is more pronounced for smaller values of $\alpha$, i.e., as the distribution of cache capacities becomes more skewed. We also observe the gap to the cut-set lower bound also diminishes in this regime. ![Delivery rate versus $M$, where the cache capacity of user $k$ is $M_k = {\alpha ^{K - k}}M$, $k=1, \ldots, K$, when $\alpha=0.97$, $N=50$, and $K=70$.[]{data-label="N50K70"}](N_50_K_70.pdf) Conclusions {#Conc} =========== In this paper, we have studied coded caching to users with distinct cache capacities, and proposed a novel decentralized coded caching scheme that improves upon the best known delivery rate in the literature. The improvement is achieved by improving the delivery of bits that have been cached by none of the users, or by only a single user. In particular, the proposed scheme exploits the group-based coded caching scheme we have introduced previously for centralized caching in a system with homogeneous cache capacities [@MohammadQianDenizITW]. Our numerical results show that the improvement upon the scheme proposed in [@WangHeterogenous] is even more pronounced when the cache capacities of the users are more skewed. We are currently aiming to improve the delivery rate for larger values of cache capacities by finding a more efficient coded delivery scheme for the delivery of the missing bits of the requested files that have been cached by more than one user. ![Delivery rate versus $\alpha \in \left[ 0.9, 1 \right]$, where $M_k = {\alpha ^{K - k}}M$, $N=30$, $K=45$, and $M=2$.[]{data-label="N_75_K_90_AlphaVary"}](N_30_K_45_AlphaVariable.pdf)	{ "pile_set_name": "ArXiv" }
--- abstract: \| [Abstract:]{} The physical history of the Universe is completed by including the quantum planckian and super-planckian phase before Inflation in the Standard Model of the Universe in agreement with observations. In the absence of a complete quantum theory of gravity, we start from quantum physics and its foundational milestone: The [universal]{} classical-quantum (or wave-particle) duality, which we extend to gravity and the Planck domain. As a consequence, classical, quantum planckian and superplanckian regimes are covered, and the usual quantum domaine as well. A new quantum precursor phase of the Universe appears beyond the Planck scale ($t_P$): $10^{-61} t_P \leq t \leq t_P$; the known classical/semiclassical Universe being in the range: $t_P \leq t \leq 10^{+61} t_P$. We extend in this way de Sitter universe to the quantum domain: [classical-quantum de Sitter duality]{}. As a result: (i) The classical and quantum dual de Sitter Temperatures and Entropies are naturally included, and the different (classical, semiclassical, quantum planckian and super-planckian) de Sitter regimes characterized in a precise and unifying way. (ii) We apply it to relevant cosmological examples as the CMB, Inflation and Dark Energy. This allows to find in a simple and consistent way: (iii) Full quantum Inflationary spectra and their CMB observables, including in particular the classical known Inflation spectra and the quantum corrections to them. (iv) A whole unifying picture for the Universe epochs and their quantum precursors emerges with the cosmological constant as the vacuum energy, entropy and temperature of the Universe, clarifying the so called cosmological constant problem which once more in its rich history needed to be revised.\ [email protected], author: - \| Norma G. SANCHEZ\ CNRS LERMA Observatoire de Paris PSL University,\ Sorbonne Université, 61, Avenue de l’Observatoire, 75014 Paris, France title: New Quantum Phase of the Universe before Inflation and its Cosmological and Dark Energy Implications --- Introduction and Results ======================== The set of robust cosmological data (cosmic microwave background, large scale structure and deep galaxy surveys, supernovae observations, measurements of the Hubble-Lemaitre constant and other data) support the Standard (concordance) Model of the Universe and place de Sitter (and quasi-de Sitter) stages as a real part of it [@WMAP1],[@WMAP2],[@Riess],[@Perlmutter],[@Schmidt][@DES],[@Planck6]. Moreover, the physical classical, semiclassical and quantum planckian and super-planckian de Sitter regimes are particularly important for several reasons: [(i)]{} The classical, present time accelerated expansion of the Universe and its associated dark energy or cosmological constant in the today era: classical cosmological de Sitter regime. [(ii)]{} The semiclassical early accelerated expansion of the Universe and its associated Inflation era: semiclassical cosmological de Sitter (or quasi de Sitter) regime (classical general relativity plus quantum field fluctuations.) [(iii)]{} The quantum, very early stage preceeding the Inflation era: Planckian and super-Planckian quantum era. Besides its high conceptual and fundamental physics interest, this era could be of realistic cosmological interest for the test of quantum theory itself at such extreme scales, as well as for the search of gravitational wave signals from quantum gravity for e-LISA [@LISA] for instance, after the success of LIGO [@LIGO],[@DESLIGO]. In addition, this quantum stage should be relevant in providing quantum precursors and consistent initial states for the semiclassical (fast-roll and slow roll) inflation, and their imprint on the observable primordial fluctuation spectra for instance. Moreover, and as a novel result of this paper, this quantum era allows to clarify the issue of dark energy as the vacuum energy or cosmological constant of the Universe. [(iv)]{} de Sitter is a simple and smooth constant curvature vacuum background without any physical singularity, it is maximally symmetric and can be described as a hyperboloid embedded in Minkowski space- time with one more spatial dimension. Its radius, curvature and equivalent density are described in terms of only one physical parameter: the cosmological constant. In despite of the simplicity of de Sitter background, the generic nature of inflation, and the relevance of dark energy, there is no satisfactory description of de Sitter background, nor of inflation in string theory, - it is fair to recall here that this was pointed out 25 years ago [@deVega1994]. \[Contrary to Anti-de Sitter, de Sitter background does not appear as a solution of the effective string equations. The lack of a full conformal invariant string de Sitter description is not an handicap for de Sitter background, but to the current formulation or understanding of string theory, [@deVega1993] (by satisfactory description we mean in particular, one not based in tailored constructions, nor in conjectures)\]. The lack of a complete theory of quantum gravity (in field theory and in strings) does not preclude to explore and describe planckian and superplanckian gravity regimes. Instead of going from classical gravity to quantum gravity by quantizing general relativity (as it was tried with its well known developpements and shortcomings, (is not our aim here to review it), we start from Quantum physics and its foundational milestone: the classical-quantum (wave-particle) duality, and extend it to include gravity and the Planck scale domain, namely, wave-particle-gravity duality, (or classical-quantum gravity duality), [@Sanchez2019], [@Sanchez2003-1]. As a consequence, the different gravity regimes are covered: classical, semiclassical and quantum, together with the Planck domain and the elementary particle domain as well. This duality is [universal]{}, it includes the known classical-quantum duality as a special case and allows a general clarification from which physical understanding and cosmological results can be extracted as shown in this paper. This is not an assumed or conjectured duality. As the wave-particle duality, this does not rely on the number of space-time dimensions (compactified or not), nor on any symmetry or isometry nor on any other [at priori]{} condition. [In this paper]{}, we link the Standard Model of the Universe to the classical-quantum duality. We complete in this way the history of the Universe beyond the Inflation era and the current picture by including the quantum precursor phase within the Standard Model of the Universe in agreement with observations. Quantum physics is more complete than classical physics and contains it as a particular case: It adds a new quantum planckian and superplanckian phase of the Universe from the Planck time $t_P$ untill the extreme past $10^{-61} t_P$ which is an upper bound for the origin of the Universe, with energy $H = 10^{61} h_P$. Besides the arguments given above, the reasons supporting such a phase are many: (i) The generic and physical existence of classical-quantum duality in Nature. (ii) The universality of it. (iii) The consistent and concordant physical results and coherent whole picture obtained from it supported by observations. We provide an unifying description of the classical, semiclassical, quantum, planckian and superplanckian stages of the Universe, their relevant physical magnitudes: size, mass, vacuum energy density, cosmological constant, gravitational entropy and temperature and the relations between them. [The main results of this paper:]{} [1.]{} The classical dilute Universe today and the highly dense very early quantum superplanckian Universe are classical-quantum duals of each other in the precise meaning of the classical-quantum duality. This means the following: The classical Universe today $U_{\Lambda}$ is clearly characterized by the set of physical gravitational magnitudes or observables (age or size, mass, density, temperature, entropy) $\equiv (L_\Lambda, M_\Lambda, \rho_\Lambda, T_\Lambda, S_\Lambda)$ $$\label{ULambda1} U_{\Lambda} = (L_\Lambda, M_\Lambda, \rho_\Lambda, T_\Lambda, S_\Lambda)$$ The highly dense very early quantum Universe $U_Q$ is characterized by the corresponding set of quantum dual physical quantities $(L_Q, M_Q, \rho_Q, T_Q, S_Q)$ in the precise meaning of the classical-quantum duality: $$\label{UQ1} U_Q = (L_Q, M_Q, \rho_Q, T_Q, S_Q)$$ $$\label{Udual1} U_Q = \frac{u_P^2}{U_\Lambda}, \qquad u_P = (l_P, m_P, \rho_P, t_P, s_P)$$ $u_P$ standing for the corresponding quantities at the fundamental constant Planck scale, the [crossing scale]{} between the two main (classical and quantum) gravity domains. The classical $U_{\Lambda}$ and quantum $U_Q$ Universe eras or regimes (classical/semiclassical eras of the known Universe and its quantum planckian and superplanckian very early phases), satisfy Eqs.(\[ULambda1\])-(\[Udual1\]). The [total]{} Universe $U_{Q\Lambda}$ is composed by their classical/semiclassical and quantum phases: $$\label{Utotal1} U_{Q\Lambda} = \left(\; U_Q + U_{\Lambda} + u_P \;\right)$$ $Q$ stands for Quantum, $\Lambda$ for classical, and $P$ for the fundamental Planck scale constant values. In particular, the quantum dual de Sitter universe $U_Q$ is generated from the classical de Sitter universe $U_{\Lambda}$ through Eqs.(\[ULambda1\])-(\[Utotal1\]): [classical-quantum de Sitter duality]{}, as we done in Section III. The [total]{} (classical plus quantum dual) de Sitter universe $U_{Q\Lambda}$ is obtained in Section IV: [classical-quantum de Sitter symmetry]{}. This includes in particular the classical, quantum and total de Sitter Temperatures and Entropies (Sections V and VI). This allows to characterize in a complete and precise way the different classical, semiclassical, quantum planckian and superplanckian de Sitter regimes (Section VII). $H$ stands for the classical Hubble-Lemaitre constant, or its equivalent $\Lambda = 3 \; \left ({H}/{c}\right)^2$. $Q$ stands for quantum dual, and $QH$ (or $Q\Lambda$) for the total or complete quantities. For instance, the size of the Universe is the gravitational length $L_\Lambda = \sqrt{ 3/\Lambda}$ in the classical regime, it is the quantum Compton length $L_Q$ in the quantum dual regime (which is the full quantum planckian and superplanckian regime), and it is the Planck length $l_P$ at the fundamental Planck scale: the [crossing scale]{}. The [total]{} (or complete) size $L_{Q \Lambda}$ is the sum of the two components. Similarly, the horizon acceleration (surface gravity) $K_{\Lambda}$ of the Universe in its classical gravity regime becomes the quantum acceleration $K_Q$ in the quantum dual gravity regime. The temperature $T_{\Lambda}$, measure of the classical gravitational length or mass becomes the quantum Temperature $T_Q$ (measure of the quantum size or Compton length) in the quantum regime. Consistently, the Gibbons-Hawking temperature is [precisely]{} the quantum temperature $T_Q$. Similarly, the classical/semiclassical gravitational area or entropy $S_\Lambda$ (Gibbons-Hawking entropy) has its quantum dual $S_Q$ in the quantum gravity (Planckian and super-Planckian) regime. The concept of gravitational entropy is [the same]{} for any of the gravity regimes: $ Area / 4 l_P^2$ in units of $k_B$. (For a classical object of size $L_\Lambda$, this is the classical area $A_\Lambda$, for a quantum object, of size $L_Q$, this is the area $A_Q$.) [2. Results for Inflation.]{} We apply these results to Inflation, Dark Energy and the cosmological constant in the framework of the Standard Model of the Universe, (Sections VIII-XI). The precursor quantum phase of the known classical/semiclassical Inflation does appear, as well as the precursors for the classical standard eras and today Dark Energy era. H-inflation means the classical known Inflation (classical H) era, Q-inflation is its quantum dual precursor, QH stands for the total Inflation era including the known classical/semiclassical Inflation and its precursor: the quantum Inflation era (in the planckian and superplanckian) phase. The [total]{} or complete QH inflationary spectra turn out expressed as $$[\;\Delta^S_{k, \;QH}\;] = [\;\Delta^S_{k, \;H}\;]\; \left(\frac{1}{[ \;1 + (H/h_P)^2 \;] }\right)\; \frac{1}{( 1 - \delta \epsilon_{QH})^{1/2}}$$ $$[\;\Delta^T_{k, \;QH}\;] = [\;\Delta^T_{k, \;H}\;] \; \left(\frac{1}{[\; 1 + (H/h_P)^2 \;] }\right)$$ where $[\Delta^S_{k,\;H}]$ and $[\Delta^T_{k, \;H}]$ are the known standard spectra of scalar curvature and tensor perturbations in classical H Inflation Eqs.(\[DeltaSO\]). Here $\delta\epsilon_{QH}$ is the first order QH slow-roll parameter (computed in Section IX) which contain in particular the classical known slow-roll $\epsilon$ parameter, and $h_P$ is the value of the Hubble constant at the Planck scale (or mass Planck value $m_P$). The total QH spectra contain both: the standard known spectra of the classical/ semiclassical Inflation including its quantum corrections of order $(H/h_P)^2 = 10^{-12}$ in the classical/semiclassical gravity phase $H = 10^{-6} h_P$, at $t = 10^{6} t_P$, (or $10^{-5} M_P$ for the reduced Planck mass $M_P = m_P/\sqrt{8\pi}$), and their quantum dual spectra in the quantum precursor Inflation era $H_Q = 10^6 h_P$, at $t = 10^{-6} t_P$. The CMB observables: scalar spectral index $n_S$, ratio $r$ and departure from scalar invariance $\Delta$ are computed (in the two Inflation phases, classical H, quantum Q, and the total QH): In the classical H known phase, it yields in a simple and direct way the same quantum corrections to the spectra, sign and magnitude, as the quantum inflaton corrections [@Boya2005], [@Boya2006] in the Ginsburg-Landau effective approach to Inflation [@CiridVS],[@BDdVS]. The [departure from scale invariance]{} $\Delta_{QH} = (n_{s\;QH} - 1)/2 + r_{QH}/8$ , gets corrected as \[Eq.(\[DeltaQH2\])\]: $$\Delta _{QH} = \Delta\;[\; 1 - 2 \; (H/h_P)^2 \;] + \sqrt{\epsilon} \;(H/h_P)^2 \; [\;2 \sqrt{\epsilon} - \frac{m_P}{\sqrt{\pi}}\;] + O\;(H/h_P)^4,$$ where $\epsilon$ and $\Delta = (\;n_s - 1\;)/2 + r/8 $ are the slow roll parameter and the [departure from scale invariance]{} of the known classical H Inflation respectively. The QH corrections to the known scalar index $n_s$ and ratio $r$ of classical/semiclassical Inflation $ (H/h_P) = 10^{-6}$ are: $$\frac{r_{QH}}{r} - 1 = - 2 \;10^{-12}, \; \qquad \frac{n_{s\;QH}}{n_s} - 1 = 2 \;10^{-12}\;[\;1 - \frac{1}{n_s} (\;1- \frac{m_P}{2}\sqrt{\frac{\epsilon}{\pi}}\;)\;].$$ The QH factor modifying the Hubble constant and the complete QH inflation spectra: $$\label{sum1} QH \equiv \frac{H}{[\;1 + (H / h_P)^2\;]} = \frac{H_Q}{[\;1 + (H_Q / h_P)^2\;]} = H \; \sum _{n=0}^{\infty} (-1)^n \left(\frac{H}{h_P}\right)^2,$$ covers the [full classical and quantum range]{} $H \leq h_P$ and $H \geq h_P$. If $H< h_P$, it covers the known classical/semiclassical range. If $H > h_P$, it changes consistently to the quantum Hubble rate $H_Q = h_P^2/H$ in the quantum domain. [3. Results for Dark Energy.]{} This framework reveals enlighting for the issue of [Dark Energy]{} as discussed in Section X, and allows clarification into the cosmological constant problem as discussed in Section XI. The classical Universe today $U_\Lambda$ is precisely a [classical dilute gravity vacuum dominated by voids and supervoids]{} as shown by observations [@VoidsHistory], [@Voids1], [@VoidsPRL] whose observed $\rho_\Lambda$ or $\Lambda$ value today [@Riess],[@Perlmutter],[@Schmidt],[@DES],[@Planck6] is [precisely]{} the classical dual of its quantum precursor values $\rho_Q,\Lambda_Q$ in the quantum very early precursor vacuum $U_Q$ as determined by Eqs.(\[ULambda1\])-(\[UQ1\]). The high density $\rho_Q$ and cosmological constant $\Lambda_Q$ are precisely the quantum particle physics superplanckian value $10^{122}$. This is precisely expressed by Eqs.(\[ULambda1\])-(\[UQ1\]) applied to this case, Section X \[Eqs.(\[LambdaHvalue\])-(\[LambdaQLambda\])\]: $$\label{LambdaHvalue1} \Lambda = 3 H^2 = \lambda_P \left (\frac{H}{h_P}\right)^2 = \lambda_P \left (\frac{l_P}{L_H}\right)^2 = (2.846 \pm 0.076) \; 10^{-122}\; m_P^2$$ $$\label{LambdaQvalue1} \Lambda_Q = 3 H_Q^2 = \lambda_P \left (\frac{h_P}{H}\right)^2 = \lambda_P \left (\frac{L_H}{l_P}\right)^2 = (0.3516 \pm 0.094) \; 10^{122}\;h_P^2$$ $$\label{LambdaQLambda1} \Lambda_Q = \frac{\lambda_P^2}{\Lambda}, \qquad \lambda_P = 3 h_P^2$$ [The quantum dual value $\Lambda_Q$ is [precisely]{} the quantum vacuum value $\rho_Q = 10^{122}\; \rho_P$ obtained from particle physics:]{} $$\label{rhoQii1} \rho_Q = \rho_P \left(\frac{\Lambda_Q}{\lambda_P}\right) = \frac{\rho_P^2}{\rho_\Lambda} = 10^{122}\; \rho_P$$ Eqs.(\[LambdaHvalue1\])-(\[rhoQii1\]) are consistently supported by the data [@Riess],[@Perlmutter],[@Schmidt],[@DES],[@Planck6] which we also [link to the gravitational entropy and temperature of the Universe]{}, as we done in Section XI and summarized by Eqs.(\[Lambda1\]) - (\[LambdaQLambdavalue\]). The [complete]{} cosmological constant $\Lambda_{Q\Lambda}$ or total vacuum energy density $\rho_{Q\Lambda}$ is the sum of its classical and quantum components (corresponding to the classical today era and its quantum planckian and super-planckian precursor): $$\label{LambdaQLambda3} \Lambda_{Q \Lambda} = \lambda_P \left (\; \frac{\Lambda}{\lambda_P} + \frac{\lambda_P}{\Lambda} + 1\;\right) = \lambda_P \;(\; 10^{-122} + 10^{+122} + 1\;)$$ The observed $\Lambda$ or $\rho_\Lambda$ today is the [classical gravity vacuum]{} value of the classical Universe $ U_\Lambda$ [today]{}. Such observed value must be consistently in such way because of the [large classical]{} size of the Universe today $L_ \Lambda = \sqrt{3/\Lambda}$, and of the empty or vacuum dilute state today dominated by [ voids and supervoids]{} as shown by the set of large structure observations [@VoidsHistory], [@Voids1], [@VoidsPRL]. This is one main physical reason for such a [low]{} $\Lambda$ value at the present age today $10^{61} t_P$. Its precursor value and density $\Lambda_Q , \rho_Q$ is a high superplanckian value precisely because this is a high density very early [quantum cosmological vacuum]{} in the extreme past $10^{-61} t_P$ of the quantum superplanckian precursor phase $U_Q$. The quantum cosmological constant and associated density $\Lambda_Q = \rho_Q = 10^{122}$ (in Planck units) in the quantum precursor superplanckian phase $U_Q$ at $10^{-61} t_P$, (the extreme past), became the classical cosmological constant and density $\Lambda = \rho_\Lambda = 10^{-122}$ in the classical Universe $U_\Lambda$ today at $10^{61} t_P$. The superplanckian value is consistently in such way because is a extreme quantum gravity vacuum in the extreme quantum past $10^{-61} t_P$ with minimal entropy $S_Q = 10^{-122} = \Lambda = \rho_\Lambda$. All physical quantities: the vacuum energy density, the cosmological constant, the gravitational entropy and gravitational temperature, both classical and quantum are consistently linked by the [classical-quantum (or wave-particle) duality through the Planck scale]{} in agreement with observations Eqs.(\[ULambda1\])-(\[Udual1\]), (Sections X-XI). Eqs.(\[LambdaHvalue1\]) to (\[rhoQii1\]),(\[LambdaQLambda3\]), \[Eqs.(\[Lambda1\])-(\[LambdaQLambdavalue\])\] concisely and synthetically express such complete set of classical-quantum dual relations and [explain why]{} the classical gravitational vacuum: cosmological constant $\Lambda$ or density $\rho_\Lambda$ [coincides]{} with such observed [low value]{} $10^{-122}$ in Planck units, and [why]{} their corresponding quantum gravity precursor vacuum has such extremely [high]{} superplanckian density [value]{} $10^{122}$ in Planck units. This is [not]{} trivial, this is simple, deep and robust, (Section X). This is [not]{} a tailored argument or construction to the dark energy /cosmological constant problem. This is a consequence of a whole general picture, (Section XI), within the Standard Model of the Universe. [4. A Whole picture.]{} Overall, a consistent unifying clear picture of the history of the Universe does emerge in terms of the gravitational classical, semiclassical and quantum phases and their relevant characterizing physical magnitudes as the size, age, vacuum density, gravitational entropy and temperature, all in terms of the cosmological constant. This sheds light in the Inflation and Dark energy eras and in the cosmological constant problem. This is summarized in the end of Section XI (the whole history), and depicted in Fig.(1). The evolution of the Universe can be described by two big phases: Classical and Quantum, that is to say, after and before the Planck time $t_P = 10^{-44}\; sec$ respectively. Each cosmological stage in the classical known Universe $t_P \leq t \leq 10^{61}\;t_P$ has a dual quantum stage in the preceding quantum phase before the Planck time: $10^{-61}\; t_P \leq t \leq t_P$. The whole duration (of the classical plus quantum phases) is precisely $10^{-61}\; t_P \leq t \leq 10^{+61}\; t_P$. That is to say, [each]{} component [naturally]{} dominates in each phase: classical time component $ 10^{+61}\; t_P $ in the classical era, quantum Planck value $t_P$ in the quantum preceding era. The present time of the Universe at $10^{+61}\; t_P$, which is [a lower bound]{} for the future (if any) age of the Universe, has a remote past quantum precursor equal to $10^{-61}\; t_P$, which is an [upper bound]{} for the origin of the Universe. The classical/semi-classical known inflation era which occurred at about $10^{+6}\; t_P, H = 10^{-6}\; h_P$ has a preceding quantum dual era at $10^{-6}\; t_P, H = 10^{6}\; h_P$ which is a semi-quantum era (’low $H$’ with respect to the extreme past quantum state $H = 10^{61}\; h_P$), and similarly, for any of the other known eras in the Classical post-planckian Universe. This appears to be the way in which the Universe has evolved. A complete picture is discussed in Section XI including the gravitational entropy and temperatures, and summarized in Fig.(1). In Planck units, is the same to express the age of the Universe $10^{61}$ in terms of time, length, mass, temperature or square root of entropy (arrow of time) to describe the complete Universe. Similarly, the vacuum energy density is the dual to the gravitational entropy. The complete quantum theory is a theory of [pure numbers]{}. The [total or complete]{} physical quantities are invariant under the classical-quantum duality: $H \leftrightarrow Q$, as it must be: This means physically that: (i) what occurred in the quantum phase before $t_P$ [determines]{} through quantum duality Eqs.(\[ULambda1\])-(\[Utotal1\]) what occurred in the classical phase after $t_P$. And: (ii) what occurred in the quantum phase before $t_P$ is the [same physical observable]{}, or event which occurred after $t_P$ in the precise meaning of the classical-quantum dual relations Eqs.(\[ULambda1\])-(\[Udual1\]). That is to say: The quantum dual quantities in the quantum phase before $t_P$, are the [quantum precursors]{} of the classical/semiclassical quantities after $t_P$. As the wave-particle duality at the basis of quantum mechanics, the wave-particle-gravity duality, is reflected in all cosmological eras and its associated quantites, temperatures and entropies as well. Cosmological evolution goes from a quantum planckian and superplanckian phase to a semiclassical accelerated era (de Sitter inflation), then to the classical eras untill the present classical de Sitter phase. The classical-quantum or wave-particle-gravity duality specifically manifests in this evolution, between the different gravity regimes, and could be view as a mapping between asymptotic (in and out) states characterized by sets $U_Q$ and $U_\Lambda$ and thus as a Scattering-matrix description. [This paper is organized as follows:]{} In Section II we describe the classical - quantum duality including gravity together with its properties covering the different gravity regimes: (classical, semiclassical and quantum gravity domains) passing through the Planck scale and the elementary particle domain as well. In Sections III and IV we describe the classical, quantum dual and complete de Sitter universe, its physical duality symmetry and its properties. Sections V and VI deal with the classical, quantum dual and complete de Sitter Temperature and Entropy. In Section VII we characterize in a precise and unifying way the different (classical, semiclassical, Planckian and super-Planckian) de Sitter regimes. Sections VIII-X illustrate the results with relevant cosmological examples and their implications for the CMB fluctuations, Inflation and Dark Energy. In Section XI we provide a clarifying unifying picture with the cosmological constant as the vacuum energy, entropy and temperature of the Universe. Section XII summarizes outlook and conclusions. Classical - Quantum Duality through the Planck scale ==================================================== Let us stand by $O_{G}$ the set of relevant physical variables or observables characteristic of the classical gravity regime, (as size, mass, surface gravity (or gravity acceleration), and usual temperature for instance), and by $O_Q$ the corresponding set of quantities in the quantum dual regime in the precise sense of the wave-particle or classical-quantum duality: The magnitudes $O_{G}$ and $O_Q$ are classical-quantum gravity duals of each other, (in the precise meaning of the classical-quantum duality) here through the Planck scale, [@Sanchez2019]: $$\label{O} O_{G}=o_{P}^2\, O_Q^{-1}$$ This relation holds in general for any quantity in the set. $O_{G}$ and $O_Q$ are the same conceptual physical quantities in the different (classical/semiclassical and quantum) gravity regimes respectively. The constant $o_{P}$ stands for the corresponding quantity at the Planck scale, ie purely depending of the fundamental constants ($\hbar ,\, c,\, G)$. $O_Q$ stands for relevant quantum concepts as quantum size $ L_Q$, quantum mass $ M_Q$, quantum acceleration $ K_Q = c^2/L_Q$, quantum temperature $T_Q$ and other physical magnitudes associated to them . This is not an assumed or conjectured duality. This duality is [universal]{}. As the wave-particle duality, this classical/semiclassical-quantum gravity duality does not relate to the number of dimensions, nor to any particular or imposed symmetry of the background manifold or space- time, nor to any other condition. Each of the sides of Eq.(\[O\]) accounts for each domain separately: classical [or]{} quantum, ie $ O_G$ [or]{} $O_Q$, and their respective associated set of magnitudes. The [total or complete]{} or QG magnitudes take into account the different gravity domains: classical and quantum, and their duality properties, passing through the Planck scale and including the elementary particle domain as well [@Sanchez2019]: $$\label{OQG} O_{QG} = (O_Q + O_G)$$ In Planck units, the complete QG magnitudes simply read $$\label{Oop} O_{QG} = o_P \; ( o + \frac{1}{o} ), \qquad o \equiv \frac{O_G}{o_P} = \frac{o_P}{O_Q}$$ The two domains $(O\geq o_P)$ and $(O \leq o_P)$ being the classical and quantum domains respectively, with the two ways of reaching the Planck scale. The QG magnitudes cover all the classical and quantum domains, with and without gravity. The two domains precisely account for the elementary particle domain: $0 \leq O \leq o_P $, and for the macroscopic gravity domain: $o_P \leq O\leq \infty$. These two domains are duals of each other in the precise sense of the classical-quantum duality through the Planck scale: we call it “Planck scale duality”. For instance: Quantum particle theory has $L_Q >> l_P$ and $L_G << l_P$. Classical gravity has $L_Q << l_P$ and $L_G >> l_P$. Quantum Gravity has $L_{QG}$ and any value of $ L_G$ and $L_Q$, and includes the Planck domain as well. We implement the classical-quantum duality in de Sitter universe in the next section. Classical and Quantum Dual de Sitter Universes ============================================== de Sitter space-time in $D$ space-time dimensions is the hyperboloid embedded in Minkowski space-time of $(D+1)$ dimensions: $$\label{dS} X^2 - T^2 + X_j X^j + Z^2 = L_H^2, \qquad j= 2, 3, ...(D-2)$$ $L_H$ is the radius or characteristic length of the de Sitter universe. The scalar curvature $ R$ is constant. Classically: $$L_H = {c}/{H}, \qquad R = H^2 D(D-1) = \frac{2D}{(D-2]}\;\Lambda, \qquad \Lambda = \frac{H^2}{2} (D-1) (D-2)$$ Moreover, a mass $M_H$ can be associated to $L_H$ or $H$, such that (we take D = 4 here for simplicity): $$\label{LH} L_H = \frac{G M_H}{c^2} \equiv L_G, \;\;\qquad M_H = \frac{c^3}{GH}$$ The corresponding quantum magnitudes $L_Q$, $M_Q$ are the quantum duals of $L_H$, $M_H$ respectively in the precise meaning of the classical-quantum (de Broglie or Compton) duality: $$\label{MH} L_Q = \frac{\hbar}{M_H c} = \frac{\hbar\; G H}{c^3}\;, \;\;\qquad M_Q = \frac{\hbar H}{c^2}$$ $$\label{MQ} {\text ie,} \quad L_Q = \frac{l_P^2}{L_H}\;, \;\; \qquad M_Q = \frac{m_P^2}{M_H}$$ where $l_P$ and $m_P$ are the Planck length and Planck mass respectively: $$\label{lp} l_P = \sqrt{\frac{\hbar\;G}{c^3}}\;, \;\; \qquad m_P = \sqrt{\frac{c\;\hbar}{G}}$$ Similarly, for the quantum dual Hubble constant $H_Q$ and the quantum curvature $R_Q$: $$\label{dualQH} H_Q = \frac{h_P^2}{H} , \qquad R_Q = \frac{r_P^2}{R}, \qquad \Lambda_Q = \frac{\lambda_P^2}{\Lambda}$$ where $ h_P, r_P, \lambda_P $ are the Planck scale values of the Hubble constant, scalar curvature and cosmological constant respectively: $$\label{P} h_P = \frac{c}{l_P} ,\qquad r_P = h_P^2 \;D (D-1) , \qquad \lambda_P = \frac{h_P^2}{2} \;(D-1) (D-2)$$ $$\label{PD4} h_P = c^2 \sqrt{\frac{c}{\hbar G}},\qquad r_P = 12 \; h_P^ 2 = 4 \;\lambda_P ,\qquad \lambda_P = 3 \; \left(\frac{c^5}{\hbar G}\right), \qquad (D=4)$$ Total de Sitter Universe and its Duality Symmetry ================================================= The classical $L_H \equiv L_G$ length and the quantum length $ L_Q$ can be extended to a more complete length $L_{QH}$ which includes both: (we call it [complete]{} or Quantum Gravity (here QH) length since it contains both: Q and H lengths): $$\label{LQH} L_{QH} = (L_H + L_Q) = l_P \; (\frac{L_H}{l_P} + \frac{l_P}{L_H}).$$ and we have then : $$\label{ZQH} X^2 - T^2 + X_j X^j + Z^2 = L_{QH}^2 = 2\;l_P^2 \left[1 + \frac{1}{2}\; [\;(\frac{L_H}{l_P})^2 + (\frac{l_P}{L_H})^2\;]\right]$$ with $j = 2,3, ...(D-3)$. Eq.(\[ZQH\]) quantum generalize de Sitter space-time including the classical, semiclassical and quantum de Sitter regimes and the Planck scale de Sitter regimes as well. It contains two non-zero lengths $(L_H, L_Q)$ or two relevant scales ($H$, $l_P$) enlarging the possibilities for the space-time regimes or phases: Quantum, semiclassical and classical de Sitter regimes. Thus, - [For $ L_H >> l_P$, ie $L_Q << L_H$, Eq.(\[ZQH\]) yields the classical de Sitter space-time. For intermediate $L_H$ values between $l_P$ and $L_Q $ it yields the semiclassical de Sitter space-time. ]{} - [For $ L_H = l_P $ ie $L_Q = l_P = L_{QH}$, Eq.(\[ZQH\]) yields the Planck scale de Sitter hyperboloid.]{} - [For $ L_H << l_P $, ie $L_Q >> L_H $ it yields the highly quantum de Sitter regime, deep inside the Planck domain.]{} $H = c / L_H $ is ($ c ^{-1} )$ times the surface gravity (or gravity acceleration) of the classical de Sitter space-time. Similarly, $ H_Q = c / L_Q $ and $H_{QH} = c / L_{QH}$ are the surface gravity in the quantum and whole QH de Sitter phases respectively. Similarly, from Eq. (\[LQH\]) and Eqs (\[LH\])-(\[MQ\]), we have for the mass: $$\label{MQH} M_{QH} = (\;M_H + M_Q\;) = m_P \;(\;\frac{M_H}{m_P} + \frac{m_P}{M_H}\;)$$ $$\label{MLQH} \frac{M_{QH}} {m_P} = m_P \;(\;\frac{L_H}{l_P} + \frac{l_P}{L_H}\;) = \frac{L_{QH}}{l_P}$$ $M_{QH}/m_P$ and $L_{QH}/l_P$ both have the same expression with respect to their respective Planck values. [The complete QH Hubble constant $H_{QH}$, curvature $R_{QH}$ and $\Lambda_{QH}$.]{} The fully quantum QH Hubble $H_{QH}$ constant, curvature $R_{QH}$ and $\Lambda_{QH}$ constant follow from the QH de Sitter length $L_{QH}$ Eq.(\[LQH\]): $$\label{QH1} H_{QH} = \frac{c}{L_{QH}}, \qquad R_{QH} = H_{QH}^2\; D\;(D-1), \qquad \Lambda_{QH} = \frac{H_{QH}^2}{2} \;(D-1) (D-2)$$ where from Eqs.(\[LQH\]) and (\[dualQH\]): $$\label{QH2} H_{QH} = \frac{H}{ [\;1 + (l_P H / c)^2\;]}, \qquad H_{QH}/ h_P = \frac{(H/h_P)}{ [\;1 + (H / h_P)^2\;]}, \qquad h_P = c / l_P$$ We see the [symmetry]{} of $H_{QH}$ under $(H/h_P) \rightarrow (h_P/H)$, ie under $ H \rightarrow H_Q = (h_P^2/H)$ : $$\label{symQH} H_{QH} (H / h_P) = H_{QH} (h_P / H)$$ That is, the classical $H$ and quantum $H_Q$ are classical-quantum duals of each other through the Planck scale $h_P$, but the complete or total $H_{QH}$ which contain both of them is [invariant]{}. And similarly, for the quantum curvature $R_{QH}$ and cosmological constant $\Lambda_{QH}$ Eq.(\[QH1\]) derived from them : $$\label{symRQH} R_{QH}(H/h_P) = R_{QH}(h_P/H), \qquad \Lambda_{QH}(H/h_P) = \Lambda_{QH}(h_P/H)$$ where: $$\label{R1} R_{QH} = \frac{R_H}{ [\;1 + R_H /r_P \;]^2} = \frac{R_Q}{[\;1 + R_Q /r_P \;]^2},\qquad r_P = 12\;h_P^2$$ $$\label{L1} \Lambda_{QH} = \frac{\Lambda_H}{[\;1 + \Lambda_H /\lambda_P \;]^2} = \frac{\Lambda_Q}{[\;1 + \Lambda_Q /\lambda_P \;]^2},\qquad \lambda_P = 3\;h_P^2$$ The classical $ H/h_P << 1$, quantum $ H/ h_P >> 1$ and Planck $H/h_P = 1$ regimes are clearly exhibited in the QH expressions Eqs (\[QH1\]), Eq.(\[QH2\]): $$\label{sHP} H_{QH \;(H << h_P)} = H \;[\; 1 - (H / h_P)^2 \;] + O\;(H / h_P)^4 = \frac{c}{L_H} \;[ 1 - (\frac{l_P}{L_H})^2 ] + O\;(\frac{l_P}{L_H})^4$$ $$\label{HP} H_{QH} \; (H = h_P) = \frac{h_P}{2}, \; \; h_P= c/l_P$$ $$\label{bHP} H_{QH \; (H >> h_P)} = (h_P^2/H)\; [ 1 - (h_P/H)^2 ] \;+\; O (h_P/H)^4 = \frac{ c\;L_H}{l_P^2}\; [\; 1 - \;(\frac{L_H}{l_P})^2 \;] \;+ \;O\;(\frac{L_H}{l_P})^4$$ The three above equations show respectively the three different de Sitter phases: - [The classical gravity de Sitter universe (with lower curvature than the Planck scale $r_P $) [outside]{} the Planck domain $(l_P < L_H < \infty)$.]{} - [The Planck curvature de Sitter state $(R_H = r_P, \;\; L_ H = l_P)$]{} - [The highly quantum or high curvature ($R_H >> r_P$) de Sitter phase [inside]{} the quantum gravity Planck domain ($0 < L_H \leq l_P$).]{} Eqs (\[sHP\])-(\[bHP\]) show the classical-quantum duality through the Planck scale: The highly quantum gravity regime $H_{QH}\;(H >> h_P)$ entirely expresses in terms of the quantum Hubble constant $H_Q$, dual through the Planck scale value $h_P$ to the classical/semiclassical Hubble constant $H$, Eq.(\[dualQH\]). Is natural to define here the dimensionless magnitudes: $$\label{dimless} {\cal L} \equiv \frac{L_{QH}}{l_P}, \qquad {\cal M}\equiv \frac{M_{QH}}{m_P}, \qquad {\cal H} \equiv \frac{H_{QG}}{h_P}, \qquad l\equiv \frac{L_H}{l_P}, \qquad {\quad h} \equiv \frac{H}{h_P} = l^{-1}$$ Then, Eqs (\[LQH\]),(\[MQH\]) and (\[QH2\]) simply reads: $$\label{dimless2} {\cal L} = (l + \frac{1}{l}) = {\cal M},\qquad {\cal H} = \frac{1}{(l + \frac{1}{l})} = {\cal L}^{-1}$$ Similarly, for $R_{QH}/r_P$ and $\Lambda_{QH}/\lambda_P$: $$\label{dimlessR} \frac{R_{QH}}{R_P} = \frac{\Lambda_{QH}}{\Lambda_P} = \left(\frac{H_{QH}}{h_P}\right)^2 \equiv {\cal H}^2 = \frac{1}{(h + h^{-1})^2}$$ In dimensionless variables, the duality symmetry endowed by $L_{QH},M_{QH}$ and Eqs (\[symQH\]), (\[symRQH\]), simply reads: $$\label{sym} {\cal L} (l^{-1}) ={\cal L} (l), \qquad {\cal M} (l^{-1}) = {\cal M} (l)$$ $$\label{sym2} {\cal H} (l^{-1}) ={\cal H} (l), \qquad {\cal R} (l^{-1}) = {\cal R} (l), \qquad {\mathbf \Lambda} (l^{-1}) = {\mathbf \Lambda} (l)$$ The QH magnitudes are complete variables covering both classical and quantum, Planckian and super Planckian domains. They are more complete magnitudes than the $ Q$ or $H$ magnitudes alone which cover only one phase or domain: classical gravity or quantum/semiclassical domain. [The complete (QH) de Sitter density.]{} Let us complete the set of physical de Sitter magnitudes with the classical, quantum and QH de Sitter densities respectively, ($\rho_P$ being the Planck density scale): ($\rho_H$, $\rho_Q$, $\rho_{QH}$): $$\label{rhoH} \rho_H = \rho_P \left(\frac{H}{h_P}\right)^2 = \rho_P \left(\frac{\Lambda}{\lambda_P}\right), \qquad \rho_P = \frac{3 \; h_P^2}{8 \pi G}, \quad \lambda_P = \frac{3\; h_P^2}{c^4}$$ $$\label{rhoQ} \rho_{Q} = \rho_P \left(\frac{H_Q}{h_P}\right)^2 = \rho_P \frac{\Lambda_Q}{\lambda_P} = \frac{\rho_P^2}{\rho_H} = \rho_P \left(\frac{h_P}{H}\right)^2 = \rho_P \left(\frac{\lambda_P}{\Lambda}\right)$$ $$\label{rhoQH} \rho_{HQ} = \rho_{H} + \rho_{Q} = \rho_P \left(\frac{H_{HQ}}{h_P}\right)^2 = \rho_P \frac{\Lambda_{HQ}}{\lambda_P}$$ From Eqs. (\[QH2\]), (\[rhoQ\]) it follows that:$$\rho_{HQ}= \frac{\rho_H}{[\; 1 + \rho_H / \rho_P\; ]^2} = \frac{\rho_Q}{[\; 1 + \rho_Q / \rho_P \; ]^2},$$ which satisfies $$\rho_{HQ}\; (\rho_H) = \rho_{HQ}\; (\rho_Q) = \rho_{HQ} \;(\rho_P^2 / \rho_H),$$ For small and high densities with respect to the Planck density $\rho_P$, the QH de Sitter density $\rho_{QH}$ behaves as: $$\label{RhoQH1} \rho_{QH} \;(\rho_H << \rho_P) = \rho_H \;[\; 1 - 2 (\rho_H / \rho_P) \;] + O\;(\rho_H / \rho_P)^2$$ $$\label{RhoQHP} \rho_{QH} \; (\rho_H = \rho_Q = \rho_P) = \frac{1}{4}\rho_P : \;\;\mbox {(Planck regime)}$$ $$\label{RhoQH2} \rho_{QH} \;(\rho_H >> \rho_P) = \rho_Q \;[\; 1 - 2 (\rho_Q / \rho_P) \;] + O\;(\rho_Q / \rho_P)^2,$$ corresponding to the classical/semiclassical de Sitter regime (and its quantum corrections) Eq.(\[RhoQH1\]), to the Planck scale de Sitter state Eq.(\[RhoQHP\]), and to the highly quantum, super Planckian, de Sitter phase Eq.(\[RhoQH2\]). In the very classical regime, $\rho_{QH}$ is proportional to the classical density $\rho_H$, as it must be. In the highly quantum regime, $\rho_{QH}$ is proportional to $\rho_Q$, as it must be too. Classical and Quantum Dual de Sitter Temperatures and Entropies =============================================================== We complete now the set of relevant intrinsic de Sitter magnitudes, by including the Temperature and the Hubble horizon area. The temperature $T_H $ of the classical de Sitter Universe and the temperature $T_Q$ of the quantum de Sitter Universe are consistently defined as ($\kappa _B$ is the Boltzmann constant) : $$\label{TH} T_H = \frac{M_H c^2}{2 \pi \kappa _B} , \qquad T_Q = \frac{M_Q c^2}{2 \pi \kappa _B},$$ which from Eqs. (\[LH\]),(\[MH\]),(\[MQ\]) yield: $$\label{TH2} T_H = t_P \left(\frac{M_H}{m_P} \right) = t_P \left(\frac{L_H}{l_P} \right) = t_P \left(\frac{h_P}{H} \right) = t_P \sqrt{\frac{\lambda_P}{\Lambda}}$$ $$\label{TQ2} T_Q = t_P \left(\frac{m_P}{M_H} \right) = t_P \left(\frac{l_P}{L_H} \right) = t_P \left(\frac{H}{h_P} \right) = t_P \sqrt{\frac{\Lambda}{\lambda_P}}$$ $t_P$ being the Planck temperature, $T_Q$ and $T_H$ satisfy: $$T_Q = \frac{t_P^2} {T_H}, \qquad t_P = \frac{m_P c^2}{2\pi \kappa _B}$$ We see that the Quantum de Sitter Temperature $T_Q$ is the Hawking-Gibbons de Sitter temperature [@GibbHawkTemp]. This is the Quantum dual of the Classical de Sitter temperature $T_H$. The classical and quantum de Sitter areas $ A_H, A_Q$ are defined as: $$\label{AH} A_H = 4 \pi L_H^2, \qquad A_Q = 4 \pi L_Q^2,$$ which, from Eqs. (\[LH\]),(\[MH\]), (\[MQ\]), (\[TH2\]),(\[TQ2\]) yield: $$\label{AHL} A_H = a_P\left(\frac{L_H}{l_P}\right)^2 = a_P\left(\frac{M_H}{m_P}\right)^2 = a_P\left(\frac{T_H}{t_P}\right)^2 = a_P\left(\frac{h_P}{H}\right)^2$$ $$\label{AQL} A_Q = a_P\left(\frac{l_P}{L_H}\right)^2 = a_P\left(\frac{m_P}{M_H}\right)^2 = a_P\left(\frac{t_P}{T_H}\right)^2 = a_P\left(\frac{H}{h_P}\right)^2$$ $a_P$ being the Planck area. $A_Q$ and $A_H$ satisfy $$A_Q = \frac{a_P^2}{A_H}, \qquad a_P = 4 \pi\; l_P^2$$ The classical and quantum areas are dual to each other through the Planck scale area $a_P$, and have the expressions: $$\label{AMTQ} \frac{A_H}{a_P} = \frac{a_P}{A_Q} = \frac{T_H}{T_Q} = \frac{1}{2\pi \kappa_B}\frac{M_H\;c^2}{T_Q}$$ $$\label{AMTH} \frac{A_Q}{a_P} = \frac{a_P}{A_H} = \frac{T_Q}{T_H} = \frac{1}{2\pi \kappa_B}\frac{M_Q\;c^2}{T_H},$$ which are entirely [symmetric]{} under the change $H\leftrightarrow Q$. Interestingly enough, the areas can be expressed as (one half) the ratio of the energy over the temperature, Eqs.(\[AMTQ\]),(\[AMTH\]), which is a typical [entropy]{} expression. The corresponding gravitational entropies $S_H$, $S_Q$ are: $$\label{SH} S_H = \frac{\kappa_B}{4} \;\frac{A_H}{l_P^2}, \qquad S_Q = \frac{\kappa_B}{4}\;\frac{A_Q}{l_P^2}$$ Eq.(\[AMTQ\]) is the Gibbons-Hawking or classical/semiclassical gravity de Sitter entropy. Eq.(\[AMTH\]) is its quantum dual gravity de Sitter entropy. From Eqs.(\[AMTQ\]),(\[AMTH\]): $$\label{SQ} S_Q = \frac{s_P^2}{S_H} , \qquad s_P = \frac{\kappa_B}{4}\;\frac{a_P}{l_P^2} = \pi \kappa_B$$ $s_P$ being the Planck entropy. $S_H$ and $ S_Q$ read: $$\label{SHL} S_H = s_P \left(\frac{L_H}{l_P}\right)^2 = s_P \left(\frac{h_P}{H}\right)^2 = s_P \left(\frac{\lambda_P}{\Lambda}\right)$$ $$\label{SQL} S_Q = s_P \left(\frac{l_P}{L_H}\right)^2 = s_P \left(\frac{H}{h_P}\right)^2 = s_P \left(\frac{\Lambda}{\lambda_P}\right)$$ The classical and quantum entropies $S_H, S_Q$ satisfy the classical-quantum duality through the Planck scale entropy $s_P$, and have the expression: $$\label{SHMT} \frac{S_H}{s_P} =\frac{1}{4 \pi \kappa_B}\;\frac{M_H}{T_Q} c^2, \;\; \qquad \frac{S_Q}{s_P} =\frac{1}{4 \pi \kappa_B}\;\frac{M_Q}{T_H} c^2,$$ which is a typical entropy expression in terms of Mass and Temperature. Total de Sitter Temperature and Entropy ======================================= From the total (QH) de Sitter magnitudes above discussed as the whole QH radius $L_{QH}$ and associated mass $ M_{QH}$, we define the corresponding total de Sitter temperature $T_{QH}$, area $A_{QH}$ of the $QH$ Hubble radius and entropy $S_{QH}$. The $QH$ temperature is defined as: $$\label{TQH1} T_{QH} = \frac{M_{QH}c^2}{2\pi \kappa_B} ,$$ which from Eqs (\[MQH\]),(\[TH2\]),(\[TQ2\]), read $$\label{TQH2} T_{QH} = \frac{m_P c^2}{2\pi \kappa_B} \; (\frac{M_H}{m_P} + \frac{m_P}{M_H}) = t_P\;(\frac{T_H}{t_P} + \frac{t_P}{T_H}),$$ $$\label{TQH3} T_{QH} = (T_H + T_Q)\;\; \mbox{invariant under $H \leftrightarrow Q$}$$ Consistently, the QH de Sitter Temperature is the sum of the classical gravity de Sitter temperature $T_H$ plus the quantum (or semiclassical) Gibbons-Hawking de Sitter temperature $T_Q$. In terms of $H_{QH}$, the QH temperature $T_{QH}$ simply reads: $$T_{QH} = t_P \left(\frac{h_P}{H_{QH}}\right), \;\;\mbox{or}\;\;\frac{T_{QH}}{t_P} \equiv {\cal T} = {\cal H}^{-1}$$ Explicitely: $$T_{QH} = t_P (\frac{h_P}{H})\; [\;1 + (\frac{H}{h_P})^2\;] = t_P (\frac{H}{h_P})\; [\;1 + (\frac{h_P}{H})^2\;],$$ which is entirely invariant under the interchange $H \leftrightarrow h_P$. The $QH$ Area of the Hubble horizon follows from the QH Hubble radius $L_{QH}^2$ Eq.(\[LQH\]): $$\label{AQH} A_{QH} = 4 \pi L_{QH}^2 = 2 \; (4 \pi l_P^2)\;\left[1 + \frac{1}{2}\; [\;(\frac{L_H}{l_P})^2 + (\frac{l_P}{L_H})^2\;]\right],$$ which from Eqs.(\[AH\]),(\[AHL\]),(\[AQL\]) expresses as: $$\label{AQH2} A_{QH} = 2\; a_P\left[ 1 + \frac{1}{2}\; [\;\frac{A_H}{a_P} + \frac{a_P}{A_H}\;] \right], \qquad a_P = 4 \pi l_P^2$$ The $QH$ area $A_{QH}$ is thus the sum of the Planck area $a_P$, the classical area $A_H$ and the quantum area $ A_Q$: $$\label{AQH3} A_{QH} = 2 a_P + A_H \;+ \; A_Q$$ $$A_{QH} (A_H = a_P = A_Q) = 4 \; a_P$$ In units of the Planck area $a_P$, each of the areas can be in turn expressed as a energy over the temperature ratio Eqs.(\[AMTH\]),(\[AMTQ\]), which yields: $$\label{AQHMT} \frac{A_{QH}}{a_P} = 2 \; \left[\; 1 + \frac{1}{4 \pi \kappa_B} [\; \frac{M_H c^2}{T_Q} + \frac{M_Q c^2}{T_H}\; ]\;\right]$$ The corresponding QH gravitational entropy $S_{QH}$ is given by $$\label{SQH5} S_{QH} = \frac{\kappa_B}{4} \frac{A_{QH}}{l_P^2} = 2 \kappa_B\; \left[\; \frac{a_P}{4 l_P^2} + \frac{1}{2}\;(\; \frac{A_H}{4 l_P^2} + \frac{A_Q}{4 l_P^2} \;)\; \right]$$ Thus, $$\label{SQH6} S_{QH} = 2\;[\; s_P + \frac{1}{2}\;( S_H + S_Q )\;], \qquad s_P = \frac{\kappa_B}{4} \frac{a_P}{l_P^2} = \pi \kappa_B$$ $$S_{QH}\;(S_H = s_P = S_Q) = 4 s_P = 4 \pi \kappa_B$$ The whole entropy $S_{QH}$ turns out to be the sum of the Planck entropy $s_P$, the classical entropy $S_H$ and the quantum entropy $S_Q$ Eqs.(\[SH\]),(\[SQ\]),(\[SHMT\]), and have the expression: $$\frac{S_{QH}}{s_P} = 2 \;\left[\; 1 + \frac{1}{4 \pi \kappa_B} \; (\;\frac{1}{8} \frac{M_H c^2}{T_Q} + \frac{1}{8} \frac{M_Q c^2 }{T_H}\;)\;\right]$$ - [We see how the concept of classical-quantum duality and the QH variables naturally accompass, unify and simplify the relationship between the classical and quantum gravity magnitudes and regimes, in particular this is well appropriated to discuss the gravitational temperature and entropy in the different, classical and quantum, gravity regimes.]{} - [ The concept of gravitational entropy is the same for any of the gravity regimes: classical, quantum, Planck scale and quantum gravity or super-Planckian regimes: $S_H, S_Q, s_P$ or $S_{QH}$, namely: $Area /4 l_P^2$ in units of $\kappa_B$.]{} - [ For a classical size, ie a large macroscopic gravitational object, or our universe of radius $L_H$, this is the classical/semiclassical area $A_H$ and so the classical/semiclassical gravitational entropy $S_H$, which is the known Gibbons-Hawking de Sitter entropy [@GibbHawkEntropy].]{} - [ For a quantum size, ie a quantum microscopic object or quantum universe, ie of size equal to the Compton length $L_Q$, this is the quantum dual area $A_Q$ and so the quantum dual entropy $S_Q$. For a Planck length object or universe this is the Planck entropy $s_P$. The whole or complete $QH$ entropy $S_{QH}$ turns to be the sum of the three components, as it must be.]{} In dimensionless variables: $${\cal T} \equiv \frac{T_{QH}} {t_P}, \qquad {\cal A } \equiv \frac{A_{QH}} {a_P}, \qquad {\cal S} \equiv \frac{S_{QH}} {s_P},$$ $$t \equiv \frac{T_H}{t_P}, \qquad a \equiv \frac{A_H}{a_P}, \qquad s \equiv \frac{S_H}{s_P},$$ Eqs.(\[TQH1\]),(\[AQH2\]),(\[SQH5\]), simply read: $${\cal T} = (t + \frac{1}{t}), \qquad {\cal A } = 2 \;[\;1 + \frac{1}{2} \;(a + \frac{1}{a})\;], \qquad \qquad {\cal S} = 2 \; [\; 1 + \frac{1}{2} \;(s + \frac{1}{s})\;]$$ $${\cal T}(t = 1) = 2,\qquad {\cal A} (a = 1) = 4, \qquad {\cal S} (s = 1) = 4$$ And their duality symmetry simply stands: $${\cal T}(t) = {\cal T}(t^{-1}), \qquad {\cal A} (a) = {\cal A} (a^{-1}), \qquad {\cal S} (s) = {\cal S} (s^{-1}),$$ which show the simplification in terms of the Planck units, natural to the problem. Classical, Semiclassical, Planckian and Super-Planckian de Sitter Regimes ========================================================================= The complete QH radius $L_{QH} = L_{QH} (L_H, L_Q) = L_{QH} (L_H, l_P)$ and their corresponding QH Hubble constant $H_{QH}$, $QH$ mass $M_{QH}$, and their constant Planck scale values $(l_P, h_P, m_P)$ only depending on $(c, \hbar, G)$, allow to characterize in a precise way the classical, semiclassical, Planckian and quantum (super-Planckian) de Sitter regimes: - [$L_{QH} = L_{QH}(L_H, L_Q) \equiv L_{QH} (H, \hbar) $ yields the [whole]{} (classical/semiclassical, Planck scale and quantum (super-Planckian) de Sitter universe.]{} - [$L_{QH} = L_H = L_Q$ yields the Planckian de Sitter state, (Planck length de Sitter radius, Planckian vacuum density and Planckian scalar curvature): $$L_H = l_P, \quad H = h_P, \quad \lambda_P = 3\; h_P^2, \quad R = r_P = 4 \;\lambda_P, \quad l_P = \sqrt{(\hbar G / c^3)}$$ ]{} - [$L_{QH} = L_H >> L_Q $, ie $ L_H >> l_P $ , $ H << h_P $, yields the classical de Sitter space-time.]{} - [$L_{QH} = L_Q >> L_H $, ie $ L_H << l_P $, $H >> h_P$, (high curvature $R >> r_P = 4 \Lambda_P,$)\ yields a full quantum gravity super Planckian (inside the Planck domain $0 < L_H \leq l_P$) de Sitter phase.]{} - [$L_{QH} >> L_Q$ ie $L_{QH} \rightarrow \infty$ for $ L_H \rightarrow \infty $, ie $H \rightarrow 0 $ ie $\Lambda \rightarrow 0$, (zero curvature) yields consistently the classical Minkowski space-time, equivalent to the limit $L_Q \rightarrow 0 $ ie $ l_P \rightarrow 0$ ($\hbar \rightarrow 0$).]{} The three de Sitter regimes are characterized in a complete and precise way: - [(i) [Classical and Semiclassical de Sitter Regimes]{}: (Inflation and more generally the whole known -classical and semiclassical- Universe is within this regime): $ l_p < L_H < \infty$, ie $0 < L_Q < l_P$, $\;$ $0 < H < h_P$,$\;$ $ m_P < M_H < \infty$.]{} - [ (ii) [Planck Scale de Sitter state with Planck curvature and Planck radius:]{} $L_H = l_P , \;\; L_Q = l_P,\;\; H = h_P = c/l_P,\;\; M_H = m_P$.]{} - [(iii) [Quantum Planckian and super-Planckian Regimes:]{} $ 0 < L_H \leq l_P $, ie $ \infty < L_Q \leq l_P $, $ h_P \leq H < \infty $, $ 0 < M_H < m_P$.]{} The two above classical and quantum de Sitter regimes (i) and (iii) are duals of each other in the precise meaning of the classical-quantum or wave-particle duality through the Planck scale de Sitter state (ii). This is the Planck scale duality or classical-quantum gravity duality Eqs.(\[O\]),(\[OQG\]),(\[Oop\]) at work. Numbers and Cosmological Implications ===================================== Let us see now some illustrative cosmological values for the relevant classical, quantum and Planck scale magnitudes above refered. Typical cosmological values are: - [[For the Universe Today]{}: $$H = 100 \; h \;\frac{Km}{sec\; Mpc}, \quad h = 0.7,\quad \rho_{crit} = 2.77 \; 10^{11} h^2 \; \frac{M_{sun}}{(Mpc)^3},\quad \rho_{\Lambda} = 0.7 \; \rho_{crit}$$]{} [Classical, Planck and Quantum Dual values of the Hubble Radius, Mass and Age of the Universe Today, are typically:]{} $$L_H = 1.2 \; 10^{28} cm = 1.2\; 10^{61}\;l_P,\quad l_P = 10^{-33} cm ,\quad L_Q = 0.8 \; 10^{-61}\;l_P$$ $$M_H = 1.5 \; 10^{48} gr = 1.5 \; 10^{53} m_P,\quad m_P = 10^{-5} gr,\quad M_Q = 0.67 \; 10^{-53} m_P$$ $$T_H = 0.4 \; 10^{18} sec = 4 \; 10^{61}\; t_P,\quad t_P = 10^{-44} sec,\quad T_Q = 0.2 \;10^{-61} t_P$$ [Classical, Planck and Quantum Dual values of the Hubble Constant, Cosmological Constant and Density of the Universe Today are typically:]{} $$H = 2.5 \; 10^{-17} sec^{-1} = 2.5\; 10^{-61} h_P,\quad h_P = 10^{44} sec^{-1},\quad H_Q = 10^{61} h_P$$ $$\Lambda = 3\; 10^{-34} sec^{-2} = 10^{-122} \lambda_P,\quad \lambda_P = 3 \; 10^{88} sec^{-2},\quad \Lambda_Q = 10^{122} \lambda_P$$ $$\rho_H = 10^{-29} \frac{gr}{cm^3} = 10^{-122} \rho_P,\quad \rho_P = 10^{93}\frac{gr}{cm^3},\quad \rho_Q = 10^{122} \rho_P$$ The values above correspond to the classical Universe today (subscript H), the Planck values (subscript P) and the Quantum dual values (subscript Q). - [[For the CMB era, Classical and Quantum values of the Age, Hubble constant, Size and Density of the Universe are typically:]{}]{} $$L_H = 10^{24}\;cm = 10^{57} l_P,\quad T_H = 10^{13} sec = 10^{57} t_P,\quad H = 10^{-57} h_P,\quad \rho_H = 10^{-114}\rho_P$$ $$L_Q = 10^{-57}l_P,\quad T_Q = 10^{-57}\; t_P,\quad H_Q = 10^{57} h_P,\quad \rho_Q = 10^{114}\rho_P$$ - [[For the Inflation era, Classical/semiclassical and Quantum dual values of the Hubble Constant, Horizon size and Inflaton Mass are typically:]{}]{} $$L_H = 10^{-27} \;cm = 10^{6} l_P,\quad T_H = 10^{6} t_P,\quad H = 10^{-6} h_P,\quad M_H = 10^{6} m_P$$ $$L_Q = 10^{-6}l_P,\quad T_Q = 10^{-6}\; t_P,\quad H_Q = 10^{6} h_P,\quad \quad M_Q = 10^{-6} m_P,$$ - [[For the Solar system]{}: $M_{sun} = 10^{33} gr = 10^{38} m_P, \quad M_{Q\; sun} = 10^{-38}\; m_P$ $$M_{moon} = 7\; 10^{25} gr = 7 \; 10^{30}\; m_P,\; \qquad M_{Q\; moon} = 0.14\; 10^{-30} \;m_P$$ $$M_{asteroid,\;comet} = 10^{15} gr = 10^{20}\; m_P,\; \qquad M_{Q\; asteroid,\;comet} = 10^{-20}\;m_P$$]{} - [[For Human scales]{}: $M_{human} = 10^{5} gr = 10^{10} \;m_P, \quad M_{Q\; human} = 10^{-15} gr = 10^{-10}\; m_P$ $$L_{human} = 1.7\; 10^{2} cm = 1.7 \; 10^{35}\; l_P,\; \qquad L_{Q\; human} = 10^{-68} cm = 10^{-35}\;l_P$$]{} - [[For atomic scales]{}: $L_{atom} = 10^{20} \;l_P, \; \qquad T_{atom} = 10^{20} \;t_P, \; \qquad M_{atom} = 10^{-20} \;m_P$ $$L_{Q\;atom} = 10^{-20}\;l_P, \; \qquad T_{Q\;atom} = 10^{-20}\; t_P, \; \qquad M_{Q\;atom} = 10^{20}\; m_P$$]{} - [[For elementary particles (ex.the electron mass)]{}: $M(eV/c^2) = 10^{-33} gr = 10^{-28}\; m_P,\quad M_Q(eV/c^2) = 10^{23} gr = 10^{28}\;m_P$]{} - [We see that the elementary particle masses do appear as the [quantum duals through the Planck scale]{} of the typical solar system objects. For instance, the quantum dual of a typical comet or asteroid mass say is a typical atomic mass. The quantum dual of the electron mass $M_Q(eV/c^2)$ is a typical moon mass of $10^{22}$ kgr.]{} - [That is to say, there is a [physical classical-quantum duality]{} through the Planck mass or correspondence between the [macroscopic or astronomical gravitational masses/sizes]{} and the [elementary particle and quantum masses/sizes]{}, the Planck scale being the [crossing or inversion scale]{} of the two mass/size domains: ${M_Q = {m_P^2}/{M}}$. These two domains are [precisely]{} connected through the classical-quantum (ie wave-particle, Compton, de Broglie) duality including gravity Eqs. (\[O\]),(\[MH\]),(\[MQ\]): namely, a Planck scale duality or classical-quantum gravity duality Eqs.(\[O\])-(\[Oop\]).]{} - [ Notice that in the [classical CMB era]{} (at about $3.8 \; 10^5 yr = 10^{13} sec$), the gravitational size $L_H$, age and temperature $T_H$ of the Universe are, as reported above, equal to $10^{57}$ (in Planck units), while their quantum duals in the quantum precursor era are $10^{-57}$. The classical gravitational entropy $S_H$ and the $H$-associated density $\rho_H$ in the classical CMB era are respectively: $$S_H = 10^{114} s_P, \;\qquad \rho_H = 10^{-114},\; \qquad s_P = \pi \kappa_B$$ Their quantum dual values in the quantum CMB precursor era $10^{-57} t_P$ being respectively $S_Q = 10^{-144} s_P$ and $\rho_Q = 10^{144} \rho_H$. The quantum entropy $S_{Q}$ in the quantum CMB precursor era is extremely low as it must be, because it will increase along the Universe evolves and classicalizes, reaching its classical gravitational value $S_{H} = 10^{114}$ in the classical known CMB era (arrow of time).]{} - [Let us recall the classical entropy $S_{cmb}$ of the CMB black body radiation contained in the classical Hubble volume: $S_{cmb} = (4/3) \pi s_{\gamma} H^{-3}$, the total number of the CMB photons being $1.5 \; 10^{89}$, the average entropy per photon $3.6 \;\kappa_B$, hence:$$S_{cmb} = 1.72 \;10^{89} s_P, \; \qquad S_{Q\; cmb} = \frac{s_P^2}{S_{cmb}} = 0.58 \; 10^{-89} s_P,$$ The gravitational entropy $S_H$ is the dominant component in the classical CMB era, and represents un [upper bound]{} for the CMB photon radiation entropy $S_{cmb}$ in it. The quantum gravitation entropy $S_Q$ is the smaller component and is a lower bound for the $S_{Q\; cmb}$ value in the quantum precursor era. Similarly, for the respective Temperatures: The classical and quantum dual gravitational temperatures $T_H$ and $T_Q$ at the CMB age are $$T_H = 10^{57} t_P, \qquad T_Q = 10^{-57} t_P, \qquad t_P = 10^{32}\; K$$ The classical and quantum dual Temperatures of the CMB radiation are: $$T_{cmb} = 2.73 K = 2.73 \; 10^{-32} \; t_{P}, \qquad T_{Q\;cmb} = 0.37 \; 10^{32} \;t_{P}$$ The gravitational temperature $T_H$ in the classical CMB era, represents un [upper bound]{} for the CMB photon radiation Temperature $T_{cmb}$ in it. The quantum gravitation temperature $T_Q$ is the smaller component and is a lower bound for the $T_{Q\; cmb}$ value in the precursor era.]{} Relevant implications for Inflation and Dark Energy, are discussed in detail in Sections IX and X below. [Quantum Inflationary Fluctuations, the Gibbons-Hawking Temperature and CMB anisotropies]{}: Interestingly enough, the power spectra of quantum primordial fluctuations of Inflation can be expressed in terms of the Quantum and Planck temperatures $T_Q$, $t_P$. As is known, the spectra of inflationary scalar curvature and tensor perturbations are given by: $$\Delta_{k, H}^S = \frac{1}{\sqrt{\pi \epsilon}} \; \frac{H}{m_P}, \qquad \Delta_{k, H}^T = \frac{4}{\sqrt{\pi}}\;\frac{H}{m_P},$$ $\epsilon$ being the slow-roll parameter. From Eq.(\[TQ2\]) they can be expressed in terms of $ T_Q/t_P$ as: $$\Delta_{k, H}^S = \frac{1}{\sqrt{\pi \epsilon}} \; \frac{T_Q}{t_P}, \qquad \Delta_{k,H}^T = \frac{4}{\sqrt{\pi}}\;\frac{T_Q}{t_P}$$ Thus: $$T_Q = t_P \; \sqrt{\pi \epsilon}\; \Delta_{k, H}^S = t_P \; \frac{\sqrt{\pi}}{4} \; \Delta_{k, H}^T$$ Or, in terms of the ratio $r$: $$T_Q= t_P \; \frac{\sqrt{\pi r}}{4} \Delta_{k, H}^S, \qquad r = \frac{[\Delta_{k, H}^T]^2}{[\Delta_{k, H}^S]^2}$$ Therefore, for the amplitude value $\Delta_{k, H}^S$ from the CMB data [@WMAP1],[@Planck6], we get for $ T_Q$: $$T_Q = \sqrt{\pi \epsilon} \; 10^{28} \; K = \sqrt{\pi \epsilon} \; 10^{-4} \; t_P \; \qquad \mbox{and} \;\qquad T_Q = \frac{\sqrt{\pi\;r}}{4} \; 10^{-4} \; t_P$$ From the last recent bound $r < 0.07$ [@Planck6]: $$T_Q < 1.169 \; 10^{-5}\; t_P = 1.169 \; 10^{27}\; K$$ We see that $T_Q$ for classical/semiclassical Inflation is constrained to be less than $10^{-5} t_P$, [consistent]{} with the semiclassical gravity character of Inflation. Interestingly, the Quantum Temperature $T_Q$, which is here precisely the [Hawking-Gibbons de Sitter temperature]{}, can be measured or constrained through the real CMB data which constrain Inflation. This is important because: [(a)]{} The conceptual quantum/semiclassical gravity nature of the Hawking-Gibbons de Sitter temperature, and [(b)]{} Contrary to the Inflation case, the equivalent Hawking temperature for astrophysical black holes cannot be experimentally measured since it is extremely low: lower than the CMB temperature. $T_Q$ could be higher for small or primordial black holes but these have not been detected. Implications for Inflation ========================== We discuss here in more detail the consequences of the Q and QH observables for Inflation. Recall that in Classical Inflation, at first order in the slow-roll expansion, the scalar curvature and tensor perturbation spectra are given by: $$\label{DeltaSO} [\Delta^S_{k,\;H}]^2 = \frac{1}{\pi \epsilon}\;\left(\frac{H}{m_P}\right)^2, \qquad [\Delta^T_{k, \;H}]^2 = \frac{16}{\pi }\;\left(\frac{H}{m_P}\right)^2$$ ($\Delta^S_{k, \;H}, \Delta^T_{k, \;H}$ stand here for the scalar and tensor fluctuations respectively). The slow roll parameters are given by: $$\label{epsilon} \epsilon = \frac{m_P^2}{4\pi} \left(\frac{H'}{H}\right)^2, \qquad \eta = \frac{m_P^2}{4\pi} \left(\frac{H''}{H}\right), \quad \xi = \frac{m_P^2}{2\pi} \left(\frac{H'\;H'''}{H^2}\right)$$ H’ and H” stand for the first and second derivatives with respect to the inflaton field. $\epsilon \approx \eta << 1 $ are first order in slow roll, $\xi$ second order slow roll parameter, with the hierarchy $\xi = O (\epsilon^2)$, and following so on in the slow roll expansion. The slow roll parameters are related to the observables ratio $r$ and spectral scalar index $n_s$ by: $$\label{parameters} \epsilon = \frac{r}{16},\qquad \eta = \frac{1}{2}(\;n_s - 1 + \frac{3}{8}\;r\;), \qquad \xi = \frac{r}{4}(\;n_s - 1 + \frac{3}{16}\;r - \frac{1}{2}\frac{dn_s}{dlnk}\;)$$ The difference $(\epsilon - \eta)$ is a measure of the [departure from scale invariance]{} at first order in slow roll: $$\label{Delta} \Delta \equiv (\epsilon - \eta) = \frac{1}{2}(\;n_s - 1\;) + \frac{r}{8}$$ From Eqs. (\[DeltaSO\]),(\[QH2\]) the Quantum (QH) generalization of the power spectrum of scalar curvature and tensor perturbations is given by: $$\label{DeltaSQH} [\Delta^S_{k, \;QH}]^2 = \frac{1}{\pi \epsilon_{QH}}\;\left(\frac{H_{QH}}{m_P}\right)^2 = \frac{1}{\pi \epsilon_{QH}}\;\left(\frac{H}{m_P\;[\; 1 + (H/h_P)^2 \;] }\right)^2$$ $$\label{DeltaTQH} [\Delta^T_{k, \;QH}]^2 = \frac{16}{\pi }\;\left(\frac{H_{QH}}{m_P}\right)^2 = \frac{16}{\pi}\;\left(\frac{H}{m_P\;[\; 1 + (H/h_P)^2 \;]}\right)^2$$ Here $\epsilon_{QH} << 1$ is the first order QH slow roll parameter, we compute it below, and $ h_P$ is the Planck Hubble constant value. Thus, the QH inflationary spectra get expressed as: $$[\;\Delta^S_{k, \;QH}\;]^2 = [\;\Delta^S_{k, \;H}\;]^2\; \left(\frac{1}{[ \;1 + (H/h_P)^2 \;]^2 }\right)\; \frac{1}{( 1 - \delta \epsilon_{QH})}$$ $$[\;\Delta^T_{k, \;QH}\;]^2 = [\;\Delta^T_{k, \;H}\;]^2 \; \left(\frac{1}{[\; 1 + (H/h_P)^2 \;]^2 }\right)$$ where $[\Delta^S_{k,\;H}]^2$ and $[\Delta^T_{k, \;H}]^2$ are the standard spectra of scalar curvature and tensor perturbations in Classical H Inflation Eqs.(\[DeltaSO\]). Each QH power spectrum $[\Delta^{S,T}_{k, \;QH}]$ is expressed in terms of each Classical H Inflation spectrum $[\Delta^{S,T}_{k, \;H}]$ and it is modified by a factor $[\; 1 + (H /h_P)^2\;]^{-2}$ arising from $H_{HQ}$. In addition, $[\Delta^S_{k,\;QH}]^2$ gets also modified by the QH factor $(1 - \delta \epsilon_{QH})^{-1}$ arising from the QH slow parameter $\epsilon_{QH}$: $$\epsilon_{QH} = \frac{m_P^2}{4\pi} \left(\frac{H'_{QH}}{H_{QH}}\right)^2 \equiv \epsilon \;(\; 1 - \delta \epsilon_{QH}\;),$$ where $\epsilon$ is the standard slow-roll parameter Eq.(\[epsilon\]) and $\delta \epsilon_{QH}$ is its QH modification given by: $$\label{deltaepV} \delta \epsilon_{QH} = 2 \;\frac{H H_{QH}}{h_P^2}\;[\; 1 - \frac{H H_{HQ}}{2 h_P^2}\;] \;= \; \frac{4 (H/h_P)^2}{[\; 1 + (H/h_P)^2\;]} \left[\; 1 - \frac{(H/h_P)^2}{[\; 1 + (H/h_P)^2\;]}\;\right]$$ The ratio $r_{QH}$ turns out precisely modified by this $\delta \epsilon_{QH}$ factor: $$r_{QH} = \frac{[\;\Delta^T_{k,\;QH}\;]^2}{[\;\Delta^S_{k,\;QH}\;]^2} = \frac{[\;\Delta^T_{k\;H} ]^2}{[\;\Delta^S_{k\; H}]^2}\; (\;1 - \delta \epsilon_{QH}\;)$$ Thus, $$r_{QH} = r \;(1 - \delta \epsilon_{QH}),\qquad r_{QH} = 16 \;\epsilon_{QH}, \qquad r = 16 \;\epsilon$$ with $\delta \epsilon_{QH}$ given by Eq.(\[deltaepV\]). The QH slow roll parameter $\eta_{QH}$ is given by: $$\eta_{QH} = \frac{m_P^2}{4\pi} \left(\frac{H_{QH}''}{H_{QH}}\right),$$ which from Eq.(\[QH2\]) can be recasted as: $$\eta_{QH} = \left[\;\eta -\frac{1}{2} m_P \sqrt{\frac{\epsilon}{\pi}}\;\frac{H H_{QH}}{h_P^2}\; \right] \left[\; 1- \frac{H H_{QH}}{h_P^2}\;\right]$$ $$\eta_{QH} = \left[\;\eta -\frac{1}{2} m_P \sqrt{\frac{\epsilon}{\pi}}\; \frac{(H/h_P)^2}{[\; 1 - (H / h_P)^2\;]}\;\right] \left[\; 1- \frac{(H/h_P)^2}{ [\; 1 - (H / h_P)^2\;]}\;\right]$$ where $\epsilon$ is the standard slow roll parameter of classical H inflation Eq.(\[epsilon\]). The QH modifications express themselves in terms of $(H/h_P)^2 = (l_P H/c)^2$ and powers of it. Is also of interest to compute the QH departure from scale invariance which is given by $$\label{DeltaQH} \Delta_{QH} \equiv (\eta_{QH} - \epsilon_{QH})= \frac{1}{2}(\;n_{s\; QH} - 1 \;) + \frac{r_{QH}}{8}$$ [Typically, for (classical) Inflation]{}: $H = 10^{-6} h_P$, ie $H << h_P$ and we can safely expand the above QH Inflation expressions in powers of $(H/h_P)^2$, namely: $$\epsilon_{QH} (H << h_P) = \epsilon\; [\; 1 - 4\;(l_P H)^2 + O (l_P H)^4 \;]$$ $$\eta_{QH}\; (H << h_P) = \eta \; [\; 1 - 2\;(l_P H)^2 ] - m_P \sqrt{\frac{\epsilon}{\pi}}\;(l_P H)^2 + O (l_P H)^4 \;]$$ Thus, at first order in $(H/h_P)^2$: $$[\;\Delta^S_{k, \;QH}\;]^2 = [\;\Delta^S_{k, \;H}\;]^2\; [\; 1 + 2\;(l_P H)^2 + O (l_P H)^4 \;]$$ $$[\;\Delta^T_{k, \;QH}\;]^2 = [\;\Delta^T_{k, \;H}\;]^2 \; [\; 1 - 2\;(l_P H)^2 + O (l_P H)^4 \;]$$ $$\label{rQH2} r_{QH} = r \;[\; 1 - 2\;(l_P H)^2 + O (l_P H)^4 \;]$$ And the [QH departure from scale invariance]{} Eq.(\[DeltaQH\]) gets corrected as: $$\label{DeltaQH2} \Delta _{QH} = \Delta\;[\; 1 - 2 \; (l_P H)^2 \;] + \sqrt{\epsilon} \;(l_P H)^2 \; [\;2 \sqrt{\epsilon} - \frac{m_P}{\sqrt{\pi}}\;] + O (l_P H)^4$$ where $\epsilon$ is the standard slow roll parameter of classical H inflation Eqs.(\[epsilon\])-(\[parameters\]), and $\Delta$ is the [departure from scale invariance]{} of the classical H Inflation Eq.(\[Delta\]) meaning that $(n_s - 1)$ [and]{} $r$ [are not zero]{}. In terms of the observables $(n_s, r)$ and $(n_{s\;QH}, r_{QH})$, Eq.(\[DeltaQH2\]) yields for $n_{s\;QH}$: $$\label{nsQH} n_{s\;QH} = n_s \;[\; 1 + 2 \;(l_P H)^2 \;] - 2 \;(l_P H)^2\; [\; 1 + m_P \sqrt{\frac {\epsilon}{\pi}} \;] + O(l_P H)^4$$ where Eqs.(\[rQH2\]),(\[DeltaQH\]) and (\[Delta\]) have been used. Typically, for classical Inflation: $L_H = 10^{6}\;l_P$, the total QH corrections are thus: $$\frac{r_{QH}}{r} - 1 = - 2 \;10^{-12}$$ $$\frac{n_{s\;QH}}{n_s} - 1 = 2 \;10^{-12}\;[\;1 - \frac{1}{n_s} (\;1- \frac{m_P}{2}\sqrt{\frac{\epsilon}{\pi}}\;)\;]$$ - [We see that the complete QH quantities allow to get [quantum corrections to Inflation and its observables in a direct, simple and consistent way]{}]{}. - [Notice the [sign]{} of the corrections: The quantum gravity QH corrections [enhance]{} the scalar curvature spectrum and [reduce]{} the tensor perturbations. The QH corrections are of the same order of magnitude and [ sign]{} as the [quantum inflaton]{} corrections computed in the Effective Theory of Inflation within the Ginsburg-Landau approach [@Boya2005], [@Boya2006], [@BDdVS]. This also shows the [robustness and reliability]{} of the slow roll approximation and the Effective Theory of inflation. \[If the reduced Planck mass $M_P$ is used, $m_P = \sqrt{8\pi} \;M_P$: $(H/M_P)_{Inflation} = 10^{-5} = 10 \;(H/m_P)$ here\].]{} - [Notice that the QH factor modifying the Hubble constant and the inflationary spectra can be written as the summation of the series: $$\label{sum} QH \equiv \frac{H}{[\;1 + (H / h_P)^2\;]} = H \; \sum _{n=0}^{\infty} (-1)^n \left(\frac{H}{h_P}\right)^2$$ The QH factor covers the [full classical and quantum range]{}, namely: If $H < h_P$, Eq.(\[sum\]) yields the usual corrections in $(l_P H)^2$. If $H >> h_P$, Eq.(\[sum\]) precisely [changes to the quantum regime]{}, ie to the quantum Hubble rate $H_Q$, which is the [super-Planckian domain]{}: $$\label{sumQ} HQ \equiv \frac{H_Q}{[\;1 + (H_Q / h_P)^2\;]}$$]{} - [In the case of classical Inflation: $H$ is about $10^{-6}\;h_P$, as we have seen. In the case of the quantum precursor Inflation era at about $10^{-6}\;t_P$: $H_Q$ is about $10^{6}\;h_P$, Whatever it be: in the classical phase (after $t_P$) or in the quantum precursor phase (before $t_P$), Inflation occurs not too far from the Planck scale: $10^{\pm 6} t_P$ or $10^{\mp 6} h_P$.]{} Implications for Dark Energy ============================ Dark energy and its more direct candidate, the cosmological constant, [@Riess],[@Perlmutter],[@Schmidt],[@WMAP1],[@WMAP2],[@DES],[@Planck6] is relevant to both modern cosmology and particle physics. Let us recall the value of the observed dark energy density today $ \rho _H \equiv \rho_{\Lambda}$: $$\label{rhovalue} \rho_\Lambda = \Omega_\Lambda \rho_c = 3. 28 \; 10^{-11} (eV)^4 = (2.39 \; meV)^4, \qquad m eV= 10^{-3} eV$$ corresponding to $ h = 0.73, \qquad \Omega_\Lambda = 0.76 , \qquad H = 1.558 \; 10^{-33} eV.$ The last Planck satellite data yield the values [@Planck6]: $$\label{Hvalue} H = 67.4 \pm 0.5 \; Km \; sec^{-1}\; Mpc^{-1}, \quad \Omega_\Lambda h^2 = 0.0224 \pm 10^{-4}$$ and $$\label{Omegavalue} \Omega_\Lambda = 0.6847 \pm 0.0073, \quad \Omega_\Lambda h^2 = 0.3107 \pm 0.0082,$$ which implies for the cosmological constant [today]{}: $$\label{Lambdavalue} \Lambda = (4.24 \pm 0.11) \; 10^{-66}\; (eV)^2 = (2.846 \pm 0.076) \; 10^{-122}\; m_P^2$$ The density $\rho_\Lambda$ associated to $\Lambda$ Eq.(\[rhovalue\]) is precisely: $$\label{rhoLambda} \rho_\Lambda = \frac{\Lambda}{8 \pi G} = \rho_P \left(\frac{\Lambda}{\lambda_P}\right),$$ where the Planck scale values $\rho_P,\lambda_P$ are: $$\rho_P = \frac{\lambda_P}{8 \pi G} \quad \; \lambda_P = \ 3 h_P^2$$ The enormous discrepancy between the large theoretical value expected from microscopic particle physics for the vacuum energy density $\approx 10^{122}$ and the small cosmological value observed today $\rho_{\Lambda}\approx 10^{-122}$ is largely known as the cosmological constant problem. [However, several clarifications are in order here]{}: [(i) The classical gravity vacuum.]{} The $\Lambda$ value Eq. (\[rhovalue\]), (\[Hvalue\]), (\[Lambdavalue\]) observed at the present era today corresponds to the [classical]{} (non quantum) value of the vacuum energy density of the [classical large Universe today]{}: large radius $L_\Lambda$ or large Age, large mass $M_\Lambda$ and large classical/semiclassical entropy $S_H$, and thus small rate $H$, low temperature $T_\Lambda$, small $\Lambda$, [small and dilute classical vacuum density $\rho_{\Lambda}$]{} described here in the above sections. [(ii) The quantum gravity vacuum.]{} The $\Lambda$ density $\rho_{\Lambda}$ observed today is the [classical]{} vacuum density of the Universe today which is a gravitationally [classical, large]{} and [dilute]{} Universe. The value of $\rho_{\Lambda}$ and $\Lambda$ Eqs.(\[rhovalue\]), (\[rhoLambda\]) is precisely the [classical dual value]{} of the [quantum]{} cosmological constant value $\Lambda_Q$ Eq.(\[dualQH\]), and therefore the classical dual value of the [quantum vacuum energy $\rho_Q$]{} Eq.(\[rhoQ\]): This is precisely and clearly expressed in the following Eqs: $$\label{LambdaHvalue} \Lambda = 3 H^2 = \lambda_P \left (\frac{H}{h_P}\right)^2 = \lambda_P \left (\frac{l_P}{L_H}\right)^2 = (2.846 \pm 0.076) \; 10^{-122}\; m_P^2$$ $$\label{LambdaQvalue} \Lambda_Q = 3 H_Q^2 = \lambda_P \left (\frac{h_P}{H}\right)^2 = \lambda_P \left (\frac{L_H}{l_P}\right)^2 = (0.3516 \pm 0.094) \; 10^{122}\;m_P^2$$ $$\label{LambdaQLambda} \Lambda_Q = \frac{\lambda_P^2}{\Lambda}, \qquad \lambda_P = 3 h_P^2 = 3 m_P^2$$ [The quantum dual value $\Lambda_Q$ is [precisely]{} the quantum vacuum value obtained from particle physics.]{} $$\label{rhoQii} \rho_Q = \rho_P \left(\frac{\Lambda_Q}{\lambda_P}\right) = \frac{\rho_P^2}{\rho_\Lambda} = 10^{122}\; \rho_P$$ [(iii) The classical and quantum dual values]{}. That is to say, the two huge different values: $10^{-122}$ and $10^{122}$ (in Planck units) refer to [two huge physically different]{} vacuum energies of the Universe corresponding to two huge different eras, to two huge different physical cosmological conditions (present time and very early eras), to two different vacuum states or regimes of the Universe, and consistently, they [must be different]{}. Such enormous difference must be in such way and is [not]{} a problem or inconsistency: Moreover and consistently, one value is the [quantum physics dual]{} of the other -or the quantum precursor of the other- as expressed by Eqs.(\[LambdaHvalue\]),(\[LambdaQvalue\]),(\[LambdaQLambda\]),(\[rhoQii\]). [(iv)]{}. This is [not fortuitous]{}, that is to say, this is not pure chance or unexplained coincidence. [(v)]{}. This is [not trivial]{}, that is to say, this is simple, deep and robust. There is no problem between the two extremely different values $\Lambda$ and $\Lambda_Q$ or equivalently between $\rho_{\Lambda}$ and $\rho_Q$, because the two values [do not]{} refer to the same vacuum or eras: one is exactly the [ classical]{} physics today vacuum energy density $\rho_{\Lambda}$, the other is [its quantum dual]{} value in the planckian and superplanckian very early phase $10^{-61}\;t_P \leq t \leq t_P$: This early phase of the Universe is exactly the [quantum precursor]{} of the today classical era in the precise meaning of the wave-particle (or classical-quantum) duality including gravity, Eqs.(\[LambdaHvalue\]) to (\[rhoQii\]). The two different values are explained by the fact that they are exactly, mathematically and physically, the classical-quantum dual of each other: [The $\Lambda_Q$ value Eq.(\[LambdaQvalue\])-(\[LambdaQLambda\]), that is to say, the vacuum value computed from particle physics is exactly the quantum dual value of the classical $\Lambda$ value observed today Eq.(\[LambdaHvalue\]).]{} [(vi) Crossing the Planck scale]{}. The two values: $\Lambda$ and $\Lambda_Q$, (or equivalently $\rho_ \Lambda$ and $\rho_Q$) refer to the same concept or nature of $\Lambda$ or $\rho_\Lambda$ as a vacuum energy density or cosmological constant but they are in two huge different vacuum states or two huge different cosmological epochs: Classical state and classical epoch today for $\Lambda$ observed today, and quantum state and quantum super-Planckian very early universe epoch for the quantum mechanical super-Planckian value $\Lambda_Q$. The classical value today $\Lambda = 3H^2$ corresponds to the classical Universe today of classical rate $H$ and classical cosmological radius $L_H = c/ H$. The quantum mechanical value $\Lambda_Q = 3 H_Q^2$ corresponds to the early quantum Universe of quantum rate $H_Q$ and quantum radius $L_Q = l_P^2/L_H = \hbar/M_H c$ which is [exactly]{} the quantum dual of the classical horizon radius $L_H$: $L_Q$ is [precisely]{} the quantum (Compton) length of the Universe for the gravitational mass $M_H = L_H c^2/G$. [(vii) Two extremely different physical conditions and gravity regimes]{}. This is a realistic, clear and precise illustration of the [physical classical-quantum duality between the two extreme Universe scales and gravity regimes]{}: the dilute state and Horizon size of the Universe today on the one largest known side, and the super-Planckian scale and highest density state on the smallest side: Length, Mass, and their associated time (Hubble rate) and vacuum energy density ($\Lambda, \rho_\Lambda$) of the Universe [today]{} are truly [classical]{}, while its extreme past at $10^{-61}\; t_P$ = $ 10^{-105}$ sec deep inside the Planck domain of extremely small size and high vacuum density value ($\Lambda_Q, \rho_Q$) are truly [quantum and super-Planckian]{}. This is the [classical-quantum or wave-particle duality]{} between the classical macroscopic (cosmological) gravity physical domain and the quantum microscopic particle physics and super-Planckian domain through the [crossing]{} of the Planck scale, [Planck scale duality]{} in short. [(viii) The true problem]{}. The huge difference between the two values $\Lambda = \rho_\Lambda = 10^{-122}$ and $\Lambda_Q = \rho_Q = 10^{122}$ is indeed [correct]{} and must be such way, precisely because the two values refer to huge different physical conditions, regimes and states which are classical-quantum duals of each other. The two values refer to two different gravitational vaccua: classical, on one side (present time era), and full quantum super-Planckian energy on the opposite physical side (past remote era), and these are [two extreme different and dual energy density components]{}, $\rho_{\Lambda}$ and $\rho_Q$, contributing to the [same total]{} vacuum energy of the Universe $\rho_{\Lambda Q}$. Namely, there is indeed a cosmological constant problem but the true problem is [not]{} the huge discrepancy between the observed value today and the computed particle physics value. The true problem is to know the origin and the nature (the type) of the predominant particle(s) associated to the vacuum energy density and how to identify and detect them. [(ix) A General framework]{}. This is not a tailored argument in order to explain solely one problem (dark energy) or one cosmological constant value. This is just one of the consequences or applications of a general clarifying simplifying framework which completes at the level of the classical and quantum observables, the classical/semiclassical gravity observables on the one hand, and the microscopic quantum particle physics, planckian and super-planckian magnitudes in the early quantum eras on the other hand, and connects them through the classical-quantum (wave-particle) duality, (and one of such observables is just the vacuum energy density). [In Summary:]{} There is a [deep concept]{} behind the cosmological vacuum energy density or cosmological constant: the classical-quantum (or wave- particle) duality through the Planck scale, or Planck scale duality. This extends to the Planckian and super-Planckian domain the classical-quantum duality of quantum theory and includes gravity in it: classical-quantum gravity duality or wave-particle-gravity duality. Interestingly enough, including the Temperature and gravitationnal Entropy of the Universe in the description [consistently]{} supports the cosmic classical-quantum duality and shed more insight in the cosmological constant/vacuum energy nature of the dark energy. We discuss it in the next section. The Cosmological Constant: Vacuum Energy, Entropy and Temperature of the Universe ================================================================================= As we have seen, a key concept in order to understand the present value of the cosmological constant value and the so-called cosmological constant problem is the [classical-quantum duality]{}, precisely when applied to gravitational masses or objects, namely the classical-quantum duality through the Planck scale, or shortly Planck scale duality. The second important concept is the gravitational entropy $S_\Lambda$, namely the area of gravitational objects in units of $\kappa_B$ which is the Hawking-Gibbons de Sitter entropy, and its quantum dual entropy $S_Q$, as we will see below. The classical/semiclassical gravitational entropy $S_\Lambda$ of the Universe today is given by Eq.(5.13). The Gibbons-Hawking de Sitter entropy is exactly $S_{\Lambda}$, while its quantum dual entropy $S_Q$ is given by Eq.(5.14). The classical $\Lambda$-temperature $T_{\Lambda}$ of the Universe today is given by Eq.(5.2). The quantum temperature $T_Q$ Eq.(\[TQ2\]) is precisely the quantum dual temperature of $T_{\Lambda}$. The Gibbons-Hawking de Sitter Temperature is exactly $T_Q$. Moreover, let us recall that the cosmic (de Sitter) gravitational entropy and temperature were first derived in the context of the euclidean (imaginary time) quantum gravity [@GibbHawkEntropy]: That is to say, the Wick rotated path integral or partition function of gravitation and matter fields which in the saddle point approximation yields the classical action as the gravitational (Bekenstein-Hawking-Gibbons) entropy [@GibbHawkTemp],[@GibbHawkEntropy]. Notice too that the semiclassical regime yields as saddle point of the euclidean path integral of gravity the inverse value of $\Lambda$ [@HawkPhysLett]: $$\label{saddlepoint} 3 m_P^2 /\Lambda:\qquad\; \mbox{saddle point of quantum gravity path integral}$$ This expression is precisely our quantum dual cosmological constant $\Lambda_Q$: $$\label{LambdaQ} \Lambda_Q = \lambda_P/\Lambda = 3h_P^2/\Lambda$$ The reason why the saddle point of the Euclidean path integral of gravity is the inverse of $\Lambda$ is simply because the cosmological constant acts in the gravitational action as a Lagrange multiplier as it is only coupled to the space-time volume of the Universe, implying the term $\Lambda L^4_\Lambda$. The quantum gravity context and the semiclassical regime in which Eq.(\[saddlepoint\]) does appear show that the classical/semiclassical gravitational entropy $S_{\Lambda}$ and the classical and quantum temperatures $T_{\Lambda}$, $T_Q$ are completely in agreement with the physical context of the classical-quantum duality including gravity in which we describe it. - [The Universe at its present age $H$, is in a [classical gravitational state or regime]{} of classical radius $L_H = c/H$ and classical cosmological constant $\Lambda = 3 H^2$. In Planck units, the gravitational entropy $S_{\Lambda}$ of the Universe today, and thus the [classical gravitational entropy]{}, is [precisely]{} the inverse of the today cosmological constant value, ie $$\label{SSLambdavalue} S_{\Lambda} /s_P = (L_H/l_P)^2 = (h_P/H)^2 = (\lambda_P / \Lambda) = 10^{122}$$ This is precisely the inverse of the today classical $\Lambda$ density $\rho_{\Lambda}$ in Planck units $\rho_P$: $ \rho_P /\rho_{\Lambda} = 10^{122}$.]{} - [The $\Lambda$ density]{} $ \rho_{\Lambda} /\rho_P = 10^{-122} = \Lambda / \lambda_P$ observed today is [precisely the quantum entropy]{} $S_Q / s_P = \rho_{\Lambda} /\rho_P$, namely the area of the Universe of quantum radius $L_Q = l_p^2/ L_H$, ie the quantum dual radius of $L_H$, which is the Compton radius $L_Q = \hbar /(c M_H)$ of the Universe of mass $M_H = L_H c ^2 /G$. That is to say: $$\label{SSQvalue} S_Q / s_P = s_P/ S_H = (\Lambda / \lambda_P) = 10^{-122}.$$ The quantum gravitational entropy $S_Q$ is [precisely]{} the quantum dual of the classical gravitational entropy $S_\Lambda$ through its Planck scale value $s_P$: $$\label{SrhoQLambda} S_{\Lambda} = s_P \;\left(\frac{\rho_Q}{\rho_P}\right) = s_P \;\left(\frac{\lambda_P}{\Lambda}\right) = s_P \; 10^{+122}$$ $$\label{SQrhoLambda} S_Q = s_P \; \left(\frac{\rho_\Lambda}{\rho_P}\right) = s_P \; \left(\frac{\Lambda}{\lambda_P}\right) = s_P \; 10^{-122}$$ The [total]{} $Q\Lambda$ gravitational entropy turns out the sum of the three components as it must be: classical (subscript $\Lambda$), quantum dual (subcript $Q$) and Planck value (subscript P) corresponding to the tree gravity regimes: $$\label{Stotalvalue} S_{Q\Lambda} = 2 \;[s_P + \frac{1}{2}(S_{\Lambda} + S_Q)] = 2 \;s_P \;[\; 1 + \frac{1}{2}(10^{+122} + 10^{-122})\;]$$ - [The gravitational entropy $S_{\Lambda}$ of the present time large [classical Universe]{} is a very [huge number]{}, consistent with the fact that the Universe today contains a very huge amount of information. In order for $S_{\Lambda}$ to be associated with a [vacuum]{} energy density this must be a [very high density]{}: This is precisely the [quantum vacuum density $\rho_Q$ or quantum cosmological constant $\Lambda_Q$, which are the quantum duals -quantum precursors- of the classical density $\rho_{\Lambda}$ and classical cosmological constant $\Lambda$]{} respectively.]{} - [The value of $\Lambda$ today, that is the [classical]{} cosmological constant value, as a classical vacuum energy density $\rho_\Lambda$ is [naturally]{} a very small value because the accelerated Universe [today]{} is a [classical]{} and [dilute vacuum]{} Universe (in contrast to the quantum and highly dense super-Planckian very early state of the Universe). This is consistent with the well established set of observational results (refs [@VoidsHistory],[@Voids1] and refs therein) showing that the Universe today is an [empty]{} Universe dominated by [voids and supervoids]{}: Observations, numerical simulations and analytic results agree in the distribution of voids and supervoids which are the large scale vacuum sites of dominance of dark energy, (see for ex refs [@VoidsHistory],[@Voids1],[@VoidsPRL] and refs therein).]{} - [On the contrary, the quantum particle physics vacuum energy is the [quantum]{} dual density $ \rho_Q$ which is a huge value $10^{122}\;m_P$ deep inside in the quantum super-Planckian precursor era within a extremely small quantum radius $ L_Q$. The density $\rho_Q$ is the quantum dual of $\rho_\Lambda$ through its Planck scale value $\rho_P$: $$\label{rhoQ2} \rho_Q = \frac{\rho^2_P}{\rho_\Lambda} = \rho_P \left(\frac{L_\Lambda}{l_P}\right)^2 = \rho_P \left(\frac{\lambda_P}{\Lambda}\right)$$ The two densities, $\rho_{\Lambda}$, and $\rho_{Q}$, are [the same concept]{}: the vacuum energy density, [in two different states]{} (early superplanckian quantum phase and present time classical stage) of the Universe. The two densities are components of the [complete]{} $\rho_{Q \Lambda}$ density.]{} - [The [complete]{} $Q\Lambda$ density (classical plus quantum density) is : $$\label{rho3} \rho_{Q \Lambda} = \rho_\Lambda + \rho_Q + \rho_P = \rho_P \left(\frac{\rho_\Lambda}{\rho_P} + \frac{\rho_P}{\rho_ \Lambda} + 1 \right) = \lambda_P \left(\frac{\Lambda}{\lambda_P} + \frac{\lambda_P}{\Lambda} + 1 \right)$$ In the case of the Universe till [today]{}, these values are: $$\label{rhototalvalue} \rho_{Q \Lambda} = \rho_P\;(\;10^{-122}\; + \;10^{122}\; + 1 \;)$$]{} [Summing up:]{} The present Universe today of large classical horizon radius $L_\Lambda$ and very low density $\rho_\Lambda$ is a empty or dilute vacuum Universe (dominated by voids and supervoids) and [not]{} a dense quantum Universe. The very early Universe is a highly quantum dense Universe. The classical dilute Universe today and the highly dense very early quantum super-Planckian Universe are classical-quantum duals of each other in the precise meaning of the classical-quantum duality including gravity: The classical Universe today $U_{\Lambda}$ is clearly characterized by the set of physical magnitudes or observables (size/age, mass, density, temperature, entropy): $U_{\Lambda} \equiv (L_\Lambda, M_\Lambda, \rho_\Lambda, T_\Lambda, S_\Lambda)$. The highly dense very early quantum Universe $U_Q$ is characterized by the corresponding set of quantum dual physical magnitudes $U_Q \equiv (L_Q, M_Q, \rho_Q, T_Q, S_Q)$ in the precise meaning of the classical-quantum duality: $$\label{Udual} U_Q = \frac{u_P^2}{U_\Lambda}, \qquad u_P \equiv (l_P, m_P, \rho_P, t_P, s_P)$$ The [total]{} Universe is composed by their classical/semiclassical and quantum phases: $$\label{Utotal} U_{Q\Lambda} = \left(\; U_Q + U_{\Lambda} + u_P \;\right)$$ The Universe at its [present age]{} is a [classical]{} Universe of huge classical radius $L_\Lambda$ and thus a huge [classical horizon area]{} $ A_\Lambda$ and so a huge value for the classical/semiclassical gravitational entropy $S_\Lambda = 10^{+122}\; \pi \kappa_B$. The entropy $S_\Lambda$ is related to the [classical]{} area $ A_\Lambda \approx L_{\Lambda}^2 \approx 1/\Lambda$ and thus to the inverse of the classical $\Lambda$. This explains [why]{} the cosmological constant has such a small value and $S_\Lambda$ a so high one. $S_\Lambda$ today is [not]{} proportional to $\rho_\Lambda$ which is a extremely small value, but to the [quantum dual]{} of $\rho_\Lambda$, ie the quantum density $\rho_Q$ which is its precursor: a extremely high (superplanckian) value in the extreme past. This is clearly seen from Eqs.(\[SSLambdavalue\]),(\[SrhoQLambda\]),(\[SQrhoLambda\]) simply summarized as: $$\label{S3} \frac{S_\Lambda}{s_P} = \left(\frac{L_\Lambda}{l_P}\right)^2 = \frac{\rho_Q}{\rho_P} = \frac{\rho_P}{\rho_\Lambda} = \frac{\lambda_P}{\Lambda} = \frac{s_P}{S_Q} = 10^{122}$$ By going back in time along the Universe evolution, from the present era today to the early Universe stages, the gravitational entropy of the Universe is decreasing from its present huge value $S_\Lambda = 10^{122}\pi\kappa_B$ today at the age $10^{61} t_P$ to its inflationary value $S_\Lambda = 10^{12}\pi\kappa_B$ during inflation at the time $10^{-6} t_P$, then falling to its Planck value $s_P = \pi\kappa_B$ at the Planck time $t_P$ and then following decreasing till reaching its extreme lowest known value $ S_Q = 10^{-122}\pi\kappa_B$ in the earliest quantum era at $10^{-61} t_P$. The extreme smallest value of the entropy is the quantum dual of the largest known classical entropy at the horizon today: $$\label{Svalue} \frac{S_\Lambda}{s_P} = \frac{s_P}{S_Q} = 10^{+ 122}.$$ The largest time and length in the Universe are its present age and horizon size: $10^{+61}\; t_P$ and $L_\Lambda = 10^{+61}\; l_P$ . The smallest time and length in the Universe are the quantum duals of them: $10^{-61}\; t_P$ and $L_Q = 10^{-61}\; l_P$ [The classical and quantum $\Lambda$ Temperatures.]{} The above results can be also seen in terms of the classical and quantum temperatures of the Universe $T_\Lambda$ and $T_Q$. Eqs (\[TH2\]),(\[TQ2\]) and Eq.(\[Udual\]) for the classical today Universe $U_{\Lambda}$ and its quantum earlier dual $U_{Q}$ yield the following enlighting summary : $$\frac{T_\Lambda}{t_P} = \left(\frac{L_\Lambda}{l_P}\right) = \frac{h_P}{H} = \sqrt{\frac{\lambda_P}{\Lambda}} = \sqrt{\frac{S_\Lambda}{s_P}} = \sqrt{\frac{s_P}{S_Q}} = 10^{61}$$ $$\frac{T_Q}{t_P} = \left(\frac{l_P}{L_\Lambda}\right) = \frac{H}{h_P} = \sqrt{\frac{\Lambda}{\lambda_P}} = \sqrt{\frac{S_Q}{s_P}} = \sqrt{\frac{s_P}{S_\Lambda}} = 10^{-61}$$ From the above results and the observed value of $\Lambda$ today Eq.(\[Lambdavalue\]), the values of the classical temperature $T_{\Lambda}$ of the Universe [ today]{}, and the temperature $T_Q$ of its quantum precursor are: $$T_{\Lambda \; today} = t_P \;\sqrt{\frac{\lambda_P}{\Lambda}} = (0.5875 \pm 0.0800) \; 10^ {61} \; t_P$$ $$T_{Q \; today} = t_P \;\sqrt{\frac{\Lambda}{\lambda_P}} = (1.6865 \pm 0.0229)\; 10^{-61} \; t_P$$ That is: $$T_{\Lambda \; today} = (0.5875 \pm 0.0800) \; 10^{93} K$$ $$T_{Q \; today}= (1.6865 \pm 0.0229)\; 10^{-29} K$$ The [total]{} or [complete]{} $Q\Lambda$ Temperature $T_{Q\Lambda}$ Eq.(\[TQH2\]) is precisely [the sum]{} of the different components (classical plus quantum): $$\label{Ttotal} T_{Q\Lambda \; today} = [\; T_\Lambda + T_Q + t_P \;]_{today} = (10^{61} + 10^{-61} + 1)\; t_P$$ In the classical large Universe today, the classical [today]{} component $T_\Lambda$ dominates, as it must be. In its quantum precursor, the quantum Planck component $t_P$ dominates, as it must be. [Comparison to the Inflation $\Lambda$ Temperatures and Entropies]{}: For comparison, the temperatures $T_\Lambda$ and $T_Q$ for the Inflation era are: $$T_{\Lambda \;inflation} = t_P \;\sqrt{\frac{\lambda_P}{\Lambda}}_{inflation} = 10^{38} K = 10^{6} \; t_P$$ $$T_{Q \;inflation} = t_P \;\sqrt{\frac{\Lambda}{\lambda_P}}_{inflation} = 10^{26} K = 10^{-6}\; t_P$$ In the classical inflation era, the classical $T_\Lambda$ component dominates as it must be, while in its quantum precursor era, $t_P $ dominates as it must be. The complete $T_{Q \Lambda \;inflation}$ is the sum of its components, $$\label{Tinftotal} T_{Q\Lambda \;inflation} = [ \; T_\Lambda + T_Q + t_P \;]_{inflation} = (10^{6} + 10^{-6} + 1)\; t_P$$ Eqs.(\[Ttotal\]) and (\[Tinftotal\]) show consistently that the difference between the classical and quantum temperatures (which is huge in the today classical Universe highly dominated by the classical $T_ \Lambda$) diminishes in the early and more quantum stages as in Inflation where the difference between the two values $T_\Lambda = 10^{6}\; t_P$ and $T_Q = 10^{-6}\; t_P$ is considerably smaller than in the present time. In addition, Eq (\[Tinftotal\]) [consistently]{} reflects the [semi-classical or semi-quantum gravity]{} character of Inflation. In other words, as well as the Planck scale $ m_P$ is from the classical side the crossing to the quantum gravity regime, the [Inflation scale $10^{-6} m_P$ in the classical phase is the typical scale for the semi-classical gravity]{} regime. And the quantum dual Inflation scale in the quantum precursor phase is consistently $10^{6} m_P$. (This last could be viewed as a “semi-quantum gravity” scale, “low” with respect to the higher superplanckian scales of the earlier quantum stages, the highest $H = 10^{61} hp$ being at the extreme quantum past. Whatever be, classical or quantum, Inflation is at $10^{\pm 6}$ from the Planck scale). Consistently, this can be also seen in terms of the classical and quantum entropies $S_{\Lambda}$ and $S_Q$ of Inflation: $$S_{\Lambda \;Inflation} = s_P \;\left(\frac{\lambda_P}{\Lambda}\right) = 10^{+12} \;s_P = 10^{+12} \; \pi \; \kappa_B$$ $$S_{Q \;Inflation} = s_P \; \left(\frac{\Lambda}{\lambda_P}\right) = 10^{-12} \;s_P = 10^{-12} \; \pi \; \kappa_B$$ $S_{\Lambda \;Inflation}$ in the classical Inflation stage at $10^{6}t_P$ is larger than its precursor value $S_{Q \;Inflation}$ in the quantum Inflation precursor stage, (arrow of time), as it must be. [Cosmological Constant Summary:]{} Summing up, the solution to the cosmological constant can be explicitely summarized in the following simple equations: $$\label{Lambda1} \frac{\lambda_P}{\Lambda_Q} = \frac{\Lambda}{\lambda_P} = \frac{\rho_\Lambda}{\rho_P} = \frac{S_Q}{s_P} = \left(\frac{T_Q}{t_P}\right)^2 = 10^{-122}$$ and $$\label{LambdaQ2} \frac{\Lambda_Q}{\lambda_P} = \frac{\lambda_P}{\Lambda} = \frac{\rho_Q}{\rho_P} = \frac{S_\Lambda}{s_P} = \left(\frac{T_{\Lambda}}{t_P}\right)^2 = 10^{+122}$$ The [complete]{} $Q\Lambda$ cosmological constant $\Lambda_{Q\Lambda}$ or complete vacuum energy density $\rho_{ Q\Lambda}$ is given by: $$\label{LambdaQLambdavalue} \Lambda_{Q \Lambda} = \Lambda + \Lambda_Q + \lambda_P = \lambda_P \left (\; \frac{\Lambda}{\lambda_P} + \frac{\lambda_P}{\Lambda} + 1 \;\right) = \lambda_P \;(\; 10^{-122} + 10^{+122} + 1 \;)$$ which is the sum of its classical and quantum [components]{}. The observed value today is the classical $\Lambda $ vacuum value $10^{-122}$ corresponding to the classical Universe today which is a large, classical and empty or vacuum dilute Universe. This is the main physical reason for such low value. The computed particle physics quantum $\Lambda_Q$ value $10^{+122}$ is the vacuum value corresponding to the very early Universe which is a extremely small, quantum and high density (superplanckian) vacuum. This is the main physical reason for such high value. All physical magnitudes: the vacuum energy density, the cosmological constant, the gravitational entropy and gravitational temperature, both classical and quantum are linked by the classical-quantum (or wave-particle) duality through the Planck scale. Eqs. (\[Lambda1\]),(\[LambdaQ2\]),(\[LambdaQLambdavalue\]) concisely and synthetically express such classical-quantum dual relations and [explain why]{} the classical vacuum cosmological constant $\Lambda$ or classical density $\rho_\Lambda$ [coincides]{} with such observed [low value]{} $10^{-122}$ in Planck units. The vacuum computed density from particle physics $10^{+122}$ is a quantum extreme vacuum value, it is precisely the [quantum dual]{} density $\rho_Q$ to the classical density $\rho_ \Lambda$ today. [Cosmological Constant Conclusion:]{} The quantum vacuum density or quantum $\Lambda_Q =\rho_ Q = 10^{+122}$ (in Planck units) [is not]{} what is observed today and must be consistently such way because the Universe today [is not]{} in a quantum super-planckian gravitational state. The Universe today is in a classical gravitational regime and dilute classical state. And what is observed today is consistently and correctly the classical low dilute value $\rho_ \Lambda = 10^{-122}$ or classical vacuum $\Lambda$ corresponding to the classical Universe today. The quantum vacuum density $\rho_ Q = 10^{+122}$ [is ]{} a quantum super-planckian value and must be consistently such way because it is the quantum precursor in a quantum gravitational super-planckian very early past state. The past Universe before the Planck time is in a quantum gravitational super-planckian regime and highly quantum superplanckian state, precursor of the observed era today of the Universe. [The Whole History. An Unifying Picture:]{} We see that going back in time along the Universe evolution from the present era to the early stages where the Universe becames more and more quantum, the classical temperature $T_{\Lambda}$ decreases, as it must be, the quantum temperature $T_Q$ becomes higher and the values of the Classical and Quantum temperatures $T_{\Lambda}$ and $T_Q $ Eqs.(\[TH2\]),(\[TQ2\]) become closer of each other, the difference dissapearing at the [Planck scale]{}: $ T_{\Lambda} = T_Q = t_P$, which is the [crossing scale]{} between the classical/semiclassical and quantum gravity regimes or eras. Similarly, going back in time, from the present era to the early quantum eras of the Universe, the classical gravitational entropy $S_{\Lambda}$ decreases from its huge value today $S_{\Lambda \;today} = 10^{122}\;s_P$ at $10^{61} t_P$ to the inflationary value $S_{\Lambda \;inflation} = 10^{12}\;s_P$ in the Inflation era (semiclassical gravity era) at $10^{6} t_P$, then after descending to its small Planckian value $s_P = \pi\kappa_B$ at the Planck time $t_P$, in which it enters the quantum and super-Planckian regime, decreasing for instance to $S_{Q\; inflation} = 10^{-12}\;s_P$ (the quantum dual phase of Inflation) at the time $ 10^{-6} t_P$, untill reaching its smallest extreme value $s_P \; 10^{-122}$ at $ 10^{-61}\; t_P$. ![ The Universe is composed of two big phases after and before the Planck scale $t_P$ (the [crossing scale]{}). The post-planckian classical gravitation phase is the Universe from $ t_P $ to the present age $10^{61} t_P$. The quantum (planckian and super-planckian) phase from the extreme past $10^{-61} t_P$ to $t_P$ is its precursor. The complete history goes from $10^{-61} t_P$ to $10^{61} t_P$. Quantum physics, Planck scale, natural to the system, and gravitation unify and clarify the whole history. See the text at the end of Section XI and the complete figure caption there.](QG06.eps){height="20" width="26"} Conversely, starting from the earliest past quantum era from $10^{-61}\; t_P$ to $t_P$, the quantum entropy $S_Q$ increases from its extreme small value $S_{Q} = 10^{-122} \;s_P$ at the earliest time $ 10^{-61}\; t_P$ till for instance its quantum inflation value $10^{-12} \;s_P$ at the time $10^{-6}\; t_P$, to its Planck value $S_Q = s_P = \pi \kappa_B$ at the Planck time $t_P$, the [crossing scale]{}, after which it goes to its semi-classical inflationary value $S_{\Lambda \;inflation} = 10^{12} \;s_P$ at the classical inflationary stage at $10^{6} \; t_P$ and it follows [increasing and classicalizes]{} till the extreme maximal classical value today $S_{\Lambda} = 10^{122} \;s_P$ at the present Universe time $10^{61}\; t_P$. And $S_{\Lambda}$ will continue increasing to higher values in the future as far as the Universe will continue expanding its horizon. The [total]{} $Q\Lambda$ gravitational entropy (for the whole history) is the sum of the three values above discussed corresponding to the three regimes: classical $\Lambda$, quantum dual $Q$ and Planck values (subcript $P$), Eq.(\[LambdaQLambdavalue\]). In the past remote and more quantum (Q) eras: $10^{-61} \;t_P \leq t \leq t_P$, the Planck entropy value $s_P = \pi \kappa_B$ dominates $S_Q$. In the classical eras: $t_P \leq t \leq 10^{61} t_P$, the today entropy value $S_\Lambda = 10^{+122} s_P$ dominates. [The whole picture is depicted in Figure (1)]{}, where: $\Lambda$ refers to the cosmological constant (or associated Hubble-Lemaitre constant H) in the Classical gravity phase. Q means quantum. P means Planck scale. Planck’s units, natural to the system, greatly simplify the history. (The complete history is a theory of pure numbers). Each stage is characterized by the set of main physical gravitational quantities: ($\Lambda$, density $\rho_\Lambda$, size $L_\Lambda$, gravitational temperature $T_\Lambda$ and entropy $S_\Lambda$). In the Quantum phase, their corresponding quantum precursors are labeled with the subscript Q. Classical and quantum precursor stages and their associated physical quantities are classical-quantum duals of each other in the precise meaning of the classical-quantum or wave-particle duality including gravity Eq.(\[Udual\]). Total means the whole history including the two phases or regimes. The present age of the Universe $10^{61}$, (with $\Lambda = \rho_\Lambda = 10^{-122} = 1/S_\Lambda$) is a [lower bound]{} to the future Universe age and similarly for the present entropy value $S_\Lambda$. While $10^{-61}$, (with $\Lambda_Q = 10^{122} = \rho_Q = 1/S_Q$ is an [upper bound]{} to the extreme past (origin) of the Universe and quantum initial entropy, (arrow of time). \[Similarly, the values given in Fig.1 (in Planck units) for the CMB are the classical CMB age ($3.8 \; 10^5 yr = 10^{57} t_P$) and the set of caracteristic gravitational properties of the Universe at this age, and their corresponding quantum precursors in the quantum preceding era at $10^{-57} t_P$. $T_\Lambda$ and $S_\Lambda$ are also un upper bound to the temperature and entropy of the CMB photon radiation.\] Conclusions =========== We have accounted in the Introduction and along the paper the main new features of the paper and will not include all of them here. We refer to Section I for a summary of the main results and the end of previous Section XI for the whole picture. - We described classical, semiclassical and quantum de Sitter regimes. A clear picture for the de Sitter background and the whole Universe epochs emerges, going beyond the current picture, both for its classical and quantum regimes, depicted in Fig (1). This is achieved by recognizing the relevant scales of the classical and quantum regimes of gravity. They turn out to be the classical-quantum duals of each other, in the precise sense of the wave-particle (de Broglie, Compton) duality extended to the quantum gravity (Planck and super-Planck) domain: wave-particle-gravity duality. - [ Concepts as the Hawking temperature and the usual (mass) temperature are shown to be precisely the same concept in the different classical and quantum gravity regimes respectively. Similarly, it holds for the Bekenstein-Gibbons and Hawking entropy.]{} - [ An unifying clarifying picture has been provided including the main physical gravitational intrinsic magnitudes of the Universe: age, size, mass, vacuum density, temperature, entropy, in terms of the cosmological constant covering the relevant gravity regimes or cosmological stages: classical, semiclassical and quantum -planckian and superplanckian- eras.]{} - [ Cosmological evolution goes from a super-planckian and planckian quantum phase to a semiclassical accelerated de Sitter era (field theory inflation), then to the classical phase untill the present de Sitter era. The wave-particle-gravity duality precisely manifests in this evolution, between the different gravity regimes, and could be view as a mapping between asymptotic (in and out) states characterized by the sets $U_\Lambda$ (or $U_H$) and $U_Q$, and thus as a Scattering-matrix description: The most early quantum super-Planckian state in the remote past being the in-state, and the very late classical dilute state being the far future or today out-state.]{} - [ Along its physical history, from the very early stages to the present time, the Universe evolved from quantum stages to classical physics stages: that is to say, the Universe [ classicalized]{}. And conversely, from the present time to the earlier stages, the Universe becomes [quantized]{}. Inflation is part of the standard cosmological model and is supported by the CMB data of temperature and temperature-E polarisation anisotropies. This points to $10^{-6} m_P $, (or $10^{-5} M_P$ for the reduced mass $M_P = m_P/\sqrt{8 \pi}$) as the energy scale of Inflation [@CiridVS],[@BDdVS], safely below the Planck energy scale $m_P$ of the onset of quantum gravity. This implies that Inflation is consistently in the [semiclassical gravity regime]{}. This in turn implies that the preceding phase of Inflation corresponds to a quantum gravity phase in the Planckian and super-Plankian quantum gravity domaine. Inflation being a de Sitter, (or quasi de Sitter) stage, it has a smooth space-time curvature [without any physical space-time singularity]{}.]{} - [ Integrating the above different pieces of knowlodge, and because the more earlier known stages of the Universe are de Sitter (or quasi de Sitter) eras, it appears as a consequence of our results that there is [no singularity]{} at the Universe’s origin. First: the so called $t = 0$ Friedman-Robertson Walker mathematical singularity is [not]{} physical: it is the result of the extrapolation without any quantum physics of the purely classical (non quantum) General Relativity theory, out of its domain of physical validity. The Planck scale is not merely a useful system of units but a physically meaningful scale: quantum gravity. The Planck scale precludes the extrapolation to zero time or length. This is precisely what is expected from quantum physics in gravity: the smoothness of the classical gravitational singularities. Second: Inflation (classical or quantum) in the very past ($10^{6} t_P$ or $10^{-6} t_P$ is mainly a de Sitter (or quasi de Sitter) smooth constant curvature era [without any curvature singularity]{}. Third: the extreme past (at $10^{-61} t_P$) is a superplanckian de Sitter state of high [bounded]{} superplanckian constant curvature and therefore [without singularity]{}. Of course, this paper is not devoted to the singularity issue but this argument and the whole picture emerging from this paper indicate the trend and insight into the problem.]{} - [ The main property used here is the classical-quantum duality, which is a universal foundational milestone of quantum theory. Further couplings, interactions and background fields can be added. The conceptual results here will not change by adding further couplings or interactions, or further background fields to the background here. Of course, this is just a first input in the construction of a complete physical theory and understanding [in agreement with observations]{}.]{} - [ The existence and present state of the Universe is physically explained because of the classical-quantum duality as a basic and universal property of Nature: Our known classical Universe does exist [precisely]{} because there existed a preceding phase or precursor: Such preceding phase is [exactly]{} the quantum dual phase of the existing known classical phase, the Planck time being precisely [the crossing]{} time between the two phases, as given by Eq.(\[Udual\]). The Planck time is the transition to the Classical/semiclassical gravity Universe from the “end” (“late” time or entropy) of the Quantum dual preceding phase.]{} - [ Besides its conceptual and fundamental physics interest, this framework revealed of deep and useful clarification for relevant cosmological eras and its quantum precusors and for problems as the cosmological constant. This could provide realist insights and science directions where to place the theoretical effort for cosmological missions and future surveys such as Euclid, DESI and WFIRST for instance, [@Euclid], [@DESI], [@WFIRST], and for the searching of cosmological quantum gravitational signals.]{} - [ The exhibit of $(c, G, h)$ helps in recognizing the different relevant scales and physical regimes. Even if a hypothetical underlying “theory of everything” could only require pure numbers (option three in [@Duff]), physical touch at some level asks for the use of fundamental constants [@Okun],[@Sanchez2003]. Here we used three fundamental constants, (tension being $c^2/G$). It appears from our study here and in ref [@Sanchez2019], that a complete quantum theory of gravity would be a theory of pure numbers.]{} [ACKNOWLEDGEMENTS]{} The author acknowledges the French National Center of Scientific Research (CNRS) for Emeritus Director of Research contract and F. Sevre for help with the figure. This work was performed in LERMA-CNRS-Observatoire de Paris- PSL University-Sorbonne University. WMAP collaboration: A. Kogut et.al. Astrophys. J. Suppl. 148, 161 (2003); D. N. Spergel et.al. Astrophys. J. Suppl. 148, 175 (2003). WMAP Collaboration: G. Hinshaw et al., Astrophys. J. Suppl. 208, 19 (2013). C. L. Bennett et.al. Astrophys. J. Suppl. 208, 20 (2013). Supernova Search Team: A. G. Riess et al., Astron. J 116, 1009 (1998). Supernova Cosmology Project Collaboration: S. Perlmutter, et al., Astrophys. J. 517, 565 (1999). B. P. Schmidt, [Measuring global curvature and Cosmic acceleration with Supernovae]{}, in “Phase Transitions in Cosmology: Theory and Observations” NATO ASI Series vol 40, pp 249-266, Eds H.J. de Vega, I.M. Khalatnikov and N.G. Sanchez, Kluwer Publ (2001). Dark Energy Survey Collaboration, T. Abbott, F.B. Abdalla, et al., MNRAS, 460, 1270 (2016); T.M.C. Abbott, S. Allam et al., arXiV:1811.02374 (2018). Planck Collaboration: N. Aghanim et al, Planck 2018 results. VI. Cosmological parameters, arXiv:1807.06209 (2018); Y. Akrami et al, Planck 2018 results.X, Constraints on inflation, arXiv:1807.06211 (2018). LISA Mission: https://www.lisamission.org/; http://sci.esa.int/lisa/; https://lisa.nasa.gov/ LIGO and Virgo Collaborations: B.P. Abbott et al., Phys. Rev. Lett. 116, 061102 (2016). DES and LIGO/Virgo Collaborations, M. Soares-Santos, A. Palmese et al., arXiv:1901.01540 (2019). H.J. de Vega and N. Sanchez, Phys Rev D50, 7202 (1994); H.J. de Vega, A.L. Larsen, and N.G. Sanchez, Nucl. Phys. B427, 643 (1994); H.J. de Vega, A.L. Larsen, and N.G. Sanchez, Phys. Rev. D51, 6917 (1995). H. J. de Vega and N. Sanchez Phys. Rev. D47, 3394 (1993); A.L. Larsen, and N.G. Sanchez, Phys. Rev. D54, 2801 (1996); M. Ramon Medrano, N.Sanchez, Phys. Rev. D60, 125014 (1999). N.G. Sanchez, Int. J. Mod. Phys. A 19, 4173, (2004); A. Bouchareb, M. Ramon Medrano, N. G. Sanchez, Int. J. Mod. Phys. D16, 1053 (2007). N. G. Sanchez, Int. J. Mod Phys D28, 1950055 (2019). N. G. Sanchez, Int. J. Mod. Phys. A19, 2011, (2003); M. Ramon Medrano, N.G. Sanchez, Mod. Phys. Lett. A18, 2537 (2003); D. J. Cirilo-Lombardo, N.G. Sanchez, Int. J. Mod. Phys. A23, 975 (2008). D. Boyanovsky , H. J. de Vega and N. G. Sanchez, Phys. Rev. D72, 103006 (2005). D. Boyanovsky, H. J. de Vega and N. G. Sanchez, Phys. Rev. D71 023509 (2005); Nucl. Phys. B747, 25 (2006). D. Cirigliano, H. J. de Vega, N. G. Sanchez, Phys. Rev. D71, 103518 (2005); H. J. de Vega and N. G. Sanchez, Phys. Rev. D74, 063519 (2006); D. Boyanovsky, H. J. de Vega, and N. G. Sanchez, Phys. Rev. D73, 023008 (2006). D. Boyanovsky, C. Destri, H. J. de Vega , N. G. Sanchez, [Review on The Effective Theory of Inflation in the Standard Model of the Universe and the CMB+LSS data analysis]{}, Int. J. Mod. Phys. A24, No.20n21, 3669 (2009); C. Burigana, C. Destri, H. J. de Vega, A. Gruppuso, N. Mandolesi, P. Natoli, N.G. Sanchez, Astrophs. J. 724, 588 (2010). L. A. Thompson and S. A.Gregory, [An Historical View: The Discovery of Voids in the Galaxy Distribution]{}, arXiv:1109.1268, (2011) and refs therein. G. Chincarini and H.J. Rood, Nature, 257, 29 (1975); G. L. Chincarini, R. Giovanelli and M. P. Haynes, ApJ, 269, 13 (1983); V. de Lapparent, M. J. Geller and J.P. Huchra, ApJ, 302, L1-L5 (1986). N. Hamaus, P.M. Sutter, and B. D. Wandelt, Phys. Rev. Lett. 112, 251302 (2014); P. M. Sutter, G. Lavaux, N. Hamaus, B. D. Wandelt, D. H. Weinberg, and M. S. Warren, Mon. Not. R. Astron. Soc. 442, 462 (2014). G.W. Gibbons and S.W. Hawking, Phys Rev D15, 2738 (1977). G.W. Gibbons and S.W. Hawking, Phys Rev D15, 2752 (1977). S.W. Hawking, Phys Lett 134B, 403 (1984). Euclid http://sci.esa.int/euclid/, DESI: http://desi.lbl.gov WFIRST: https://wfirst.gsfc.nasa.gov N.G. Sanchez, Int. J. Mod. Phys. A18, 2011 (2003). M.J. Duff, L. B. Okun and G. Veneziano, JHEP 03, 023 (2002). L. Okun, Sov. Phys. Usp.34, 818 (1991); L. B. Okun, physics/0310069 and references therein; K. Kuchar, Phys Rev D22, 1285 (1980); N. G. Sanchez, Chalonge School Lectures, Erice 1996; G.W. Gibbons, Found. Phys. 32, 1891 (2002).	{ "pile_set_name": "ArXiv" }
--- abstract: 'Understanding the hierarchical self-organization of living systems is one of the biggest conceptual challenges of the present century. A generically observed mechanism that drives such organization is interaction among the individual elements—which may represent cells, bacteria, or even enzymes—via chemical signals. We use dynamical renormalization group approach to study a stochastic model for chemotactic particles in which we introduce polarization effects in chemotaxis. We find exact dynamic scaling exponents that represent superdiffusive behavior of the particles. The number fluctuations within sub-regions of the system show a hyperuniform structure or exhibit giant number fluctuations, depending on whether or not the noise is conserved. We expect our results to shed light on how molecular regulation of chemotactic circuits can determine large-scale behavior of cell colonies and tissues through emergent properties that result from a subtle interplay between nonequilibrium fluctuations and long-range interactions that are constrained by exact symmetries.' author: - Saeed Mahdisoltani - Riccardo Ben Alì Zinati - Charlie Duclut - Andrea Gambassi - Ramin Golestanian title: Controlled dynamics and number fluctuations with two strategies for quorum sensing --- [^1] [^2] [^3] Introduction {#sec:intro} ============ Characterizing the emergence of macroscopic properties in colonies of prokaryotic [@adler1966; @ben-jacob2000] and eukaryotic [@levine2013] cells based on the complicated chemical interactions among the individuals in the colony is a long-standing endeavor in various areas of biology such as morphogenesis [@hogan1999; @crick1970; @friedl2009], tissue growth and homeostasis [@tzur2009], wound healing [@schneider2010], and cancer metastasis [@hanahan2011; @bockhorn2007]. The term chemotaxis is often used to describe the ability of bacteria and cells to detect the changes in the concentrations of specific chemical molecules in their surrounding media [@iijima2002] and to respond to them by adjusting their polarization or direction of motion [@roussos2011; @iglesias2008; @alon1999]. Although the detailed mechanisms responsible for chemotaxis in cells are rather complex [@Barkai1997; @sourjik2004; @wadhams2004; @Tu2008; @Emonet2008; @Tu2013; @westendorf2013], the phenomenon seems to emerge generically in nature and, in fact, it has also been observed in smaller and more primitive systems such as enzymes [@agudo-canalejo2018] and synthetic catalytically active colloids [@illien2017; @stark2018]. Phenomenological models describing chemotaxis as a directed motion guided by the chemical gradients [@keller1971; @tsori2004] and undergoing stochastic fluctuations [@grima2005; @golestanian2009; @sengupta2009; @taktikos2012] have proven useful in studying chemotactic collapse of bacteria [@brenner1998; @chavanis2004] and collective behavior of active colloids [@golestanian2012; @cohen2014]. Generalizations of these models which incorporate the polarity of the active particle and its active alignment have been used to study collective properties of synthetic active Janus particles [@saha2014] as well as chemotaxis for trail–following bacteria [@kranz2016; @gelimson2016]. The common trait of these models is the long-range nature of the interactions, which typically act for relatively long times, as the transmitting molecules decay very slowly. Accordingly, it is not surprising that chemotactic interactions share some of the features of other long-range interactions, such as the electrostatic and gravitational ones, and the self-organization of chemotactic species resembles the formation of galaxies in astrophysics, as well as the large-scale vortices in two-dimensional turbulence [@chavanis2010; @chavanis2008]. Owing to the large number of degrees of freedom usually involved in a colony, coarse-grained descriptions are particularly useful in studying these and more general active systems [@marchetti2013]. When the correlations become long-ranged and collective phenomena emerge, the standard field-theoretical approaches have been applied to a wide range of models of biological or synthetic colonies such as flocks of birds, schools of fish, aggregations of molecular motors, chemotactic particle, and active phase separation [@vicsek1995; @toner1995; @toner2012; @risler2004; @gelimson2015; @caballero2018a]. Here, we investigate the large-scale behavior of systems consisting of particles with chemotactic interactions (see Fig. \[fig:schematic\]) using a dynamical renormalization group (RG) treatment of the Langevin equation governing the particle density. Noting that the Keller–Segel description of chemotaxis [@keller1971] resembles an expansion in terms of the spatial gradient of the chemical signals, we extend this expansion and find a previously overlooked chemotactic nonlinear term which, based on dimensional analysis and symmetry considerations, has to be included in the Langevin description. This term represents a polarization-induced chemotactic interaction and happens to break detailed balance (Appendix \[app\_detailedbalance\]). The resulting stochastic field theory is invariant under a “Galilean” transformation, which implies that the two relevant nonlinearities of the model are not affected by the RG flow. Moreover, the noise strength is not renormalized due to the structure of the vertex function in the hydrodynamic limit. These simplifications enable us to obtain exact scaling exponents as functions of the spatial dimensionality $d$ of the system. Our one-loop calculations show that the chemotactic term introduced here is necessary in order to obtain stable RG fixed-points and therefore to observe scaling behavior. The rest of the paper is organized as follows. In Sec. \[sec:newclass\] we summarize the main physical characteristics of the self-organization phenomenon that we study here and put our work in the context of existing studies in the literature. Section \[sec:bio\] provides a mechanistic view of the biological context that our work is addressing and the phenomenology that arises from the new mechanism of quorum sensing. In Sec. \[sec:stoch\] the stochastic framework of our chemotaxis model is introduced, and followed by a description of the RG calculation in Sec. \[sec:RG\]. Section \[sec:expo\] develops the symmetry considerations that afford us exact exponents, and Sec. \[sec:phase\] summarizes some of the features of the resulting phase diagram of the system. Finally, Sec. \[sec:conc\] concludes the paper with some discussions. There are four appendices that elaborate further on some of the details of the calculations regarding the derivation of our stochastic framework from the generic approach of gradient expansion and power counting (Appendix \[app\_gradexpansion\]), the question of detailed-balance in our chemotactic field theory (Appendix \[app\_detailedbalance\]), the RG flows in all dimensions (Appendix \[app\_rgflowsdim\]), and the perturbative RG calculation (Appendix \[app\_rgprocedure\]). ![Schematics of a system of chemotactic particles. The particles (shown as light green blobs) secret chemicals that are pictured as light blue dots. The red arrows at each point represent the chemotactic force corresponding to the $\mu_1$ term in Eq. . The white arrows show forces stemming from polarization mechanisms and correspond to the $\mu_2$ term in the same equation. []{data-label="fig:schematic"}](schematic_with_dots.pdf){width="48.00000%"} A new class of nonequilibrium self-organization {#sec:newclass} =============================================== To put our work in the appropriate perspective, it is helpful to start by summarizing the new conceptual insights and technical advances that it entails. We investigate the emergent macroscopic properties of a chemotactic system based on the interactions between its individuals. These systems are often studied through various extensions of Keller–Segel models, which are treated in the mean-field approximation by ignoring correlations between the particles. We focus on the frequently observed cases where the chemical signals created by the chemotactic particles decay very slowly, resulting in long-range correlations between the particles. In these situations, the fluctuations determine the behavior of the system and an RG treatment is required to identify the large-scale and long-time properties such as the (anomalous) diffusion exponent and number fluctuations. The RG analysis, along with power counting and symmetry considerations, has helped us identify a new chemotactic interaction, whose mechanistic origin relies on the induced polarization of the particles. We discover a generalized Galilean symmetry in the spirit of what has been studied in studies of the Kardar–Parisi–Zhang equation, which is realized when the diffusion of the chemical signals is considerably faster than that of the particles, and therefore, the chemical concentration field is effectively quasi-stationary. This symmetry, which had been overlooked in all previous studies of chemotaxis, boosts our understanding of the formal structure of the stochastic field-theoretical description of the system, as it provides nonperturbative relationships between correlation functions (Ward identities) as well as exponent identities, which are highly valuable in such physical formulations. Our findings provide a broad new picture about how to qualitatively rationalize the complex interplay between nonequilibrium fluctuations and long-range interactions. The polarization-induced nonlinearity is a purely nonequilibrium interaction, which indicates that the system cannot reach an equilibrium state in the long-time limit. This is in contrast with the traditional Keller–Segel chemotactic drift term that is essentially an equilibrium-like interaction, as it satisfies detailed-balance. The fluctuations are most relevant at the critical state when the correlation length diverges and the mean-field treatment breaks down. We focus on this case and use the Dean–Kawasaki formalism that encapsulates all the statistical correlations, which are neglected in the Keller–Segel equation, and then use RG techniques to relate the microscopic theory to the effective theory at macroscopic spatial and temporal scales. Since the interactions may include non-analytical parts, the coarse-graining step of RG for such systems can become cumbersome. However, implementing the newly found Galilean symmetry greatly simplifies the calculations by imposing constraints on the RG flows, which eventually result in exact identities. Note that the symmetry identified here is expected to be realized in a wide range of chemical, electrostatic and gravitational systems with long-range interactions. The exact analysis of the scaling properties of the stochastic framework provides us with exact exponents that predict superdiffusion due to quorum sensing. The existence of polarization-induced chemotactic coupling, which represents a nonequilibrium interaction, is essential for the emergent superdiffusive behavior with the reported exponents, and is relatively insensitive to the symmetry of the noise. On the other hand, the fluctuations in the particle-number depend on the nature of the noise: in the case of a conserved noise the fluctuations are suppressed and the distribution becomes hyperuniform, while a nonconserved noise enhances the fluctuations leading to the emergence of giant number fluctuations. Biological significance of polarization-induced quorum sensing {#sec:bio} =============================================================== Before elaborating on the details of the work, we highlight the biological implications of our new findings and describe the resulting phenomenological framework that can be used in the relevant quantitative studies of living systems. The chemotactic response of a [single]{} cell in a medium with a concentration field $\phi(\mathbf{x},t)$ is commonly described by a drift velocity $$\mathbf{v}_{\rm KS}=M_1 \nabla \phi,\label{eq:KS-1}$$ as first studied by Keller and Segel [@keller1970; @keller1971]. This biased motion can be a result of temporal sensing mechanisms [@sourjik2004; @wadhams2004], as observed in prokaryotes such as E. coli, or spatial sensing [@iglesias2008; @levine2013; @westendorf2013], as observed in eukaryotes. To understand this phenomenology, let us consider the stochastic motion of a motile cell (or an active self-propelled colloid), which can be described by instantaneous velocity $$\mathbf{v}_{\rm sp}=v_0 \mathbf{n}, \label{eq:sp-1}$$ where $v_0$ is the self-propulsion speed and $\mathbf{n}$ is a unit vector that describes the instantaneous orientation of the cell. The cell orientation is randomized over the time scale of $D_r^{-1}$ (where $D_r$ is an effective rotational diffusion coefficient). An isolated single cell undergoes unbiased diffusion at time scales longer than $D_r^{-1}$. In the presence of a gradient, a feedback mechanism modulates the motility apparatus such that a net bias is generated towards (or away from) the gradient [@alon1999]. Mechanistically, this type of response is expected to arise when the cell can undergo a polarization change in response to an external chemical gradient. This can be achieved through shape changes, alignment via reorientation, or redistribution of surface receptors [@roussos2011; @iglesias2008], which are assumed to occur over a time scale of $D_r^{-1}$ by generalizing the notion of the randomization of the orientation. The alignment tendency can be described by an effective angular velocity $\bm{\omega}=\chi \bf{n} \times \nabla \phi$, which is characterized by a polarization coupling $\chi$. This type of response is known to be prevalent in eukaryotic cells [@iglesias2008; @levine2013], and reported in the context of chemotactic response of surface-moving bacteria [@kranz2016; @gelimson2016], chemically active colloids [@saha2014], and enzymes [@AdelekeLarodo2019b]. Beyond the time scale of rotational diffusion, the angular velocity leads to a net bias in the average orientation of the cell along the direction of the gradient that reads [@RG-LesHouches] $$\mathbf{p}=\langle \mathbf{n} \rangle=\frac{\chi}{d D_r} \nabla \phi,\label{eq:polar-1}$$ in $d$ dimensions. This suggests that averaging Eq. (\[eq:sp-1\]) over time scales beyond $D_r^{-1}$ and inserting the degree of polarization from Eq. (\[eq:polar-1\]), we obtain the Keller–Segel form for the net drift velocity as given in Eq. (\[eq:KS-1\]), where the response coefficient is obtained as $$M_1=\nu_1+\frac{v_0 \chi}{d D_r}.\label{eq:nu1-micro-1}$$ Note that here we have included an additional microscopic drift term of the form $\nu_1 \nabla \phi$ that might exist, as it is known in the case of enzymes [@agudo-canalejo2018]. There exists another, independent, mechanism by which cell polarization can influence chemotactic response. To illustrate this mechanism, we consider a chemotactic cell for which the distribution of the chemical sensory units locally determines the feedback onto the motility machinery. For instance, it is has been reported that the distribution of the chemical sensing units on the membrane of neutrophils is a function of the chemical gradient in the surrounding [@servant2000]. A manifestation of this response has been shown to arise in surface-moving bacteria due to a coupling between the asymmetric geometry and the spatial distribution of sensors [@kranz2016; @gelimson2016]. By assuming that the overall movement of the cell is the resultant of the combined influence of such local couplings between gradient sensing and motility, we need to augment the general expression for drift velocity to include higher order moments of the shape orientation as follows $$\mathbf{v}=v_0 \mathbf{n}+\nu_1 \nabla \phi+\nu_2 \mathbf{n}\cdot\nabla\nabla\phi+\cdots,\label{eq:v-nu1-nu2-1}$$ where the new coupling $\nu_2$ represents the dipolar contribution of the sensing-motility coupling. Averaging over time scales beyond the reorientation time and inserting Eq. (\[eq:polar-1\]) in the new term, we obtain the following expression for the polarization-induced quorum sensing (QS) drift term $$\mathbf{v}_{\rm pi-QS}=M_2 \nabla(\nabla\phi)^2,\label{eq:pi-QS-1}$$ where $$M_2=\frac{\chi \nu_2}{2 d D_r}.\label{eq:M2-1}$$ Note that the two terms depend differently on the properties of the chemical gradient vector field, as schematically represented in Fig. \[fig:schematic\]. These two microscopic mechanisms can be incorporated into an effective field theoretical description to study the large-scale collective properties of such systems (see below). In the resulting framework that is represented by Eq. , which governs the time evolution of the fluctuations in the stochastic density (denoted by $\rho$), the Keller–Segel term leads to the $\mu_1 \nabla(\rho\nabla\phi)$ term, while the new polarization-induced mode results in the $\mu_2 \nabla^2 (\nabla\phi)^2$ term. One can obtain microscopic values for the corresponding coupling constants, in terms of the calculations described above. This yields the following expressions $$\begin{aligned} \mu_1^{\rm mic}&=&M_1=\nu_1+\frac{v_0 \chi}{d D_r},\label{eq:mu1-mic}\\ \mu_2^{\rm mic}&=& M_2 \rho_0=\left(\frac{\chi \nu_2}{2 d D_r}\right) \rho_0,\label{eq:mu2-mic}\end{aligned}$$ where $\rho_0$ is the average concentration of the chemotactic cells in the medium. The above microscopic derivation helps to clarify the fundamental difference between the two mechanisms: while the standard term exists already at the single-particle level, the new term is proportional to the mean density of the cells in the environment, and hence, represents quorum sensing. This means that while such a polarization-induced response is of higher orders at the single cell level, when interacting collectively this term provides a new mechanism that becomes equally important as the Keller–Segel mechanism. Moreover, as we will demonstrate below, the new term represents a fundamentally nonequilibrium mechanism that can lead to emergent properties that cannot be captured by the traditional Keller-Segel model of chemotaxis. Stochastic model for chemotaxis {#sec:stoch} =============================== We consider a system of mobile particles which sense and respond to the gradients of chemical signals in their surrounding medium, as sketched in Fig. \[fig:schematic\], and study it in the overdamped regime, typical of biological situations. Chemical signals are secreted at a constant rate by all particles in the system, creating a chemical concentration field whose spatial average is increasing in time. The fluctuations of the chemical concentration are responsible for non-zero gradients which, in turn, result in the motion of the chemotactic particles. Consequently, the fluctuations of the particle density $\rho$, act as sources for $\phi$. We assume that the diffusion of the chemical signals occurs on a time scale that is negligible compared to the dynamics of the particles, and thus the diffusion equation can be used in the steady-state, which results in the Poisson equation $$\begin{aligned} -\nabla^2\phi(\mathbf{x},t) = \rho(\mathbf{x},t).\label{eq:poisson}\end{aligned}$$ Note that the proportionality constant has been set to unity without possible loss of generality. In what follows we shall continue to refer to $\rho$ and $\phi$ as the density of particles and the chemical concentration, respectively, bearing in mind that, being fluctuations, they can assume also negative values. The Langevin equation governing the dynamics of $\rho$ can be derived from the microscopic equations of motion of chemotactic particles by using the Dean–Kawasaki formalism [@dean1996; @kawasaki1994]. We obtain an extension of the stochastic Keller–Segel model [@chavanis2010; @gelimson2015], which reads: $$\label{eq:langevin} \begin{split} ( \partial_t - D \nabla^2 + \sigma) \, \rho (\mathbf{x},t) = &- \mu_1{\nabla} \cdot (\rho\nabla\phi)\\ &- \mu_2\nabla^2(\nabla\phi)^2 + \zeta(\mathbf{x},t)\,, \end{split}$$ where $D$ is the diffusivity of the particles and $\sigma$ accounts for the decay of the particles as well as the linear terms arising from chemotaxis (see Appendix \[app\_gradexpansion\]). Furthermore, $\mu_1$ represents the Keller–Segel chemotactic response of the particles, while $\mu_2$ is a novel chemotactic interaction resulting from the polarization of the particles, which we explain in more details below. The stochastic nature of the dynamics, due to the microscopic degrees of freedom that have not fully decoupled from the macroscopic variables, is accounted for by the Gaussian noise $\zeta$, whose correlation function in $d$ spatial dimensions is $$\left\langle\zeta(\mathbf{x},t)\zeta(\mathbf{x}',t')\right\rangle=2(\mathcal{D}_0-\mathcal{D}_2\nabla^2)~\delta^d(\mathbf{x}-\mathbf{x}')\delta(t-t')\,, \label{eq:noise0}$$ where $\mathcal{D}_0$ and $\mathcal{D}_2$ are the strengths of the noise. For $\mathcal{D}_0 \neq 0$, the noise does not conserve locally the number of particles which fluctuates around a mean value. This nonconserved noise is relevant to describe colonies of bacteria or cells that have reached a homeostatic state in which the death and division rates of the individuals compensate each other only on average, with some stochastic fluctuations. When $\mathcal{D}_0 = 0$, on the other hand, the resulting conserved noise induces fluctuations only in the particle current and the particle number is locally conserved, as it happens in systems with strictly fixed number of particles, e.g., enzymes or active colloids. In Eq. , the nonlinear term $- \mu_1 \nabla \cdot (\rho\nabla\phi)$ accounts for the current $\mu_1 \rho\nabla\phi$ of the particles as they chemotax up ($\mu_1>0$) or down ($\mu_1<0$) the chemical gradient, assuming a linear form of their response functions. Note that this term can be derived from a free energy functional $\mathcal{F}_{\rm KS} = \int \rho(\mathbf{x},t) \phi(\mathbf{x},t)$ and thus satisfies the detailed balance condition (see Appendix. \[app\_detailedbalance\]). An investigation of chemotaxis, as a generic motion of particles caused by non-uniform chemical fields, using gradient expansion approaches reveals that the term $-\mu_2 \nabla^2(\nabla\phi)^2$ is relevant for particles whose sensitivity to chemical signals in the surrounding is amplified along the gradient of the chemical field [@iglesias2008; @servant2000], and it has to be included in the Langevin equation (see Appendix \[app\_gradexpansion\] for details). As outlined in Appendix. \[app\_detailedbalance\], this term cannot be derived from a free energy and breaks the detailed balance condition, introducing purely nonequilibrium interaction between the particles. The characteristic time and length scales of the dynamics depend on the chemotactic response of the particles, as well as their linear birth and decay processes, through the linear term $-\sigma \rho$ in Eq. (see Appendix \[app\_gradexpansion\]). At the critical state, all these effects are fine-tuned such that $\sigma \to 0$ and the system becomes scale-free. This will also hold when the decay length as set by $\sigma$ is considerably larger than the system size. RG techniques can then be utilized to gain valuable information about the macroscopic properties and the corresponding phase transitions for different values of the chemotactic couplings. Renormalization group analysis {#sec:RG} ============================== The solutions to the nonlinear Langevin dynamics Eq. in the critical regime become scale-invariant and therefore are represented by homogeneous functions. In particular, the long-time and large-scale particle density correlations take the scaling form [@medina1989] $$\begin{aligned} \left\langle \|\rho(\mathbf{x},t) - \rho(\mathbf{x}',t')\|^2 \right\rangle \sim \|\mathbf{x}-\mathbf{x}'\|^{2\chi} \, F\left( \frac{\|t-t'\|}{\|\mathbf{x}-\mathbf{x}'\|^z} \right), \label{eq:rhorhoscalingform}\end{aligned}$$ where $F$ is a scaling function and $\chi$ and $z$ are scaling exponents. In the absence of the nonlinear terms in Eq. , these exponents can be obtained by requiring the invariance of the equation under a change of spatial and temporal scales according to $\mathbf{x}' = \mathbf{x}/b$ and $t' = t/b^z$ (where $b>1$), combined with a rescaling of the density field as $\rho' = \rho/b^\chi$. For the conserved noise, the dynamics is then made scale-invariant by the choice of (Gaussian) mean-field values $z_{\rm MF}=2$ and $\chi^{\rm con}_{\rm MF}=-d/2$, whereas for a nonconserved noise the exponents read as $z_{\rm MF}=2$ and $\chi^{\rm non}_{\rm MF}=1-d/2$. Furthermore, similar dimensional analysis reveal that with conserved noise, the nonlinear terms scale as $\propto b^{2-d/2}$ and hence they increase upon successive application of rescaling for $d<d^{\rm con}_c=4$. For the nonconserved noise, on the other hand, the nonlinearities scale as $\propto b^{3-d/2}$ and grow in $d<d^{\rm non}_c=6$ spatial dimensions. Accordingly, below the critical dimension $d_c$ the nonlinearities are relevant in determining the scaling exponents and the macroscopic behavior of the system and will be examined via the RG analysis in the following. Upon performing Fourier transformations according to $\rho(\mathbf{x},t)=\int_{\hat{k}} e^{-i\omega t+i\mathbf{k}\cdot\mathbf{x}} \rho(\hat{k})$ with $\hat{k}=(\mathbf{k},\omega)$ and $\int_{\hat{k}}=\int d\omega \, d^d k/(2\pi)^{d+1}$, and using Eq. to represent $\phi$ in terms of $\rho$, Eq. reads as $$\begin{aligned} \rho(\hat{k})= G_0(\hat{k})\left[\zeta(\hat{k}) + \int_{\hat{q}} \Gamma_0(\mathbf{k},\mathbf{q}) \rho(\hat{k}-\hat{q}) \rho(\hat{q}) \right], \label{eq:fourierlangevin}\end{aligned}$$ where we have introduced the so-called bare (i.e. microscopic) propagator $G_0$ and the bare vertex function $\Gamma_0$ as $$\begin{aligned} \label{eq:barepropagator} G_0(\hat{k})=\left(\sigma-i\omega+D\mathbf{k}^2\right)^{-1}\end{aligned}$$ and $$\begin{aligned} \label{eq:barevertex} \Gamma_0(\mathbf{k},\mathbf{q})=\frac{\mu_1}{2} \left(\frac{\mathbf{k}\cdot\mathbf{q}}{\mathbf{q}^2} + \frac{\mathbf{k}\cdot(\mathbf{k}-\mathbf{q})}{(\mathbf{k}-\mathbf{q})^2}\right)-\mu_2 \frac{\mathbf{k}^2 \mathbf{q}\cdot(\mathbf{k}-\mathbf{q})}{\mathbf{q}^2(\mathbf{k}-\mathbf{q})^2}, \end{aligned}$$ respectively. In addition, we also define the bare dynamic correlation function as $$\begin{aligned} \label{eq:barecorr} C_0(\hat k)=2(\mathcal{D}_0+k^2\mathcal{D}_2) \|G_0(\hat k)\|^2. \end{aligned}$$ Note that in the following, we are focusing on the critical state for which $\sigma=0$. In the standard procedure [@medina1989; @tauber2014], the scale-invariant behavior of the system in the presence of nonlinearities (i.e., below $d_c$) is captured by using perturbative RG techniques. A series expansion of Eq. in terms of the couplings $\mu_{1,2}$ is constructed according to the Feynman diagrams in Fig. \[fig:diagrams\]. Firstly, the loop integrals appearing in the perturbation series are computed by integrating out fluctuations with wavevector $\mathbf{k}$ within the momentum shell $\|\mathbf{k}\|\in \left[\Lambda/b,\Lambda\right]$, where $\Lambda$ is the cutoff set by the microscopic length-scale of the particles. This step corresponds to coarse-graining the short-distance fluctuations. In the second step, the variables are rescaled similar to the mean-field case, such that the original cut-off $\Lambda$ is restored and the same Langevin equation holds, but with renormalized (i.e., effective) couplings. By choosing an infinitesimal scaling parameter $b=e^\ell$ with $\ell \to 0$, this coarse-graining procedure can be cast into a differential form. The lowest order, one-loop form of the RG flow equations in the case of conserved noise therefore read as (see Appendix \[app\_rgprocedure\]) ![One-loop Feynman diagrams describing the propagator $G$ (top), the dynamic correlation function $C$ (centre), and the vertex function $\Gamma$ (bottom) renormalization to leading order.[]{data-label="fig:diagrams"}](dcNEW2.pdf){width="48.00000%"} \[eq:cnoiseflows\] $$\begin{aligned} \partial_{\ell} \mu_{1,2} &= \left[ z+\chi \right]\mu_{1,2}, \label{eq:cnoiseflowmu12}\\[1.5mm] \partial_{\ell} D \,\, &= \Big[ z-2 - \big( a_{11} U_1^2 + a_{12} U_1 U_2 + a_{22} U_2^2\big)\Big] D,\label{eq:cnoiseflowD} \\[1.5mm] \partial_\ell \mathcal{D}_2 &= \left[-2-d+z-2\chi\right] \mathcal{D}_2, \label{eq:cnoiseflowD2}\end{aligned}$$ with coefficients $a_{11} = 3/4 - 3/(2 d)$, $a_{12} = 2+3/d-6/(d+2)$, $a_{22} = 1-4/d$, and where we have defined the combined dimensionless chemotactic couplings $U_{1,2}^2 = \mu_{1,2}^2 \, \mathcal{D}_2 \, K_d \Lambda^{d-4}/D^3$ with $K_d=2/[(4\pi)^{d/2}\Gamma(d/2)]$. Notice that $\sigma$ is not renormalized and remains zero at the critical state (see Appendix \[app\_rgprocedure\]). The previous flow equations then imply $$\begin{aligned} \label{eq:cnoiseflowU1U2} \frac{\partial_\ell U_{1,2}}{U_{1,2}} = 2-\frac{d}{2} +\frac{3}{2} \big( a_{11} U_1^2 + a_{12} U_1 U_2 + a_{22} U_2^2 \big),\end{aligned}$$ the solution of which traces rays with fixed $U_2/U_1$ in the $U_1-U_2$ plane, as shown in Fig. \[fig:U1U2flows\]. A similar analysis for the nonconserved noise, as outlined in Appendix \[app\_rgprocedure\], leads to RG flows analogous to Eqs. and, upon introducing the suitable dimensionless chemotactic couplings $U_{1,2}^2 = \mu_{1,2}^2 \, \mathcal{D}_0 \, K_d \Lambda^{d-6} / D^3$, they imply the flow $$\label{eq:nnoiseflowU1U2} \frac{\partial_\ell U_{1,2}}{U_{1,2}} = 3-\frac{d}{2} +\frac{3}{2} \big( b_{11} U_1^2 + b_{12} U_1 U_2 + b_{22} U_2^2 \big),$$ with coefficients $b_{11} = 3/4 - 1/d - 3/[d(d+2)]$, $b_{12} = 2 + 6/d - 9/(d+2)$, and $b_{22} = 1-6/d$, which has the same structure as Eq. . Exact scaling exponents {#sec:expo} ======================= The simple flow structure for the coupling constants $\mu_{1,2}$ resembles that of the nonlinearity in Kardar–Parisi–Zhang equation [@kardar1986; @medina1989] which is not renormalized. An in-depth examination reveals that Eq. is invariant under a Galilean transformation given by $$\begin{aligned} \label{eq:galileantransform} \phi'(\mathbf{x},t) &= \phi\left(\mathbf{x}+ t\left(\mu_1-2\mu_2\right)\mathbf{v} ,t\right)-\mathbf{v}\cdot\mathbf{x},\\ \rho'(\mathbf{x},t) &= \rho\left(\mathbf{x}+t\left(\mu_1-2\mu_2\right)\mathbf{v} ,t\right),\end{aligned}$$ where $\mathbf{v}$ is an arbitrary $d$-dimensional vector. To maintain this microscopic symmetry, which involves $\mu_1$ and $\mu_2$ as its transformation parameters, the RG flow does not affect the chemotactic couplings (this is further corroborated by Ward identities as detailed in Appendix \[app\_rgprocedure\]). This condition gives the exact exponent identity $$\begin{aligned} \label{eq:exponidentity} z + \chi = 0 \,,\end{aligned}$$ to all orders of the perturbative expansion. Moreover, we note that since $\Gamma_0(\mathbf{k},\mathbf{q})$ defined in Eq. is proportional to $\mathbf{k}^2$ in the hydrodynamic limit ($\|\mathbf{k}\|/\|\mathbf{q}\| \ll 1$), the leading order correction to the noise strength obtained from Fig. \[fig:diagrams\] does not affect the conserved or the nonconserved parts. We expect this structure to remain valid at all orders of perturbation and, thus, the RG flow of the noise to be entirely determined by the rescaling in the second step of RG. Consequently, by setting $\partial_\ell \mathcal{D}_2 = 0$ in Eq. and using Eq. , we obtain the exact scaling exponents for $d<d^{\rm con}_c=4$ as $$\begin{aligned} \label{eq:cnoiseexpons} z^{\rm con} = - \chi^{\rm con} = (d+2)/3,\end{aligned}$$ in the case of conserved noise. A similar analysis for the nonconserved noise shows that for $d<d^{\rm non}_c=6$, the exact values of scaling exponents are $$\begin{aligned} \label{eq:nnoiseexpons} z^{\rm non}=-\chi^{\rm non}= d/3.\end{aligned}$$ To make a comparison with the case of simple diffusion, it is convenient to introduce the exponent $\alpha$ as $\Delta L^2 \sim t^\alpha$ with $\alpha =1$ for diffusion. With the chemotactic interactions, on the other hand, we have: $$\begin{aligned} \label{eq:alphaexponent} \alpha=\frac{2}{z}= \begin{cases} 6/(d+2) \quad &\textrm{for conserved noise}, \\ 6/d \quad &\textrm{for nonconserved noise}. \end{cases}\end{aligned}$$ For both conserved and nonconserved noise, one has $\alpha > \alpha_{\rm MF}=1$ as shown in Fig. \[fig\_exponents\], indicating that the chemotactic interactions result in collective superdiffusion of the particles in the colony. As Eq. implies, one can obtain this dynamic exponent in practice by measuring the spatial spreading of the density correlations in time. The number $N$ of particles within a sub-region of volume $V$ has fluctuations given by $\Delta N \sim \rho V \sim N^\gamma$ where $\gamma=1+\chi/d$ and $\gamma=1/2$ corresponds to Poissonian fluctuations. In the presence of chemotaxis, this exponent is given by $$\begin{aligned} \label{eq:gammaexponent} \gamma=1+\frac{\chi}{d}= \begin{cases} (2/3)(1-1/d) \quad &\textrm{for conserved noise}, \\ \, 2/3 \quad &\textrm{for nonconserved noise}. \end{cases}\end{aligned}$$ For the conserved noise, $\gamma<\gamma^{\rm con}_{\rm MF}=1/2$ denotes a hyper-uniform density distribution. For the nonconserved noise, one has $\gamma = \gamma^{\rm non} =2/3 <\gamma^{\rm non}_{\rm MF}=1/2+1/d$: the fluctuations are weaker than the corresponding mean-field case but giant number-fluctuations are still present since $\gamma^{\rm non}>1/2$. Interestingly, the number fluctuations in the nonconserved case appear to be [super-universal]{}, as the exponent $\gamma$ does not depend on the dimensionality of space. As the definition suggests, in experimental setups or simulations the exponent $\gamma$ can be calculated by measuring the extent of population fluctuations in sub-regions of the system. ![ Exact dynamic exponents $\alpha$ (top) and $\gamma$ (bottom) which characterize the anomalous diffusion and number fluctuations (Eqs. and ) as functions of the dimension $d$, in the case of a conserved (green circles) and nonconserved (blue triangles) noise.[]{data-label="fig_exponents"}](gammaz_vertical-crop.pdf){width="48.00000%"} -1cm ![image](u1u2cond2v5L.pdf){width="\linewidth"} ![image](u1u2nond2v5L.pdf){width="\linewidth"} Phase diagrams {#sec:phase} ============== The RG flows at one-loop order for the effective couplings $U_{1,2}$, given by Eqs. and , are plotted in Fig. \[fig:U1U2flows\] for $d=2$ (see Appendix \[app\_rgflowsdim\] for the analysis of the RG flows in different dimensions). At this order of perturbation, the fixed points of the RG equations for both conserved and nonconserved noise take the shape of hyperbolas (red curves) whose asymptotes (blue lines) mark transitions between two different behaviors; the red arrows indicate RG flows toward the fixed-point curves and, hence, show regions in the parameter space whose macroscopic behavior at criticality is described by the scaling exponents discussed in the previous section. The blue arrows, instead, show runaway flows toward strong-coupling regions, with possible first-order phase transitions. Examining these flow equations (see Appendix \[app\_rgflowsdim\]) shows that there are stable fixed points located along the $U_2$-axis (i.e. with $U_1=0$) for all $d$. On the contrary, fixed points with $U_2=0$, i.e., on the $U_1$-axis, exist only in $d=1$ in the case of conserved noise and in $d \lesssim 2.27$ in the presence of a nonconserved noise (such that fixed-point solutions for Eq. with $U_2=0$ are available). These stable fixed-points with finite values of $\mu_1$ refer to non-trivial macroscopic physics of particles with (only) Keller-Segel chemotactic interactions. In $d=1$, one finds that the hyperbolas become fully attractive parallel straight lines (see Appendix \[app\_rgflowsdim\]) and hence generic scaling behavior is expected throughout the $U_1-U_2$ space. Note that $d=1$ is a special case as the chemotactic interactions proportional to $\mu_1$ and $\mu_2$ are linearly dependent. Furthermore, the two hyperbolas of fixed points become straight lines—coinciding with their asymptotes—at the upper critical dimension $d_c$. For $d>d_c$, the Gaussian fixed-point is fully stable, with the hyperbolas of fixed-point marking possible phase-transitions to strong coupling regimes (see Appendix \[app\_rgflowsdim\]). Note that in contrast to the exact scaling exponents, the RG flows discussed here are only valid up to one-loop in a perturbative expansion around the upper critical dimension, while higher-order terms may be required to complete the picture of the phase diagram. Conclusions {#sec:conc} =========== We analyzed the large-scale and long-time dynamical behavior of a system of chemotactic particles by using the dynamical RG approach. The new nonlinear term, $\mu_2 \nabla^2(\nabla^2\phi)$, introduced in the Langevin description is allowed by symmetry considerations and has the same scaling properties as the Keller–Segel term $\mu_1 \nabla\cdot(\rho \nabla\phi)$. These considerations are reminiscent of the approaches employed in the study of phase transitions [@kardarfields] although the resulting interactions in the present nonequilibrium arrangement need not be derivable from an underlying free energy (Appendix \[app\_detailedbalance\]). Mechanistically, this type of response is expected to arise when the cell can undergo a polarization change—achieved through shape changes or redistribution of surface receptors—in response to an external chemical gradient [@roussos2011; @iglesias2008]. This type of response is known to be prevalent in eukaryotic cells [@iglesias2008; @levine2013], and reported in the context of chemotactic response of chemically active colloids [@saha2014] and enzymes [@AdelekeLarodo2019b]. A manifestation of this response can arise in bacteria as well due to a coupling between the asymmetric geometry and the spatial distribution of sensors [@kranz2016; @gelimson2016]. As a result of the underlying Galilean symmetry in Eq. , the chemotactic couplings $\mu_{1,2}$ are not affected by the RG flow, resulting in an exact exponent identity (see Eq. ). With the non-renormalization of the noise strength, which is observed at one-loop order and is expected to hold at all orders of perturbation, these findings enable us to obtain the dynamical scaling exponents exactly. These exponents imply super-diffusive behavior with either hyper-uniform populations (conserved noise) or giant number-fluctuations (nonconserved noise), see Eqs. (\[eq:cnoiseexpons\]) and (\[eq:nnoiseexpons\]) as well as Fig. \[fig\_exponents\]. The fixed points of the RG flows for the effective chemotactic couplings $U_{1,2}$ that (unlike the exact exponents) are only one-loop results, represent a pair of hyperbolas with identical scaling exponents throughout (see Fig. \[fig:U1U2flows\] and Appendix \[app\_rgflowsdim\] for details). The existence of superdiffusion in living colonies due to quorum sensing is a very desirable trait, as it allows the cells to efficiently explore their surroundings [@harris2012generalized; @viswanathan1999optimizing]. The anomalous diffusion of chemotactic particles had been reported in the presence of cell division [@gelimson2015] or repellent chemical signals [@grima2005]. We have shown that within our model, the $\mu_2$ coupling results in a super-diffusive behavior for a wide range of parameters. In fact, in the absence of the Keller–Segel interaction this superdiffusion emerges in all dimensions (below $d_c$). While hyperuniformity provides a natural representation of dynamic size regulation in colonies or tissues, which is an expected property at homeostasis, a state with giant number fluctuations and considerably stronger superdiffusion can represent an invasive metastatic phase. Our analysis reveals how a transition between these two states can be triggered using the noise terms. It is worth mentioning that the analysis here has been carried out for the system at its critical state, when the (average) number of particles in the system is constant. In this case, the characteristic length and time scales of the Langevin dynamics, arising from the chemotactic interactions or possible decay of the particles (Appendix \[app\_gradexpansion\]), diverge and the system becomes scale-invariant. This indicates the breakdown of the mean field approaches of Keller-Segel model and its variants, as they are derived by neglecting the statistical correlations between the particles [@newmangrima]. In contrast, the Dean–Kawasaki formalism encodes all such correlations into Eq. , which can then be treated using RG techniques to investigate the properties of the system at the macroscopic level. Note that the relevance of the new nonlinear effect, which is initially introduced via scaling considerations, is further corroborated in the RG calculations. The scaling behavior described here is particularly relevant for polarizable particles and can be searched for by measuring the scaling exponents, as outlined by Eqs. and and the discussion thereof. The interplay between chemical signals and generic growth processes of the particles, which in many cases are asymmetrical processes accompanied by polarization of the cells [@jan2000; @neumuller2009], adds another level of complexity to the collective properties of growing colonies [@kruse2005; @toner2012a; @gelimson2015; @malmi-kakkada2018] which we plan to investigate in future works. The growth of individuals is limited by, for example, the availability of nutrients in an environment [@wang2017] and cell homeostatic regulations [@tzur2009]. The complex internal machinery determining the size and dynamic structure of a colony, however, remains largely unknown. Such self-regulations are crucial in the development of different organs in the body and show signs of failure when, for instance, tumor cells acquire increased proliferation by breaking away from them [@preston-martin1990; @hanahan2011]. Input from powerful physical considerations such as scaling properties and symmetry transformations are crucial for picking out the most relevant interactions from an otherwise large number of possibilities. An understanding of different phases of the system in the presence of both chemical signals and growth processes will help us shed light on these regulatory mechanisms. R.B.A.Z. and C.D. thank B. Delamotte for arousing their curiosity about this model, and for stimulating discussions. S.M. acknowledges the support of the Clarendon Fund Scholarships. Derivation of the Langevin equation for chemotactic particles {#app_gradexpansion} ============================================================= In this section, we take a closer look into the mathematical representation of single-particle chemotaxis and show how the two nonlinear terms on the r.h.s. of the Langevin dynamics for the particle density in Eq. are obtained from a gradient expansion approach and by utilizing dimensional analysis and symmetry considerations. Note that we ignore the dependence of the chemotactic response of the particles on the absolute concentration of chemicals in the environment. Gradient expansion and the new chemotactic interaction ------------------------------------------------------ The velocity, $\dot{\mathbf{r}}(t)$, of a chemotactic particle located at a position $\mathbf{r}(t)$ within the medium with a concentration field $\phi(\mathbf{x},t)$ in principle depends on all the vectorial expressions constructed from combinations of $\phi$ and gradient operators, namely $$\begin{aligned} \label{eq:gradexpansion} \dot{\mathbf{r}}(t) &= A_1 \nabla \phi + A_2 \nabla (\nabla^2 \phi) + \cdots \nonumber\\ &\,\, + B_1 \nabla (\nabla\phi)^2 + B_2 \nabla (\nabla^2 \phi^2) + \cdots \, \\ &= \mathbf{\nabla} f(\phi, \phi^2, \ldots, \nabla^2\phi, \nabla^2\phi^2,\ldots) \, ,\end{aligned}$$ where $f$ is a scalar function of its arguments. In the following sections, we will show that among the various terms arising in this gradient expansion, only those proportional to $A_1$ and $B_1$ are relevant to describe the critical behavior of chemotactic systems. Here, we discuss the physical meaning of these terms in more details. The first term, $A_1 \nabla \phi$, is the well-known form of chemotaxis first studied by Keller and Segel [@keller1970; @keller1971] and represents the motion of chemotactic particles along the gradient of the chemical field. In Sec. \[sec:bio\], we have argued that the term $B_1 \nabla(\nabla\phi)^2$ in Eq. represents this polarization mode of chemotaxis. The resulting equation of motion of the cell $$\begin{aligned} \dot{\mathbf{r}}(t) &= M_1 \nabla \phi + M_2 \nabla(\nabla\phi)^2 + \bm{\xi}(t) \, , \label{eq:microscopic}\end{aligned}$$ where the Gaussian white noise $\bm\xi(t)$ accounts for the diffusive behavior of the cell at the microscopic level. The corresponding Dean–Kawasaki description [@dean1996; @kawasaki1994] for the density function of a system of such chemotactic cells can then be cast into the Langevin form of Eq. displayed in the main text. In particular, the $M_1$ term is the familiar Keller–Segel gradient-sensing term [@chavanis2010] that leads to the $\mu_1 \nabla(\rho\nabla\phi)$ term in Eq. , while the second term proportional to $M_2$ represents the polarization mode leading to the $\mu_2 \nabla^2 (\nabla\phi)^2$ term. Power counting and upper critical dimension ------------------------------------------- From a renormalization group (RG) perspective, higher-order terms in the expansion may still be important in determining the macroscopic properties of the Langevin equation and, therefore, one must examine their scaling dimension in order to be able to include all the relevant terms in the theoretical formulation. The scaling dimensions of the fields $\rho$ and $\phi$ depend on the nature of the noise, which we assume here to be nonconserved with $\mathcal{D}_0\neq0$ (similar analysis can be carried out for the conserved noise). Following a rescaling of lengths given by $\mathbf x'=\mathbf x/b$, we assume a rescaling of time and fields according to $t'= t/ b^z$ and $$\begin{aligned} \rho' (\mathbf{x},t)= b^{-\chi} \rho (\mathbf{x},t) \, , \quad \phi' (\mathbf{x},t) = b^{-\psi} \phi(\mathbf{x},t) \,.\end{aligned}$$ The exponents $z$ and $\chi$ are chosen such that the form of the equation remains unchanged after the rescaling procedure. In the absence of nonlinearities ($\mu_{1,2}=0$), the model is Gaussian and the equation is made scale-invariant if the fields have the following scaling dimensions: $$\begin{aligned} \chi^{\rm non}_{\!_{\rm MF}}=\frac{2-d}{2} \, , \quad \psi^{\rm non}_{\!_{\rm MF}}=\frac{6-d}{2} \, , \label{eq_fieldDimensions}\end{aligned}$$ where the subscript ${\rm MF}$ indicates that we are referring to the mean-field, engineering dimensions. Based on the engineering dimensions, we can now systematically determine the relevance of any interaction terms that may be generated by the RG flow of the Langevin equation. The most general form of an interaction term reads, symbolically, $$g_{npq}~\rho^n ~ (\nabla\phi)^p ~ \nabla^q \, , \label{generalCoupling}$$ where we assume $n,p\geq0$ (and $n+p>0$, otherwise no field is involved in the coupling). Moreover, since we consider conserved dynamics, interaction terms come as the divergence of a vector field and, hence, we must have $q\geq1$. In addition, $\phi$ can only appear together with a gradient to enforce the symmetry $\phi \to \phi + \mathrm{const.}$ and, finally, $p+q$ must be even to have a scalar interaction term. According to Eq. and using the Langevin equation , the coupling constant $g_{npq}$ turns out to have the dimension $$\begin{aligned} [g_{npq}]_0=1+n+2p-q-\frac{d}{2}(n+p-1) \, , \label{eq_generalDimension}\end{aligned}$$ and the corresponding interaction is relevant whenever this quantity is positive. The upper critical dimension $d^{\rm non}_c$ is therefore obtained by maximizing the following expression $$d = 2\left(\frac{n+2p+1-q}{n+p-1}\right) \, , \label{eq_maxd}$$ with $n+p\neq1$. One obtains the upper critical dimension $d^{\rm non}_c=6$ in the case of a nonconserved noise. A similar procedure yields $d^{\rm con}_c=4$ in the case of a conserved noise ($\mathcal{D}_0=0$). Galilean symmetry and final form of the chemotaxis equation {#sec_galilean} ----------------------------------------------------------- [17.5cm]{}[p[0.4cm]{} p[0.4cm]{} p[0.4cm]{} p[3.7cm]{} p[1.6cm]{} p[10cm]{}]{} $n$ & $p$ & $q$ & Form of the coupling & Galilean invariant & Comments\ & 1 & 1 & $\nabla^2\phi $ & yes & Equivalent to the linear term $\rho$.\ & 1 & 3 & $\nabla^3\nabla\phi $ & yes & Equivalent to the diffusion term $\nabla^2\rho$ (marginal in all dimensions)\ & 2 & 2 & $\nabla^2 (\nabla\phi)^2 $ & yes & New chemotactic term.\ & 3 & 1 & $\nabla (\nabla\phi)^3$, $\nabla \left(\nabla\phi (\nabla\phi)^2\right)$ & no & Higher-order single-particle chemotactic terms.\ & 0 & 2 & $\nabla^2\rho $ & yes & Diffusion term (marginal in all dimensions).\ & 1 & 1 & $\nabla(\rho \nabla\phi) $ & yes & Chemotactic term.\ The derivation of the Langevin equation is finally concluded by identifying all the terms that are marginal or relevant at the upper critical dimension (see Table \[table\_relevantd6\]). Some of those terms, however, are not allowed in the effective equation since they do not respect the symmetries of the microscopic chemotactic dynamics. Consider the Galilean transformation defined by $$\begin{aligned} \phi'(\mathbf{x},t) &= \phi\big(\mathbf{x}+a t\mathbf{v},t\big)-\mathbf{v}\cdot\mathbf{x},\\ \rho'(\mathbf{x},t) &= \rho \big(\mathbf{x}+a t\mathbf{v},t\big) \, , \label{eq_galileanTransformation}\end{aligned}$$ where $\mathbf{v}$ is an arbitrary $d$-dimensional vector and $a$ a parameter to be specified. The diffusion term $D\nabla^2\rho$ is clearly invariant under this transformation. The invariance of the chemotactic terms $\mu_1 \nabla(\rho \nabla\phi)$ can be established by considering it together with $\partial_t \rho$ in the Langevin equation with the choice $a=\mu_1$. Similarly, the chemotactic term $\mu_2 \nabla^2 (\nabla\phi)^2$ is also Galilean invariant when considered together with $\partial_t \rho$ and for $a=-2\mu_2$. The last marginal term appearing in Table \[table\_relevantd6\], i.e., $\nabla (\nabla\phi)^3$, however, cannot be made Galilean invariant and therefore this is a term that will not be generated under RG flows if not already present in the original equation (breaking the Galilean symmetry). The final form of the Langevin equation describing a chemotactic system with a fixed number of particles close to criticality therefore reads: $$\begin{aligned} \label{eq:langevin_supp} ( \partial_t - D \nabla^2 + \sigma) \, \rho (\mathbf{x},t) = &- \mu_1{\nabla} \cdot (\rho\nabla\phi) \\ &- \mu_2\nabla^2(\nabla\phi)^2 + \zeta(\mathbf{x},t)\,, \nonumber\end{aligned}$$ where the correlations of the noise are $\langle\zeta(\mathbf{x},t)\zeta(\mathbf{x}',t')\rangle=2 \mathcal{D}_0~\delta^d(\mathbf{x}-\mathbf{x}')\delta(t-t')$. This Langevin equation is Galilean symmetric and thus invariant under the transformation with $a=\mu_1-2\mu_2$. Notice that the same reasoning can be applied to the conserved noise case, yielding the very same Langevin equations but where the noise correlations read $\langle\zeta(\mathbf{x},t)\zeta(\mathbf{x}',t')\rangle=-2 \mathcal{D}_2\nabla^2~\delta^d(\mathbf{x}-\mathbf{x}')\delta(t-t')$. Note finally that this symmetry remains valid as long as the noise is delta-correlated in time [@medina1989]. Note that the linear mass term $\sigma \rho$ appears in the derivation of Eq. from Eq. using Dean–Kawasaki formalism [@dean1996], and also arises if linear birth and death processes are considered. The system becomes scale free when $\sigma$ is fine-tuned to its critical value $\sigma=0$. Note that this mass term is not renormalized and there is thus no mass-shift to be performed (see below). Detailed balance {#app_detailedbalance} ================ In this section we show that the new polarization-induced chemotactic term breaks the detailed-balance in the dynamics and it is therefore a purely nonequilibrium interaction term. In a conserved system for which the interactions are derivable from an underlying free energy $\mathcal{F}$, one can in general use model B dynamics [@hohenberg1977] and write $$\begin{aligned} \label{eq:modelB} \partial_t \rho(\mathbf{x},t) = \nabla\!\cdot\!\left(\rho(\mathbf{x},t)\nabla\frac{\delta\mathcal{F}[\rho,\phi]}{\delta \rho(\mathbf{x},t)}\right),\end{aligned}$$ where $\rho(\mathbf{x},t)$ is the density of the particles and $\delta\mathcal{F}/\delta\rho$ is a chemical potential. The $\mu_1$ term in Eq. that represents the Keller–Segel chemotaxis can be written in this form using the functional $\mathcal{F}_{\!{\rm KS}} = \int \rho(\mathbf{x},t) \phi(\mathbf{x},t)$ [@chavanis2010]. Note that in $d=1$, $\nabla\cdot(\rho\nabla\phi)$ and $\nabla^2 (\nabla\phi)^2$ are linearly dependent and thus the detailed balance is incidentally satisfied. Assuming a free energy $\mathcal{F}_2$ whose functional derivative gives the $\mu_2$ interaction term, i.e., $\delta \mathcal{F}_2 / \delta\rho(\mathbf{x})= \left(\nabla\phi(\mathbf{x})\right)^2$, we show that its second derivatives do not commute which amounts to the breakdown of the symmetric form imposed by Onsager relations. By taking another derivative we arrive at $$\begin{aligned} \frac{\delta^2\mathcal{F}_2}{\delta\rho(\mathbf{x}') \delta\rho(\mathbf{x})} = 2 \left(\nabla\phi(\mathbf{x})\right)\cdot \frac{\delta \nabla\phi(\mathbf{x})}{\delta\rho(\mathbf{x}')}.\end{aligned}$$ To proceed further, we use the fact that the particles are sources of chemicals and hence one can write the $\phi$ field at each time as an integral over the $\rho$ field by $$\begin{aligned} \phi(\mathbf{x}) = \int d\mathbf{y} K(\mathbf{x}-\mathbf{y}) \rho(\mathbf{y})\end{aligned}$$ where the Coulomb kernel $K$ satisfies the condition $\nabla_x^2 K(\mathbf{x}-\mathbf{y})=-\delta(\mathbf{x}-\mathbf{y})$ imposed by the Poisson equation. Note that in analogy to electrostatics, $K(\mathbf{x}-\mathbf{y})$ is the potential at point $\mathbf{x}$ due to a unit source at point $\mathbf{y}$ and $-\nabla_x K(\mathbf{x}-\mathbf{y})$ is the corresponding chemotactic force at position $\mathbf{x}$ which is parallel to $\mathbf{x}-\mathbf{y}$. Now we can write $$\begin{aligned} \frac{\delta^2 \mathcal{F}_2}{\delta\rho(\mathbf{x}') \delta\rho(\mathbf{x})} &= 2 \left(\nabla\phi(\mathbf{x})\right)\cdot\nabla K(\mathbf{x}-\mathbf{x}'),\\ \frac{\delta^2 \mathcal{F}_2}{\delta\rho(\mathbf{x}) \delta\rho(\mathbf{x}')} &= 2 \left(\nabla'\phi(\mathbf{x}')\right)\cdot\nabla' K(\mathbf{x}'-\mathbf{x}).\end{aligned}$$ Eventually, using the fact that $ \nabla' K(\mathbf{x}'-\mathbf{x}) = - \nabla K(\mathbf{x}-\mathbf{x}'), $ we get $$\begin{aligned} &\frac{\delta^2\mathcal{F}_2}{\delta\rho(\mathbf{x}') \delta \rho(\mathbf{x})} - \frac{\delta^2\mathcal{F}_2}{\delta \rho(\mathbf{x}) \rho(\mathbf{x}')} = \nonumber\\ &\quad 2 \nabla K(\mathbf{x}-\mathbf{x}') \cdot \int d\mathbf{y} \rho(\mathbf{y}) \big[\nabla K(\mathbf{x}-\mathbf{y}) - \nabla' K(\mathbf{x}'-\mathbf{y})\big]. \end{aligned}$$ For a general source density $\rho(\mathbf{y})$, the r.h.s. is zero only if $\nabla K(\mathbf{x}-\mathbf{x}') \cdot \big[\nabla K(\mathbf{x}-\mathbf{y}) - \nabla' K(\mathbf{x}'-\mathbf{y})\big]$ is identically zero. When $\mathbf{x}\neq\mathbf{x}'$, this implies $\nabla K(\mathbf{x}-\mathbf{x}') \perp \big[\nabla K(\mathbf{x}-\mathbf{y}) - \nabla' K(\mathbf{x}'-\mathbf{y})\big]$ which is not satisfied for arbitrary position vectors $\mathbf{x}, \mathbf{x}'$ and $\mathbf{y}$. We therefore conclude that in contrast to the $\mu_1$ term, the $\mu_2$ term cannot be derived from an underlying functional form needed to establish the detailed balance, and hence is a nonequilibrium interaction. Flow diagrams in various dimensions {#app_rgflowsdim} =================================== In this section, we look into the structure of the renormalization group flows in various spatial dimensions $d$. Note that despite the scaling exponents that are obtained exactly, the RG flow equations and the corresponding analysis are only valid to one-loop order and a higher order calculation will be required to establish a more conclusive picture of different phases of the system in the parameter space. ![image](FPtable.pdf){width="99.00000%"} The RG flows for the effective couplings $U_{1,2}$ in the presence of conserved and nonconserved noise are given by Eq. and Eq. , respectively. Note that in both cases the r.h.s is the same for both $\partial_\ell U_1$ and $\partial_\ell U_2$, indicating that the flows occur along the rays with a fixed ratio of $U_2/U_1$. The fixed points are obtained by setting $\partial_\ell U_{1,2}=0$ which, besides the Gaussian fixed point $U_1=U_2=0$, gives a quadratic equation $A U_1^2 + B U_{1} U_{2} + C U_2^2 + D = 0$ where the coefficients $A, B, C, D$ are defined below Eqs. and in each case. This equation defines conic sections in the $U_1-U_2$ plane whose shape can be determined based on the sign of its discriminant $\Delta$ defined as $$\begin{aligned} \Delta = B^2 - 4 A C.\end{aligned}$$ Moreover, the stability of each curve can be analyzed by considering small displacement from a fixed-point (the displacement should be considered along the rays so that for an attractive fixed point the final state is the same as the initial one). In table \[fig:rgflowvardim\], we summarize the results of these analysis. Note that for conserved noise, $d_c=d_c^{\rm con}=4$ and for nonconserved noise $d_c=d_c^{\rm non}=6$. Renormalization group procedure and exact scaling exponents {#app_rgprocedure} =========================================================== Renormalization group --------------------- As discussed in the main text, we implement a perturbative momentum-shell renormalization group (RG) procedure to study the critical behavior of the chemotactic particles [@tauber2014; @medina1989]. The perturbative expansion is performed at first order in $\varepsilon=d_c-d$, where $d_c$ is the upper critical dimension of the model: $d^{\rm con}_c=4$ for conserved noise, $d^{\rm non}_c=6$ for nonconserved noise. Using the Poisson equation , the field $\phi$ can be systematically eliminated and the RG treatment of the Langevin equation is done considering only the concentration field $\rho$. This allows us to define the following diagrams for the bare propagator $G_0$, the bare (chemotactic) interaction vertex $\Gamma_0$ and the bare dynamic correlation function $C_0$ defined in Eqs. to $$\begin{aligned} G_0(\hat k) &= \begin{tikzpicture} [baseline={([yshift=-.5ex]current bounding box.center)},vertex/.style={anchor=base, circle,fill=black!25,minimum size=18pt,inner sep=2pt}] \draw [line width=0.3mm, -{Latex[length=2.5mm]}] (-0.7,0) -- (0,0); \draw [line width=0.3mm] (-0.2,0) -- (0.5,0); \node at (-0.15,0.4) {$\hat{k}$}; \node at (0,-0.6) {$ $}; \end{tikzpicture} \, , \\ \Gamma_0(\mathbf{k},\mathbf{q}) &= \begin{tikzpicture} [baseline={([yshift=-.5ex]current bounding box.center)},vertex/.style={anchor=base, circle,fill=black!25,minimum size=18pt,inner sep=2pt}] \draw [line width=0.3mm] (-1,0) -- (0,0); \draw [line width=0.3mm, -{Latex[length=2.5mm]}] (-1,0) -- (-0.4,0); \draw [line width=0.3mm] (0,0) -- (45:1.0cm); \node at (-0.6,.4) {$\mathbf{k}$}; \draw [line width=0.3mm, -{Latex[length=2.5mm]}] (0,0) -- (45:.75cm); \node at (.8,1) {$\mathbf{q}$}; \draw [line width=0.3mm] (0,0) -- (-45:1.0cm); \draw [line width=0.3mm, -{Latex[length=2.5mm]}] (0,0) -- (-45:.75cm); \node at (.8,-1) {$\mathbf{k}-\mathbf{q}$}; \end{tikzpicture} \, ,\label{eq_bareVertex} \\ C_0 (\hat k) &= \begin{tikzpicture} [baseline={([yshift=-.5ex]current bounding box.center)},vertex/.style={anchor=base, circle,fill=black!25,minimum size=18pt,inner sep=2pt}] \draw [line width=0.3mm] (-.5,0) -- (0,0); \draw [line width=0.3mm, -{Latex[length=2.5mm]}] (-1,0) -- (-.4,0); \draw [line width=0.3mm] (0,0) -- (1,0); \draw [line width=0.3mm, {Latex[length=2.5mm]}-] (0.4,0) -- (1,0); \node[circle,draw=black, fill=white, inner sep=0pt,minimum size=5pt] (b) at (0,0) {}; \node at (-0.6,0.4) {$\hat{k}$}; \node at (.6,0.4) {$-\hat{k}$}; \node at (-1,-0.6) {$ $}; \node at (.9,-0.6) {$ $}; \end{tikzpicture} \, ,\end{aligned}$$ where we have used the following Fourier transform convention: $$\begin{aligned} f(\mathbf x,t)=\int_{\hat k} f(\hat k) \, e^{i \mathbf k\cdot \mathbf x - i \omega t} \, .\end{aligned}$$ We use the same name for a function and its Fourier transform and we introduced the shorthand notation $\int_{\hat{k}}=1/(2\pi)^{d+1}\int \mathrm{d}^d \mathbf k \int \mathrm{d}\omega$ and $\hat{k}=(\mathbf k,\omega)$. Galilean symmetry and Ward identity ----------------------------------- In Appendix \[sec\_galilean\], the Galilean symmetry of the Langevin equation was discussed. We noted that the dynamics is invariant under the transformation with $a=\mu_1-2\mu_2$, and since the chemotactic couplings $\mu_1$ and $\mu_2$ appear as parameters in this transformation, we concluded that they should remain unchanged under the RG flow to preserve the symmetry. This could also be seen as the corollary of the Galilean invariance of the combinations $\partial_t \rho + \mu_1\nabla\!\cdot\!(\rho\nabla\phi)$ and $\partial_t\rho + \mu_2\!\nabla^2(\nabla\phi)^2$ [@canet2011a]. This symmetry yields the exact exponent identity given by Eq. , reminiscent of a similar exponent identity in the Kardar–Parisi–Zhang equation [@kardar1986]. The derivation of this exponent identity can be made more formal by looking at the Ward identity between the vertex function and the propagator that results from the Galilean symmetry [@canet2011a]. It reads $$\begin{aligned} i\,\left(\mu_1-2\mu_2\right) \, \mathbf{q} \, \partial_{\omega} G(\hat{q}) = \partial_{\mathbf{k}} \, \Gamma (\hat{k} \,;\hat{q})\big\rvert_{\hat{k}=0}, \end{aligned}$$ where $G(\hat{q})$ and $\Gamma(\hat{k},\hat{q})$ are defined similarly to their bare value (Eq. ) by replacing the parameters $\mu_{1,2}$ and $D$ by their renormalized values. Non-renormalization of the noise term ------------------------------------- The renormalization of the noise term takes the diagrammatic form given in Fig. \[fig:diagrams\]. One notices that because all the diagrams contributing to the renormalization of the noise strength must include at least two bare vertices with external momenta $\mathbf k$ and $-\mathbf k$, respectively. As a bare vertex scales as $\Gamma_0(\mathbf k,\mathbf q)\sim k^2$ for $k\to0$ (see Eq. ), each diagram will scale at least as $k^4$ when $k\to0$. Since the noise term scale as $k^0$ in the nonconserved case, and as $k^2$ in the conserved case, we conclude that the noise term is not renormalized in either case. This non-renormalization of the noise term provides a second exponent identity, which, combined with the exponent identity Eq. , provides us with the exact exponents given in Eq. for the conserving noise case. Similarly, the exact exponents in the nonconserved noise case are given by Eq. . Renormalization of the propagator --------------------------------- We have seen that the non-renormalization of the vertex and of the noise strength are valid at all order in perturbation theory. A similar argument to the one given for the non-renormalization of the noise tells us that the mass is also not renormalized. Conversely, the diffusive part of the propagator does get renormalized and we compute the perturbative contribution at one-loop order, following the diagrams given in Fig. \[fig:diagrams\]. Renormalization group equations ------------------------------- Using standard RG procedure and using the diagrammatic representation reported in Fig. \[fig:diagrams\], we have derived the RG equations of the chemotactic model in the case of a conserved noise ($\mathcal{D}_0=0$), which are shown in the main text, see Eqs. to . The case of nonconserved noise ($\mathcal{D}_0\neq0$) can be treated similarly and the corresponding RG equations read: \[eq:nnoiseflows\] $$\begin{aligned} \partial_{\ell} \mu_{1,2} &= \left[ z+\chi \right]\mu_{1,2}, \label{eq:nnoiseflowmu12}\\[1.5mm] \partial_{\ell} D \,\, &= \Big[ z-2 - \big( b_{11} U_1^2 + b_{12} U_1 U_2 + b_{22} U_2^2\big)\Big] D,\label{eq:nnoiseflowD} \\[1.5mm] \partial_\ell \mathcal{D}_0 &= \left[-d+z-2\chi\right] \mathcal{D}_0, \label{eq:nnoiseflowD0}\end{aligned}$$ where $U_{1,2}^2 = \mu_{1,2}^2 \, \mathcal{D}_0 \, K_d \Lambda^{d-6} / D^3$ and with coefficients $b_{11} = 3/4 - 1/d - 3/[d(d+2)]$, $b_{12} = 2 + 6/d - 9/(d+2)$, and $b_{22} = 1-6/d$. [69]{}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 [**, ()](https://doi.org/10.1126/science.153.3737.708) [, ()](https://doi.org/10.1080/000187300405228) [, ()](https://doi.org/10.1063/PT.3.1884) [, ()](https://doi.org/10.1016/S0092-8674(00)80562-0) [, ()](https://doi.org/10.1038/225420a0) [, ()](https://doi.org/10.1038/nrm2720) [, ()](https://doi.org/10.1126/science.1174294) [, ()](https://doi.org/10.1159/000276562) [, ()](https://doi.org/10.1016/j.cell.2011.02.013) [, ()](https://doi.org/10.1016/S1470-2045(07)70140-7) [, ()](https://doi.org/10.1016/S1534-5807(02)00292-7) [, ()](https://doi.org/10.1038/nrc3078) [, ()](https://doi.org/10.1016/j.ceb.2007.11.011) [, ()](https://doi.org/10.1038/16483) [, ()](https://doi.org/10.1038/43199) [, ()](https://doi.org/10.1038/nature02406) [, ()](https://doi.org/10.1038/nrm1524) [, ()](https://doi.org/10.1073/pnas.0807569105) [, ()](https://doi.org/10.1073/pnas.0705463105) [, ()](https://doi.org/10.1146/annurev-biophys-083012-130358) [, ()](https://doi.org/10.1073/pnas.1216629110) [, ()](https://doi.org/10.1021/acs.nanolett.8b00717) [, ()](https://doi.org/10.1039/C7CS00087A) [, ()](https://doi.org/10.1021/acs.accounts.8b00259) [, ()](https://doi.org/10.1016/0022-5193(71)90050-6) [, ()](https://doi.org/10.1209/epl/i2003-10237-5) [, ()](https://doi.org/10.1103/PhysRevLett.95.128103) [, ()](https://doi.org/10.1103/PhysRevLett.102.188305) [, ()](https://doi.org/10.1103/PhysRevE.80.031122) [, ()](https://doi.org/10.1103/PhysRevE.85.051901) [, ()](https://doi.org/10.1016/S0006-3495(98)77880-4) [, ()](https://doi.org/10.1103/PhysRevE.69.016116) [, ()](https://doi.org/10.1103/PhysRevLett.108.038303) [, ()](https://doi.org/10.1103/PhysRevLett.112.068302) [, ()](https://doi.org/10.1103/PhysRevE.89.062316) [, ()](https://doi.org/10.1103/PhysRevLett.117.038101) [, ()](https://doi.org/10.1103/PhysRevLett.117.178102) [, ()](https://doi.org/10.1016/j.cnsns.2008.09.002) [, ()](https://doi.org/10.1140/epjb/e2008-00142-9) [, ()](https://doi.org/10.1103/RevModPhys.85.1143) [, ()](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.75.1226) [, ()](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.75.4326) [, ()](https://doi.org/10.1103/PhysRevE.86.031918) [, ()](https://doi.org/10.1103/PhysRevLett.93.175702) [, ()](https://doi.org/10.1103/PhysRevLett.114.028101) [, ](https://doi.org/10.1088/1742-5468/aaf321) [, ()](https://doi.org/10.1016/0022-5193(70)90092-5) [, ()](https://doi.org/10.1063/1.5081717) @noop (), [, ()](https://doi.org/10.1126/science.287.5455.1037) [, ()](http://iopscience.iop.org/article/10.1088/0305-4470/29/24/001/meta) [, ()](https://doi.org/10.1016/0378-4371(94)90533-9) [, ()](https://doi.org/10.1103/PhysRevA.39.3053) @noop []{} (, , ) [**, ()](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.56.889) @noop []{} (, ) [**, ()](https://www.nature.com/articles/nature11098) [, ()](https://www.nature.com/articles/44831) [, ()](https://doi.org/10.1103/PhysRevE.70.051916) [, ()](https://www.cell.com/fulltext/S0092-8674(00)80695-9) [, ()](http://genesdev.cshlp.org/content/23/23/2675.short) [, ()](https://doi.org/10.1140/epje/e2005-00002-5) [, ()](https://doi.org/10.1103/PhysRevLett.108.088102) [, ()](https://doi.org/10.1103/PhysRevX.8.021025) [, ()](https://doi.org/10.1088/1367-2630/aa983f) [, ()](https://cancerres.aacrjournals.org/content/50/23/7415.short) [, ()](https://doi.org/10.1103/RevModPhys.49.435) [, ()](https://doi.org/10.1103/PhysRevE.84.061128) [**, ()](https://doi.org/10.1103/PhysRevE.50.1024) [^1]: These three authors contributed equally. [^2]: These three authors contributed equally. [^3]: These three authors contributed equally.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We proposes an alternative model of duality symmetry, based on the previously obtained divergence theory, including an scalar field, an internal vector and a metric signature. At some small scale an effective scalar field equation has appeared whose potential acts like a Higgs one, where the metric signature plays the role of an order parameter. Non-vanishing Vacuum condensation of this Higgs field occurs once a signature change from Euclidean to Lorentzian is formed. The mass scale of Higgs field excitations around this vacuum may contribute, in the Lorentzian sector, to the cosmological constant, in agreement with observations.' author: - \| F. Darabi[^1]\ [Department of Physics, Azarbaijan University of Tarbiat Moallem , Tabriz, 53714-161 Iran.]{} title: \| Signature change, vacuum condensation\ and\ cosmological constant --- -0.75in Introduction ============ The initial idea of signature change was due to Hartle, Hawking and Sakharov [@HHS]. This idea would make it possible to have both Euclidean and Lorentzian metrics in the path integral approach to quantum gravity. However, it was later shown that signature change may happen, as well, in classical general relativity [@CSC]. Most of the works regarding the signature change dealt with situations where the signature changing metric is defined [apriori]{} on the manifold and one looks for the effects of the assumed signature change on the Einstein equations or propagation of particles in such a manifold. However, there are some other viewpoints in which the signature generation of the space-time is studied and considered to be a dynamical phenomenon at quantum gravity regime. This is because a quantum formulation of gravitation should accommodate geometries with degenerate metrics and nontrivial topologies. However, the quantum nature of gravity is quite different from the one of quantum field theory. Therefore, in these viewpoints the question is how best to model such a dynamical effect in the language of quantum field theory. Percacci has offered a formalism in which one can dissociate the conventional geometrical interrelations between the metric tensor components and a field of co-frames, and work in close analogy with the Higgs model in non-Abelian gauge theories [@Pre]. Classical geometry is then regarded as an interpretation of certain expectation values which minimize an effective action. Greensite, on the other hand, has developed this idea further by assuming that a particular pattern of signature arises dynamically as a result of dynamical phase field which interpolates between signatures [@Gre]. He has argued that, at least for the free scalar field theory interacting with such a dynamical phase field, the Lorentzian signature of a four-dimensional manifold can be predicted as a ground state expectation value. Odintsov [et al]{} have obtained the effective potential for the dynamical phase field ( dynamical signature ) induced by the quantum effects of massive fields on a topologically non-trivial $D$-dimensional background, and shown that the Lorentzian signature is preferred in both $D=6$ and $D=4$ [@Odin]. From a different viewpoint, it was shown that the signature change phenomena can be studied in a variant of the [divergence theory]{}, from the viewpoint of a [time asymmetric law]{} in vacuum which breaks an specific duality symmetry [@Sal]. It was based on the realization that a duality breaking must be connected with the emergence of a preferred arrow of time in vacuum, and this provides an indication that the time asymmetric law may act to produce the condensation of vacuum. In the present paper, following the above idea, we propose a new duality symmetry different from the one introduced in [@Sal]. We remark that at small distances, an effective Higgs type potential may arise and show that the duality breaking manifests directly as the condensation of vacuum due to a signature change from Euclidean to Lorentzian. This model as an alternative to [@Sal] has no specific advantage, but may deserve to be studied, as well, to show the potential of divergence theory to accommodate two different models of duality symmetry. Both models interrelate the signature change problem to the vacuum condensation of a scalar field, however, in two different ways of duality breaking. Unlike [@Sal], here, we introduce a duality breaking as a [spontaneous symmetry breaking]{} and add a new element as the cosmological constant originating from field excitations around the vacuum and show that the cosmological constant does not essentially change due to the signature change. The model ========= It is well known that an exact Lorentz invariant vacuum of the quantum field theory has, by definition, zero energy density. On the other hand, the vacuum in quantum field theory is related to the condensation of scalar fields leading to constant vacuum expectation values for these fields. The appearance of these non-zero values and the resulting mass scales contribute to the energy content of the vacuum to increase its value from zero. The resulting non-vanishing value accounts for a principle violation of Lorentz invariance. Such a violation of Lorentz invariance may have origin in the unification of gravity and quantum physics, and is expected to occur at ultrashort distance regime described by an absolute length scale, $l_0$. It is obvious that the existence of this universal length scale is in sharp contrast with the universal requirement of Lorentz invariance. In fact, in the Minkowski space-time, as a framework for Lorentz invariance, there is no absolute line of demarcation between large and small scales. However, if a positive definite measure of distance is defined, then the Lorentz invariance will be broken down. In so doing, we may follow the Blokhintsev point of view by associating a time-like vector $N_{\mu}$ (the so called internal vector) to the Minkowski space-time [@Blo]. In this way, one may distinguish between small and large scales by taking the positive definite interval $$R^2=(2N_{\mu}N_{\nu} -\eta_{\mu \nu})x^{\mu}x^{\nu}, \label{1}$$ where $\eta_{\mu \nu}=diag(1, -1, -1, -1)$ is the Minkowski metric and $N_{\mu}=(1, 0, 0, 0)$ is a time-like vector. Given such a positive definite metric as $$\bar{\eta}_{\mu \nu}=2N_{\mu}N_{\nu}-\eta_{\mu \nu}, \label{2}$$ it is then possible to determine the absolute size of a distance by comparing $R$ with the universal length $l_0$. One may then define the pair ($l_0, N_{\mu}$) whose physical interpretation is as follows [@Blo; @Sal]: $l_0$ is a characteristic size in the ultrashort distance regime which acts as a sort of universal length that determines a lower bound on any scale of length probed in a realistic measurement. $N_{\mu}$ serves as a four-velocity of a preferred inertial observer in vacuum by which the concept of the universal length $l_0$ makes sense. Therefore, the pair ($l_0, N_{\mu}$) arises as the physical feature of a Lorentz non-invariant vacuum. We now introduce a divergence theory developed in Ref.[@Sal]. This theory begins with a current $$J_{\mu}=-\frac{1}{2}\phi \stackrel{\leftrightarrow}{\partial_{\mu}}\phi^{-1}, \label{3}$$ for which we obtain $$\partial_{\mu}J^{\mu}=\phi^{-1}[\Box \phi -\phi^{-1}\partial_{\mu}\phi \partial^{\mu}\phi], \label{4}$$ and $$J_{\mu}J^{\mu}=\phi^{-2}\partial_{\mu}\phi \partial^{\mu}\phi, \label{5}$$ where $\phi$ is a real scalar field. Combining these relations leads to $$\Box \phi + \Gamma\{ \phi \} \phi=0, \label{6}$$ where $\Gamma\{ \phi \}$ is called the [dynamical mass term]{} $$\Gamma\{ \phi \}=-J_{\mu}J^{\mu}-\partial_{\mu}J^{\mu}. \label{7}$$ It is important to note that Eq.(\[6\]) is a formal consequence of the definition (\[3\]), in the form of an identity, and is not a dynamical equation for $\phi$. However, a large class of dynamical theories may be obtained in the form of a [divergence theory]{} by taking various currents $J^{\mu}$ in the dynamical mass term. For example, a simple divergence theory is developed by assuming $$\partial_{\mu}J^{\mu}=0, \label{8}$$ which leads, through the field redefinition $\sigma=\ln \phi$, to $$\Box \sigma=0. \label{9}$$ The dynamical mass term vanishes for this divergence theory and the field $\sigma$ becomes massless. One may allow for a dynamical coupling of $\phi$ with the internal vector $N_{\mu}$ by taking a more complicate dynamical mass term. This is done by a divergence equation of the type $$\partial_{\mu}J^{\mu}=N_{\mu}N_{\nu}J^{\mu}J^{\nu}+g{l_0}N_{\mu}N_{\nu}N_{\sigma}J^{\mu}J^{\nu}J^{\sigma},\label{10}$$ which leads to the field equation $$\Box \phi -(J_{\mu}J^{\mu}+N_{\mu}N_{\nu}J^{\mu}J^{\nu}+g{l_0}N_{\mu}N_{\nu}N_{\sigma}J^{\mu}J^{\nu}J^{\sigma}) \phi=0,\label{11}$$ where $g$ stands for the metric signature which is positive for Euclidean and negative for Lorentzian domains[^2]. Now, the basic point is that the source of the divergence is invariant under the following [dual]{} transformations[^3] $$\phi \rightarrow -\phi \:\:\:,\:\:\:N_{\mu}\rightarrow -N_{\mu} \:\:\:,\:\:\: g\rightarrow -g,\label{12}$$ where the latter transformation accounts for a signature change from Lorentzian to Euclidean $$\eta_{\mu \nu} \rightarrow \bar{\eta}_{\mu \nu}= 2N_{\mu}N_{\nu}-\eta_{\mu \nu},$$ or vice versa. We note that the identity (\[6\]) holds no matter what signature is used in its derivation. Therefore, the field equation (\[11\]) holds for both Lorentzian and Euclidean signatures. One may then assume a dynamical symmetry between the dual configurations $$(\phi, N_{\mu}, g)\Longleftrightarrow(-\phi, -N_{\mu}, -g),\label{13}$$ so that the field equation (\[11\]) makes no distinction between these dual configurations. In fact, both signatures are related by the dynamical symmetry of the field equation under the dual transformations (\[12\]). It is important to note that the emergence of this duality in the field equation (\[11\]) is considered as reflecting the essential feature of the broken phase of Lorentz invariance at small distance $\sim l_0$. Symmetry breaking ================= In physics there are many models in which there are some symmetries at the dynamical level leading to equivalent vacua in the theory. No distinction between these equivalent vacua is possible unless we require some symmetry breaking conditions to be imposed on the physically realizable configuration of vacuum. Usually, It is interesting to realize a symmetry breaking directly by means of boundary conditions. But other approaches are also possible and deserve to be studied, as well. For example, in [@Sal] the authors try to establish a duality symmetry between two configurations $$(\phi, N_{\mu}, \eta_{\mu \nu})\Longleftrightarrow(\phi^{-1}, -N_{\mu}, \bar{\eta}_{\mu \nu}),\label{14}$$ where $\eta_{\mu \nu}$ and $\bar{\eta}_{\mu \nu}$ stand for Lorentzian and Euclidean metrics, respectively. They introduce duality breaking by resorting to a [time asymmetric law]{} to be imposed on a specific source of the divergence. The time asymmetric law is then related to the vacuum condensation of the quantum scalar field. The present model, although follows the same purpose to take into account the signature change problem, but takes different source of divergence (\[10\]) and different equivalent vacuum configurations (\[13\]). Therefore, a different duality breaking condition is required to realize the physical vacuum configurations. One alternative aspect of duality breaking seems to be a preferred background value $\bar{\phi}$ and correspondingly a preferred background direction of $N_{\mu}$, as average values taken over large distances, which may be interrelated via a relation of the type [@Sal] $$\bar{J}_{\mu}=\frac{\lambda}{l_0}N_{\mu}, \label{15}$$ where $\lambda$ measures the ratio of the universal length and the spatial size of the universe as $$\lambda=\frac{l_0}{R}.\label{15'}$$ In this way, the duality breaking and correspondingly the preference of a signature is considered to be a cosmological effect. If, however, one is interested in realizing a preferred signature, it is not unreasonable to expect that such a duality breaking should have its origin at small distance regime rather than large cosmological distances. This is because, the metric is, in principle, defined to produce the most small distance appropriate for a realistic measurement process. Therefore, it is more natural to think about the origin and preference of Lorentzian over Euclidean metric as a small distance effect. In this regard, unlike Ref.[@Sal], we consider Eq.(\[15\]) not as a duality breaking but as a trick to effectively linearize the source of divergence equation (\[10\]) at ultrashort distance regime. This linearization is necessary in order to compute the effective form of the dynamical mass term in (\[11\]) by obtaining an approximate solution for $J_{\mu}$. Having this in mind, we can linearize the quadratic term in $J_{\mu}$ in the source to find the approximations $$\partial_{\mu}J^{\mu}=N_{\mu}N_{\nu}\bar{J}^{\mu}J^{\nu}+g{l_0}N_{\mu}N_{\nu}N_{\sigma}\bar{J}^{\mu}J^{\nu}J^{\sigma}, \label{16}$$ and $$\Box \phi -(\bar{J}_{\mu}J^{\mu}+N_{\mu}N_{\nu}\bar{J}^{\mu}J^{\nu}+g{l_0}N_{\mu}N_{\nu}N_{\sigma}\bar{J}^{\mu}J^{\nu}J^{\sigma}) \phi=0.\label{17}$$ We now truncate the nonlinear term in $J_{\mu}$ from the source of the divergence (\[16\]) and use Eq.(\[15\]) to obtain $$\partial_{\mu}J^{\mu}\simeq \frac{\lambda}{l_0}N_{\mu}J^{\mu}. \label{18}$$ Using Eq.(\[3\]), we can write the divergence equation (\[10\]) in terms of $\phi$ as $$\partial_{\mu}J^{\mu}\simeq \frac{\lambda}{l_0}N_{\mu}\frac{\partial^{\mu}\phi}{\phi}.\label{19}$$ If we linearize this equation by inserting the average background value of $\phi$ in the dominator we find an approximate solution for $J^{\mu}$ as $$J_{\mu}\simeq \frac{\lambda}{l_0}N_{\mu}\frac{\phi}{\bar{\phi}}. \label{20}$$ By using this solution for $J^{\mu}$ in (\[11\]) we arrive at the following field equation $$\Box \phi -(2\frac{\lambda^2}{l_0^2}\frac{\phi}{\bar{\phi}}+g \lambda \frac{\lambda^2}{l_0^2}\frac{\phi^2}{\bar{\phi}^2}) \phi=0.\label{21}$$ If we replace ${\phi}$ by $\bar{\phi}$ in the linear term, we can get a more effective form of this equation as $$\Box \phi -(2\frac{\lambda^2}{l_0^2}+g \lambda \frac{\lambda^2}{l_0^2}\frac{\phi^2}{\bar{\phi}^2}) \phi=0.\label{22}$$ This equation can be derived from the following effective potential $$V(\phi)=-\frac{\lambda^2}{l_0^2}\phi^2-\frac{1}{4}g \lambda \frac{\lambda^2}{l_0^2}\frac{\phi^4}{\bar{\phi}^2}. \label{23}$$ This potential, has one minimum at $\phi_0=0$ and two degenerate minima at $\phi_0=\pm \sqrt{\frac{2}{-g\lambda}}\bar{\phi}$ provided $g>0$ and $g<0$, respectively. This means, as far as the signature of the metric is Euclidean, no duality breaking happens to distinguish between dual configurations $\phi$ and $-\phi$. When, however, a signature change happens from Euclidean to Lorentzian the scalar field then condensates in one of the two vacua $\phi_0=\pm \sqrt{\frac{2}{-g\lambda}}\bar{\phi}$ and one configuration $(\phi, N_{\mu}, \eta_{\mu \nu})$ or $(-\phi, -N_{\mu}, \bar{\eta}_{\mu \nu})$ is singled out, permanently[^4]. Therefore, the condensation of $\phi$ corresponds to a signature change from Euclidean to Lorentzian. The ground state value $\phi_0$ of $\phi$ is obtained by minimizing $V(\phi)$ which leads to the condition $$\phi_0^2=\frac{2}{-g\lambda}\bar{\phi}^2.\label{24}$$ If the potential $V(\phi)$ is expanded around $\phi_0$ we obtain ( neglecting constant terms ) $$V(\phi)= 4\left(\frac{\lambda}{l_0}\right)^2(\phi-\phi_0)^2+{\it O }(\phi-\phi_0)^3, \label{25}$$ from which one infers that physical excitations of $\phi$ around $\phi_0$ have a preferred mass scale $m \sim \frac{\lambda}{l_0}$. According to the quantum field theory considerations, any massive particle excitation around a given scalar field vacuum can contribute to the total value of the cosmological constant. But this leads to the well-known cosmological constant problem. A solution of this problem is to reduce the issue of the cosmological constant to a picture in which the consistent contribution to the total value of that constant comes from a [preferred mass scale]{} of the vacuum. In this regard, the preferred mass scale $m \sim \frac{\lambda}{l_0}$ may contribute to the cosmological constant, in the Lorentzian sector, as $\Lambda \sim (\frac{\lambda}{l_0})^2 \sim \frac{1}{R^2}$, where use has been made of Eq.(\[15’\]). This is in agreement with the present observational bound for the cosmological constant as a remarkable consequence of the well-known empirical fact that the present universe has just the characteristic size $R \sim 10^{29}cm$. Note that according to Eq.(\[15’\]) this agreement with the observational bound for $\Lambda$ is obtained merely by the ratio $\lambda/l_0$, and is independent of the specific numerical values of $\lambda$ or $l_0$. The understanding of the relation of $l_0$ to the Planck length is an elusive task of quantum gravity. However, one must take a physically reasonable estimation for the universal length. For example, $l_0$ may act as the length scale that measures the size of the regime on which a significant nonlinear self coupling like the Higgs potential $V(\phi)$ can occur. This may constrain $l_0$ to be smaller than electroweak or even supersymmetry scale. It is worth noting that the vacuum energy as the minimum of potential vanishes for the Euclidean metric, but the field excitations around this minimum have almost the same mass scale as that of Lorentzian sector. Therefore, they may contribute to the cosmological constant in the Euclidean sector in the same manner as in the Lorentzian sector. In other words, the cosmological constant is almost invariant under the dual transformations (\[13\]). Conclusion ========== In this paper, we have studied a model of scalar field coupled to an internal vector, in the presence of a universal length, having a special duality symmetry in which the signature change appears as a natural symmetry. When this dynamical symmetry is combined with a duality breaking, a fixed background metric signature is singled out. Contrary to similar models already introduced [@Sal] in which the duality breaking arises due to large cosmological scale, we have argued that this duality breaking may be established, as well, at small distance regime by spontaneous symmetry breaking in an effective potential which has been derived in the study of a principle of duality invariance of the dynamical mass term of $\phi$ at a universal length in the small distance regime. It is shown that the duality breaking as the condensation of the scalar field occurs once a signature change from Euclidean to Lorentzian is happened at small distance regime. This signature change leads to the emergence of a preferred mass scale $\sim \frac{1}{R^2}$ of the vacuum which contributes to the cosmological constant, in the Lorentzian sector. The same mass scale is also appeared in the Euclidean sector which leads to the same cosmological constant. In the future, we hope to study and report more on the possible deep relation between the signature change and spontaneous symmetry breaking scenarios, since both of them are seriously assumed to be happened at early universe. Acknowledgment {#acknowledgment .unnumbered} ============== This work has been supported by the Research Department of Azarbaijan University of Tarbiat Moallem, Tabriz, Iran. [99]{} J. B. Hartle, S. W. Hawking, Phys. Rev. D.[28]{} (1983), 2960. T. Dereli, R. W. Tucker, Class. Quantum Grav.[10]{} (1993), 365; G. F. R. Ellis, A. Sumruk, D. Coule and C. Hellaby, Class. Quantum Grav.[9]{} (1992), 1535; S. A. Hayward, Class. Quantum Grav.[9]{} (1992), 1851; M. Kossowski, M. Kriele Proc. R. Soc. Lond. A.[446]{} (1995), 115; C. Hellaby, T. Dray, Phys. Rev. D. [49]{} (1994), 5096; F. Darabi, H. R. Sepangi, Class. Quantum Grav.[16]{} (1999), 1565; F. Darabi, Phys. Lett. A[259]{} (1999), 97; M. Mohseni, Phys. Lett. A[267]{} (2000), 240. R. Percacci, Nucl. Phys. B[353]{} (1991), 271. J. Greensite, Phys. Lett. B[300]{} (1993), 34; A. Carlini, J. Greensite, Phy. Rev. D[49]{} (1994), 866. S. D. Odintsov, A. Romeo and R. Tucker, Class. Quantum Grav.[11]{} (1994), 2951; E. Elizalde, S. D. Odintsov and A. Romeo, Class. Quantum Grav.[11]{} (1994), L61. H. Salehi, Phys. Lett. A[251]{} (1999), 95; Int. J. Theor. Phys [36]{} (1997), 2035. D. I. Blokhintsev, Physics Letters [12]{} (1964), 272. [^1]: e-mail: [email protected] [^2]: The excess of plus signs over minus signs is called the signature. [^3]: The source (\[10\]) and the dual transformations (\[12\]) are modifications of those introduced in Ref.[@Sal]. [^4]: This is what we meant by the duality breaking at small distance regime. In fact, the smallness of $\lambda$ leads the two degenerate vacua to be too far apart from each other, so the possibility of tunneling from one vacuum to the other one almost vanishes.	{ "pile_set_name": "ArXiv" }
--- abstract: \| The cosmic infrared background (CIB) radiation is the cosmic repository for energy release throughout the history of the universe. The spatial fluctuations of the CIB resulting from galaxy clustering are expected to be at least a few percent on scales of a degree, depending on the luminosity and clustering history of the early universe. Using the all-sky data from the COBE DIRBE instrument at wavelengths 1.25 - 100 we attempt to measure the CIB fluctuations. In the near-IR, foreground emission is dominated by small scale structure due to stars in the Galaxy. There we find a strong correlation between the amplitude of the fluctuations and Galactic latitude after removing bright foreground stars. Using data outside the Galactic plane ($\|b\| > 20^\deg$) and away from the center ($90\deg< l < 270\deg$) we extrapolate the amplitude of the fluctuations to cosec$\|b\|=0$. We find a positive intercept of $\delta F_{\rm rms} =15.5^{+3.7}_{-7.0}, 5.9^{+1.6}_{-3.7}, 2.4^{+0.5}_{-0.9}, 2.0^{+0.25}_{-0.5}$ 2sr at 1.25, 2.2, 3.5 and 4.9 $\mu$m respectively, where the errors are the range of 92% confidence limits. For color subtracted maps between band 1 and 2 we find the isotropic part of the fluctuations at $7.6^{+1.2}_{-2.4}$ 2sr . Based on detailed numerical and analytic models, this residual is not likely to originate from the Galaxy, our clipping algorithm, or instrumental noise. We demonstrate that the residuals from the fit used in the extrapolation are distributed isotropically and suggest that this extra variance may result from structure in the CIB. We also obtain a positive intercept from a linear combination of maps at 1.25 and 2.2 $\mu$m. For $2^\deg < \theta < 15^\deg$, a power-spectrum analysis yields firm upper limits of $(\theta/5^\deg) \times\delta F_{\rm rms} (\theta) < $ 6, 2.5, 0.8, 0.5 2sr at 1.25, 2.2, 3.5 and 4.9 $\mu$m respectively. From 10 - 100 , the dominant foregrounds are emission by dust in the Solar system and the Galaxy. There the upper limits on the CIB fluctuations are below 1 2sr and are lowest ($\leq$ 0.5 2sr ) at 25 . author: - \| A. Kashlinsky$^{1, 3}$, J. C. Mather$^2$, S. Odenwald$^4$\ $^1$NORDITA, Blegdamsvej 17, DK-2100 Copenhagen, Denmark\ $^2$Code 685, NASA Goddard Space Flight Center, Greenbelt, MD 20771\ $^3$Raytheon STX, Code 685,\ NASA Goddard Space Flight Center, Greenbelt, MD 20771\ $^4$Raytheon STX, Code 630,\ NASA Goddard Space Flight Center, Greenbelt, MD 20771\ title: 'Clustering of the Diffuse Infrared Light from the COBE DIRBE maps. III. Power spectrum analysis and excess isotropic component of fluctuations.' --- epsf \#1[=]{} \#1 2sr [Wm$^{-2}$sr$^{-1}$ ]{} 2m4sr2 [nW$^2$m$^{-4}$sr$^{-2}$ ]{} 2sr [nWm$^{-2}$sr$^{-1}$ ]{} 2m4sr [nW$^2$m$^{-4}$sr$^{-1}$ ]{} \#1[10$^{#1}$]{} Introduction {#s1} ============ Diffuse backgrounds and their structure contain important information about the history of the early Universe, when discrete objects either do not exist or are not accessible to telescopic studies. The formation and early evolution of galaxies should have generated radiation redshifted into the infrared bands (e.g. Partridge and Peebles 1967; Bond et al.1986). This cosmic infrared background (CIB) may come from the entire history of the Universe between the epoch of last scattering, mapped by the microwave background, and the present day. The COBE Diffuse InfraRed Background Experiment (DIRBE) (Boggess et al.1992; Silverberg et al. 1993; Weiland et al., DIRBE Explanatory Supplement, 1998) mapped the entire sky with 0.3 pixels and 0.7 resolution in ten bands between 1.25 and 240 . The results have been published (Hauser et al. 1998; Kelsall et al. 1998; Arendt et al. 1998; Dwek et al., 1998), based on a zodiacal light model believed accurate within a few percent (Kelsall et al. 1998). Their Galactic dust emission model (Arendt et al. 1998) is derived from fits to the DIRBE data and comparison to hydrogen maps, and accounts for the variation of dust temperature. Their model for Galactic starlight is derived from external star count models, with no free parameters. The residuals from the modeling were not significantly above the uncertainties except at 140 and 240 . The FIRAS instrument on COBE gave similar answers, (Fixsen, Mather, Bennett, and Shafer, 1998), using three different methods for removing the foregrounds. The agreement is significant because the instruments were calibrated independently. The measured values of the far-IR background are comparable to the total observed Galactic emission at visible and near-IR wavelengths, and imply that about half of the luminosity of the universe has been obscured by dust and converted into far-IR radiation. Some distant galaxies might be more reddened or obscured by dust than otherwise expected, and the CIB and its fluctuations might be brighter than predicted from visible band galaxy counts. In this paper we continue our search for a CIB by trying to detect its spatial structure in the DIRBE data. The method is similar to that first suggested by Gunn and later applied in the visible (Shectman 1973, 1974) and UV (Martin and Bowyer 1989). Recently, Vogeley (1998) applied the method to the Hubble Deep Field constraining the amount of the visible cosmic background, concluding that there is very little visible cosmic background from unresolved sources. They considered arcsecond scales where fluctuations are dominated by random parts of individual galaxy (Vogeley 1998) or cluster (Shectman 1974) profiles. We apply the method to degree scales where the dominant contribution comes from galaxy clustering. Our previous results are described by Kashlinsky, Mather, Odenwald and Hauser (1996a; Paper I) and Kashlinsky, Mather and Odenwald (1996b; Paper II). If we assume that the galaxies producing the CIB are clustered, the amplitude of the resulting CIB 2-point correlation function, $C(\theta)$, depends on the rate of the flux emission, $dF/dz$, and on the galaxy 2-point correlation function, $\xi(r; z)$. If the latter is known on the linear scales subtended by a given angle, measurements or upper limits on $C(\theta)$ can constrain the levels of the CIB emitted by clustered material. Measuring or limiting the structure of the CIB can be especially valuable in the mid- to far-IR bands where the foreground emission is very bright, but smooth. The plan of the paper is as follows: In Sec. \[s2\] we define the quantities used and provide mathematical background. We show that for the measured galaxy correlation function we expect fluctuations of about 5 to 10% in the CIB flux on 0.5 scales, almost independently of the particular mechanism of the CIB production. We discuss our search for the CIB structure in the DIRBE data between 1 and 100 . The last two bands of DIRBE - 140 and 240 - are not useful for this because of the large instrument noise there. Section \[s3\] deals with the beam profile and the spatial window function for power spectrum analysis. Section \[s4\] discusses the data for $\lambda \leq 5$ , where the foregrounds are dominated by stellar emission from the Galaxy. We find a significant trend in the amplitude of the measured fluctuations with Galactic latitude. When our star removal algorithm is applied to simulated data, we recover the measured slope of the correlation with Galactic latitude, and extrapolation of the simulated fluctuations to $\cosec\|b\|=0$ leads to a zero intercept. We derive a formula relating the measured fluctuations to star counts and show that for a plane-parallel Galaxy there should be a simple relationship between the $b$ dependence of the fluctuations and the slope of the star counts function at the Galactic pole. Then we show that for $ 90\deg < l < 270\deg$ and $\|b\|>20\deg$ the measured structure of the Galaxy fits the plane-parallel model. Furthermore, the scatter in the $C(0)$ - cosec$\|b\|$ plot is low enough to allow an extrapolation to $\cosec\|b\|=0$, with positive intercepts in all four near-IR bands. We perform a power spectrum analysis but most of the measured structure is due to fluctuating star counts. If the star removal algorithm removes stars fainter than 5 times the confusion noise limit, there are too few pixels left to calculate a power spectrum. Nevertheless, limits are sufficiently low to be of interest in testing theoretical predictions. We test our methods with numerical and analytic models of the Galaxy and possible CIB contributions and find good agreement with the data. Results from the mid- to far-IR bands, where foreground emission structure is dominated by the large scale gradients from dust emission in the Solar System and the Galaxy, are presented in Sec. \[s5\]. The upper limits on these fluctuations are $<$ 1 2sr , depending on wavelength, and provide much stronger tests of the predictions of galaxy counts than do direct measurements from inside the zodiacal dust cloud. We summarize our results in Sec. \[s6\]. Theoretical preliminaries {#s2} ========================= Here we provide a general mathematical basis without specific cosmological and galaxy evolution models. We will use the data to constrain the CIB properties in a model-independent way, and then to constrain models of galaxy evolution. We extend our previous work (Paper I,I; Jimenez & Kashlinsky 1999) to show the power spectrum of the fluctuations and to estimate the typical amplitude of the fluctuations in a model-independent way. We start with definitions. The surface brightness in the CIB per unit wavelength will be denoted as $I_\lambda$, per unit frequency as $I_\nu$, and per logarithmic wavelength interval $F=\lambda I_\lambda = \nu I_\nu$, and we call them all “flux.” The fluctuation in the CIB flux is then $\delta F(\mbox{\boldmath$x$})= F(\mbox{\boldmath$x$}) - \langle F \rangle$, where $\mbox{\boldmath$x$}$ is the two dimensional coordinate on the sky and $\langle ... \rangle$ denotes ensemble averaging. We also use the two-dimensional Fourier transform, $\delta F(\mbox{\boldmath$\theta$})= (2\pi)^{-2} \int \delta F_q \exp(-i\mbox{\boldmath$q$}\cdot \mbox{\boldmath$\theta$}) d^2\mbox{\boldmath$q$}$. If $\delta F(\mbox{\boldmath$x$})$ field is a random variable, then it can be described by the moments of its probability distribution function. The first non-trivial moment is the [projected 2-dimensional]{} correlation function $C(\theta) = \langle \delta F(\mbox{\boldmath$x$}+\theta) \delta F(\mbox{\boldmath$x$})\rangle$. The 2-dimensional power spectrum is $P_2(q) \equiv \langle \|\delta F_q\|^2\rangle$, where the average is performed over all phases. The correlation function and the power spectrum are a pair of 2-dimensional Fourier transforms and for isotropically distributed signal are related by $$C(\theta)= \frac{1}{2\pi} \int_0^\infty P_2(q) J_0(q\theta) q dq, \label{e1}$$ $$P_2(q)= 2\pi \int_0^\infty C(\theta) J_0(q\theta) \theta d\theta, \label{e2}$$ where $J_n(x)$ is the $n$-th order cylindrical Bessel function. If the phases are random, then the distribution of the flux field is Gaussian and the correlation function, or its Fourier transform the power spectrum, uniquely describe its statistics. In measurements with a finite beam radius $\vartheta$ the intrinsic power spectrum is multiplied by the window function $W$ of the instrument. Another useful quantity is the mean square fluctuation within a finite beam, or zero-lag correlation signal, which is related to the power spectrum by $$\begin{aligned} \langle (\delta F)^2 \rangle_\vartheta = \frac{1}{2\pi} \int_0^\infty P_2(q) W_{TH}(q\vartheta) q dq \nonumber \\ \sim \frac{1}{2\pi} q^2P_2(q)\|_{q\sim {\pi}/{2} \vartheta} . \label{e3}\end{aligned}$$ For a top-hat beam the window function is $W_{TH}=[2J_1(x)/x]^2=0.5$ at $x\simeq \pi/2$ where $x=q \vartheta$, and hence the values of $q^{-1}$ correspond to fluctuations on angular scales of diameter $\sim \pi/q$. The CIB flux and its structure are measured in projection on the celestial sphere and reflect both the 3-dimensional clustering pattern of the galaxy distribution and the rate of emission at redshift $z$. We introduce the 3-dimensional two-point correlation function of galaxy clustering, $\xi(r)$, and its 3-dimensional power spectrum, $P_3(k)$. These are related via 3-dimensional Fourier transforms, and assuming isotropy are related by $$\xi(r) = \frac{1}{2\pi^2} \int_0^\infty P_3(k) j_0(kr) k^2 dk . \label{e4}$$ The projected CIB correlation function is related to the underlying two-point correlation function of the galaxy distribution and the rate of the CIB flux emission via the Limber equation $$C(\theta) = \int_{z_1} \int_{z_2} \frac{dF}{dz_1}\frac{dF}{dz_2} \xi(r_{12}; z) dz_1dz_2, \label{elimber}$$ where $r_{12}$ is the [proper]{} length subtended by the angle $\theta$ and redshifts $z_1, z_2$. In the limit of small angles, $\theta \ll 1$, the Limber equation becomes (e.g. Peebles 1980) $$C(\theta) = \int_0^\infty dz \left(\frac{dF}{dz}\right)^2 \int_{-\infty}^\infty \xi(r_{12}; z) du. \label{e5}$$ For Robertson-Walker metrics the proper separation is given by $$\begin{aligned} r_{12}^2 = (\frac{dx/dz}{ \sqrt{1-kx^2} } u)^2+\frac{x^2(z)\theta^2}{(1+z)^2}\\ = c^2(\frac{dt}{dz})^2 u^2 +d_A^2(z)\theta^2.\end{aligned}$$ Here $d_A(z)$ is the angular diameter distance, $x(z)\equiv d_A(z)/(1+z)$, $u$ is the integration variable and $t$ is the cosmic time. The comoving volume occupied by a unit solid angle in the redshift interval $dz$ is $dV/dz = (1+z) x^2(z) cdt/dz$, and the power received per unit wavelength and collecting area in band $\lambda$ from each galaxy with absolute bolometric luminosity $L$ at redshift $z$ is $[L/({4\pi x^2(1+z)^3})]f_\lambda(\frac{\lambda}{1+z}; z)$. Here $f_\lambda d\lambda$ is the fraction of the total light emitted in the wavelength interval $[\lambda; \lambda +d\lambda]$ and the extra factor of $(1+z)$ in the denominator accounts for the fact that the flux received in band $\lambda$ comes from a redshifted galaxy. The contribution to the total CIB flux from the redshift interval $dz$ is given by $$\frac{dF}{dz} = \frac{R_H}{4\pi} \frac{1}{(1+z)^2} \frac{d(H_0t)}{dz} \sum_i {\cal L}_i(z) [\lambda f_{\lambda, i}(\frac{\lambda}{1+z}; z)], \label{e6}$$ where the sum is taken over all galaxy populations contributing flux in the observer rest-frame band at $\lambda$, and $f_\lambda$ characterizes the spectral energy distribution (SED) of galaxy population $i$. Here $R_H=cH_0^{-1}$ and ${\cal L}(z)= \int \Phi(L; z) L dL$ is the comoving luminosity density from galaxies with the luminosity function $\Phi(L; z)$ at epoch $z$. It is illustrative to study the redshift dependence of the flux rate production, Eq. (\[e6\]). At small redshifts the factor $(1+z)^{-2}dt/dz$ varies little with $z$, and the rate of flux production is governed by the comoving bolometric luminosity density ${\cal L}(z)$ and the SED of the galaxy emission $f_\lambda$. If the luminosity evolution at these redshifts is small, the rate of flux production is governed by the SED shape. If $f_\lambda(\lambda)$ increases towards shorter wavelengths then $dF/dz$ increases with $z$. For $\lambda f_\lambda=$ const, and no luminosity evolution, the rate is roughly constant with small $z$. In addition, there is observational evidence for an increase of ${\cal L}(z)$ out to $z\simeq 1$ in the galaxy rest-frame UV to near-IR (1 ) bands (Lilly et al. 1996). At sufficiently high redshifts, the evolution in Eq. (\[e6\]) would be offset by the factor $(1+z)^{-2}dt/dz$, so that the rate of production would be cut off at sufficiently large $z$. This factor is responsible for resolving Olbers’ paradox even for a flat SED. The SED for rest-frame $\lambda < 10$ is dominated by stellar emission, with a peak at visible wavelengths and a decrease for $\lambda > 0.7$ . Consequently, most of the predicted J band CIB comes from redshifts $z\sim 0.3-1$, which shifts the visible emission of normal stellar populations into the J band; cf. Yoshii and Takahara (1988). In the M band at 5 , most of the predicted CIB comes from $z>1-2$. At $\lambda > 10$ , the emission is dominated by galactic dust and the situation is reversed, so $f_\lambda$ increases with wavelength roughly as $\lambda^\alpha$ with $\alpha \sim 1.5$. Hence, the dusty star-burst galaxies observed by IRAS at low redshifts should make the dominant contribution to the 10 CIB. In the far-IR, the measured CIB found by DIRBE can have large contributions from high redshifts. Measurements of the correlation function $C(\theta)= \langle \delta F({\mbox{\boldmath$x$}} +\theta) \delta F({\mbox{\boldmath$x$}}) \rangle$ have an important advantage over direct determinations of the CIB spatial power spectrum, because they are immune to the discontinuities in the maps created by point source removal. (Because, in practice, $C(\theta)$ is evaluated from masked data, the uncertainties of the $C(\theta)$ points are themselves correlated.) However, interpreting $C(\theta)$ in terms of $dF/dz$ and $\xi(r)$ can be cumbersome because the right-hand-side of (\[e5\]) contains a double integral and the 3-dimensional correlation function is not always positive. By definition, $\xi(r)$ must be negative on large scales so that $\int_0^\infty \xi(r) r^2 dr=0$. A simpler method is to work with the Fourier transform of the correlation function, $P_2(q)$, which contains the same information as $C(\theta)$ and is easier to interpret. As for $C(\theta)$ the effects of the mask will produce correlations of the power spectrum points with each other. Also, the measured power spectrum is the convolution of the Fourier transform of the mask with the true power. Although the uncertainties in $C(\theta)$ and $P_2(q)$ can be evaluated from theory, we evaluated them from the data by comparing multiple measurements. The Limber equation (\[e5\]) can be rewritten directly in terms of the power spectra, substituting Eq. (\[e4\]) into (\[e5\]) and using $\int_{-\infty}^\infty j_0(\sqrt{x^2 +y^2})dy = \pi J_0(x)$ to obtain $$\begin{aligned} C(\theta) = \frac{1}{2\pi} \int_0^\infty dz \left( \frac{dF}{dz} \right)^2 \frac{1}{cdt/dz} \nonumber \\ \int_0^\infty P_3(k; z) J_0(k d_A \theta) k\ dk . \label{e7}\end{aligned}$$ Substituting this expression into (\[e2\]) and using the orthogonality relations for the Bessel functions,\ $\int_0^\infty J_0(\alpha \theta) J_0(\beta \theta) \theta d\theta = \alpha^{-1} \delta_D(\alpha-\beta)$, leads to $$P_2(q) = \int_0^\infty \left(\frac{dF}{dz}\right)^2 \frac{(1+z)^2}{c\frac{dt}{dz} x^2(z)} P_3(qd_A^{-1}(z); z) dz. \label{e8}$$ Equation(\[e8\]) involves only one integration and the kernel is always positive. Hence it may be easier to derive cosmologically interesting quantities such as $F$ and $\xi(r)$ or $P_3(k)$. Determining $P_2(q)$ directly from the data is difficult owing to the masking created by point source removal, and it seems best to compute $C(\theta)$ and then derive $P_2(q)$ from it. Figure \[f1\]a plots the argument, $q(1+z)/x(z)$, of $P_3$ in the right-hand-side of Eq. (\[e8\]) for the largest wavenumbers (smallest scales) probed by DIRBE at various redshifts $z$. The dotted line corresponds to $\Omega=1, \Omega_\Lambda=0$; the solid to $\Omega=0.1, \Omega_\Lambda=0$ and the dashes to $\Omega=0.1, \Omega_\Lambda=0.9$, where $\Omega_\Lambda=3H_0^2(1-\Omega)$ denotes the contribution of the cosmological constant. For redshifts contributing most of the CIB flux ($z>$0.1-0.2, cf. Paper I), the DIRBE instrument probes scales which are in the quasi-linear or linear regime, and which thus can be approximated as having evolved at the same rate with time. Thus for DIRBE scales, one can write $P_3(k; z) \simeq P_3(k; 0) \Psi^2(z)$, (e.g. Peebles 1980), where $\Psi^2(z)$ accounts for the evolution of the clustering and is normalized to $\Psi^2(0)=1$. Then Eq. (\[e8\]) can be rewritten in a more compact way: $$q^2P_2(q)= \int_0^\infty \left(\frac{dF}{dz}\right)^2 \frac{\Psi^2(z)}{H_0\frac{dt}{dz}} \Delta^2(\frac{q}{d_A(z)}) dz, \label{e9}$$ with $$H_0\frac{dt}{dz} =\frac{1}{(1+z)^2\sqrt{1+\Omega z + \Omega_\Lambda[(1+z)^{-2}-1]}} . \label{e10}$$ The left-hand-side of (\[e9\]) is the same order of magnitude as the mean square fluctuation (\[e3\]), and we have defined $$\Delta^2(k) \equiv R_H^{-1}k^2 P_3(k; 0). \label{e11}$$ The quantity $\Delta(k)$ is roughly the fluctuation over the line-of-sight cylinder of length $R_H$ and diameter $k^{-1}$. For relevant scales and spectra of density fluctuations, $\Delta(k)$ increases with $k$. Hence, the CIB fluctuation on a scale of $k^{-1}$ is $\sim \sqrt{k^2 P_2(k)} \sim F \Delta(k/R_H)$. The present day power spectrum of galaxy clustering has been measured on scales corresponding to at least the smallest angular scales probed by DIRBE. The most accurate measurement comes from the APM survey data on the projected angular correlation function (Maddox et al. 1990). Baugh and Efstathiou (1993) deprojected the APM data to obtain the underlying power spectrum of galaxy clustering, $P_3(k)$. Kashlinsky (1998) used the current large scale data, and the abundance of objects at high redshifts, to reconstruct the pregalactic density field over six orders of magnitude in mass. His result requires significant fluctuation power on small scales, an early epoch of galaxy formation, and high levels of the CIB and its fluctuations. Fig. \[f1\]b plots the data from the Baugh and Efstathiou (1993) deprojection. On small scales (large $k$), $\Delta^2(k) \propto k^{0.7}$, so the integrand in Eq. (\[e9\]) behaves as $\propto z^{-0.7}$ with an integrable singularity at $z\rightarrow 0$. While a non-negligible part of the clustering part of the CIB fluctuations comes from nearby galaxies, much of it arises from galaxies at $z>1$. In Papers I and II we estimated $\sim$ 5-10% CIB fluctuations on a scale of 0.5. One can also see this in a more intuitive way from Eq. (\[e9\]). Fig. \[f1\]a plots $q(1+z)/x(z)$, the largest wavenumber that enters on the right hand side of (\[e9\]), for $q^{-1} = 0.5$. In the Friedman-Robertson-Walker Universe, $q/d_A(z)$ reaches a minimum at $z\sim 1-3$, and its value at the minimum depends weakly on cosmological parameters. Thus (\[e9\]) can be rewritten as an inequality: $$q^2P_2(q) \geq \Delta^2\left(\min[\frac{q}{d_A(z)}]\right) \int_0^\infty \left(\frac{dF}{dz}\right)^2 \frac{\Psi^2(z)}{H_0\frac{dt}{dz}} dz . \label{e12}$$ On linear and quasi-linear scales over the range of redshifts that contribute most to the CIB, the quantity $\Psi^2(z)/\frac{H_0dt}{dz}$ depends weakly on $z$ (e.g. on linear scales $\Psi^2=(1+z)^{-2}$ if $\Omega=1$). The integral is of the same order of magnitude as $F^2$ because the term $dF/dz$ is a peaked function with a full width at half maximum of order unity. Hence, the relative fluctuation in the CIB is $\sim \Delta(\min[\frac{q}{d_A(z)}]) $. Fig. \[f1\]b plots $\Delta(k)$ vs. $k$ using the power spectrum of APM galaxies from Baugh & Efstathiou (1993); it is an increasing function of $k$. At the wavenumber of the minimum in Fig. \[f1\]a, the CIB fluctuation can be as large as $\sim$ (5-10)%. In models where the bulk of the CIB comes from higher redshifts, this number may be somewhat smaller. Because this method measures a two-point process, the results constrain a measure related to the mean square of the CIB emission rate, i.e.$\sim \int \left(\frac{dF}{dz}\right)^2 dz$. Many models have calculated the expected CIB over the 1 - 100 range (e.g. Partridge and Peebles 1967; Stecker et al. 1977; Bond et al. 1986; Fall et al. 1996; Wang 1991; Beichman and Helou 1991; Franceschini et al. 1991; Cole et al. 1992; Malkan and Stecker, 1997; Jimenez and Kashlinsky 1999; Dwek et al. 1998). The models are normalized to galaxy counts and predict a typical flux of $F \sim 10$ 2sr . Deep K band counts of galaxies (e.g. Cowie et al. 1994; Djorgovski et al. 1995) or at 12 - 100 (Hacking and Soifer 1991) suggest [minimal]{} fluxes of at least a few 2sr . Therefore the CIB fluctuations on 0.5 scales may be $\sim 1$ 2sr . The APM measurements of the galaxy correlation function in the blue band might not apply to the infrared. However, the $r$ band Palomar survey (Picard 1991), and the R-band Las Campanas survey (Shectman et al. 1996, Lin et al. 1996) give results identical to the APM survey. On the other hand, for IRAS galaxies on small (non-linear) scales the correlation function has a lower coherence length. Saunders et al. (1992) show that on small scales IRAS galaxies at 60 cluster with $\xi_{IRAS} = (r/r_{, IRAS})^{-1-\gamma_{IRAS}}$ where $r_{, IRAS} \simeq 4h^{-1}$Mpc and $\gamma_{IRAS}\simeq 0.6$, as opposed to $r_=5.5h^{-1}$Mpc and $\gamma\simeq 0.7$ for the APM galaxies (Moore et al. 1994). The slight difference could be due a tendency of the IRAS dusty star-burst galaxies to avoid the central regions of rich clusters of galaxies. However, on larger (linear) scales where the galaxy clustering pattern presumably traces the pregalactic density field, IRAS galaxies as measured by the QDOT counts-in-cells analysis (Saunders et al. 1990) are consistent with the APM galaxies power spectrum. These linear to quasi-linear scales are relevant for the DIRBE beam size, so we use the APM numbers for all infrared wavelengths. In surveys with smaller beams, the differences will be more pronounced, but the differences between the IRAS and APM correlation functions do not lead to appreciably smaller mid- to far-IR CIB fluctuations. In addition to the clustering term, there is a white-noise component in the correlation signal due to individual galaxies (e.g. Peebles 1980). Its amplitude at zero lag is $$\begin{aligned} C_{\rm WN}(0) = \theta_{\rm beam}^{-2} \frac{R_H}{16\pi^2} \int \frac{\int \Phi(L; z) L^2 dL}{x^2(z)(1+z)^8} \frac{H_0dt}{dz} \nonumber \\ \left[ \lambda f_\lambda (\frac{\lambda}{1+z}; z) \right]^2 dz . \label{e13}\end{aligned}$$ Like white noise from discrete stars, the white noise from galaxies is dominated by the nearest objects, unlike the clustered component (\[e5\]), because of the presence of $x^2(z)$ in the denominator of the integrand in (\[e13\]). Because the galaxies are almost undetectable against the star fluctuations, the galaxy white noise is also negligible. DIRBE data and beam profile {#s3} =========================== The COBE DIRBE instrument provided an all-sky 41 week survey with a ten-band photometer (Boggess et al. 1992). The DIRBE bolometer measurements at 140 and 240 are too noisy for our purposes. The remaining eight bands are centered on wavelengths 1.25, 2.2, 3.5, 4.9, 12, 25, 60 and 100 , and are labeled Bands 1-8 respectively. We obtained the 41 weekly-averaged DIRBE maps from the National Space Science Data Center (NSSDC), subtracted the zodiacal light model developed by the DIRBE team (Kelsall et al. 1998; Weiland et al, 1998, DIRBE Explanatory Supplement), and averaged the weeks together. The maps are stored in a cube format and pixelized into 6 faces of 256$^2$ approximately square pixels $\sim 0.3$ on a side. For a finite beam the ideal map is convolved with the beam window function. For a circular top-hat beam similar to DIRBE’s square top hat, the Fourier transform of the window function is $W(x) = [2J_1(x)/x]^2$, where $J_1(x)$ is the first-order cylindrical Bessel function. The measured power spectrum is then the product of the underlying power spectrum and the beam window function, i.e. $P_{\rm measured}(q) = P_2(q) W(q\vartheta)$ where $\vartheta$ reflects the beam size. On large scales (small $q$), $W(x)\sim 1$. Throughout the rest of the paper, $P(q)$ will refer to the 2-dimensional power spectrum of the diffuse emission computed from DIRBE maps after deconvolution from the beam profile, i.e. $P(q)= P_{\rm measured}(q)/W(q\vartheta)$. We determined the effective $W$ and the effective beam size from maps of the beam response function archived at the NSSDC with 181$\times$225 0.0065 square pixels, which we embedded in 256$^2$, 512$^2$, and 1024$^2$ pixel arrays. The results for Band 1 are shown in Fig.\[f2\], and are similar for all eight bands. The solid line shows a top-hat beam profile with $\vartheta=0.4$ which is slightly lower than the value of $\vartheta=0.45$ adopted in Paper I. From eye fits we adopted $\vartheta=0.4$ in Band 1, 0.37 in Bands 2 and 3, and 0.35 in Bands 4 to 8. The combined effects of beam smearing from the motion of the beam during sampling, pixelization, and pointing error would increase the effective beam size by about 10%, and agree better with Paper I. The beam response function drops below 10% at $q^{-1} <$0.15 and the pixelization prevents measurement below $q^{-1}$ = 0.1. Near-IR analysis {#s4} ================ Foregrounds and point source removal {#ss41} ------------------------------------ Foreground emission from the Galaxy and the solar system is the main problem in unveiling the expected CIB. At wavelengths less than 10 , the dominant foreground after removing the zodiacal light model is emission from stars in our Galaxy. Using the SAO Bright Star Catalog magnitudes and colors, we simulated J band maps. We found that aside from the large-scale shape of the galaxy and a few star clusters this foreground has an almost uncorrelated spatial distribution. At 12 and 25 (DIRBE Bands 5 and 6) the zodiacal dust is so bright that residuals from the model subtraction dominate the map structure outside the Galactic plane. At wavelengths of 60 and 100 (DIRBE bands 7 and 8) most foreground residual fluctuations come from cirrus dust clouds in the Galaxy, and stars contribute little. The measured fluctuations contributed by point sources can be reduced substantially by identifying and removing them down to near the confusion noise limit. We used the point source finding routine developed by the DIRBE team and adopted in Paper I, and call it the “clipping” algorithm. The data from a selected region (patch) are first used to construct a smoothed model of the sky background emission for the whole patch, as follows. Surrounding each pixel, a window of size $f_{\rm size}\times f_{\rm size}$ is searched for the minimal flux value. We use the minimal value rather that the median because the stellar brightness distribution is highly skewed. The map of this minimal (or lower envelope) flux is then fitted to a 2-dimensional polynomial surface of order $n_{\rm fit}$. This polynomial is in turn subtracted from the sky map, and the standard deviation $\sigma_0$ is calculated. Then, each of the pixels with flux above $N_{\rm cut}\sigma_0$ is masked out along with the 8 adjacent pixels, about twice the DIRBE beam area. This process of fitting a polynomial, identifying bright pixels, and masking them out, is iterated about 5 times, until no more fluxes above $\sigma_0$ are found. In the near-IR no noticeable improvement is reached above $n_{\rm fit}=2$. We discuss the dependence on $f_{\rm size}$ in Sec.\[ss45\]; for the results presented below we used $f_{\rm size}=5$. The polynomial fit was used only for point source masking, no gradients were subtracted from the maps analyzed in this (near-IR) section, and the results at high galactic latitude ($\|b\|>20$) were independent of this choice. For the near-IR, where foreground stars are the dominant source of fluctuations, there should be a clear correlation between the fluxes at different wavelengths. Furthermore, because most of the flux and fluctuations come from K and M stars, and the DIRBE pixels are large enough to include many stars, the dispersion in the color diagrams for the four near-IR bands should be small. In Paper I we used these properties to search for CIB fluctuations that have a different color due to redshifts or a different source spectrum. We construct a linear combination of maps at two different bands: $$\Delta_{12} = \delta F_1 -\beta_{12} \delta F_2 . \label{e14}$$ Note that this $\Delta$ is unrelated to the $\Delta$ of Eq. (\[e9\]). The variance of this map, $C_\Delta(0)\equiv \langle \Delta_{12}^2\rangle$, is minimized for $$\beta_{12} = \frac{ \langle \delta F_1\delta F_2\rangle }{ \langle \delta F_2^2\rangle}. \label{e15}$$ Because the dispersion in the color index $\beta = \delta F_1/\delta F_2$ is small, i.e. $\sigma_\beta = \sqrt{\langle \beta^2 \rangle -\langle \beta \rangle^2 } \ll \beta_{12}^2$ with the average taken over all the pixels in the patch, the foreground contribution to $C(0)$ will be reduced by a factor $\sim (\beta_{12}/\sigma_\beta)^2 \gg 1$. If the fluxes in adjacent bands do not correlate, the emission in the two bands comes from different sources or from measurement errors (noise). Since most of the predicted CIB comes from galaxies at $z>0.2$ with typical stellar populations, its color should differ from that of the foreground stars, and the color subtracted maps should retain some of the CIB structure. If most of the CIB comes from high redshifts so that $$\beta_{CIB}=\langle \delta F_{1, CIB} \delta F_{2, CIB}\rangle/\langle \delta F_{2, CIB}^2\rangle \leq 2 \beta_{12},$$ the CIB fluctuations in the color subtracted maps will be larger than in the single band maps. We used Eq. (\[e15\]) to make all-sky color-subtracted maps for all the adjacent band pairs, in order to search for a coordinate-independent part of the fluctuations. This method can be generalized to a multi-color subtraction method, e.g. minimize $\Delta_{123} = \delta F_1 -\beta \delta F_2 - \alpha \delta F_3$ with respect to $\alpha, \beta$. We applied this to the DIRBE maps, but without significantly different results. All sky variance analysis: $C(0)$ {#ss42} --------------------------------- We divided the sky into 384 patches, each with $32\times 32$ pixels, and clipped each patch individually using the above procedure, with = 7, 5, 3.5 and 3. Because the foreground emission at these wavelengths is dominated by point sources, very few pixels are left for $<3$. We performed the same analysis on 96 patches of $64\times 64$ pixels, with similar results. Our star finding algorithm cuts deeper into the distribution than a simple interpretation might suggest. Each bright star is observed in about 5 pixels, depending on the position of the star. The noise distribution of these pixels in the absence of the bright star still has a non-Gaussian distribution with a non-zero mean. Therefore a star near the clipping threshold will be identified in the pixel where the background noise fluctuations are greatest, rather than at the true star location. Hence, the clipping algorithm finds stars about 1 $\sigma$ fainter than , as we confirmed with simulations. After clipping the 384 patches, we computed the color indices according to Eq. (\[e15\]), the single band $C(0)$, and the color-subtracted $C_\Delta(0)$ for each patch. Figure \[f3\] shows histograms of the numbers of the remaining pixels after clipping. In the near-IR bands, about 350-450 pixels out of 1024, or about uncorrelated 75-90 beams per patch, remain for $N_{\rm cut}=3.5, 3$, making the intrinsic uncertainty of $C(0)$ less than 15%. In each patch there is a clear correlation between the fluxes in the adjacent bands. Paper I gave flux correlation plots for selected patches and an earlier model of the zodiacal light. The current data give similar results. For bands 1-4 all the patches have correlation coefficient $R >0.9$ between all the pairs of bands. The color indices have very small dispersion, $\sigma_\beta/\beta < 10\%$. Because the dispersion in $\beta$ is so small, and most of the predicted CIB emission at these wavelengths comes from redshifts $z > 0.1$, the CIB fluctuations should not be removed by the color subtraction. Fig. \[f4\] plots the histograms of $\beta$ for $N_{\rm cut}=3$. The maps for the band pair \[4, 5\] have color index $\beta \simeq 0$ because the dust in the solar system is not strongly correlated with the stars seen in Band 4. Between Bands 1 and 2, most of the patches have a color index of $\beta_{12} \simeq 2$, typical of K-M giants. In band pairs \[2, 3\] and especially \[3, 4\] the spread in $\beta$ is substantial, suggesting that star fluctuations are not the only source, or that their color is different in different regions. The range of color indices for the band differences $[2-3]$ and $[3-4]$ is reduced for $\|b\|>20$, but nevertheless remains much wider than for the $[1-2]$ maps. Most of the map peaks are due to stars, as we found them in simulated maps derived from the SAO Bright Star Catalog, using wavelength extrapolations appropriate to the tabulated spectral types. The lowest limits on $C(0)$ for single bands are similar to those in Paper II, and for the color subtracted bands they are not very different from Paper I. However, we can now address the dependence of $C(0)$ on Galactic coordinates. Fig. \[f5\] shows a strong correlation of $\sqrt{C(0)}$ with $\cosec \|b\|$. Fig. \[f6\] plots $\sqrt{C(0)}$ vs.cosec$\|b\|$ for the color subtracted maps \[1-2\], \[2-3\], and \[3-4\]. Although the color subtracted $C(0)$ is a factor of $\sim 10$ lower than for single bands, the residual fluctuations still depend strongly on Galactic latitude. Near-IR fluctuations from Galactic stars {#s43} ---------------------------------------- We would now like to extrapolate the data to estimate the extragalactic contribution to the fluctuations. We use analytic and numerical models to support a power-law dependence of the stellar contribution to $C(0)$ on cosec$\|b\|$. We also show how to extrapolate the fluctuations due to Galactic stars to cosec$\|b\|=0$, and that we expect no isotropic residual for $C(0)$ due to stars. We assume that Galactic stars are distributed with a white noise power spectrum, i.e. they are spatially uncorrelated except for large scale distribution functions such as those in the Bahcall and Soneira (1980) and Wainscoat et al. (1992) models. Thus $C(0)$ in the direction $b$ is given by the following (Galactic) version of the Limber equation: $$\begin{aligned} \sigma^2(x) = \sum_i \int \frac{L_1}{4\pi r_1^2}\frac{L_2}{4\pi r_2^2} dP_{12} \nonumber \\ = \sum_i \frac{L_i^2}{4\pi} \omega \int_{R_i}^\infty \frac{n_i(r;x)}{r^2} dr. \label{e16}\end{aligned}$$ Here the sum is taken over stars of type $i$ with intrinsic luminosity $L_i$, $\omega$ is the pixel solid angle, $x = {\rm cosec}\|b\|$, $r$ is the distance to each star, and $n_i$ is the number density of stars at that distance in the direction of $b$. The distance to the star is related to the height, $Z$, above the Galactic plane by $r= Z$cosec$\|b\|$. In our clipping algorithm, we remove sources exceeding the flux limit of $\times$(standard deviation of the background polynomial fit). This standard deviation is approximately $\sigma$ from Eq. (\[e16\]) above, if the integration limits are chosen to match the threshold for star detection and clipping. We find a very good correlation between the deviation of the fit and the value of the residual fluctuations in the patch; if all the histograms had the same non-Gaussian shape the two should be proportional. Hence, the lower limit radius can be approximated to depend on $\sigma$ itself via $$R_i =[L_i/(4\pi N_{\rm cut} \sigma)]^{1/2}. \label{e17}$$ Equations (\[e16\]), (\[e17\]) then allow us to determine $\sigma(b)$ for a given distribution of stars. We now derive the latitude dependence implied by these equations. As an example, we use a plane parallel exponential model for stars in the Galaxy: $n_i = n_{0,i} \exp(-\frac{r}{h_ix})$. This should be valid if our clipping radius $R$ is small compared to the exponential scale length for the Galactic disk $\alpha^{-1} \simeq $ (3 - 4) Kpc, and is in rough agreement with the disk in hydrostatic equilibrium with the Galactic gravity field (Mihalas & Binney 1981). Now Eq. (\[e16\]) becomes $$\sigma^2(x) = \sum_i R_i^{-1} \frac{ L_i^2 n_{0,i}\omega}{4\pi} E_2(\frac{ R_i }{ h_i x }), \label{e18}$$ where $E_n(u)\equiv \int_1^\infty t^{-n} \exp(-ut) dt$. We express (\[e18\]) in terms of the total number of stars in the $dR_i$ interval: $$\frac{dN_i}{dR_i} = n_{0,i} \omega R_i^2 \exp(-\frac{ R_i }{ h_i x }). \label{e19}$$ Then the equation for $\sigma(x)$ becomes: $$1=N_{\rm cut}^2 \sum_i \left(R_i \frac{ dN_i }{ dR_i }\right) [\exp(q)E_2(q)]\|_{q= { R_i }/{ h_ix } }. \label{e20}$$ Consider first the limit when $R_i \ll h_i$. In this case $q<1$ and the term in square brackets is near unity. The number of clipped stars per beam-width then becomes $N_> = \omega \sum_i \int_0^{R_i} n(r;x) r^2 dr \simeq 3^{-1} \sum_i (R_i dN_i/dR_i)$, leading to $N_> =N_{\rm cut}^{-2}/3$. This is much smaller than we find in the data. Our clipping algorithm removes 9 pixels per star and cuts deeper than =3.5 would indicate, and leaves only about 40% of the pixels. This proves that $q>1$, i.e. bright stars are seen out to the scale height of the disk and beyond. We now consider the variations in star counts with magnitude in a plane-parallel Galaxy. The differential counts in the magnitude interval $dm$ are given by $dN/dm = \omega [d\ln r_m/dm] n(r_m) r_m^3$, where $r_m \propto 10^{0.2m}$ is the radial distance to the star of apparent magnitude $m$. In the limit of $r_m < h$, or $n\simeq$ constant, the counts converge to the uniform distribution limit, $dN/dm \propto 10^{0.6m}$. In a plane-parallel Galaxy in which the radial structure can be neglected, the number density of stars depends only on $\|Z\| \propto 10^{0.2m} \sin \|b\|$, and the differential counts in the direction $x={\rm cosec}\|b\|$ can be related to those at the Galactic pole by $$\frac{dN}{dm}\|_x = x^3 \frac{ dN(m-5\log_{10} x)}{dm}\|_{\rm Pole}. \label{e21}$$ Hence, once the counts at the pole are measured, we can evaluate the expected number counts in any direction $b$ and then compute the fluctuations in the flux they produce via $$\sigma^2(x) \propto \int_m^\infty 10^{-0.8 m} \frac{dN}{dm}\|_x dm. \label{e22}$$ This formulation has two advantages: 1) the data on star counts in the polar regions can be used to evaluate the dependence of star fluctuations on cosec$\|b\|$; and 2) the only assumption is plane parallelism. We tested this assumption with both models and observations. We now determine $dN/dm$ from the DIRBE data, and compare with prior observations. Figure \[f7\] shows a Band 2 histogram of $\log_{10}(\|F-\langle F \rangle\|)$ for a 128$^2$ pixel (38.4$\times$ 38.4) region at the North Galactic Pole (NGP). This logarithmic form is useful for displaying both large and small deviations, and we have normalized it to the equivalent number of stars per magnitude per square degree. The equivalent K magnitude is plotted on the top horizontal axis. With the DIRBE beam of $1.42\times 10^{-4}$ sr (4.5 pixels), $F =1$ 2sr corresponds to $m_K$ = 9.45. Because the CIB level and the smooth foreground contributions from dust and faint stars are not known in advance, we must subtract the mean of this confusion noise before seeking to measure brightnesses of individual bright stars. We define $\langle F \rangle$ as the mean flux of the map clipped with = 3.5; in this case it is 66 2sr or $m_K=5$, near the turnover of the distribution. The total number of pixels with positive values of the flux after the subtraction is 12,359; the remaining 24.6% of the pixels have negative fluxes. We have plotted both positive and negative deviations to illustrate the asymmetry of the histogram far from the peak, as well as its symmetry near the peak. Negative deviations are statistical fluctuations, while very bright pixels are individual cataloged stars. For the clipped map at = 3.5, a total of 6,380 pixels remain. Out of these, 2,982 pixels (47%) have positive fluxes, showing the symmetry of the flux distribution for the clipped map. Figure \[f7\] also shows a Gaussian distribution with the variance $\sigma\simeq10$ 2sr computed with = 3.5, and for $\sigma=9$ 2sr . The difference between these two Gaussians is not large on this plot, showing the difficulty of direct detection of a sub-population of Gaussian fluctuations with dispersion of 5 2sr . The NGP star counts were observed directly by Elias (1978). We show his data at $K= 1$, 2.5, 3.25 and 8 with $N^{1/2}$ error bars and our binning of his data. Further NGP data were obtained by the 2MASS survey in K$_s$ band, almost identical to the DIRBE Band 2, and were kindly provided to us by Tom Jarrett (1998). The cumulative counts from these measurements were shown in Fig. 1 of Beichman (1996) out to $K_s >15$, who found that they follow $dN/dm \propto 10^{0.3 m}$ (cf. his Table 4). Actual 2MASS star counts from a region of 5 square degrees centered on the NGP are plotted in Fig. \[f7\]. The agreement between the DIRBE counts, the Elias (1978) and Jarrett (1998) data, and the $B=0.3$ extrapolation is excellent over 15 magnitudes, or six decades in flux. South Galactic Pole counts from Minezaki et al. (1998) are also shown, and confirm the slope as well as the north-south symmetry of the Galaxy. At $m_K < 1.5$, the counts tend to the slope of $B=0.6$ coming from stars much closer than the scale height; if $B$ were less than 0.6, the integrated star brightness would diverge at the bright end. The star counts agree with model predictions. Both Beichman (1996) in his Fig. 1 and Minezaki et al. (1998) in their Fig. 1 show that the counts are fitted well by extensions of either the Bahcall & Soneira (1980) or Wainscoat et al. (1992) models. An eyeball fit to their data gives $B=0.3-0.32$ at $K=11$. The Wainscoat et al. model at $K=11$ shown in Fig. 1 of Minezaki et al. (1998) gives $\log dN/dm \simeq 1.35$, whereas continuation of the solid line in our Fig.\[f7\] to $K = 11$ gives $\log dN/dm = 1.3$ if $B=0.3$ and 1.4 if $B=0.33$. The agreement between the two slopes and normalizations is thus very good. Even the large-beam DIRBE instrument sees far beyond the scale height of the bright K band stars. We can now estimate the star brightness fluctuations. We begin with a power law for star counts in a plane-parallel Galaxy, $dN/dm\|_{\rm Pole} \propto 10^{Bm}$. Then the fluctuation due to stars fainter than $m$ at the Galactic pole becomes: $$\sigma^2\|_{\rm Pole} \propto \int_m^\infty 10^{-0.8m^\prime +Bm^\prime} dm^\prime . \label{e23}$$ We now derive the latitude scaling law implied by these assumptions. In our clipping method the lower limit on $m$ in the integral is given by $10^{-0.4 m} \propto N_{\rm cut} \sigma$. Combining this with Eq.(\[e21\]) we find that the fluctuation in the total star flux at latitude $b$ is given by the power law $$\sigma \propto (\cosec\|b\|)^{{1.2\over B}-2}. \label{e24}$$ A uniform, infinite star cloud has $B=0.6$, so there is no latitude dependence in that case. If $B=0.4$, then we expect $\sigma \propto \cosec\|b\|$, while if $B=0.3$ then $\sigma \propto \cosec^2 \|b\|$. The dependence of $C(0)$ on cosec$\|b\|$ can also be inverted via Eqs.(\[e21\]), (\[e22\]) to give a model for $dN/dm$. There is a unique relation between the two functions, for a plane-parallel Galaxy and our method of bright star extraction. The value of enters indirectly, since it determines the magnitude $m$ of the cutoff of detected stars and hence the magnitude at which $B$ is important. Another measure of the star distribution function is the number of clipped stars, $N_>$. In the power law case we have a very simple result: $$N_> = {{0.8-B}\over B N_{\rm cut}^2}= \frac{2p+1}{3N_{\rm cut}^2}, \label{e25}$$ where $p = (1.2/B)-2$ is the logarithmic slope in Eq. (\[e24\]). The dependence of $N_> \propto N_{\rm cut}^{-2}$ agrees with our results from DIRBE. Evaluating (\[e25\]) for $B=0.6$ (or $r_m <h$) we find $N_>=1/3N_{\rm cut}^2$, in agreement with the exponential distribution of Eq. (\[e20\]) with $h=\infty$, while for $B=0.3$ we find $N_>=5/3N_{\rm cut}^2$. The fraction of pixels clipped in the near-IR DIRBE maps at = 3.5 is $\sim 60\%$. Because the clipping algorithm uses a mask with about twice the beam area, this corresponds to $N_> = 0.35$ so the effective = 2.2 to 2.4 for $p=$ 2.0 to 2.5. The effective is less than the input value of 3.5 because each star is observed in many pixels, each with its own noise. Similarly, Eqs. (\[e21\]), (\[e22\]) allow us to evaluate the expected variation of the residual fluctuation with , $\sigma \propto N_{\rm cut}^{-1 + 0.8/B}$. In the limit of a spatially uniform distribution, $B=0.6$, and $\sigma \propto N_{\rm cut}^{1/3}$. For $B=0.3$ it follows that $\sigma \propto N_{\rm cut}^{1.7}$, and reducing from 7 to 3.5 should decrease $C(0)$ by a factor of 10, in agreement with the DIRBE data. To summarize this section, we have shown that, for $dN/dm \propto 10^{Bm}$ with $B=0.3$ out to $K<15$, and a plane-parallel Galaxy, we should recover a unique dependence for Galactic star fluctuations, $\sqrt{C(0)} \propto ({\rm cosec}\|b\|)^2$. This value of $B$ is consistent with the dependence of the fraction of clipped pixels and the amplitude $\sigma$ of the remaining fluctuations in the DIRBE data. Extrapolating to ${\rm cosec}\|b\|=0$ {#ss44} ----------------------------------- The scatter in Figs. \[f5\] and \[f6\] is large, because the plane-parallel model does not describe the large scale structure of the Galaxy. We divided the sample of 384 patches into four latitude bins and plot them against Galactic longitude $l$ in Figs. \[f8\] and \[f9\]. The dependence on longitude is significant for latitudes even as high as $\|b\| \sim 60\deg\ $. However, for Galactic longitudes between 90 and 270 and high latitudes there is almost no longitude dependence. We therefore selected data with $90\deg \leq l \leq 270\deg $, and plot the dependence of $\sqrt{C(0)}$ on $\cosec \|b\|$ in Figs. \[f10\] and \[f11\]. The scatter is much reduced. We tried several fitting functions for the 127 patches for which $\|b\|>20\deg$, $90\deg < l < 270 \deg$: $$\sqrt{C_{\rm fit}(0)} = a + A ({\rm cosec}\|b\|)^{p}, \label{e26}$$ where $a$ presumably contains the cosmological and instrumental parts of the signal, $$C_{\rm fit}(0) = a^2 +A^2 ({\rm cosec}\|b\|)^{2p}, \label{e27}$$ $$\sqrt{C_{\rm fit}(0)} = a+\sum_{i=1}^2A_i({\rm cosec}\|b\|)^i , \label{e28}$$ and $$C_{\rm fit}(0) = a \exp(A {\rm cosec}\|b\|), \label{e29}$$ minimizing, for example, $ \langle [(C(0) - C_{\rm fit})/C(0)]^2\rangle$ with respect to $a, A, p$. The fits for Eq. (\[e27\]) are shown in Figs. \[f10\] and \[f11\] and in Table 1. Except for the color-subtracted $[2-3]$ maps, all fits give positive values of $a$. The color-subtracted maps $[2-3]$ and $[3-4]$ have large scatter in the color indices at which the variance $C_\Delta(0)$ is minimized, and do not allow for a robust determination of $a$. Note that $p>1$ as expected from the model. The star and CIB fluctuations add in quadrature, as we demonstrated by simulation, so functional forms such as Eq. (\[e27\]) are better justified than the other fits, although all give consistent results. The last row in Table 1 summarizes our limits on $a$ with errors corresponding to the extreme range from 92% confidence levels from all the fits, around our preferred central value from Eq. (\[e27\]). We estimated the statistical uncertainty as follows. We define a relative variance $\sigma_0^2\equiv \min \langle [({\rm data} - {\rm fit})/{\rm data}]^2\rangle$ and a normalized $\chi_N^2(a, A, p) \equiv \langle [({\rm data} - {\rm fit})/{\rm data}]^2\rangle/\sigma_0^2$. Fig. \[f12\] plots deviation histograms vs. $(\delta_{\rm fit}/\sigma)^2$ where $\delta_{\rm fit} \equiv (C(0) - a^2 - A^2 x^{2p})/\sqrt{C(0)}$. For purely Gaussian deviations the histograms would be straight lines of slope $-1/2$. We plot contours for $\Delta \chi^2 = 7$ which for 3 parameters $(a,p,A)$ corresponds to a confidence level of 92%. Fig. \[f13\] shows thus determined confidence contours projected onto the $(a, p)$ plane for Bands 1 to 4, according to Eq. (\[e27\]). The uncertainties shown in Table 1 correspond to the largest span of $a$ in the panels in Fig. \[f12\]; for any given value of $p$ the corresponding uncertainty levels are reduced significantly. This is the reason for the smaller formal uncertainties on $a$ when an exponential fit - Eq. (\[e29\]) - is assumed. Excluding from the analysis the patches that lie close to the Ecliptic plane further reduces the scatter in Figs. \[f10\], \[f11\], but the contours are almost identical to those in Fig. \[f13\]. Likewise, keeping only the patches at $120 \deg < l < 240 \deg $ produces the same result. The results for = 7 and 5 are consistent but with a larger uncertainty, whereas for = 3 the results are close to those shown in Table 1. The values for $p$ are in good agreement with the expectations from the star counts. For Band 2, the maximal spread is $1.5 \leq p \leq 2.6$, which corresponds to $0.26 \leq B \leq 0.34$ in good agreement with Fig.\[f7\]. To summarize, we have justified simple power law star count models and a plane parallel model of Galactic fluctuations by comparing the predictions to the measurements. The model applies only far from the Galactic center and Galactic plane. There is a statistically significant residual term in the DIRBE data after extrapolation to zero cosec$\|b\|$, with consistent values for four different fitting functions and for different . Plate 1 of our Paper II shows visually that there is no obvious structure in the selected anti-center $\|b\|>20\deg$ data set. We checked the validity of our simple models by studying the residuals of the fits. Fig. \[f14\] shows the residuals for Eq. (\[e27\]) for each patch fluctuations, $\sqrt{C(0)}$ in Band 2. There is no apparent correlation with Galactic longitude or latitude and the residuals are similarly independent of the ecliptic coordinates. To consider whether the instrument itself or the zodiacal light could be a source for this residual, we review the estimates in Paper I. There we constructed maps of the differences between different weeks of observation, and found noise levels $\sqrt{C(0)}\sim$ 1.5, 0.3, 0.1 2sr in Bands 1 through 3 for = 3.5. These are much smaller than $a$ in Table 1, and it is unlikely that $a$ is due to noise. Similar arguments apply to errors in the zodiacal light modeling, which varies from week to week, and would contribute to the measurement error calculation. In any case, in the near-IR bands the contribution of the zodiacal light is small. It should be largest at 12 and 25 , where the residual fluctuations are below 1 2sr (Paper II and Sec.\[s5\] of this paper), or $< 1 \%$ of the total foreground. Therefore, the near-IR zodiacal modeling errors should be quite negligible compared to $a$. Furthermore, zodiacal light has a sharp cusp near the ecliptic plane, and zodiacal model errors would reflect this spatial dependence. We tried excluding patches in or near the ecliptic plane and found no change in $a,p$. The spatial correlation function that we find for = 3 .5 (cf. Fig. \[f19\]) is significantly different from that expected for zodiacal light. The zodiacal light is smooth except near the ecliptic plane, where dust resonances and asteroid family debris are found. We conclude that the single band plots all indicate a positive and approximately isotropic residual term that is unlikely to be produced by either instrumental noise or errors in the zodiacal modeling. Since independent contributions add linearly to the combined variance, such a component would contribute only $\sim10\%$ to the total dispersion of the confusion noise at the faint end of the flux distribution plotted in Fig. \[f7\], and would not be detectable there. The Figure shows two Gaussians differing in dispersion by 10%, and they both seem reasonable fits. Analytic and Numerical Modeling {#ss45} ------------------------------- Since the measured residual $a$ can not be explained by known errors, we must investigate the stellar foreground fluctuations more carefully. To do so, we simulated the confusion noise process both analytically and numerically. Our analytical approach assumes a power law for $dN/dm$, with a Fourier method to simulate the histogram of Fig. \[f7\]. Let ${\cal P}(F)$ be the probability distribution function to find a single star in the line of sight with flux $F$. The probability distribution function and its characteristic function $G(f)$ are a pair of one dimensional Fourier transforms, $G(f) = \int {\cal P}(F) \exp(ifF) dF$. The probability distribution function of finding two stars is the convolution of ${\cal P}$ with itself and the characteristic function for it is $G^2(f)$. Similarly, for $n$ stars the characteristic function is $G^n(f)$. For sufficiently large $n$ the characteristic function tends to a Gaussian, which can be seen from expanding $G(f) \simeq 1 - f^2 \langle F^2\rangle$, (Chandrasekhar 1954). For a Poisson distribution with $m$ stars per pixel on the average, the probability to find $n$ stars in a given pixel is ${\cal P}(n)=m^n \exp(-m)/n!$. Then the complex Fourier transform of the Poisson distribution of stars with an average of $m$ stars per pixel is $\exp(m(G-1))$. For many stars and sufficiently small $f$ or large fluctuations $G(f) \simeq 1-\frac{1}{2}f^2 \langle F^2 \rangle$, and the previous expression converges to a Gaussian distribution whose width $\propto m^{-1/2}$. This prescription was implemented numerically assuming ${\cal P}(F) \propto 10^{- 2.5B\log_{10} F}$ with $B=0.3$. Convolution of the predicted star histogram with Gaussian measurement noise and Gaussian cosmic background fluctuations can be included by multiplying $G(f)$ by a Gaussian. The effective beam size sets the maximum value of the apparent $dN/dm$ in the confusion noise region. We are able to reproduce Fig. \[f7\] very well, including the negative fluctuations and the transition from the confusion noise region to direct star detection. The result is robust in that the effective value of $B$ measured at the bright end of the distribution is not altered by subtracting the mean value, and clipping with various has no effect on the value of $B$ at the bright end. The apparent number of bright stars is about 10% larger than the input value, owing to the bias introduced by undetected stars of medium brightness near the detected bright stars. We also want to know whether there is any feature of our processing algorithm that could produce a spurious residual fluctuation $a$. For this, we need a numerical simulation of the sky, including both stars and possible cosmic terms. We used the DIRBE star model of Arendt et al. (1998), who implemented the 96-component star population model of Wainscoat et al. (1992), including spatial distribution models for the disk, bulge, halo, ring and arm populations. These were evaluated using the K-band $dN/dm$ for the central pixel of each patch, and a 2-dimensional uniform random number generator, and the stars out to $K=20$ were placed at the centers of the pixels in the map. The flux from each star was then distributed among 9 pixels according to the measured or assumed beam shape, using delta-function, uniform, and Gaussian beam profiles. If the correct beam profile is used, the simulated maps processed with our algorithms match the DIRBE histograms in each patch very well. We also simulated CIB fluctuations, starting with Gaussian fluctuations in Fourier space with power spectrum $P(k)=k^{-1.3}$, as expected on the smallest DIRBE scales, and then multiplied by the top-hat beam window function determined in Sec. \[s3\]. The resultant field was then Fourier transformed to real space and normalized to the modeled cosmic variance, $\sigma_{\rm sky}$. We added these maps to the simulated star fluctuation maps and examined the results. As was discussed in Sec. \[ss42\], our clipping algorithm clips effectively about $\sim 1\ \sigma$ below the nominal value of , because each star is seen in multiple pixels, each with its own noise fluctuations. Furthermore, according to Eq. (\[e25\]), there should be no variation with cosec$\|b\|$ in the number of stars clipped in each patch if the star counts follow the same $B$ over the relevant range of magnitudes. To test this we computed the effective $N_{\rm cut,eff}=\sqrt{\frac{5}{3}N_>^{-1}}$ according to Eq. (\[e25\]) assuming $B=0.3$. The number of clipped stars, $N_>$, is half the number of clipped pixels since our mask has twice the beam area. The results are plotted against cosec$\|b\|$ in Fig. \[f15\] for both the real sky and the model Galaxy. The data show that in the anti-center quadrants outside the Galactic plane, there is no trend with cosec$\|b\|$, and the effective values of have very small dispersion. For the real sky, the mean in this range is $N_{\rm cut,eff}=2.26$ and the dispersion is $\sigma_{\rm cut} = [\langle N_{\rm cut,eff}^2 \rangle - \langle N_{\rm cut,eff} \rangle^2]^{1/2}=0.06$. For the model sky, the mean $N_{\rm cut,eff}=2.27$, and the dispersion is 0.07. The effective is independent of the simulated beam properties. We found a simple way to test the assumption of a plane parallel Galaxy. As Eq. (\[e21\]) shows, the quantity $x^{-3}({dN}/{dm})\|_x$ plotted against $m-5\log_{10} x$ should be independent of $x= \cosec\|b\|$ and be equal to the counts at the Pole. We computed histograms for 14 patches of $64 \times 64$ pixels at $\|b\|>20 \deg$, $\|\beta_{\rm ecl}\| > 30\deg$ and $90\deg < l < 270\deg$. The patches were clipped at = 3.5 and the average flux for the remaining pixels of each patch was subtracted from the map. The resultant distribution of $x^{-3} dN/dm$ is plotted in Fig. \[f16\] for K band, with Poisson error bars for those points that contain at least 25 pixels (or 5.5 stars). The figure confirms that the stars are distributed in a plane-parallel way for this data set, with $B\simeq 0.3$. The plot includes lines for the model Galaxy, with error bands. They agree with the observations within the errors except at the bright end, where the model is slightly low. We simulated skies with a cosmic term of $\sigma_{\rm sky} = 5$ 2sr for the same $64 \times 64$ patches. The observations and the simulated model are plotted in Fig. \[f17\], with fluxes measured in units of $\sqrt{C(0)}$. If all the fluctuations are due to stars drawn from a power law distribution of the same slope, then all of the patches should follow the same line on this diagram. The DIRBE data and the synthetic Galaxy model look nearly identical, and there is no noticeable difference between $\sigma_{\rm sky} =5$ 2sr and 0. The DIRBE data and the models could be matched even better by adjusting the beam shape for the synthetic maps. The simulated confusion noise does not affect the amplitude and slope of the star counts at $m_K < 5.5$. Using our simulated models we also tested the scaling of the amplitude of the residual variance at given as a function of the input isotropic component $\sigma_{\rm sky}$. This is important because our clipping algorithm might remove both cosmic and star fluctuations in a complicated and possibly non-linear way. We find that even at = 3.5, the residual variance $C(0)$ is to good accuracy the sum of the variances from stars and the simulated CIB contribution. This shows that Eq. (\[e27\]) is better justified than the other choices for extrapolating to cosec$\|b\|=0$. The choice of beam shape - Gaussian, delta-function, or uniform - has no effect on the effective , but does lead to a systematic change in the residual $C(0)$, with the largest beam area having the highest variance. Similarly, changing the size $f_{\rm size}$ of the lower envelope region used by the point source recognition algorithm has no affect on $N_{\rm cut,eff}$, but leads to a systematic dependence of $C(0) \propto 1/f_{\rm size}$. Clearly, the larger the value of $f_{\rm size}$, the more diligently the algorithm recognizes point sources. We used $f_{\rm size}=5$, corresponding to about 5 DIRBE beams. We can now predict what the Galactic star counts ought to be in order to reproduce the $C(0)$ vs. $x = {\rm cosec}\|b\|$ relation found in the previous section. Combining Eqs. (\[e21\]) and (\[e22\]) shows that for a plane-parallel model the star contribution to the fluctuation is related to $dN_P/dm$ at the Pole via $$x\sigma^2(x) = F_0^2 \int_{m_L - 5 \log_{10} x}^\infty 10^{-0.8y}\frac{dN_P}{dy} dy, \label{e30}$$ where $F_0$ is the zero magnitude flux; for Band 2 it is 6050 2sr per DIRBE pixel. The lower magnitude for our clipping method is given by $m_L = -2.5 \log_{10}[f_m N_{\rm cut}\sigma(x)/F_0]$. Here $f_m$ is a factor accounting for the beam and the lower envelope used; based on the discussion in the previous paragraph we expect $f_m \sim 0.5-0.6$. Differentiating both sides of Eq. (\[e30\]) leads to $$\frac{dN_P}{dm} = 10^{0.4m} \;\; \frac{0.4\ln10}{f_m N_{\rm cut} F_0} \;\; \frac{\partial[x\sigma^2(x)]/\partial x}{\partial[x^2\sigma(x)]/\partial x}. \label{e31}$$ Here the right-hand-side should evaluated for $x$ given by: $$m(x)=-2.5\log_{10}[f_m N_{\rm cut}x^2\sigma(x)/F_0]. \label{e32}$$ Eqs. (\[e31\]), (\[e32\]) form a closed set to determine the star counts required to reproduce the $\sigma(x)$. They thus provide an important consistency check between the measured effective , the $C(0)$ - $x$ relation, and the star counts at the Galactic Poles. They also show whether the cosec$\|b\|$-independent part of $C(0)$ can be produced by the observed stars. We can measure the $f_m$ factor by normalizing the recovered star counts to the DIRBE data at $m_K=4$, where the confusion noise is negligible, and find $f_m \simeq 0.6$. The solid line in Fig. \[f7\] shows the recovered star counts at the Pole from $C(0)$ given by Eq. (\[e27\]) with $a=0$ with parameters ($A,p$) taken from Table 1. The agreement between the star counts inverted from the observed $C(0)$ - $x$ relation and the actual data is remarkably good, considering the statistical uncertainty in $A,a$ and $p$, and given that all three ($C(0)$ vs. $x$, and $dN_P/dm$) were determined independently. The dashed line in Fig. \[f7\] shows the star counts required to reproduce Eq. (\[e26\]) with $a = 4.8$ 2sr according to Eqs.(\[e31\]), (\[e32\]). The line overshoots the data by several standard deviations at $m_K >5$ and shows that this value of $a$ cannot be produced by the observed Galactic stars. We can now use the simulated sky maps to test the extrapolations to $\cosec\|b\|=0$. We constructed simulated data for 384 patches of $32 \times 32$ pixels which contained both the K Band Galaxy synthetic model, and a contribution from the CIB with $\sigma_{\rm sky}$ varying from 0 to 20 2sr . Processing the simulated maps with the standard algorithm, we find that the model without the CIB term is a good match to the DIRBE data at intermediate latitudes, but is significantly steeper as cosec$\|b\| \rightarrow 0$ at $\|b\| > 45\deg$, and has a zero intercept within the statistical errors. On the other hand the simulated data curve at $\|b\| > 45\deg$ flattens out for positive values of $\sigma_{\rm sky}$. We also tried fitting the simulated sky fluctuations to the observations, using $C(0)= A_m \sigma^2(\sigma_{\rm sky}) + a_m$, and found the effective gain and offset $A_m,a_m$. This approach has the advantage of including all that is known about the geometrical shape of the Galaxy, since the simulated maps use the Wainscoat et al. (1992) shapes for the disk, bulge, halo, ring and arm. If the model includes a cosmic term of the correct amplitude, then we should find $A=1$, $a_m=0$. We find that for a model $\sigma_{\rm sky}=0$ the value of $A_m = 1.4\pm 0.3$, consistent with unity, and the numbers for $A_m$ for positive values of $\sigma_{\rm sky}$ are similar. We could achieve $A=1$ by better modeling of the DIRBE beam size, since we have already shown that relationship. We also confirm the values for the residual fluctuations $a$ in Table 1. If the modeled sky has $\sigma_{\rm sky}=0$, we find $\sqrt{a_m} = 7.3^{+2.0}_{-2.8}$ 2sr in agreement with Table 1. Conversely, if we choose a model $\sigma_{\rm sky}$ similar to $a$ in Table 1, then $a_m$ should be consistent with zero. We find $a_m = 0$ with the 92% confidence level for 2.8 2sr $< \sigma_{\rm sky} <$ 8.3 2sr with the central value lying at $\sigma_{\rm sky}=5.2$ 2sr . We can imagine only one possible feature in the DIRBE data processing that could lead to a false conclusion about the latitude dependence. We have shown that for a range of fixed beam size and shape, the sky models all predict a dependence of star fluctuations on latitude that extrapolates to zero fluctuations at zero $\cosec\|b\|$. There is however some possibility that the effective beam size might depend on the line of sight and hence possibly on Galactic latitude. To explain the measurements, we would have to find an effect that systematically changes the effective beam area by an amount of the order of 30%, comparable to the fraction of the fluctuations at the Galactic pole that seem to be of cosmic origin. The DIRBE team measured the actual beam size very carefully from transits of bright stars through the beam, and did not find a significant dependence of beam profile on time or direction. However, as noted elsewhere, the effective beam profile for measuring fluctuations depends on several other effects. The effective beam area is approximately $\Omega = 2 \pi (\theta_{\rm beam}^2/2 + \theta_{\rm pixels}^2 + \theta_{\rm pointing}^2)$, where $\theta_{\rm beam}$ is the measured beam radius, $\theta_{\rm pixels}$ is the rms radius of the pixels, and $\theta_{\rm pointing}$ is the rms (vector) pointing error. The rms pixel size is $d_{\rm pix}/6^{1/2} = 0.132\deg$ for a square pixel of side $d_{\rm pix}$, but for a rhombus of the same area and a 60corner angle, it is increased by a factor of $(2/\sqrt{3})^{1/2} = 1.075$ to 0.142. Ignoring pointing error, this increases the effective beam area by 3.2%, a negligible amount in this context. The measured DIRBE pointing error is 1.5 arcmin (1 $\sigma$), and increases the effective beam area by a fixed 3.9%. If, however, the pointing error were much larger than indicated by the statistics of the residuals from the pointing solution fits, and in addition were strongly dependent on Galactic latitude, the effect could be important for us. The pointing solution was the subject of extraordinary scrutiny by the COBE team, and such errors would have been noticed. We conclude that the latitude fitting method is not subject to errors due to changes of the effective beam size with latitude. Power spectrum {#ss46} -------------- Although the dominant spatial structure of the near IR maps is simply the white noise of stars, it is interesting to see whether large scale averages could reveal a CIB component. We computed both power spectra and angular correlation functions, and the results are shown in Figs.\[f18\] and \[f19\]. We describe the power spectra first. We used 96 patches of 64$\times$64 pixels (or 19$\times$19), and computed spatial power spectra without star clipping, and with $N_{\rm cut} =$ 7 and 5, leaving over 90% and 80% of the pixels remaining at high latitudes. Smaller left too few pixels for reliable power spectra, and showed significant effects of the masking. We computed power spectra for single bands and for the color-subtracted maps, $\delta_1 - \beta \delta_2$, with $\beta$ evaluated for each patch according to Eq. (\[e15\]). After each power spectrum was calculated, it was divided by the beam window function discussed in Sec. \[s3\] to take out the instrument signature. The fact that the power spectra are approximately flat confirms that stars are the dominant sources of fluctuations for = 5, and that the effective window function, accounting for pixelization, map distortion, beam smearing, and pointing errors, is correct. The single band power spectra also show large-scale gradients produced by the Galactic structure. The amplitude of the power spectrum decreases with $N_{\rm cut}$, but the overall shape does not change appreciably, indicating that the beam mask effects are negligible for these values of . The amplitudes of the power spectrum for all angular scales are too large, and the shape too wrong, to allow for detection of the CIB structure. However, for some patches at $N_{\rm cut}=5$ the spectrum is close to that expected from Table 1. For of 5 or more, we can not extrapolate these power spectra to cosec$\|b\| =0$ for most of the scales probed by $P(q)$, and for smaller there are too few pixels left. Fig. \[f18\] shows the power spectrum for patch No. 7, at Galactic $(l,b)=(115\deg, 61\deg)$, and ecliptic $(\beta_{\rm Ecl}, \lambda_{\rm Ecl})$ = $(56\deg, 163\deg)$. In the single bands, the amplitude of the power spectrum decreases strongly after point source removal, but the shape remains approximately the same. The solid line in each panel shows the CIB signal according to Table 1, assuming that the CIB power spectrum has $P \propto q^{-1.3}$. This is a valid approximation for scales below one degree, but on large scales the power index of the CIB power spectrum may be different. The power spectrum of the foreground exceeds the estimated CIB by only a modest factor, particularly in Band 4. An instrument with a smaller beam might detect these fluctuations directly. Color subtraction significantly reduces the foreground structure, as shown in Fig. \[f18\]. Without clipping, we find a reduction by a factor of $\sim 30$ between adjacent bands, although the spectra are still flat from star fluctuation noise. Galactic stars are dominant in Band 4 and zodiacal and cirrus emission are dominant in Band 5, so these bands are not strongly correlated. For this patch, color subtraction of Bands 4 and 5 shows structure dominated by cirrus dust, which has $P \propto k^n$ with a steep $n=-2$ to $-3$; see Sec. \[s5\]. The solid line in the lower panels shows the estimated small-scale CIB between Band 1 and Band 2 with logarithmic slope of $-1.3$ and $C(0)$ taken from Table 1. In Bands $[2-3]$, we plot $\sqrt{C(0)}=$2 2sr . For lower values of $N_{\rm cut}$ and $q^{-1} > 1\deg $, the amplitude of $P(q)$ is comparable to that expected from the CIB according to Table 1. Furthermore, the slope of the power spectrum for this patch flattens significantly in the color-subtracted maps, and at some scales and bands approaches the logarithmic slope expected from the CIB. However, inspection of the power spectra shows that even in the color-subtracted maps, we do not detect a cosmological signal for $\geq 5$. The spectra are all consistent with the white-noise stellar distribution and large-scale Galactic gradients. To check for a cosmological power spectrum consistent with Table 1 we would have to clip to much lower levels of where beam masking problems prevent reliable determination of the power spectrum. For $<5$, it is better to compute angular correlation functions, which are unbiased by masking effects, even for = 3.5 where less than 50% of pixels remain. We evaluated $C(\theta)= \langle \delta F({\mbox{\boldmath$x$}} +\theta) \delta F({\mbox{\boldmath$x$}}) \rangle$, averaging over the remaining pairs of pixels separated by angular distance $\theta$. (The points at different $\theta$ are not statistically independent of each other. For = 7 and 5 the correlation function is that of white noise, a delta function at zero lag, with a level consistent with the power spectrum plotted in Fig. \[f18\]. It remains close to the zero-lag value on scales inside the beam, $\theta < 0.5^\deg$ and rapidly drops to very small (positive or negative) numbers on larger scales. For = 3.5 for patches with low values of $C(0)$ the correlation function flattens significantly, which would be consistent with CIB structure at the levels of Table 1. Fig. \[f19\] shows the correlation function for Patch 7 at = 3.5, for 10 bins of 1 degree width. The straight line illustrates the slope of a CIB correlation function $\propto \theta^{-0.7}$. The correlation function amplitude and slope are consistent with Table 1 and a CIB interpretation. Large-scale gradients in the star and dust populations are responsible for the positive correlation function at very large angles. A further comparison between the results in Table 1 and the power spectrum analysis can be made in the following way. For = 5 at each angular scale $q^{-1}$, we evaluate the minimal value of $P(q)$ for all the patches, and plot the value of a typical fluctuation, $\sqrt{q^2P(q)/2\pi}$, versus $\theta= \pi/q$ in Fig. \[f20\]. The slope of the data is roughly that of white noise, $\sqrt{q^2P(q)/2\pi} \propto q$, implying that we are still seeing Galactic stars. The upper limits over angular scales of 2 $< \theta <$ 15 are: $$\delta F_{rms}(\theta) \leq A (\frac{\theta}{5\deg})^{-1}, \label{e33}$$ with $A$ = 6, 2.5, 0.8, 0.5 2sr for Bands 1 to 4 respectively. The process of finding an all-sky minimum of $P_2(q)$ produces a very smooth curve whose uncertainties are predominantly systematic. These are strong constraints on galaxy evolution. The shaded areas represent the power spectra according to Table 1, assuming that the CIB $P(q) \propto q^{-1.3}$. Table 1 is consistent with the upper limits given by Eq. (\[e33\]), except in Band 4. In Band 4, Table 1 would require a small scale for the turn-over in the CIB power spectrum, implying that much of it comes from high redshifts. Results in mid- to far-IR Bands {#s5} ================================ Foregrounds and $C(0)$ analysis {#ss51} ------------------------------- At wavelengths greater than 10 , dust in the Solar system and the Galaxy produce most of the foreground emission. These sources are smooth on small scales, so do not necessarily prevent detection of CIB fluctuations. Odenwald, Newmark and Smoot (1998) detected $< 100$ nearby galaxies in the DIRBE data at wavelengths greater than 10 , but with the exception of M31 and the Large and Small Magellanic clouds, the galaxies are unresolved. Because there are few discrete sources in the mid and far-IR DIRBE data, we can clip the DIRBE maps to lower values of $N_{\rm cut}$ and keep the same number of pixels as for the near-IR bands. Clipping to low values of $N_{\rm cut}$ would remove some fluctuations in the CIB as well as the foreground. If the CIB zero-lag signal were as high as the upper limit found in the maps, clipping at $N_{\rm cut }=2$, 2.5 or 3 levels would decrease the real $C(0)$ by 20%, 10% and 0.001% respectively. Given that we find only upper limits, clipping down to $N_{\rm cut}=2$ is safe but lower levels would require interpretation. Fig. \[f21\] shows histograms of the number of pixels remaining in each patch of 32$\times$32 pixels after removing point sources with $N_{\rm cut}=2$. Because of the extended nature of the foreground emission in these bands, we removed large-scale gradients first, using polynomials of order up to 4. Fig. \[f22\] shows the variation of the residual $C(0)$ with cosec$\|b\|$ at = 2 for $90\deg < l < 270$. A significant fraction of the foreground emission still comes from the zodiacal light, even after subtraction of the DIRBE zodiacal light model, and contributes to the large scatter in the plots. Fig.\[f22\] also shows the dependence of $C(0)$ on Galactic latitude for the 111 patches that also have Ecliptic latitude $\|\beta_{\rm Ecl}\| > 25\deg$. The scatter is reduced but not enough to extrapolate to cosec$\|b\|$=0. In Bands 5 and 6 there is only a weak dependence on $b$, and the amplitude of the typical fluctuations seen in these bands changes by only $\sim 50\%$ between cosec$\|b\| = 1$ and 3, showing approximate isotropy for $\|b\|\geq 20\deg$. We interpret the lowest values of $\sqrt{C(0)}$ as upper limits on the CIB fluctuations. They are slight improvements over Paper II, and are shown in Table 2. In Bands 7 and 8 the dependence on Galactic latitude is more prominent. Extrapolation to cosec$\|b\|=0$ with Eq. (\[e26\]) gives values of $a$ in agreement with the lowest $\sqrt{C(0)}$ shown in Table 2, but with significant error bars. Therefore we again interpret the derived $C(0)$ as upper limits on the CIB fluctuations. The slope of $C(0)$ with Galactic latitude is consistent with a plane parallel Galaxy distribution, i.e. $\sqrt{C(0)} \propto {\rm cosec}\|b\|$. The fluctuations are approximately proportional to the total brightness. The color diagrams at these wavelengths are not as clean as in the near-IR. In many patches no color correlations between the adjacent bands exist, and the color indices show large variations across the sky. Local variations in the parameters of the Galactic and solar system dust (e.g. density and temperature) are expected on large angular scales. No improvement in the CIB limits was achieved with mid-IR color subtraction. Mid- and Far-IR Power Spectra {#ss52} ----------------------------- We computed power spectra for the same 96 patches of 64$\times$64 pixels used for the near IR bands. Both single band power spectra and color-subtracted maps show a weak decrease with clipping threshold, a consequence of the extended character of the foreground emission. The reduction of the foreground with the color subtraction is not as large as in the near-IR bands. In order to preserve information about $P(q)$ at all angular scales, no gradients were removed before power spectrum analysis. The power spectra have a shape typical of the known cirrus distribution (cf. Gautier et al. 1992). The spectrum $P(q)$ is steep: $P(q) \propto q^{-n}$ with $n\simeq (2.5-3).$ This is consistent with little small scale structure and, hence, no strong dependence in the resultant $C(0)$ on $N_{\rm cut}$. It is also consistent with the power spectrum of cirrus emission measured on arc-minute and degree scales (Gautier et al. 1992; Wright 1998). Fig. \[f23\] shows power spectra for the same Patch 7 used in the near-IR analysis. It shows the steep power spectrum typical of the cirrus distribution with $P(q) \propto q^{-3}$ for most of the scales. Note that some of the fluctuations at 12 and 25 can be due to errors in modeling the zodiacal light. As in the near-IR analysis, we evaluated the minimal values of $P(q)$ for each angular scale $q^{-1}$ across the entire sky. Fig. \[f23\] shows these minimal values of the fluctuation $\sqrt{q^2P(q)/2\pi}$ as a function of angular scale $\pi/q$. The minimal values come close to, but are not as small as, those in Table 2, which were evaluated at =2 after gradient subtraction. Even in the patches with the least fluctuations, the power spectrum is still as steep as $P\propto q^{-2}$, leaving the fluctuation $\sqrt{q^2P(q)/2\pi}\simeq$ const. This is indicative of cirrus emission in all areas of the sky. Infrared sky surveys with higher angular resolution should be able to reduce this contribution, and possibly uncover the CIB fluctuations. Conclusions {#s6} =========== Fig. \[f24\] summarizes our results. At 1-5 , we find a positive residual fluctuation by extrapolating $C(0)$ to zero $\cosec\|b\|$. Based on detailed numerical and analytic models, this residual is not likely to originate from the Galaxy, our clipping algorithm, or instrumental noise. We conclude that this extra variance may result from structure in the CIB. The variance found in this way from individual DIRBE Bands 1-4 is plotted with diamonds, with 92% uncertainty levels from Fig.\[f13\]. The results for color-subtracted maps are plotted with triangles at wavelengths halfway between the bands used. We find a positive residual in the color-subtracted map \[1-2\] between Bands 1 and 2, but not in other color subtracted maps. The color-subtracted map \[1-2\] has a unique color of $\beta_{12}\simeq 2$ with little variation across the sky and the limit measures $[\langle(\delta F_1 - 2 \delta F_2)^2\rangle]^{1/2}$. Taken at face value, these high values of the near-IR CIB fluctuations, if produced by evolving normal galaxy populations, would require substantial CIB fluxes. These would have to be above the estimates from the K-band galaxy counts, but are below the upper limits found by Hauser et al. (1998). The upper limits on CIB fluctuations at 10 - 100 are plotted with arrows, and are below 1 2sr , with $\sqrt{C(0)} < 0.5-0.7$ 2sr at 25 . These limits are lower than those in Paper II and imply strong constraints on how and when the early galaxies formed and evolved. On larger scales, $2^\deg < \theta < 15^\deg$, we obtain upper limits on the CIB fluctuations from the all-sky power-spectrum analysis: $(\theta/5^\deg)\times \delta F_{\rm rms}(\theta) < $ 6, 2.5, 0.8, 0.5 2sr in Bands 1-4 respectively. These limits, when taken in conjunction with our possible detection of the zero-lag CIB signal, limit the turn-over scale in the spectrum of the primordial density field to not much more than $\sim 100 h^{-1}$Mpc. We fully recognize the difficulty of finding small fluctuations in the presence of larger fluctuations from foregrounds. While we have found no local explanation for our results, it is still quite possible that the fluctuations are not of cosmic origin, but come from some fault of the instrument, the data processing, or an unexpected feature of the Galactic foreground. The best way to resolve this uncertainty is to get better data, such as from a higher resolution sky survey with exceptionally good attention to flat fielding. This may be possible with satellites like SIRTF or the proposed NGSS, or even from 2MASS data or rocket data. We acknowledge support from the NASA Long Term Space Astrophysics grant 399-20-61-02. We particularly thank the COBE DIRBE team, led by the principal investigator Michael Hauser and by Thomas Kelsall, for developing an exceptionally well calibrated and stable instrument, and for producing public data archives with the complex zodiacal light foreground models removed. We have benefited greatly from conversations with Richard Arendt on applications of star models to the DIRBE data and for a careful reading and constructive comments on the manuscript of this paper. Our special thanks go to Tom Jarrett for providing the 2MASS star counts data. We thank the referee, Michael Vogeley, for careful reading of the paper. Arendt, R. et al. 1998, Ap.J., [508]{},74. Bahcall, J. and Soneira, R., 1980, Ap.J.Suppl, [44]{}, 73. Baugh, C. & Efstathiou, G. 1993, MNRAS, [265]{}, 145. Beichman, C.A. 1996, in “Diffuse Infrared Radiation and the IRTS", eds. Okuda, H., Matsumoto, T. and Roellig, T., p.82. Beichman, C.A. & Helou, G. 1991, Ap.J., [370]{}, L1. Boggess, N. et al. 1992, Ap.J., [397]{}, 420. Bond, J.R., Carr, B.J. and Hogan, C. 1986, Ap.J., [306]{}, 428. Chandrasekhar, S. 1954, in “Noise and stochastic processes", ed. N. Wax, Dover. Cowie, L. et al. 1994, Ap.J., [434* ]{}, 114. Djorgovski, G. et al. 1995, Ap.J., [438]{}, L13. Dwek, E. et al. 1998, Ap.J., [508]{},106. Elias, J.H. 1978, A.J., [83]{}, 791. Fall, S.M., Charlot, S. & Pei, Y.C. 1996, Ap.J., [464]{}, L43. Fixsen, D.J., Dwek, E. , Mather, J.C., Bennett, C.L., & Shafer, R.A. 1998, Ap.J., [508]{}, 123. Franceschini, A. et al. 1991, Ap.J.Suppl., [89]{}, 285. Gautier, N. et al. 1992, A.J., [103]{}, 1313. Hauser, M. et al. 1998, Ap.J.,[508]{}, 25. Hacking, P.B. and Soifer, B.T. 1991, Ap.J., [367]{}, L49. Jarrett, T. 1998, IPAC/Caltech, private communication. Jimenez, R. & Kashlinsky, A. 1999, Ap.J., [511]{}, 16. (astro-ph/9802337). Kashlinsky, A. 1998, Ap.J., [492]{}, 1. Kashlinsky, A., Mather, J., Odenwald, S. and Hauser, M. 1996, Ap.J., [470]{}, 681. (Paper I) Kashlinsky, A., Mather, J., and Odenwald, S. 1996, Ap.J., [473]{}, L9. (Paper II) Kelsall, T. et al. 1998, Ap.J., [508]{}, 44. Loveday, S. et al. 1992, Ap.J., [390]{}, 338. Lilly, S. et al. 1996, Ap.J., [460]{}, L1. Lin, H. et al. 1996, Ap.J., [464]{}, L60. Maddox, S. et al. 1990, MNRAS, [242]{}, 43P. Malkan, M.A. and Stecker, F. 1998, Ap.J., [496]{}, 13. Martin, C. and Bowyer, S. 1989, Ap.J., [338]{}, 677. Mihalas, D. and Binney, J. 1981, “Galactic Astronomy", Freeman & Co., San Francisco. Minezaki, T. et al. 1998, A.J., [115]{}, 229. Moore, B. et al. 1994, MNRAS, [269]{}, 742. Odenwald, S., J. Newmark, & Smoot, G. 1998, Ap.J., [500]{}, 554. Partridge, B. & Peebles, P.J.E. 1967, Ap.J., [148]{}, 377. Peebles, P.J.E. 1980, Large Scale Structure of the Universe, Princeton University Press. Picard, A. 1991, Ap.J., [368]{}, L7. Saunders, W. et al. 1991, Nature, [349]{}, 32. Saunders, W. et al. 1992, MNRAS, [258]{}, 134. Shectman, S. 1973, Ap.J., [179]{}, 681. Shectman, S. 1974, Ap.J., [188]{}, 233. Shectman, S. et al. 1996, Ap.J., [470]{}, 172. Silverberg, R.F. et al. 1993, Proc. SPIE, [2019]{}, 180. Stecker, F., Puget, J-L., & Fazio, G.G. 1977, Ap.J., [214]{}, L51. Vogeley, M. 1998, astro-ph 9711209. Wainscoat, R.J., et al. 1992, Ap.J.Suppl., [83]{}, 111. Wang, B. 1991, Ap.J., [374]{}, 465. Weiland, J. et al. 1998, DIRBE Explanatory Supplement, http://www.gsfc.nasa.gov/astro/cobe/\ dirbe\_exsup.html Wright, E.L. 1998, Ap.J., [496]{}, 1. Yoshii, Y. & Takahara, F. 1988, Ap.J., [326]{}, 1. [c c c c c c c c c]{} Fit & Band 1 & Band 2 & Band 3 & Band 4 & Band 1-2 & Band 2-3 & Band 3-4\ & 1.25 & 2.2 & 3.5 & 4.9 & & &\ $\sqrt{C(0)}=a+A x^p$\ $a$ & $12.5^{+4.3}_{-5.7}$ & $4.8^{+1.8}_{-2.4}$ & $1.9^{+0.6}_{-0.7}$ & $2.0^{+0.2}_{-0.2}$ & $6.7^{+1.6}_{-2.5}$ & $--$ & $1.4^{+1.2}_{-1.6}$\ $p$ & $2.15^{+0.41}_{-0.41}$ & $2.18^{+0.41}_{-0.41}$ & $2.31^{+0.45}_{-0.45}$ & $2.79^{+0.61}_{-0.58}$ & $1.92^{+0.55}_{-0.54}$ & $--$ & 1.47\ $C(0)=a^2+A^2 x^{2p}$\ $a$ & $15.5^{+3.7}_{-7.0}$ & $5.9^{+1.6}_{-3.7}$ & $2.4^{+0.5}_{-0.9}$ & $2.0^{+0.25}_{-0.5}$ & $7.6^{+1.2}_{-2.4}$ & $--$ & $$\ $p$ & $1.78^{+0.27}_{-0.27}$ & $1.79^{+0.28}_{-0.29}$ & $1.89^{+0.33}_{-0.33}$ & $2.03^{+0.90}_{-0.78}$ & $1.49^{+0.34}_{-0.33}$ & $--$ & 1.23\ $\sqrt{C(0)}=a+ \sum_{i=1}^2A_ix^i$\ $a$ & 14.9$\pm5.4$ & 5.6$\pm2.2$ & 2.5$\pm1.2$ & 2.7$\pm1.2$ & $6.6\pm0.8$ & $--$ & $0.9\pm 0.9$\ $\sqrt{C(0)}=a\exp(A x)$\ $a$ & $9.7\pm 0.6$ & $3.7\pm 0.2$ & $1.3\pm 0.1$ & $1.3\pm 0.1$ & $5.3\pm 0.2$ & $3.6 \pm 0.2$ & $1.5\pm 0.1$\ $x = [\ln\sqrt{C(0)}$-$\ln a]/A $\ $a$ & $8.8\pm 1.3$ & $3.3\pm 0.3$ & $1.1\pm 0.1$ & $0.8\pm 0.1$ & $4.8\pm 0.6$ & $1.8 \pm 0.4$ & $1.2\pm 0.1$\ Summary\ \ $a$ & $15.5^{+3.7}_{-7.0}$ & $5.9^{+1.6}_{-3.7}$ & $2.4^{+0.5}_{-0.9}$ & $2.0^{+0.25}_{-0.5}$ & $7.6^{+1.2}_{-2.4}$ & $--$ & $$\ Note: $x = {\rm cosec}\|b\|$ [c c c c c]{} & Band 5 & Band 6 & Band 7 & Band 8\ & 12 & 25 & 60 & 100\ \ $\sqrt{C(0)}$ & 1.0 & 0.5 & 0.8 & 1.1\ Fig. \[f1\]: (a) Left: Linear scales of 0.5 in Eq. (\[e9\]) vs. z. The minimal scale is almost independent of cosmology. (b) Right: $\Delta(k) \equiv ({R_H^{-1}k^2 P_3(k; 0))}^{1/2}$ vs. $k$ for APM spectrum. The CIB minimal relative fluctuations on the DIRBE beam scale are $\sim \Delta$ at the minimal value of the linear scale. Fig. \[f2\]: Window function and beam profile for DIRBE Band 1 beam. Plus signs show embedding in a 256$^2$ pixel field, asterisks 512$^2$ pixels, and diamonds 1024$^2$ pixels. Solid line is a top-hat profile with a beam radius of $\vartheta=0.4$. Fig. \[f3\]: Histogram of pixels surviving clipping at = 3.5 in the near-IR DIRBE bands for 384 patches of $32\times32$ pixels. Solid line is Band 1, dotted Band 2, dashed Band 3, and dashed-dotted Band 4. Fig. \[f4\]: Histograms of color indices $\beta$ from Eq. (\[e15\]) for 384 patches. Thin solid lines are all-sky; thick lines are patches with $\|b\|>20$. Fig. \[f5\]: $\sqrt{C(0)}$ vs. cosec$\|b\|$ for Bands 1-4 and $N_{\rm cut}=3.5$. Fig. \[f6\]: Same as Fig. \[f5\] but for color subtracted maps. Fig. \[f7\]: K band $dN/dm$ and DIRBE pixel histogram for 128$^2$ pixels at NGP. Flux $F$ is absolute value of deviation from mean of patch after clipping with = 3.5, measured in 2sr . Positive deviations are + signs, diamonds negatives. Poisson error bars for DIRBE data assume 4.5 DIRBE pixels per star. Dotted lines are a Gaussian fit to data after clipping, and a 10% lower dispersion. Dash-dots are positive pixels remaining after clipping and dash-dot-dot-dot shows remaining negative pixels. Solid line shows counts inverted according to Eqs. (\[e31\]) and (\[e32\]) from Eq.(\[e27\]) without an isotropic component; long dashes are inverted counts if $a$ comes from stars. Filled triangles are differential counts from Elias (1978) NGP measurements. Filled circles are differential NGP counts from 2MASS (Jarrett 1998). Open triangles are cumulative 2MASS counts multiplied by $0.3\ln 10$ to convert to differential counts for $dN/dm \propto 10^{0.3m}$. Filled diamonds with error bars are South Galactic Pole counts from Fig. 1 of Minezaki et al. (1998). Fig. \[f8\]: Longitude dependence of $\sqrt{C(0)}$ for individual DIRBE bands at various Galactic latitudes. The increase of fluctuations towards the Galactic Center can be seen for any Galactic latitude. Fig. \[f9\]: Same as Fig. \[f8\] only for color-subtracted maps. Fig. \[f10\]: Plots of $\sqrt{C(0)}$ vs. cosec$\|b\|$ for J, K, L, M bands for $90\deg < l < 270\deg$. Solid lines are fits of Eq. (\[e27\]) using data for $\|b\|>20 \deg$. Fig. \[f11\]: Same as Fig. \[f10\] for color-subtracted maps. Fig. \[f12\]: Histogram of fit residuals from Eq. (\[e26\]) for Band 2 $\delta^2_{\rm fit}= [(C(0)-a^2 -A^2 {\rm cosec}^{2p} \|b\|)/C(0)]^2$ in units of $\sigma_0$. Fig. \[f13\]: 92% confidence limits on J, K, L, M and J-K for the fits to Eq. (\[e27\]). The plus sign is the most likely value of $(a, p)$ from Eq. (\[e27\]). Fig. \[f14\]: Scatter diagram for fit residuals $\delta_{\rm fit}\equiv [\sqrt{C(0)}-\sqrt{C_{\rm fit}(0)}]/\sqrt{C(0)}$ for Eq. (\[e27\]) for Band 2 versus $l$ and $b$. Fig. \[f15\]: Effective clipping vs. cosec$\|b\|$, according to Eq. (\[e25\]) for $p=2$ or $B=0.3$. Left is DIRBE data and right is simulated Galaxy model. Fig. \[f16\]: K-band DIRBE star counts in coordinates where a plane-parallel Galaxy would be a single line; $x=\cosec\|b\|$. Data are for $64 \times 64$ pixel patches with $\|b\| \geq 20 \deg$ and $90\deg < l < 270\deg$, with Poisson errors shown for $N_{\rm pix} \geq 25$. Confusion noise affects counts at $m_K >5.5$. Lines show the model Galaxy: solid is mean $x^{-3}dN/dm$ and dashes are the $\pm$1-sigma spread. Fig. \[f17\]: K-band star counts in the $64 \times 64$ patches outside the Galactic disk and away from the center, as a function of the absolute value of the flux deviation from the mean in units of $\sqrt{C(0)}$, for = 3.5. Left shows DIRBE data and right shows the simulated Galaxy and a CIB fluctuation of $\sigma_{\rm sky} = 5$ 2sr from Table \[t1\]. Plus signs are for negative $F$ and diamonds for positive $F$; they overlap for small $\|F\|$. Fig. \[f18\]: Near IR power spectra for Patch 7, well above the Galactic and Ecliptic planes. Plus signs are before point source removal, asterisks show $N_{\rm cut}=7$, and diamonds show $N_{\rm cut}=5$. Solid lines show $P(q)$ expected if the CIB power spectrum scales as $q^{-1.3}$ for $a$ in Table \[t1\]. Fig. \[f19\]: Angular correlation function evaluated for Patch 7 for = 3.5. The zero-lag value is plotted at $\theta=0.15^\deg$, and numbers below 0.1 2m4sr are not shown. The filled circle shows $a$ from Table \[t1\], normalizing a CIB correlation function with $C(\theta) \propto \theta^{-0.7}$. Diamonds show the absolute value of the correlation function in the negative range. Fig. \[f20\]: All-sky minimum fluctuation $\sqrt{q^2P(q)/2\pi}$ versus $\pi/q$. Shaded areas show the range expected from Table \[t1\], assuming the CIB power spectrum scales as $q^{-1.3}$. Fig. \[f21\]: Histogram of pixels surviving clipping at = 2 in the mid- to far-IR DIRBE bands for 384 patches of $32\times32$ pixels. Solid line is Band 5, dots Band 6, dashes Band 7, and dash-dots Band 8. Fig. \[f22\]: $\sqrt{C(0)}$ vs. cosec$\|b\|$ for Bands 5-8 and = 2. Top panels are 90 $< l <$ 270. Lower panels are $90\deg < l < 270\deg$ and $\|\beta_{\rm Ecl}\| > 25\deg$. Fig. \[f23\]: Upper panels are near-IR power spectra in for Patch 7. Lower panels are all-sky minimum fluctuations $\sqrt{q^2P(q)/2\pi}$ plotted vs. the scale $\pi/q$. Plus signs are without clipping, asterisks show $N_{\rm cut}=7$, and diamonds show $N_{\rm cut}=5$. Fig. \[f24\]: Summary. Diamonds are values for $a$ with 92% uncertainties from Table 1. Triangles are for the color subtracted maps, shown at the mean wavelength for the two bands. The $1-2$ limits are the left triangle and $3-4$ the right. Dashes with arrows are upper limits in Bands 4-8. =1.0 z =1.0 =1.0 =1.0 =1.0 =1.0 =1.0 =1.0 =1.0 =1.0 =1.0 =1.0 =1.0 =1.0 =1.0 =1.0 =1.0 =1.0 =1.0 =1.0 =1.0 =1.0 =1.0 =1.0	{ "pile_set_name": "ArXiv" }
--- abstract: 'A new effect is described by which primordial gravity waves leave a permanent signature in the large scale structure of the Universe. The effect occurs at second order in perturbation theory and is sensitive to the order in which perturbations on different scales are generated. We derive general forecasts for the detectability of the effect with future experiments, and consider observations of the pre-reionization gas through the 21cm line. It is found that the Square Kilometre Array will not be competitive with current cosmic microwave background constraints on primordial gravity waves from inflation. However, a more futuristic experiment could, through this effect, provide the highest ultimate sensitivity to tensor modes and possibly even measure the tensor spectral index. It is thus a potentially quantitative probe of the inflationary paradigm.' author: - Kiyoshi Wesley Masui - 'Ue-Li Pen' bibliography: - 'spires.bib' - 'externalrefs.bib' date: 'August 25, 2010' title: Primordial gravity waves fossils and their use in testing inflation --- [ [Introduction.]{}— ]{} \[s:intro\] It has been proposed that redshifted 21cm radiation, from the spin flip transition in neutral hydrogen, might be a powerful probe of the early universe. The era before the first luminous objects reionized the universe–around redshift 10–contains most of the observable volume of the universe, and 21cm radiation is the only known probe of these so called dark ages (see @astro-ph/0608032 for a review). The density of the hydrogen could be mapped in 3D analogous to how the cosmic microwave background (CMB) is mapped in 2D. The wealth of obtainable statistical information may allow for the detection of many subtle effects which could probe the early universe. In particular, the primordial gravity wave background, also referred to as tensor perturbations, are of considerable cosmological interest. Inflation robustly predicts the production of tensor perturbations with a nearly scale-free spectrum, however, their amplitude is essentially unconstrained theoretically. The amplitude of the tensor power spectrum is quantified by $r$, the tensor to scalar ratio. The current upper limit is $r<0.24$ at 95% confidence [@arXiv:1001.4538], however upcoming CMB measurements will be sensitive down to $r$ of a few percent [@arXiv:1003.6108]. The current limits on $r$ correspond to characteristic primordial shear on the order of $10^{-5}$ per logarithmic interval of wavenumber. Several probes of gravity waves using the pre-reionization 21cm signal have been proposed. These include polarization [@astro-ph/0702600] and redshift space distortions [@arXiv:0901.3655]. @astro-ph/0301177 considered the weak lensing signature of gravity waves and found that the signal is sensitive to the so called metric shear. This is closely related to the present work. Here we describe a mechanism by which primordial gravitational waves may leave an imprint in the statistics of the large scale structure (LSS) of the universe. This signature becomes observable when the gravity wave enters the horizon and begins to decay. [ [Mechanism.]{}— ]{} \[s:mech\] In the following, Greek indices run from 0 to 3 and lower case Latins from 1 to 3. Latin indices are always raised and lowered with Kronecker deltas. Commas denote partial derivatives, and an over-dot ($\dot{\#}$) represents a derivative with respect to the cosmological conformal time. Finally, we adopt a mostly positive metric signature $(-1,1,1,1)$. We start with an inflating universe with some distribution of previously generated tensor modes that are now super horizon. Scalar, vector and smaller scale tensor modes may exist but their contribution to the metric is ignored. The line element is given by $$\label{e:mech:CFmetric} ds^2 = a(\eta)^2\left[ -d\eta^2 + (\delta_{ij}+h_{ij})dx^i dx^j\right].$$ where $a$ is the scale factor, $\eta$ the conformal time and a spatially flat background geometry has been assumed. The metric perturbations $h_{ij}$ are assumed to be transverse and traceless and thus contain only tensor modes. The elements of $h_{ij}$ are also assumed to be small such that only leading order terms need be retained. The assumption that all tensor modes under consideration are super horizon implies that $k_h \ll \dot{a}/a$, where $k_h$ denotes the wave numbers of tensor modes. The frame in which the line element takes the form in Eq. \[e:mech:CFmetric\] will hereafter be referred to as the cosmological frame (CF). By the equivalence principle, it is possible to perform a coordinate transformation such that the space-time appears locally Minkowski at a point. New coordinates are defined in which the tensor modes are gauged away at the origin: $$\label{e:mech:LFFcoords} {\tilde{x}}^{\alpha}= (x^{\alpha}+ \frac{1}{2}h^{\alpha}_{{\phantom{\alpha}}{\beta}}x^{\beta}),$$ where the elements $h_{0{\alpha}}$ are taken to be zero. The metric now takes the form (up to first order in $h_{ij}$) $$\label{e:mech:LFFmetric} ds^2 = a^2\left[-d\eta^2+\delta_{ij}d{\tilde{x}}^i d{\tilde{x}}^j -{\tilde{x}}^c \partial_\alpha h_{\beta c} d{\tilde{x}}^{\alpha}d{\tilde{x}}^{\beta}\right].$$ This frame will be loosely referred to as the locally Friedmann frame (LFF), because in these coordinates the metric is locally that of an unperturbed FLRW Universe. We will give quantities in these coordinates a tilde (${\tilde{\#}}$) to distinguish them from their counterparts in the CF. It is seen from Eq. \[e:mech:LFFmetric\] that the local effects of gravity waves are suppressed not only by the smallness of $h_{ij}$ but also by $k_h/k$ where $k=L^{-1}$ and $L$ is some length scale of interest. This will be important in justifying some later assumptions. Note that for super horizon gravity waves, temporal derivatives are much smaller than spacial ones. On small scales, inflation generates scalar perturbations which are then carried to larger scales by the expansion. By the equivalence principle, physical processes on small scales can not know about the long wavelength tensor modes. As such these small scale scalar modes must be uncorrelated with the long wavelength tensor modes. We assume statistical homogeneity and isotropy in the LFF as would be expected from inflation. The power spectrum of scalar perturbations can then be written as a function of only the magnitude of the wave number, i.e., ${\tilde{P}}({\tilde{k}}_a)={\tilde{P}}({\tilde{k}})$. This applies only within the local patch near the point where the tensor mode was gaged away. The average in the definition of the scalar power spectrum is over realizations of the scalar map, but not the tensor map. In the CF, the isotropy is broken. Transforming back to cosmological coordinates maps ${\tilde{k}}_i \to k_i - k_j h_i^{{\phantom{\alpha}}j}/2$. The power spectrum becomes sheared: $$\label{e:mech:P} P(k_a) = {\tilde{P}}(k)-\frac{k_i k_j h^{ij}}{2k} \frac{d {\tilde{P}}}{dk} + O(\frac{k_h}{k}h_{ij})+O({h_{ij}}^{2}).$$ If the metric perturbations are not assumed to be traceless, the right hand side of this equation gains an additional term proportional to this trace. This deviation from isotropy is not observable since any possible observation would take place in the LFF. It is noted that the leading order correction to CF power spectrum is not suppressed by $k_h/k$. It is therefore not expected that the residual terms in the LFF metric (Eq. \[e:mech:LFFmetric\]) can break isotropy to undo CF anisotropy. However it was the CF in which the power spectrum should be isotropic, then there would be observable anisotropy in the LFF. This would be a violation of the equivalence principle, since an experiment local in both space and time would be able to detect the super horizon tensor modes by measuring the power spectrum of the locally generated scalar perturbations. We would now like to evolve the system to some later time when observations can be made. Ignoring the internal dynamics of the scalar perturbations, we solve for their evolution as if they were embedded in a sea of test particles. This is trivial since an object at coordinate rest in the CF will remain at rest for any time dependence of $h_{ij}$ (this is true at all orders). At some point well after inflation, when the universe is in its deceleration phase, the horizon will become larger than the length scale of the tensor modes. The tensor modes will then decay by redshifting, and after some period of time the metric perturbations $h_{ij}$ become negligible. The CF and LFF then become equivalent and both correspond to the frame in which observations can be made. The distribution of test particles is the same as it initially was in the CF. As such, the initially physically isotropic power spectrum now contains a measurable local anisotropy given by Eq. \[e:mech:P\]. The values of the initial metric perturbations can be determined by measuring this distortion at any time in the future, constituting a fossil of the initial tensor modes. The scalar perturbations remain Gaussian but become non-stationary, and the trispectrum gains the corresponding terms. This is analogous to the apparent distortions expected in the CMB and [21cm]{} fields induced by gravitational lensing. Similarly the bispectra of mixed scalars and tensors were calculated in @astro-ph/0210603, employing similar methodology to that presented here. The effect described here is a second order perturbation theory effect, in that it is a small effect due to tensor modes on the already small scalar perturbations. This coupling occurs in the initial conditions, not between the dynamics of the scalars and tensors. The simple argument presented above avoided the complication of a full second order calculation, but it is expected that such calculations would yield the same results. Specifically, an expression agreeing with Eq. \[e:mech:P\], to relevant order, was derived in @arXiv:1005.1056 [Eq. 4.5] as part of a longer calculation. [ [Tests of inflation.]{}— ]{} The above arguments relied on perturbations on large scales being generated before perturbations on small scales. This is the case in any conceivable model of inflation, however it is not be the case in all scenarios. As an illustrative example, in the cosmic defect scenario perturbations are generated on small scales and then causally transported to larger scales as the universe evolves. It is argued that in this scenario, tensor perturbations leave no fossils. A detection of primordial tensors by another means (CMB B-modes for example) with an observed lack of the corresponding fossils would provide a serious challenge to inflation. The most specific prediction of single field inflation is the power spectrum of tensor modes, defined by $$\label{e:tests:PTdef} (2\pi)^3\delta({k_a}-{k_a}')P_h({k_a}) \equiv \langle h_{ij}({k_a})h^{ij}({k_a}')\rangle.$$ Given the amplitude of the scalar power spectrum $A_s$, the tensor power spectrum is fixed by a single parameter, the tensor to scalar ratio $r$. The shape of the spectrum is then nearly scale-free: $$\label{e:tests:PT} P_h = \frac{2 \pi^2 r A_s}{k^3} \left(\frac{k}{k_0}\right)^{n_t}.$$ We follow the WMAP conventions for defining $P_h$, $A_s$ and $r$ [@arXiv:0803.0547]. The spectral index fixed by the consistency relation, $n_t = -r/8$ [@2000cils.book.....L]. The pivot scale is taken to be $k_0= 0.002 \mathrm{\,Mpc}^{-1}$ and we assume the WMAP7 central value for $A_s$ of $2.46 \times 10^{-9}$. Because $r$ is likely small, any deviation from a scale-free spectrum will be difficult to measure, making the verification of the consistency relation correspondingly difficult. The CMB is sensitive primarily to large scale tensor modes, with smaller scale modes having decayed by recombination. Cosmic variance and lensing contamination will likely prevent a measurement of $n_t$ from the CMB, unless the lensing can be cleaned from the signal [@arXiv:0902.1851]. Conversely, the amplitude of the fossil signal does not decay as the universe expands. It may thus be possible to make a measurement of the spectral index, provided $r$ is sufficiently large. [ [Statistical detection in LSS.]{}— ]{} In practice, the tensor gravity wave fossils could be reconstructed by applying quadratic estimators to the density field. Aside from the increased dimensionality, this is identical to the manner in which lensing shear is reconstructed [@astro-ph/9810257; @arXiv:0710.1108]. Rather than considering the statistics of such estimators, here we follow a simpler line of reasoning to approximate the accuracy to which the tensor parameter can be measured. We begin by asking how well a long wavelength, tensor mode can be reconstructed from its effects on the scalar power spectrum (Eq. \[e:mech:P\]). The metric perturbations are assumed to be spatially constant and take the form $$h_{ij}=h_+e^+_{ij}(\hat{z}) + h_\times e^\times_{ij}(\hat{z})$$ where $e^+_{ij}$ and $e^\times_{ij}$ are the polarization tensors and the $\hat{z}$ direction of propagation is chosen for convenience. The uncertainty on the scalar power spectrum is $$\label{e:fore:deltaP} \left[\Delta P({k_a})\right]^2 = 2 \left[P({k_a})+N\right]^2,$$ where $N$ is the noise power. We use a Fisher Matrix analysis to sum this information over all ${k_a}$ to determine the corresponding uncertainty on the shear $h_+$ and $h_\times$. Assuming an experiment whose noise is sub dominant to sample variance ($N \ll P$), the resulting variance is inversely proportional to the number of modes surveyed: $$\left(\Delta h^C\right)^2 \sim \left[V(k_{max}/2\pi)^3\right]^{-1},$$ where $h$ stands for either $h_+$ or $h_\times$ (the superscript $C$ indicates that the formula applies for spatially constant $h$), $V$ is the volume of the survey and $k_{max}$ is set by the resolution of the survey. The constant of proportionality depends on the shape of the unsheared power spectrum ${\tilde{P}}(k)$, but to within a few tens of percent it is unity. [21cm]{} emission will be difficult to observe on large scales [@astro-ph/0608032], however it is small scales that dominate the number of modes and thus the reconstruction. It is only the coherence of small scale anisotropy that must be measured on large scales. Given the reconstruction uncertainty on a spatially constant shear, and the fact that reconstruction noise is scale independent (white) [@astro-ph/9810257], the noise power spectrum for spatially varying tensor modes is then $$N_h = 4V (\Delta h^{C})^2 = 4\left(\frac{2\pi}{k_{max}}\right)^3.$$ The factor of four comes from the definition of the power spectrum in Eq. \[e:tests:PTdef\], noting that $\langle h_{ij}h^{ij} \rangle =4 \langle h^2 \rangle$. We now sum over ${k_a}$[^1] to determine the signal to noise as a function of tensor power spectrum amplitude $r$. The signal to noise ratio squared is then $$\begin{aligned} \textit{SNR}^2 &= \sum_{{k_a}, \{+,\times\}} \frac{P_h^2}{2(N_h+P_h)^2}\\ &\approx V\int_{k_{lower}}^{k_{upper}} \frac{dk\,k^2}{2\pi^2 } \frac{P_h^2(k)}{(N_h+P_h)^2}.\end{aligned}$$ It is seen from the redness of the spectrum $P_h$ (Eq. \[e:tests:PT\]) that the result is completely independent of the upper limit of integration. The same redness makes the final result extremely sensitive to the lower limit. As described above, the fossil of a primordial tensor mode can only be observed once the mode has decayed. This begins to happen when the scale of the gravity wave becomes comparable to the horizon scale, and as such, the largest scale observable mode has wavelength $k_{lower}\approx aH$. For an initial detection, we assume that noise dominates sample variance at each ${k_a}$, i.e., $N_h\gg P_h$. Setting the signal to noise ratio to be 2, for a 95% confidence detection, yields a minimum detectable amplitude of $$r_{min} = \frac{32 \pi^2}{A_s k_{max}^3}\left(\frac{6}{VV_H(z)}\right)^{1/2}$$ where $V_H\equiv (aH)^{-3}$. While the observability of [21cm]{} radiation depends on the reionizaton model, one regime in which a strong signal may exist is near redshift 15 [@astro-ph/0608032]. The planned Square Kilometer Array (SKA) will aim to probe this era with 10km baselines [@2009IEEEP..97.1482D]. Assuming a survey volume of $200\,(\textrm{Gpc}/h)^3$ and a noiseless measurement, the limit on $r$ achievable with SKA will be $$r_{min} \approx 7.3\left(\frac{1.2\,\textrm{Mpc}/h}{k_{max}}\right)^3 \left[\frac{200\,(\textrm{Gpc}/h)^3}{V} \frac{3.3\,(\textrm{Gpc}/h)^3}{V_H}\right]^{1/2}.$$ While this constraint is not competitive with current constraints from the CMB, it is a strong function of the resolution of the experiment. The Low Frequency Array (LOFAR) for instance, has baselines extending to 400km. However LOFAR will not have the sensitivity to probe the dark ages [@astro-ph/0307240]. It is the physical shear due to gravity waves at the source that is being measured, and all light propagation effects, such as the lensing considered in @astro-ph/0301177, have been ignored. Similar arguments are used to find the achievable error on the spectral index $n_t$. Properly considering the degeneracy with $r$, the error on $n_t$ is: $$\Delta n_t = F \left[\left(\frac{2\pi}{k_{max}}\right)^{3}\frac{1}{r A_s V}\right]^{1/2},$$ where $F$ is a function of the combination of parameters $V_H/(k_{max}^3rA_s)$. In the limit that $P_h(k=aH)\gg N_h$, which is the limit in which a measurement of $n_t$ is possible, $F$ is approximately 6. Assuming the same volume and redshift as above, and that $r=0.1$, the consistency relation is tested at the 2 sigma level for $k_{max}=168h/\textrm{Mpc}$. The tensor power spectrum and error bars for this scenario are shown in Fig. \[f:tensorpower\]. Such a measurement is very futuristic indeed, requiring a nearly filled array with greater than thousand kilometre baselines. Note that such an experiment would be sensitive to $r$ down to the $10^{-6}$ level. Also, higher redshifts contain even more information, though their observation is technically more challenging. ![\[f:tensorpower\] Primordial tensor power spectrum obeying the consistency relation for $r=0.1$. The solid line is the tensor power spectrum. Error bars represent the reconstruction uncertainty on the binned power spectrum for a perfect experiment, surveying $200\,(\textrm{Gpc}/h)^3$ and resolving scalar modes down to $k_{max}=168h/\textrm{Mpc}$. The dashed, nearly vertical, line is the reconstruction noise power. The non-zero slope of the solid line is the deviation from scale-free.](tensorpower) [ [Discussion.]{}— ]{} Aside from the technical challenge of mapping the 21cm signal over hundreds of cubic gigaparsecs and down to scales smaller than a megaparsec, there may be other competing effects that could hinder a detection. Of primary concern is weak lensing which also shears observed structures, creating apparent local anisotropies. The weak lensing shear is of order a few percent, and is thus many orders of magnitude greater than gravity wave shear. However, the 3D map of gravity wave shear will be transverse, transforming intrinsically as a tensor. To linear order, the lensing pattern is the gradient of a scalar. Even at higher order, lensing always maps one point in space to another and is thus at most vector like. This test does not exist for the CMB or lensing due to the lower dimensionality of these probes. Also of concern is the preservation of the anisotropy on small scales. The scale corresponding to $k=168h/\textrm{Mpc}$ is still larger than the Jeans length at these redshifts, and as such hydrogen should trace the dark matter. However, the evolution of scalar perturbations is mildly nonlinear, and it is possible that this evolution will erase the anisotropy. Detailed analysis of the nonlinear erasure of the anisotropy is deferred to future investigation. There has been much recent interest in searching for anisotropy, and this has some implications for the fossil signal. The constraints on quadrupolar isotropy in LSS by @arXiv:1003.0673 should already imply a weak constraint at the $r\lesssim10^6$ level. Constraints from the CMB are not relevant however, since modes spanning the surface of last scatter remain super horizon today. CMB B-modes will be the most sensitive probe of primordial gravity waves in the next generation of experiments. However, fossils may eventually be sensitive well below the limits of the CMB. We would like to thank Patrick McDonald, Latham Boyle, Adrian Erickcek, Neil Barnaby, Neal Dalal, Chris Hirata and Eiichiro Komatsu for helpful discussions. KM is supported by NSERC Canada. [^1]: From this point forward, ${k_a}$ will refer to the wave number of a tensor mode, not a scalar mode. The exception will be $k_{max}$ which is the smallest scale at which a scalar can be resolved.	{ "pile_set_name": "ArXiv" }
--- abstract: 'It is argued that a certain kind of short-range interacting system exhibits nonadditivity when several time scales are well separated. Under the condition of separated time scales, the system is described by the elastic spin model. We find that it is extensive but nonadditive, which is directly confirmed by the work measurement and also indicated by ensemble inequivalence. Further, we estimate the effective Hamiltonian for the spin variables, and it is clarified that the effective interaction is long ranged. Remarkably, the so-called Kac prescription, which is usually regarded as a mathematical operation to make the system extensive, naturally holds.' author: - Takashi Mori title: Nonadditivity in Quasiequilibrium States of Spin Systems with Lattice Distortion --- Let us consider a system consisting of two macroscopic subsystems $A$ and $B$. In a usual macroscopic system, the total amount of energy is given by the sum of internal energies of the two subsystems because the interaction energy between $A$ and $B$ is negligible compared to the bulk energy. This property is called additivity (the precise definition will be given later). Additivity is regarded as a fundamental property of macroscopic systems [@Callen_text]. It ensures concavity or convexity of the thermodynamic function. In statistical mechanics, it leads to the ensemble equivalence; i.e., several statistical ensembles yield identical thermodynamic quantities [@Ruelle_text]. However, not all the macroscopic systems possess additivity. Long-range interacting systems are representative of nonadditive and physically relevant systems [@Campa_review2009; @Les_Houches2008]. Because of the lack of additivity, long-range interacting systems can exhibit unfamiliar and peculiar macroscopic properties, e.g., negative specific heat [@Lynden-Bell-Wood1968; @Thirring1970], ensemble inequivalence [@Thirring1970], macroscopic inhomogeneity [@Mori2011_instability], and no thermalization in an isolated system [@Antoni-Ruffo1995]. Apparently, a short-range interacting system is unlikely to be nonadditive since the interaction energy will be very small compared to the bulk energy. In this Letter, however, it is pointed out that in a [quasi]{}equilibrium state, a certain kind of short-range interacting system can exhibit nonadditivity. Interestingly, in spite of its nonadditivity, the system is extensive; the energy is proportional to the system size if we make the system large uniformly. Now we explain the elastic spin model studied in this Letter. The model itself has already been known in studies on spin-crossover materials [@Nishino2007]. See Ref. [@Gutlich_review1994] and references therein for a detailed account of spin-crossover materials. $N$ molecules are aligned on the two-dimensional triangular lattice of side $L$, where $N=L^2$.[^1] Each molecule is composed of a metal ion and surrounding ligands. Each molecule $i$ has two different stable internal states, that is, the high-spin (HS) state $\sigma_i=+1$ and the low-spin (LS) state $\sigma_i=-1$. Electron configurations in a metal ion are different between the HS and the LS states and the HS state has a higher spin than the LS state. For example, in $\rm Fe^{II}$, the HS state has $S=2$ and the LS state has $S=0$. As a result, the HS state with $S=2$ has degeneracies of $S^z=2,1,0,-1,-2$. In general, degeneracies of the HS state are denoted by a parameter $g$ and thus $\sigma_i\in\{\underbrace{+1,+1,+1,\dots,+1}_{g},-1\}.$ Although $\sigma_i$ is not a genuine spin, we call it a spin variable and $M=\sum_i^N\sigma_i$ the magnetization. The magnetization density is denoted by $m\equiv M/N$. -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------- (a) (b) ![(a) Typical shape of the interaction potential. (b) Triangular lattice and the label of the molecules.[]{data-label="fig:setting"}](potential.eps "fig:"){width="4cm"} ![(a) Typical shape of the interaction potential. (b) Triangular lattice and the label of the molecules.[]{data-label="fig:setting"}](triangle.eps "fig:"){width="4cm"} -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------- The intermolecular interaction is given by some short-range potential $V_{ij}(r)$ such as Fig. \[fig:setting\](a), which decays faster than $1/r^2$ in a long distance. Because of the short-range nature of the interaction, it is sufficient to consider only the nearest-neighbor interactions. The important point is that the equilibrium distance of $V_{ij}(r)$ is given by $R_{ij}\equiv R(\sigma_i)+R(\sigma_j)$, which depends on the molecular internal states. The values of $R(\pm1)$ represent the molecular radius at state $\sigma_i=\pm1$, respectively. This size difference is actually observed in experiments [@Wiehl1990] and plays an important role for spin-crossover transitions [@Miyashita2008]. When the potential depth $V_0$ in Fig. \[fig:setting\](a) is much larger than the thermal energy $k_{\rm B}T$, where $k_{\rm B}$ is the Boltzmann constant and $T$ is the temperature, $V_{ij}(r)$ is approximated as a quadratic form: V\_[ij]{}(r)(r-R\_[ij]{})\^2={r-\[R(\_i)+R(\_j)\]}\^2. \[eq:approx\] The condition of the applicability of this approximation is discussed later. In this way, the Hamiltonian of the elastic spin model is given by H=\_[i=1]{}\^N +\_[ i,j]{}(\|\_i-\_j\|-R\_[ij]{})\^2 +D\_[i=1]{}\^N\_i. \[eq:elastic\_H\] Here $\bm{q}_i$ and $\bm{p}_i$ are the coordinate and the momentum of the $i$th molecule. The symbol $\<i,j\>$ denotes all the nearest-neighbor pairs. The last term represents the effect of the ligand field. When $D=0$, there are two ground states: i.e., $\sigma_i=1$ $\forall i$ and $\sigma_i=-1$ $\forall i$. Throughout this Letter, according to the previous work [@Miyashita2008], we fix the parameters as $R(-1)=1$, $R(1)=1.1$, and $k=40.$ For convenience, we also label the molecules by the two-dimensional vectors $\bm{r}_i=(x_i,y_i)$ with $x_i,y_i\in\{1,2,\dots,L\}$. The molecule at the $x$th column and $y$th row is labeled by $(x,y)$: see Fig. \[fig:setting\](b). We distinguish $\bm{q}_i$ from $\bm{r}_i$; the former is a dynamical variable denoting the position of the $i$th molecule, but the latter is the label of the site denoting its positional relation on the lattice. Here we should mention the condition under which we can consider the equilibrium state of Eq. (\[eq:elastic\_H\]); the observation time $t_{\rm obs}$ should satisfy \_[eq]{}< t\_[obs]{}\_V, \[eq:condition\] where $\tau_{\rm eq}$ is the equilibration time of this model and $\tau_V$ is the activation time necessary to get over the potential depth $V_0$. If this condition were violated, fracture of the solid would occur and the approximation (\[eq:approx\]) is not necessarily valid; it becomes important to consider molecules which are far away from the nearest neighbors. When $V_0\gg k_{\rm B}T$, such slow things have not happened in the time scale $t_{\rm obs}$ given above, and hence the model (\[eq:elastic\_H\]) is appropriate. In other words, [an equilibrium state of the elastic spin model is regarded as a quasiequilibrium state of the original short-range interacting system]{}. ![(color online). lhs (red circles) and rhs (green triangles) of Eq. (\[eq:additivity\]) vs $1/L$. In the former, we put $m_{A}=-1$ and $m_{B}=+1$. In the latter, there is no restriction on $m_{A}$ and $m_{B}$.[]{data-label="fig:ext_add"}](ext_add_tri.eps){width="6cm"} It is noted that the Hamiltonian given by Eq. (\[eq:elastic\_H\]) is nonlocal and possesses long-range nature. The essential point is that (i) the variables $\{\bm{q}_i\}$ are not independent since two nearest-neighbor molecules cannot be far apart within the time scale (\[eq:condition\]) and (ii) equilibrium positions of molecules depend on the spin variables $\{\sigma_i\}$ nonlocally, which is due to the size difference between the HS and the LS states. In order to investigate extensivity and additivity, let us consider the system in contact with a thermal bath at the temperature $T$. We virtually divide the system into two subsystems $A$ and $B$; the subsystem $A$ is composed of the molecules with $x_i\leq L/2$ and the subsystem $B$ with $x_i>L/2$. If the molecule $i$ belongs to the subsystem $A$ ($B$), we write $i\in A (B)$. We divide the Hamiltonian into three parts, $H_{AB}(\lambda)=H_{A}+H_{B}+\lambda H_{I}$. Here $H_{X}\equiv\sum_{i\in X}(\bm{p}_i^2/2+D\sigma_i)+(k/2)\sum_{\< i,j\> \in X}(\|\bm{q}_i-\bm{q}_j\|-R_{ij})^2$, where $X$ is $A$ or $B$. The parameter $\lambda$ controls the strength of the interaction between subsystems and we choose it so that $H_{AB}(\lambda=1)$ is equal to Eq. (\[eq:elastic\_H\]). The concept of additivity is related to the probability that the magnetization of $A$ is equal to $M_{A}$ and that of $B$ is equal to $M_{B}$. In this Letter, the system is said to be additive if $P_{AB}(T,M_{A},M_{B})\simeq P_{A}(T,M_{A})P_{B}(T,M_{B})$. Here $P_{AB}(T,M_{A},M_{B})$ denotes the probability that the magnetization of $A$ at temperature $T$ is equal to $M_{A}$ and that of $B$ is equal to $M_{B}$ in a composite system $H_{AB}(1)$. On the other hand, $P_{X}(T,M_{X})$ ($X$ is $A$ or $B$) denotes the probability that the magnetization of $X$ at temperature $T$ is equal to $M_{X}$ [in a decoupled system]{} $H_{X}$. According to statistical mechanics, such probabilities are related to the thermodynamic functions with some restriction [@Landau_stat]. That is, $P_{AB}(T,M_{A},M_{B})=\exp\{[F_{AB}(T)-F_{AB}(T,M_{A},M_{B})]/k_{\rm B}T\}$, where $F_{AB}(T)$ is the free energy of the composite system $H_{AB}(1)$ at temperature $T$ and $F_{AB}(T,M_{A},M_{B})$ is that under the constraint that the magnetizations of $A$ and $B$ are fixed to be $M_{A}$ and $M_{B}$, respectively. Similarly, $P_{X}(T,M_{X})=\exp\{[F_{X}(T)-F_{X}(T,M_{X})]/k_{\rm B}T\}$, where $F_{X}(T)$ is the free energy associated with the Hamiltonian $H_{X}$ and $F_{X}(T,M_{X})$ is that under the constraint that its magnetization is fixed to be $M_{X}$. Thus the condition of additivity is rewritten as $$\begin{aligned} F_{AB}(T,M_{A},M_{B})-F_{A}(T,M_{A})-F_{B}(T,M_{B}) \nonumber \\ \simeq F_{AB}(T)-F_{A}(T)-F_{B}(T) \label{eq:additivity}\end{aligned}$$ for any values of $M_{A}$ and $M_{B}$. In particular, when the right-hand side (rhs) of Eq. (\[eq:additivity\]) is zero, the system is said to be extensive. According to thermodynamics, the left-hand-side (lhs) of Eq. (\[eq:additivity\]) is equal to the amount of work done by the system in a quasistatic (reversible) isothermal process of changing $\lambda$ from 1 to 0. During the process, $M_{A}$ and $M_{B}$ should be individually conserved. In Fig. \[fig:ext\_add\], we calculated the work per molecule, $w=-\int\<{\partial}H(\lambda)/{\partial}\lambda\>\dot{\lambda}dt/L^d$, in such a quasistatic isothermal process. In the red circles, we put $m_{A}=M_{A}/L^d=-1$, $m_{B}=M_{B}/L^d=+1$, and $T=0.1$. The value of $w$ corresponds to the lhs of Eq. (\[eq:additivity\]). It is observed that $w$ does not vanish as $L$ increases. On the other hand, the green triangles are measured values of $w$ without restriction on the magnetizations, which correspond to the rhs of Eq. (\[eq:additivity\]). In this case $w$ tends to zero as $L$ increases. Therefore, it is concluded that the system is extensive but nonadditive. ------------------------------------------- ------------------------------------------- ------------------------------------------- (a) (b) (c) ![image](ela_eff_TH_tri.eps){width="5cm"} ![image](ela_eff_TM_tri.eps){width="5cm"} ![image](ela_eff_EH_tri.eps){width="5cm"} ------------------------------------------- ------------------------------------------- ------------------------------------------- Because of nonadditivity, it is expected that the elastic spin model exhibits the peculiar properties observed in long-range interacting systems. We performed Monte Carlo simulations for (a) the canonical ensemble, (b) the restricted canonical ensemble (the canonical ensemble with restriction on the value of the magnetization), and (c) the microcanonical ensemble. Numerical results are shown by the dark gray (red) circles in Fig. \[fig:elastic\]. In the microcanonical ensemble, we subtracted $2T$, which represents the contribution of the lattice vibration, from the total energy density. The parameters are set to be $D=0$ and $g=1$ for (a) and (b), and $D=0.15$, $g=20$ for (c) [^2]. The system size is $L=40$. In the canonical ensemble, spin variables are changed according to the usual Metropolis algorithm and the positions of molecules move according to the Hamilton dynamics (by the leapfrog algorithm). We can see that the system undergoes a second order phase transition at $T_{\rm c}\simeq 0.35$ in Fig. \[fig:elastic\](a). The critical behavior belongs to the mean-field universality class [@Miyashita2008] (see also Ref. [@Nakada2012]). The algorithm used in the restricted canonical ensemble is the same as in the canonical ensemble except that we prepare an excess degree of freedom referred to as the “demon”. The demon keeps the magnetization $m_{d}=\pm 1$. Only if $m_{d}\sigma_i\leq 0$, the spin flip is accepted according to the Metropolis transition probability. After the flip, we change $\sigma_i$ and $m_{d}$ to $-\sigma_i$ and $-m_{d}$, respectively. The advantage of this method is that we can easily measure the magnetic field $h={\partial}F(T,M)/{\partial}M$ from the average value of $m_{d}$ by $h=\frac{k_{\rm B}T}{2}\ln\frac{1+\<m_d\>}{1-\<m_d\>}$. In Fig. \[fig:elastic\](b), we can clearly see the region where the susceptibility $\chi={\partial}m/{\partial}h$ is negative. The susceptibility is always positive in the canonical ensemble and this discrepancy shows that the canonical ensemble is inequivalent to the restricted canonical ensemble. In the microcanonical ensemble, we used the leapfrog algorithm for the time evolution of $\{\bm{q}_i,\bm{p}_i\}$ and the Creutz algorithm [@Creutz1983] for the spin flip, in which an excess degree of freedom also called the demon, which has a positive energy, is prepared and the total energy of the system and the demon is conserved. The temperature of the system can be measured simply by $T=\<E_{d}\>$. Figure \[fig:elastic\](c) clearly demonstrates that there is a region of the negative specific heat, which also shows the ensemble inequivalence. Although there is no direct interaction between $\sigma_i$ and $\sigma_j$ in long distance, it will be a plausible consideration that an [effective interaction]{} arises between spin variables via lattice distortion. In general, the effective Hamiltonian for $\{\sigma_i\}$ obtained by eliminating $\{\bm{q}_i,\bm{p}_i\}$ contains many-body interactions and will be very complicated. Here we assume that the effective Hamiltonian is written as $H_{\rm eff}=-(1/2)\sum_{ij}\hat{J}_{ij}\sigma_i\sigma_j$ and we try to estimate $\hat{J}_{ij}$ from the numerical data of correlation functions $\hat{C}_{ij}\equiv\<\sigma_i\sigma_j\>$. We put $g=1$ and $D=0$ and consider the canonical ensemble in the high-temperature phase ($T=0.5$). In this case, we can show that $\hat{J}_{ij}$ is obtained from $\hat{C}_{ij}$ by (L)=\[-(L)/(k\_[B]{}T)\]\^[-1]{}, \[eq:C\_J\] where the identity matrix is denoted by $\hat{I}$. The relation (\[eq:C\_J\]) is approximate one in general, but it becomes exact if the effective interaction is long ranged: i.e., \_[ij]{}(L)=L\^[-d]{}(L\^[-1]{}\_[ij]{}), \[eq:J\_scaling\] where $\bm{r}_{ij}=(x_{ij},y_{ij})\equiv\bm{r}_i-\bm{r}_j$ and $d$ is the spatial dimension (here $d=2$). It is assumed that $\int_{\|\bm{x}\|<\delta}\phi(\bm{x})d^dx<+\infty$ for an arbitrary fixed $\delta>0$. It includes the power-law interaction $\phi(\bm{x})\sim 1/x^{\alpha}$ with $\alpha<d$. The scaling $L^{-1}\bm{r}_{ij}$ means that the interaction range is comparable with the system size $L$. The scaling $L^{-d}$ in Eq. (\[eq:J\_scaling\]) ensures that the energy is proportional to the system size, which corresponds to the so-called Kac prescription [@Campa_review2009; @Les_Houches2008]. ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- (a) (b) ![(color online). (a) Global image of the interaction potential at $T=0.5$ and $L=60$. The site $j$ is fixed at the center of the system ($x_j=y_j=30$). In the white region, the interaction is almost zero; $-0.00005<\hat{J}_{ij}\leq 0$. In the black region, the interaction is antiferromagnetic; $\hat{J}_{ij}\leq -0.00005$. In the gray (red) region, the interaction is ferromagnetic and the depth of the color expresses the strength of the ferromagnetic interaction. (b) Graphs of the estimated effective interactions along the diagonal direction, $y_{ij}=L-x_{ij}$ for $L=40,60,80,$ and 100. The transverse axis is the scaled distance $x_{ij}/L$ and the longitudinal axis is the scaled interaction matrix $L^2\hat{J}_{ij}$. Only the points with $\hat{J}_{ij}>0$ are plotted.[]{data-label="fig:interaction_L"}](interaction_map_L60.eps "fig:"){width="5cm"} ![(color online). (a) Global image of the interaction potential at $T=0.5$ and $L=60$. The site $j$ is fixed at the center of the system ($x_j=y_j=30$). In the white region, the interaction is almost zero; $-0.00005<\hat{J}_{ij}\leq 0$. In the black region, the interaction is antiferromagnetic; $\hat{J}_{ij}\leq -0.00005$. In the gray (red) region, the interaction is ferromagnetic and the depth of the color expresses the strength of the ferromagnetic interaction. (b) Graphs of the estimated effective interactions along the diagonal direction, $y_{ij}=L-x_{ij}$ for $L=40,60,80,$ and 100. The transverse axis is the scaled distance $x_{ij}/L$ and the longitudinal axis is the scaled interaction matrix $L^2\hat{J}_{ij}$. Only the points with $\hat{J}_{ij}>0$ are plotted.[]{data-label="fig:interaction_L"}](eff_interaction.eps "fig:"){width="5cm"} ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- In Fig. \[fig:interaction\_L\](a), the structure of the estimated interaction matrix $\hat{J}_{ij}$ is depicted for $L=60$. In the figure, the central site $\bm{r}_j=(L/2,L/2)$ is chosen as the site $j$. It is found that the interaction is highly anisotropic and a long-range ferromagnetic interaction emerges. Importantly, its spatial average does not vanish; the long-range ferromagnetic interaction is not screened. The characteristic system size dependence of the estimated interaction is shown in Fig. \[fig:interaction\_L\](b), where the graphs of $L^d\hat{J}_{ij}(L)$ along the diagonal direction, $y_{ij}=L-x_{ij}$, are depicted as a function of $L^{-1}x_{ij}$ for several values of $L$ only in the region of $\hat{J}_{ij}>0$. We can see that these graphs are collapsed well into a single curve (and this collapse occurs for any direction of $\bm{r}_{ij}$). It means that the effective interaction is actually of the form of Eq. (\[eq:J\_scaling\]) and the use of Eq. (\[eq:C\_J\]) is indeed justified. Usually, Kac’s prescription is regarded as a purely mathematical operation to extract nontrivial thermodynamic properties of long-range interacting systems [@Les_Houches2008]. However, in the present model, such a mathematical operation is not necessary; the scaling of Eq. (\[eq:J\_scaling\]) naturally emerges in the effective interaction. This is a remarkable characteristic of this model. In Fig. \[fig:elastic\], we compared some equilibrium quantities of the elastic model with those of the effective Ising model, $H_{\rm eff}=-\sum_{ij}\hat{J}_{ij}\sigma_i\sigma_j-D\sum_i^N\sigma_i$. In all the simulations presented in Fig. \[fig:elastic\], we used the effective interaction $\hat{J}_{ij}$ estimated at $T=0.5$ in the canonical ensemble. The numerical results of the effective Ising model are plotted by green triangles. Although the effective interaction is estimated by the data at the single point in the canonical ensemble, the elastic model is indistinguishable from the effective Ising model for all the values of parameters and for all the statistical ensembles. To summarize, we have argued that a certain short-range interacting system displays nonadditivity when the system is in a quasiequilibrium state, although a genuine equilibrium state would be additive. Within the time window (\[eq:condition\]), the system relaxes to a quasiequilibrium state, which is described by an equilibrium state of the elastic spin model (\[eq:elastic\_H\]). It has been found that the elastic spin model exhibits ensemble inequivalence. In addition, it has been numerically shown that as far as equilibrium states of spin degrees of freedom are concerned, the elastic spin model is indistinguishable from the effective Ising model with a long-range interaction. It implies that statistical mechanics of long-range interacting systems is relevant for understanding quasiequilibrium states of some short-range interacting systems. The limitation of the present study is that several time scales should be separated as Eq. (\[eq:condition\]), which requires the large potential depth $V_0$ compared to $k_{\rm B}T$; otherwise considering an equilibrium state of the Hamiltonian (\[eq:elastic\_H\]) would not be justified. Finally, it is pointed out that the present study will give some insight into experimental attempts to realize a long-range interacting system and observe its peculiar properties in laboratory [@ODell2000; @Chalony2013]. Such attempts are not satisfactory yet, but we hope that this work stimulates different experimental approaches toward it. The author thanks Taro Nakada for fruitful discussion and Seiji Miyashita for continual discussion and careful reading of the manuscript. He acknowledges the financial support provided by the Sumitomo Foundation. The computation in this work has been done using the facilities of the Supercomputer Center, Institute for Solid State Physics, University of Tokyo. [17]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , (, ). , (, ). , , , (, ). , , , , (). , , (). , , (). , , (). , , (). , , , , , (). , , , , (). , , , , , (). , , , , , , (). , (, ). , , , , , , , **, (). , , (). , , , , , (). , , , , , **, (). [^1]: The conclusion of this work does not change in three dimensions. In the one-dimensional chain, however, it is shown that the elastic spin model does not exhibit nonadditivity; see K. Boukheddaden, S. Miyashita, and M. Nishino, Phys. Rev. B 75, 094112 (2007). [^2]: In the present model, the negative specific heat does not appear at $D=0$ and $g=1$. In order to demonstrate nontrivial behavior of this model, we chose parameters as $D=0.15$ and $g=20$ in (c).	{ "pile_set_name": "ArXiv" }
--- author: - \| [O. Civitarese$^1$ [^1], M. Gadella $^2$ [^2], and G. P. Pronko$^3$ [^3]]{},\ \ [$^{1}$ Departamento de Física, Universidad Nacional de La Plata, ]{}\ [c.c. 67 1900, La Plata, Argentina. ]{}\ \ [$^{2}$Departamento de Física Teórica, Facultad de Ciencias, E-47011, Valladolid, Spain]{}\ \ [$^{3}$Institute for High Energy Physics, Protvino, 142284,Moscow Region. Russia]{}\ [and Institute of Nuclear Physics, N.C.S. “Demokritos". ]{} $\;\;\;\;\;\;\;\;\;\;$\ title: 'The Friedrichs-Model with fermion-boson couplings II.' --- [: In this work we present a formal solution of the extended version of the Friedrichs Model. The Hamiltonian consists of discrete and continuum bosonic states, which are coupled to fermions. The simultaneous treatment of the couplings of the fermions with the discrete and continuous sectors of the bosonic degrees of freedom leads to a system of coupled equations, whose solutions are found by applying standard methods of representation of bound and resonant states.]{}\ [ ]{} Introduction ============ The Friedrichs model [@1] describes the coupling of discrete and continuum sectors of a boson Hamiltonian. The model has the advantage of being exactly solvable and it has been applied to a variety of systems of physical interest [@2]. The mathematical aspects of the solutions which are based in the use of rigged Hilbert spaces have been introduced in [@3]. A distinctive feature of the model is the appearance, among the solutions, of Gamow states [@4; @5; @6]. It is the only completely solvable model which describes resonance phenomena [@3] in a broad context. Generally speaking, the solutions show the importance of the discrete-continuum couplings, a fact that has been overlooked for quite a while and that certainly has crucial physical consequences. In fact, the need for the inclusion of bound as well as resonant states in the description of composite systems has been demonstrated (see [@7] and references quoted therein) in connection with the analysis of data obtained in experiments measuring the decay of clusters. The properties of the standard version of the Friedrichs model, in connection with Gamow resonances, have been revisited in [@5]. More sophisticated versions of the Friedrichs model have been presented in [@6; @8; @10]. Along the standard discrete-continuous couplings for bosons, one has to consider also the boson-boson couplings [@2] and fermi-boson couplings [@6]. In a recent paper [@6], we have proposed an extended version of the model to accommodate for fermion-boson couplings. The structure of the extended Hamiltonian is inspired in standard field theory models of the coupling between fermion and boson fields [@11; @12]. In a previous publication [@6], hereafter referred to as I, the feasibility of the extended model was discussed and its solution were explored within a schematic model situation. By considering a very reduced space consisting of one discrete fermion state in addition to the bosonic degrees of freedom, it was found a resonant behavior also in the fermion sector, as expected. In this paper, we focus our attention in the analysis of the solution within a more general framework. We base our method on the use of the resolvent for the characteristic coupled equations which describes the amplitude of the fermionic solution. The coupled equations are cast in a form which very much resembles the T-matrix formalism [@13], which is very convenient for finding out the resonance behavior. The present paper is organized as follows: In section 2, we review the essentials of the extended Friedrichs model, which in full detail has been introduced in [@6], and focus our attention on the elements needed for the present discussion. In section 3, we analyze the structure of the general solution which has been found by applying the T-matrix formalism. Finally, we draw our conclusions on section 4. Formalism {#s1} ========= We shall begin with a review of the basic ideas of the Friedrichs Model [@1; @5]. Details about the formalism concerning the standard Friedrichs model can be found in [@AGPP]-[@AGP] and references quoted therein. The simplest form of the Friedrichs model [@1] includes a free Hamiltonian $H_0$ with a real positive continuous spectrum coinciding with the positive semiaxis, $\omega > 0$, and a positive discrete eigenvalue, $\omega_0 >0$. An interaction is acting between the continuous and discrete parts of $H_0$ by means of a potential $V$. As a result of the action of $V$, the bound state of $H_0$ is dissolved in the continuous and a resonance is produced. The spectrum of the total Hamiltonian $H=H_0+ \lambda V$, where $\lambda$ is a real coupling constant, is purely continuous and coincides with the real positive semiaxis as well [@1; @3]. As shown in I, poles of the solutions appear into complex conjugate pairs, each of them represents a resonance. Then, Gamow vectors are solutions of the eigenvalue equation [@2; @4], $$(H-x)\,\Psi(x)=0\,. \label{1}$$ with the total Hamiltonian $H$, whose eigenvalues coincide with resonance poles [@3; @AGP]. The extended Friedrichs Model. {#ss21} ------------------------------ The present version of the Friedrichs model, as an extension of the standard Friedrichs model, includes [@6]: i\) The unperturbed fermion and boson Hamiltonian $H_I$ $$\begin{aligned} H_{I}=\omega_0\|1\rangle \langle 1\| + \int_0^\infty d\omega \omega \|\omega \rangle \langle \omega \| +\sum_k c_k \|k\rangle \langle k\|\,,\label{2}\end{aligned}$$ where the index $k$ runs out the set of Fermion kets $\|k\rangle$. ii\) The interaction between fermions, $\|k\rangle$, and the discrete boson, $\|1\rangle$, Hamiltonian $H_{II}$: $$\label{3} H_{II}=\sum_{k,l}\left[h_{k,l}\|k,1\rangle \langle l\|+ h^_{k,l}\|l\rangle \langle k,1\|\right]\,.$$ iii\) The interaction between fermions and the boson field, $\{\|\omega\rangle\}$, $H_{III}$: $$\label{4} H_{III}=\sum_{k,l}\int_0^\infty d\omega \left[f_{k,l}(\omega) \|k,\omega \rangle\langle l\|+ f^_{k,l}(\omega) \|l\rangle\langle k,\omega\|\right]\,.$$ The standard Friedrichs model includes the boson-boson coupling $V$ [@3]. This coupling $V$ can be generalized to include fermion-boson interactions in the following manner: $$\label{5} H_{IV}=\sum_{k,k'}\int_0^\infty d\omega \left[g_{kk'}(\omega) \|k,1 \rangle\langle k',\omega\|+ g^_{kk'}(\omega) \|k',\omega\rangle\langle k,1\|\right]\,.$$ Alternatively, $H_{IV}$ can be taken as a second order interaction mediated by the fermion state $\|k\rangle$, provided that: $$\label{6} \sum_l f^_{kl}(\omega) h_{lk'}=g_{kk'}(\omega)\,\delta_{kk'}= g_k(\omega)\,\delta_{kk'}\,, \label{24}$$ is fulfilled, as it is customarily done in the field theoretical treatment of fermion-boson couplings [@17]. In addition, it is worthy to note that $H_{IV}$ reduces to the standard Friedrichs model interaction, $V$, if $g_{k}(\omega)=g(\omega)$ for all $k$. To obtain the solution of the eigenvalue problem $$(H-E)\Psi(E)=0\,,\label{7}$$ where $H= H_I+H_{II}+H_{III}+H_{IV}$, we write $\Psi(E)$ in the form $$\label{8} \Psi(E)=\sum_k\varphi_k(E)\|k\rangle+\sum_k \phi_{k,1}(E)\|k,1\rangle +\sum_k\int_0^\infty d\omega \,\psi_k(E,\omega)\|k,\omega\rangle\,.$$ Consequently, (\[7\]) gives $$\begin{aligned} &&(H-E)\Psi(E)\equiv \sum_k\varphi_k(E)(c_k-E)\|l\rangle +\sum_{k,l} h_{kl}^\phi_{k,1}(E)\|k\rangle \nonumber \\ &+&\sum_k (c_k+\omega_0-E)\phi_{k,1}(E)\|k,1\rangle +\sum_k\int_0^\infty d\omega \psi_k(E,\omega)g_k(\omega)\|k,1\rangle \nonumber \\ &+&\sum_k \phi_{k,1}(E)\int_0^\infty d\omega g_k^(\omega)\|k,\omega\rangle +\sum_k\int_0^\infty d\omega \psi_k(E,\omega)(c_k+\omega-E)\|k,\omega\rangle\nonumber \\ &+&\sum_{k,l}\int_0^\infty d\omega \psi_k(E,\omega)f_{kl}^(\omega)\|l\rangle +\sum_{i,k}\varphi_k(E)h_{ik}\|i,1\rangle\nonumber \\ &+&\sum_{i,k}\int_0^\infty d\omega \varphi_k(E)f_{ik}(\omega)\|i,\omega\rangle=0\,. \label{9}\end{aligned}$$ Due to the linear independence of the vectors $\|k\rangle$, $\|k,1\rangle$, and $\|k,\omega\rangle$, the above equation can be written as a system of the following three coupled equations: $$\varphi_k(E)(c_k-E)+\sum_lh_{lk}^\phi_{l1}(E)+\sum_l\int_0^\infty d\omega \psi_l(E,\omega)f_{lk}^(\omega)=0 \label{10}$$ $$\sum_l\varphi_l(E)h_{kl}+(c_k+\omega_0-E)\phi_{k1}(E)+\int_0^\infty d \omega \psi_k(E,\omega)g_k(\omega)=0 \label{11}$$ $$\psi_k(E,\omega)(c_k+\omega-E)+\sum_l\varphi_l(E)f_{kl} (\omega)+\phi_{k1}(E)g_k^(\omega)=0 \label{12}$$ for $\|k\rangle$, $\|k,1\rangle$, and $\|k,\omega\rangle$, respectively. using (\[10\]) and (\[11\]) in (\[12\]), we obtain the following equation: $$\psi_k(E,\omega)=c\delta(c_k-\omega-E)- \sum_l\frac{ \varphi_l(E) f_{kl}(\omega)}{c_k-\omega-E}-\frac{\phi_{k1}(E)g_k^(\omega)}{c_k-\omega-E}\,. \label{13}$$ Consequently, if $\psi_k(E,\omega)$ is reinserted into (\[10\]) and (\[11\]), it results a system of infinite coupled equations. The formal properties of its solutions becomes obvious at glance: i.) the resonant behavior of the fermionic solutions results from the coupling to the continuum-continuum boson sector; ii.) the resonant boson sector is mostly unaffected by the fermion sector, since the latter solely depends on $V$. The solutions of the system depend on the choices of the form factors $h_{kl}$, $f_{kl}(\omega)$ and $g_k(\omega)$. After a rearrangement, (\[10\]) and (\[11\]) yields: $$\begin{aligned} \left[(c_k-E)\delta_{km}- \sum_m A_{km}(E)\right] \varphi_m(E)\nonumber\\[2ex]+\sum_m\left( h_{mk}^\right. -\left. B_{km}(E)\right)\phi_{m1}(E)=-c\sum_m f_{mk}^(E-c_m)\label{14}\end{aligned}$$ and $$\begin{aligned} &&\sum_l [h_{kl}-\widetilde B_{kl}(E)]\varphi_l(E)\nonumber\\[2ex]&&+ (c_k+\omega_0-E-C_k(E))\phi_{k1}(E)=-cg_k(E-c_k)\,.\label{15}\end{aligned}$$ Note that the functions, $A_{km}(E)$, $B_{km}(E)$, $\widetilde B_{kl}(E)$ and $C_k(E)$ are explicit functions of the form factors. These functions have been given in I, (see Equations 32-35 of [@6]). The factor $c$ in (\[15\]) is an irrelevant constant. Choice of form factors. ======================= In I, we have shown that the system of coupled equations which determine the amplitudes $\varphi_k(E)$, $\phi_{l1}(E)$ and $\psi_k(E,\omega)$ can be solved under certain simplifications like the assumption that the form factors do not depend on the fermionic indices and/or that the fermionic subspace be severely reduced. Here, we propose a more general setting for finding the solution, which consists of the study of the separate couplings of the fermions with the discrete and continuous bosons. For the form factors, we here propose the following choice: $$f_{lk}(\omega)=\alpha_l\alpha_k^ \rho(\omega) \,, \hskip1cm h_{lk}=\beta_l\beta_k^\,,\hskip1cm g_k(\omega)=0\,, \label{16}$$ This choice does not contradict (\[6\]) because we are considering the couplings (\[2\]), (\[3\]) and (\[4\]) as independent couplings, of which (\[6\]) is a special case only. Here, $\rho(\omega)$ is an unspecified function, whose explicit structure is not needed for the present discussion. In actual calculations it is defined by the model Hamiltonian. Thus, with this choice of the form factors, Equation (\[15\]) reads: $$\label{17} \phi_{k1}(E)= -\sum_s \frac{\varphi_s(E)\,h_{ks}}{c_k-\omega_0-E} \,. \label{17}$$ Then, Equations (32-35) in I become $$\begin{aligned} A_{km}(E)=\int_0^\infty d\omega\sum_l \frac{\alpha_l^\alpha_k\alpha_l\alpha_m^\,\|\rho\|^2}{c_l+\omega-E}\nonumber\\[2ex] =\alpha_k\alpha_m^ \int_0^\infty d\omega\sum_l \frac{\alpha_l^\alpha_l\,\|\rho\|^2}{c_l+\omega-E}= \alpha_k\alpha_m^\,A(E)\,.\label{18}\\[2ex] B_{km}=0\,,\hskip1cm \widetilde B_{km}=0 \,,\hskip1cm C_{km}=0\,.\label{19}\end{aligned}$$ Note that (\[18\]) defines $A(E)$. This yields the following form for (\[14\]): $$\begin{aligned} \sum_m\left[(c_k-E)\,\delta_{km}-\alpha_k\alpha_m^\,A(E)\right]\,\varphi_m(E)-\sum_lB(E)\,\beta_k\beta_l^ \varphi_l\nonumber\\[2ex] =-c\sum_m f_{mk}^(E-c_m)\,,\label{20}\end{aligned}$$ where $B(E)$ is $$\label{21} B(E)=\sum_m\frac{\beta_m^\beta_m}{c_m+\omega_0-E}\,.$$ Equation (\[20\]) can be written in a more compact form as $$\sum_s\varphi_s(E)[(c_k-E)\delta_{ks}-\alpha_k^\alpha_s\, A(E)- \beta_k^\beta_s\, B(E)] = -c\sum_l f_{lk}^(c_l-E)\,, \label{22}$$ where the functions on the variable $E$, $A(E)$ and $B(E)$, do not depend either on $k$ and $s$. Equation (\[22\]) can be also written in matrix form: $$T\left(\begin{array}{c} \cdots\\ \varphi_s(E)\\ \cdots \end{array}\right)= \left(\begin{array}{c} \cdots\\ -c\sum_l f_{lk}^(c_l-E)\\ \cdots \end{array}\right)\,, \label{23}$$ where the entries in the matrix $T$ are given by $$\begin{aligned} T_{ks}= (c_k-E)\delta_{ks}-\alpha_k^\alpha_s\, A(E)- \beta_k^\beta_s\, B(E) \nonumber\\ [2ex] = (c_k-E)\left\{\delta_{ks}-\frac{\alpha_k^\alpha_s\, A(E)}{(c_k-E)} - \frac{\beta_k^\beta_s\, B(E)}{(c_k-E)} \right\}\,. \label{24}\end{aligned}$$ To some extent, we have borrowed the notion frequently encountered in Theory of Reactions (see Newton [@13]) in dealing with the formal solution of systems of coupled equations, where $T$ stands for the transfer matrix. A solution to our problem with the choice (\[16\]) can be obtained provided that T be invertible. To prove this claim, we use the matrix identity $(1-A)^{-1}=\sum_n A^n$. In consequence, the entries of $T^{-1}$ are the following: $$\begin{aligned} (T^{-1})_{ks}= \frac{1}{c_k-E}\left\{\delta_{ks}- \frac{\alpha_k^\alpha_s\, A(E)}{c_k-E} -\frac{\beta_k^\beta_s\, B(E)}{c_k-E} \right\}^{-1} \nonumber\\ [2ex] = \frac{1}{c_k-E}\left\{\delta_{ks} +\left( \frac{\alpha_k^\alpha_s\, A(E)}{c_k-E} + \frac{\beta_k^\beta_s\, B(E)}{c_k-E} \right)\right. \nonumber\\[2ex]+ \left.\left( \frac{\alpha_k^\alpha_{k'}\, A(E)}{c_k-E} + \frac{\beta_k\beta_{k'}\, B(E)}{c_k-E} \right) \left( \frac{\alpha_{k'}^\alpha_s\, A(E)}{c_{k'}-E} + \frac{\beta_{k'}\beta_s\, B(E)}{c_{k'}-E} \right) +\cdots \right\}\nonumber\\[2ex] = \frac{\delta_{ks}}{c_k-E} + \alpha^_k\alpha_s a_{ks}+ \beta^_k\beta_s b_{ks}+ \alpha^_k\beta_s c_{ks}+ \beta_k^\alpha_s d_{ks}\,. \label{25}\end{aligned}$$ Then, it remains to determine the coefficients $a_{ks}$, $b_{ks}$, $c_{ks}$, $d_{ks}$, with the use of the obvious relation $\sum_{k'} T_{kk'}^{-1}T_{k's}=\delta_{ks}$, which gives a expression of the form: $$\begin{aligned} \alpha^_k\alpha_s C_1(a_{ks},b_{ks},\dots)+ \beta^_k\beta_s C_2(a_{ks},b_{ks},\dots) &+&\alpha^_k\beta_s C_3(a_{ks},b_{ks},\dots) \nonumber\\[2ex] &+& \beta_k^\alpha_s C_4(a_{ks},b_{ks},\dots)=0\,. \nonumber\\ \label{26}\end{aligned}$$ If we assume linear independence of the variables $\alpha_k$ and $\beta_s$, the coefficients of (\[26\]) should vanish for all values of the indices $r$ and $s$: $$C_p(a_{ks},b_{ks},\dots)=0\,,\hskip2cm p=1,2,3,4\,.$$ By implementing in (\[25\]) condition (\[26\]) one obtains $$\begin{aligned} (c_s-E) a_{ks}-\frac{A(E)}{c_k-E} -A(E)\sum_{k'}[ \alpha_{k'}\alpha_{k'}^* a_{k's} +\beta_{k'}\alpha_{k'}^* c_{kk'}] =0 \,. \label{27} \\ [2ex] (c_s-E) b_{ks}-\frac{B(E)}{c_k-E} -B (E)\sum_{k'}\left[\beta_{k'}\beta_{k}^* b_{kk'} + \alpha_{k'}\beta_{k'}^* d_{kk'}\right]=0\,. \label{28} \\[2ex] c_{ks}(c_s-E)-B(E)\sum_{k'}[\alpha_{k'}\beta_{k'}^* a_{kk'}+\beta_{k'}\beta_{k'}^* c_{kk'}]=0\,. \label{29} \\[2ex] d_{ks}(c_s-E) -A(E)\sum_{k'}[\beta_{k'}\alpha_{k'}^* b_{kk'}+ \alpha_{k'}\alpha_{k'}^* d_{kk'}]=0\,. \label{30}\end{aligned}$$ Let us start by solving Equation (\[29\]). It can be written as $$c_{ks}= \frac{B(E)R_k}{c_s-E} \,, \label{31}$$ where $$R_k= \sum_{k'}[\alpha_{k'}\beta_{k'}^* a_{kk'}+\beta_{k'}\beta_{k'}^* c_{kk'}]\,. \label{32}$$ In order to simplify subsequent equations, we shall introduce the following notation: $$\label{33} \tau(\alpha,\beta,E):= \sum_k\frac{\alpha_k\beta_k}{c_k-E}\,.$$ Using this definition in (\[31\]-\[32\]) gives $$\label{34} \frac{B(E)\,R_k}{c_s-E}=\frac{B(E)}{c_s-E}\left(\sum_{k'}\alpha_{k'}\beta_{k'}^* a_{kk'}+\tau(\beta,\beta^,E)\,B(E)\,R_k\right)\,,$$ and from it the quantity $R_k$ can be expressed as $$R_k=\frac{\sum_{k'} a_{k'}\beta_{k'}^ a_{kk'}}{1-B(E)\,\tau(\beta,\beta^,E)}\,.\label{35}$$ In a similar manner $$c_{ks}=\frac{B(E)}{c_s-E}\,\sum_{k'}\frac{\alpha_{k'}\beta_{k'}^a_{kk'}}{1-B(E)\,\tau(\beta,\beta^,E)}\,.\label{36}$$ and $$a_{ks}=\frac{A(E)}{c_s-E}\,\left[\frac{1}{c_k-E}+ \sum_{k'}(\alpha_{k'}\alpha_{k'}^a_{kk'}+\beta_{k'}\alpha_{k'}c_{kk'})\right]\,.\label{37}$$ Note that the inverse matrix $T^{-1}$ does exist if and only if the $R_k$ are nonsingular for all $E$, i.e., the denominator in (\[35\]) does not vanish: $$\label{38} 1-B(E)\,\tau(\beta,\beta^,E)\ne 0\,.$$ If one proceeds with the expression of $a_{ks}$ in (\[27\]) as done with $c_{ks}$, one finds $$a_{ks}=\frac{A(E)\,\Pi_k}{c_s-E}\,,\label{39}$$ with $$\Pi_k:= \frac{1}{c_k-E}+\sum_{k'}\alpha_{k'}\alpha_{k'}^a_{kk'}+\frac{B(E)\,\tau(\beta,\alpha^,E)}{1-B(E)\, \tau(\beta,\beta^,E)}\,\sum_{k'}\alpha_{k'}\beta_{k'}^a_{kk'}\,.\label{40}$$ From (\[39\]) and (\[40\]), one gets $$\begin{aligned} \frac{A(E)\,\Pi_k}{c_s-E}=\frac{A(E)}{c_s-E}\left[\frac{1}{c_k-E}+A(E)\Pi_k\, \tau(\alpha,\alpha^,E)\right.\nonumber\\[2ex] \left.+ B(E)\,\frac{\tau(\beta,\alpha^,E)\,\tau(\alpha,\beta^,E)}{1-\tau(\beta,\beta^,E)\,B(E)} +A(E)\Pi_k\right]\,, \label{41}\end{aligned}$$ from which we obtain a consistent equation for $\Pi_k$: $$\begin{aligned} \Pi_k=\frac{1}{c_k-E}\left[1-A(E)\,\tau(\alpha,\alpha^,E)-A(E)B(E)\,\frac{\tau(\beta,\alpha^,E)\, \tau(\alpha,\beta^,E)}{1-\tau(\beta,\beta^,E)\,B(E)}\right]\,,\label{42}\end{aligned}$$ leading to $$a_{ks}=\frac{A(E)\,(1-B(E)\,\tau(\beta,\beta^,E))}{(c_k-E)(c_s-E)}\,\Delta^{-1}\,,\label{43}$$ with $$\begin{aligned} \Delta:= 1-A(E)\,\tau(\alpha,\alpha^,E)-B(E)\,\tau(\beta,\beta^,E)\nonumber\\[2ex]+ A(E)B(E)\,[\tau(\alpha,\alpha^,E)\,\tau(\beta,\beta^,E)-\tau(\alpha,\beta^,E)\,\tau(\beta,\alpha^,E)]\,.\label{44}\end{aligned}$$ Accordingly, nontrivial solutions of $a_{ks}$ ($a_{ks}\ne 0$) exist if and only if $\Delta\ne 0$ for all values of the energy $E$. Finally, the expressions for the remaining coefficients are $$b_{ks}=\frac{B(E)\,(1-A(E)\,\tau(\alpha,\alpha^,E))}{(c_s-E)(c_k-E)}\,\Delta^{-1}\,, \label{45}$$ $$d_{ks}= \frac{A(E)B(E)\,\tau(\beta,\alpha^,E)}{(c_s-E)(c_k-E)}\,\Delta^{-1}\,. \label{46}$$ With these elements, we are now in the position of writing the explicit form for the inverse matrix $T^{-1}$ $$\begin{aligned} T_{ks}^{-1}=\frac{\delta_{ks}}{c_k-E}+\frac{\Delta^{-1}}{(c_k-E)(c_s-E)} \,\left\{\alpha_k^\alpha_s\,A(E)\,[1-B(E)\,\tau(\beta,\beta^,E)]\right.\nonumber \\[2ex] +\beta_k^\beta_s \,B(E)\,[1-A(E)\,\tau(\alpha,\alpha^,E)]+\alpha_k^\beta_s^\,A(E)\,B(E)\,\tau(\alpha,\beta^,E) \nonumber\\[2ex] +\beta_s^\alpha_s\,A(E)\,B(E)\,\tau(\beta,\alpha^,E)\}\,. \label{47}\end{aligned}$$ In consequence, the formal solution of Equation (\[23\]) is given by $$\varphi_k(E)=-c\sum_k T^{-1}_{kn}\left(\sum_l f_{ln}^(c_l-E)\right) \label{48}$$ and $$\phi_{k1}(E)=c\frac{\beta_k}{c_k+\omega_0-E}\sum_{n,m}\beta^_m\,T^{-1}_{mn}\left(\sum_l f_{lk}^(c_l-E)\right)\,.\label{49}$$ Thus, inserting (\[48\]) and (\[49\]) in (\[13\]) and then in (\[8\]) the solution to our problem is completely determined. Note that the structure of the solutions is preserved by form factors which may not been separable as those of the choice (\[16\]), since they are absorbed in the definition of the functions $\tau$. Concluding remarks. =================== In this paper, we have described the solutions of the extended Friedrichs model. The proposed model includes fermion-boson couplings in addition to the boson-boson couplings. From the physical point of view, this extension may become a suitable tool for the treatment of decaying states which participate in the observable total decay width of composite nuclear systems. From the mathematical point of view, we have generalized the schematic solution proposed in a previous article [@6] and formulated the solution in a more compact and formal way. The main features of the proposed method can be summarized as follows: i.) The resonant behavior of the fermionic sector depends on the resonant structure of the boson field, which participate in the fermion solution via the fermion-boson coupling. ii.) The set of coupled equations is greatly simplified by the use of a T-matrix like formalism, which allows us to identify the resonant behavior of the solutions. Further applications of the set of Equations (\[47\]-\[49\]) to situations of physical interest are in progress. Acknowledgements. {#acknowledgements. .unnumbered} ================= Financial support is acknowledged to the Ministry of Education of Spain, grants PR2004-0080, MTM2005-09183, FIS2005-03989, SAB2004-0169, the Junta de Castilla y León Project VA013C05, the Russian Science Foundation Grant 04-01-00352 and the CONICET (Argentina). [99]{} K. O. Friedrichs, Commun. Pure Appl. Math., [1]{}, 361 (1948). A.Likhoded, G.P. Pronko, Int. J. Theor. Phys., [36]{}, 2335 (1997). I.E. Antoniou, M. Gadella, J. Mateo, G.P. Pronko, Int. J. Theor. Phys., [42]{}, 2389 (2003). I.E. Antoniou, M. Gadella, J. Mateo, G.P. Pronko, J. Phys A: Math. Gen, [36]{}, 12109 (2003). O. Civitarese, M. Gadella, Phys. Rep., [396]{}, 41-113 (2004). O. Civitarese, M. Gadella, International Journal of Modern Physics E, [15]{}, 1273-1290 (2006). A.T. Kruppa, P.H. Heenen, Phys. Rev. C, [63]{} (2001) 034304. G. Rudin, M. Gadella, Int. J. Quan. Chem., [58]{}, 441 (1996). I.E. Antoniou, M. Gadella, [Irreversibility, Resonances and Rigged Hilbert Spaces]{} in [Irreversible Quantum Dynamics]{}, F. Benatti and R. Floreanini, Eds., Springer Lecture Notes in Physics, vol [622]{}, Springer Verlag, Berlin, Heidelberg (2003), pp. 246-302. H. Baumgärtel, Reviews in Mathematical Physics, [18]{} (2006) 61. S. Weinberg, [The Quantum Theory of Field. Volume I: Foundations]{}, Cambridge University Press, Cambridge (1995). A. Bohr, B. Mottelson, Nuclear Structure, vol. II (Benjamin, Reading, M. A. 1975). R.G. Newton, [Scattering Theory of Waves and Particles]{} (Springer, New York, Heidelberg, Berlin, 1982). I.E. Antoniou, M. Gadella, I. Prigogine, G.P. Pronko, J. Math. Phys., [39]{}, 2995 (1998). L. Horwitz, P. Marchand, Rocky Mt. J. Math., [1]{}, 225 (1971). I.E. Antoniou, M. Gadella, G.P. Pronko, J. Math. Phys., [39]{}, 2459-2475 (1998). O. Civitarese, A.G. Dumrauf, R.J. Liotta, Phys. Rev. C, [47]{}, 1463 (1993). [^1]: e-mail: [email protected] [^2]: e-mail: [email protected] [^3]: e-mail: [email protected]	{ "pile_set_name": "ArXiv" }
--- abstract: 'Let $\pi=(\pi_1,\pi_2,\hdots,\pi_n)$ be permutation of the elements $1,2,\hdots,n. $ Positive integer $k\leq2^{n-1}$ we call index of $\pi,$ if in its binary notation as $n$-digital binary number, the 1’s correspond to the ascent points. We study behavior and properties of numbers of permutations of $n$ elements having index $k.$' address: \| Departments of Mathematics\ Ben-Gurion University of the Negev\ Beer-Sheva 84105, Israel. e-mail:[email protected] author: - Vladimir Shevelev title: 'On the basis polynomials in the theory of permutations with prescribed up-down structure' --- Introduction ============ D.Andre [@2] first considered (1881) the problem of the enumerating the alternating permutations $\pi=(\pi_1,\ldots \pi_n)$ of the numbers $1,2,\ldots, n$ for which ups and downs are alternating: $$\pi_1<\pi_2>\pi_3< \ldots$$ This problem has a highly aesthetic solution: the exponential generating function of such permutations is the sum of tangent and secant. But only after a century (1968) I.Niven [@11] considered a general problem of the enumerating the permutations with given up-down structure. For permutation $\pi=(\pi_1,\ldots,\pi_n)$, the sequence $(q_1,q_2,\ldots,q_{n-1})$, where $$\label{1} q_j=sign(\pi_{j+1}-\pi_j)=\begin{cases}1,& if \;\; \pi_{j+1}>\pi_j \\-1,& if\;\; \pi_{j+1}<\pi_j\end{cases},$$ is called a Niven’s signature.For example, $a=(2,1,5,4,3)$ has the signature $(-1,1,-1,-1)$. Denote $[q_1,q_2,\ldots,q_{n-1}]$ the number of permutations having the Niven’s signature $(q_1,q_2,\ldots,q_{n-1})$. In view of the symmetry, we have $$\label{2} [q_1,q_2,\ldots,q_{n-1}]=[-q_{n-1},-q_{n-2},\ldots,-q_1].$$ Niven obtained the following basic result. \[t1\] [@11]. Let in the signature $(q_1,q_2,\ldots,q_{n-1})$ the indices of those $q_i$ which are $+1$ be $k_1<k_2<\ldots<k_m$ (if such $q_i$ do not exist then assume $m=0$). Put in addition $k_0=0,\;\;k_{m+1}=n$. Then $$\label{3} [q_1,q_2,\ldots,q_{n-1}]= det N,$$ where $N=\{n_{ij}\}$ is the square matrix of order $m+1$ in which $$\label{4} n_{ij}=\begin{pmatrix} k_i\\k_{j-1}\end{pmatrix},\;\;i,j=1,2,\ldots, m+1.$$ After this celebrated Niven’s result and until now there has been a series of articles by many authors. We mention only eleven papers in chronological order: N.G.Bruijn, 1970 [@5], H.O.Foulkes, 1976 [@7], L.Carlitz, 1978 [@6], G.Viennot, 1979 [@18], C.L.Mallows and L.A.Shepp, 1985 [@9], V.Arnold, 1990 [@3], V.S.Shevelev, 1996 [@14], G.Szpiro, 2001 [@17],B.Shapiro, M.Shapiro and A.Vainshtein, 2005 [@12], F.C.S.Brown, T.M.A. Fink and K.Willbrand, 2007 [@4],R. Stanley, 2007 [@16]. According to the de Bruijn-Viennot algorithm [@5], [@18], the calculation of $[q_1,q_2,\ldots,q_{n-1}]$ could be realized using the following numerical triangle. At the top of the triangle put 1. Write 0 to the right (left) if $q_{n-1}$ is $1(-1)$. For example, if $q_{n-1}=1$ then the first two elements of the triangle are: $\begin{matrix}& 1 \\ && 0 \end{matrix}$. Summing these elements we write the sum to the left: $\begin{matrix}& 1 \\1 && 0 \end{matrix}$. Now the following $0$ we write to the right of the last element (to the left of the first element) if $q_{n-2}$ is $1(-1)$. For example, if $q_{n-2}=1$ then we have $\begin{matrix}& 1 \\1 && 0 \\&&& 0\end{matrix}$. Now the third row is obtained by summing each element in the third row from the right to the left, first element being $0$, with the elements left and above it in the second row: $$\begin{matrix}&& 1 \\&1 && 0 \\1 && 0 && 0 \end{matrix}$$ In case of $q_{n-2}=-1$ we have $\begin{matrix}&& 1 \\&1 && 0 \\0 \end{matrix}$ and the third row is obtained by summing each element in the third row from the left to the right, the first element being $0$, with the elements right and above it in the second row: $$\begin{matrix}&& 1 \\&1 && 0 \\0 && 1 && 1 \end{matrix}$$ The process is continued until the n-th row which corresponds to $q_1$. Now the sum of elements of the n-th row is equal to $[q_1,q_2,\ldots,q_{n-1}]$. . For signature $(-1,1,1,-1,1)$ we have the triangle $$\setcounter{MaxMatrixCols}{21}\begin{matrix}&&&&& 1 \\&&&&1 && 0 \\&&&0 && 1 && 1 \\&&2&&2&&1&&0\\&5&&3&&1&&0&&0\\0&&5&&8&&9&&9&&9\end{matrix}$$ Therefore, $[-1,1,1,-1,1]=0+5+8+9+9+9=40$. In order to obtain a weight generalization, let us consider a matrix function which we call “alternant” (cf.[@13]). If a permutation $\pi$ has the signature $(q_1,q_2,\ldots,q_{n-1}),$ then we write $\pi\in (q_1,q_2,\ldots,q_{n-1})$. Furthermore, if $\pi_i=j,$ then to the two-dimensional point $(i,j)$ assign the “weight” $a_{ij}.$ Let $A=(a_{ij})$ be an $n\times n$ matrix. Denote $$\label{5} alt_{(q_1,\ldots,q_{n-1})}A=\sum_{\pi\in(q_1,\ldots,q_{n-1})}\prod^n_{i=1}a_{i\pi_i}.$$ Let $A_{1j},\; j=1,\ldots,n,$ be $(n-1)\times(n-1)$ matrix which is obtained from $A$ by the deletion of the first row and the j-th column. Denote $A^{(+1)}_{1j}(A^{(-1)}_{1j})$ the matrix which is obtained from $A_{1j}$ by replacing the $j-1$ first (the n-j last) elements of its first row by 0’s. Then, by the Viennot’s algorithm, we deduce the following expansion of the alternant by the first row of the matrix. \[t2\] $($cf.[@13]$)$ $$\label{6} alt_{(q_1,\ldots,q_{n-1})}A=\sum^n_{j=1}a_{1j}alt_{(q_2,\ldots,q_{n-1})}A^{(q_1)}_{1j}.$$ Note that, if $A=J_n-n\times n$ matrix composed of 1’s only, then $$\label{7} alt_{(q_1,\ldots,q_{n-1})}A=[q_1,\ldots,q_{n-1}].$$ In case of arbitrary (0,1) matrix $A$, Theorem 2 gives enumeration the permutations having signature $(q_1,\ldots,q_{n-1})$ with restriction on positions. For example, if $I_n$ is $(n\times n)$ identity matrix, then $alt_{(q_1,\ldots,q_{n-1})}(J-I)$ gives the number of such permutations without fixed points. Note that, by (\[5\]) $$\label{8} alt_{(1)}\begin{pmatrix}a_{11}&a_{12}\\a_{21}&a_{22}\end{pmatrix}=a_{11}a_{22},\;\; alt_{(-1)}\begin{pmatrix}a_{11}&a_{12}\\a_{21}&a_{22}\end{pmatrix}=a_{12}a_{21}.$$ \[2\]. $$alt_{(1,-1,1)}J_4=alt_{(1,-1,1)}\begin{pmatrix}1&1&1&1\\1&1&1&1\\1&1&1&1\\ 1&1&1&1\end{pmatrix}=alt_{(-1,1)}\begin{pmatrix}1&1&1\\1&1&1\\1&1&1 \end{pmatrix}+$$ $$alt_{(-1,1)}\begin{pmatrix}0&1&1\\1&1&1\\1&1&1 \end{pmatrix}+alt_{(-1,1)}\begin{pmatrix}0&0&1\\1&1&1\\1&1&1 \end{pmatrix}+alt_{(-1,1)}\begin{pmatrix}0&0&0\\1&1&1\\1&1&1 \end{pmatrix}=$$ $$\left(alt_{(1)}\begin{pmatrix}0&0\\1&1 \end{pmatrix}+alt_{(1)}\begin{pmatrix}1&0\\1&1 \end{pmatrix}+alt_{(1)}\begin{pmatrix}1&1\\1&1 \end{pmatrix}\right)+$$ $$\left(alt_{(1)}\begin{pmatrix}1&0\\1&1 \end{pmatrix}+alt_{(1)}\begin{pmatrix}1&1\\1&1 \end{pmatrix}\right)+alt_{(1)}\begin{pmatrix}1&1\\1&1 \end{pmatrix}=0+1+1+1+1+1=5,$$ while $$alt_{(1,-1,1)}(J_4-I)=alt_{(1,-1,1)}\begin{pmatrix}0&1&1&1\\1&0&1&1\\1&1&0&1\\ 1&1&1&0\end{pmatrix}=alt_{(-1,1)}\begin{pmatrix}0&1&1\\1&0&1\\1&1&0 \end{pmatrix}+$$ $$alt_{(-1,1)}\begin{pmatrix}0&0&1\\1&1&1\\1&1&0 \end{pmatrix}+alt_{(-1,1)}\begin{pmatrix}0&0&0\\1&1&0\\1&1&1 \end{pmatrix}=$$ $$\left(alt_{(1)}\begin{pmatrix}1&0\\1&0 \end{pmatrix}+alt_{(1)}\begin{pmatrix}1&0\\1&1 \end{pmatrix}\right)+alt_{(1)}\begin{pmatrix}1&1\\1&1 \end{pmatrix}=2.$$ Note that alternant is also useful for enumeration the permutations with some additional conditions. For example, if it is necessary to enumerate the permutations $\pi$ with signature $(q_1,q_2,\ldots,q_{n-1})$ for which $\pi_1=l,\;\;\pi_n=m$ then we should calculate $alt J_n^{(l,m)}$ where $J_n^{(l,m)}$ is obtained from $J_n$ by replacing all 1’s of the first and the last rows by 0’s except the l-th 1 and the m-th 1 correspondingly. For example, there are only 2 such permutations in the case of the signature $(-1,1,1,-1,1),\;n=6,\;l=2,\;m=6$. Let us now introduce an index of the Niven’s signature in the following way: the integer $k=k_n$ is called the index of the signature $(q_1,q_2,\ldots,q_{n-1})$ if $(n-1)$-digit binary representation of $k$ is $$\label{9} k=\sum^{n-1}_{i=1}q_i^{\shortmid} 2^{n-i-1},$$ where $$\label{10} q^\shortmid_i=\begin{cases} 1,\;\;if\;\;q_i=1,\\0,\;\;otherwise\end{cases}.$$ (cf [@13], [@14], where we accepted $k+1$ as an index). Denote $S^{(k)}_n$ the set of permutations of elements $1,2,\ldots,n$ having the index $k$, and put $$\label{11} \left\{\begin{matrix} n \\ k \end{matrix}\right\}=\left\|S_n^{(k)}\right\}.$$ Let $k\in[2^{t-1}, 2^t)$ and the $(n-1)$-digit binary expansion of $k$ (\[9\]) has a form: $$\label{12} k=\underbrace{0 \ldots 0}_{n-t-1}\; 1 \;\underbrace{0 \ldots 0}_{s_2-s_1-1}\; 1 \;\underbrace{0 \ldots 0}_{s_3-s_2-1}\; 1\ldots 1 \;\underbrace{0 \ldots 0}_{s_m-s_{m-1}-1}\; 1 \;\underbrace{0 \ldots 0}_{t-s_m}\enskip,$$ where $$\label{13} 1=s_1< s_2 <\ldots < s_m$$ are places of 1’s after $n-t-1$ 0’s before the first 1. From Theorem 1 follows an important formula. \[t3\] For index $k\in [2^{t-1}, 2^t)\enskip(12),$ we have $$\label{14} \left\{\begin{matrix} n \\ k \end{matrix}\right\}=$$ $$\begin{vmatrix}1 & 1& 0& \hdots & 0\\1&\begin{pmatrix}n-t+s_2-1 \\ s_2 -1 \end{pmatrix} & 1& \hdots & 0 \\1&\begin{pmatrix}n-t+s_3-1 \\ s_3 -1 \end{pmatrix} & \begin{pmatrix}n-t+s_3-1 \\ s_3 -s_2 \end{pmatrix}& \hdots & 0\\\hdots & \hdots & \hdots & \hdots & \hdots \\ 1 & \begin{pmatrix}n-t+s_m-1 \\ s_m -1 \end{pmatrix}& \begin{pmatrix}n-t+s_m-1 \\ s_m -s_2 \end{pmatrix} & \begin{pmatrix}n-t+s_m-1 \\ s_m -s_3 \end{pmatrix} & 1\\ 1 & \begin{pmatrix}n \\ t \end{pmatrix} & \begin{pmatrix}n \\ t+1-s_2 \end{pmatrix}& \begin{pmatrix}n \\ t+1-s_3 \end{pmatrix} & \begin{pmatrix}n \\ t+1-s_m \end{pmatrix}\end{vmatrix}.$$ In [@14], using a techniques of permanents, the following explicit formula was proved. \[t4\] [@14]. For $k\in[2^{t-1}, 2^t)$ we have $$\left\{\begin{matrix} n \\ k \end{matrix}\right\}=(-1)^m+$$ $$\label{15} \sum^m_{p=1}(-1)^{m-p}\sum_{1\leq i,< i_2< \hdots< i_p\leq m} \left( \begin{matrix} n \\ t+1-s_{i_p}\end{matrix}\right)\prod ^p_{r=2}\begin{pmatrix} n-t+s_{i_r}-1 \\ s_{i_r}-s_{i_{r-1}}\end{pmatrix}.$$ Most likely, (\[15\]) is the first non-determinant formula in a closed form for the number of permutations with prescribed up-down structure (cf.[@16], [@4]). The Thue-Morse sequence [@10], [@8] is defined by $$\label{16} \tau_n=(-1)^{\sigma(n)},$$ where $\sigma(n)$ denotes the number of 1’s in the binary expansion of $n$. Thus, from (\[15\]) immediately follows an interesting arithmetical property of $\left\{\begin{matrix} n \\ k \end{matrix}\right\}$. \[t5\] [@14]. If all different from 1 divisors of $n$ are larger than $\lfloor\log_2 2k\rfloor, $ then $$\label{17} \left\{\begin{matrix} n \\ k \end{matrix}\right\}\equiv \tau_k \;(\mod{n}).$$ \[1\] It is evident that the validity of $(\ref{17})$ does not depend on the fact whether $k$ is a constant or a function of $n$. \[2\] From Theorem $5$ and $(\ref{16})$ it follows that, if $n$ has only sufficiently large different from 1 divisors, then the number of permutations of $n$ elements with arbitrary prescribed up-down structure is $\pm 1(\mod{n})$. Recently, a special case of this result for a prime $n$ was reproduced in [@4]. Basis polynomials ================= As it follows from (\[15\]), if $k$ does not depend on $n,$ then $\left\{\begin{matrix}n\\ k \end{matrix}\right\}$ is a polynomial in $n$ of the degree $$\label{18} t=\lfloor\log_2(2k)\rfloor.$$ Indeed, the degree of the polynomial in the interior sum of (\[15\])equals to $$t+1-s_{i_p}+(s_{i_2}-s_{i_1})+(s_{i_3}-s_{i_2})+\hdots+(s_{i_p}-s_{i_{p-1}}) =t+1-s_{i_1}\leq t$$ and the equality is attained in those summands of the sum in which $i_1=1$. Let us draw an analogy with the binomial coefficients $\begin{pmatrix}n\\ k \end{pmatrix}.$ 1a. $\begin{pmatrix}n\\ k \end{pmatrix}$ is the number of subsets of the cardinality $k$ of a set of $n$ elements. 1b. $\left\{\begin{matrix} n\\k \end{matrix}\right \}$ is the number of permutations of $n$ elements having the up-down index $k$. 2a. Each subset of a set of $n$ elements is contained in the number of $\begin{pmatrix}n\\ k \end{pmatrix}$ subsets for some value of $k$. 2b. Each permutation of $n$ elements is contained in the number of $\left\{\begin{matrix}n\\ k \end{matrix}\right\}$ permutations for some value of the up-down index $k$. 3a. $\sum^n_{k=0}\left(\begin{matrix}n\\ k \end{matrix}\right)=2^n$. 3b. $ \sum^{2^{n-1}-1}_{k=0}\left\{\begin{matrix}n\\ k \end{matrix}\right\}=n! $ 4a. $\begin{pmatrix}n\\ n-k \end{pmatrix}=\begin{pmatrix}n\\ k \end{pmatrix}$. 4b. In view of (\[2\]), $\left\{\begin{matrix}n\\ k \end{matrix}\right\}=\left\{\begin{matrix}n\\ 2^{n-1}-1-k \end{matrix}\right\}$. 5a.$\begin{pmatrix}n\\ 0 \end{pmatrix}=\begin{pmatrix}n\\ n \end{pmatrix}=1$. 5b. $\left\{\begin{matrix}n\\ 0 \end{matrix}\right\}=\left\{\begin{matrix}n\\ 2^{n-1}-1 \end{matrix}\right\}=1$. The latter equality corresponds to the identity permutation. 6\. The central binomial coefficients and the “central” numbers $\left\{\begin{matrix}n\\ k \end{matrix}\right\}$ are equal one to another. Indeed, below (section 3) we prove that $$\left\{\begin{matrix}2n\\2^{n-1}-1 \end{matrix}\right\}=\begin{pmatrix}2n-1\\n-1 \end{pmatrix},$$ $$\left\{\begin{matrix}2n+1\\2^n-1 \end{matrix}\right\}=\begin{pmatrix}2n\\n \end{pmatrix}.$$ In view of this analogy, we call $\left\{\begin{matrix}n\\k \end{matrix}\right\},\;\;k=1,2,\hdots$, the basis polynomial in theory of permutations with prescribed up-down structure. Note that $\left\{\begin{matrix}n\\k \end{matrix}\right\}$, just as $\begin{pmatrix}n\\k \end{pmatrix}$, generally is not a polynomial if $k$ is a function of $n$. . In case of alternating permutations $\pi_1<\pi_2>\pi_3<\hdots, $ we have the sequence of indices $\{k_{n-1}\}$ such that $$k_1=1,\;\;k_2=2,\;\;k_3=5,\;\;k_4=10,\;\;k_5=21,\hdots\enskip.$$ Here $$k_n-k_{n-2}=2^{n-1},\;\; n\geq 3,$$ whence $$\label{19} k_{n-1}=\frac{2^{n+1}-3+(-1)^{n}}{6},\;\;n=1,2,\hdots\enskip$$ Thus, from the classical Andre’s result we obtain $$\label{20} \sum^{\infty}_{n=0}\left\{\begin{matrix}n\\\frac{2^{n+1}-3+(-1)^{n}}{6} \end{matrix}\right\}\frac{x^n}{n!}=\tan{x}+\sec{x},$$ where we put $\left\{\begin{matrix}0\\0 \end{matrix}\right\}=1.$ From (\[20\]) we have some values of $\left\{\begin{matrix}n\\m \end{matrix}\right\}:$ $$\left\{\begin{matrix}1\\0 \end{matrix}\right\}=1,\;\left\{\begin{matrix}2\\1 \end{matrix}\right\}=1,\;\left\{\begin{matrix}3\\2 \end{matrix}\right\}=2,\;$$ $$\label{21} \left\{\begin{matrix}4\\5 \end{matrix}\right\}=5,\;\left\{\begin{matrix}5\\10 \end{matrix}\right\}=16,\;\left\{\begin{matrix}6\\21 \end{matrix}\right\}=61,\;\hdots\enskip .$$ It is well-known that these values are explicitly expressed with help of the absolute values of the Bernoulli and Euler numbers. Another general formula for basis polynomials ============================================= Here we use formulas (\[12\])-(\[14\]) for obtaining more simple general explicit formula. It is important,for the subsequent development of our theory, to present the results in the form of linear combination of the binomial coefficients $\begin{pmatrix}n\\i \end{pmatrix}$. Firstly consider several special cases. a\) $m=1,\; k=2^{t-1}$. Then $$\label{22} \left\{\begin{matrix}n\\2^{t-1} \end{matrix}\right\}=\left\|\begin{matrix}1& 1\\1 &\begin{pmatrix}n\\t \end{pmatrix} \end{matrix}\right\|=\begin{pmatrix}n\\t \end{pmatrix}-1;$$ b\) $m=2,\; k=2^{t-1}+2^{u-1}$. Here $s_1=1,\;s_2=t-u+1$ and we have $$\left\{\begin{matrix}n\\2^{t-1}+2^{u-1} \end{matrix}\right\}=\left\|\begin{matrix}1& 1 & 0\\1 &\begin{pmatrix}n-u\\t-u \end{pmatrix}& 1\\1 &\begin{pmatrix}n\\t \end{pmatrix}& \begin{pmatrix}n\\u \end{pmatrix}\end{matrix}\right\|=\begin{pmatrix}n-u\\t-u \end{pmatrix}\begin{pmatrix}n\\u \end{pmatrix}-\begin{pmatrix}n\\t \end{pmatrix}-\begin{pmatrix}n\\u \end{pmatrix}+1.$$ Notice that, $$\label{23} \begin{pmatrix}n-u\\t-u \end{pmatrix}\begin{pmatrix}n\\u \end{pmatrix}=\begin{pmatrix}t\\u \end{pmatrix}\begin{pmatrix}n\\t \end{pmatrix}.$$ Therefore, $$\label{24} \left\{\begin{matrix}n\\2^{t-1}+2^{u-1} \end{matrix}\right\}=\left(\begin{pmatrix}t\\u \end{pmatrix}-1\right)\begin{pmatrix}n\\t \end{pmatrix}-\begin{pmatrix}n\\u \end{pmatrix}+1.$$ c\) $m=3,\;k=2^{t-1}+2^{u-1}+2^{v-1}$. Here $s_1=1,\; s_2=t-u+1,\; s_3=t-v+1$ and we have $$\left\{\begin{matrix}n\\2^{t-1}+2^{u-1}+2^{v-1} \end{matrix}\right\}=\left\|\begin{matrix}1 & 1 & 0 & 0\\1 & \begin{pmatrix} n-u\\ t-u \end{pmatrix}& 1 & 0\\1 & \begin{pmatrix} n-v\\ t-v \end{pmatrix}& \begin{pmatrix} n-v\\ u-v \end{pmatrix} & 1\\1 & \begin{pmatrix} n\\t \end{pmatrix} & \begin{pmatrix} n\\u \end{pmatrix} & \begin{pmatrix} n\\v \end{pmatrix} \end{matrix}\right\|=$$ $$=\left\|\begin{matrix} \begin{pmatrix} n-u\\ t-u \end{pmatrix} & 1 & 0\\\begin{pmatrix} n-v\\ t-v \end{pmatrix}& \begin{pmatrix} n-v\\ u-v\end{pmatrix}& 1\\\begin{pmatrix} n\\ t \end{pmatrix}& \begin{pmatrix} n\\u \end{pmatrix}&\begin{pmatrix} n\\ v \end{pmatrix}\end{matrix}\right\|-\left\|\begin{matrix} 1 & 1 & 0\\1& \begin{pmatrix} n-v\\ u-v\end{pmatrix}& 1\\1& \begin{pmatrix} n\\u \end{pmatrix}&\begin{pmatrix} n\\ v \end{pmatrix}\end{matrix}\right\|=$$ $$=\begin{pmatrix} n-u\\ t-u\end{pmatrix}\begin{pmatrix} n-v\\ u-v\end{pmatrix}\begin{pmatrix} n\\ v\end{pmatrix}-\begin{pmatrix} n-u\\ t-u\end{pmatrix}\begin{pmatrix} n\\ u\end{pmatrix}-\begin{pmatrix} n-v\\ t-v\end{pmatrix}\begin{pmatrix} n\\ v\end{pmatrix}-$$ $$\label{25} -\begin{pmatrix} n-v\\ u-v\end{pmatrix}\begin{pmatrix} n\\ v\end{pmatrix}+\begin{pmatrix} n\\ t\end{pmatrix}+\begin{pmatrix} n\\ u\end{pmatrix}+\begin{pmatrix} n\\ v\end{pmatrix}-1.$$ Notice that, $$\label{26} \begin{pmatrix} n-u\\ t-u\end{pmatrix}\begin{pmatrix} n-v\\ u-v\end{pmatrix}\begin{pmatrix} n\\ v\end{pmatrix}=\begin{pmatrix} t\\ u\end{pmatrix}\begin{pmatrix} u\\ v\end{pmatrix}\begin{pmatrix} n\\ t\end{pmatrix},$$ $$\begin{pmatrix} n-u\\ t-u\end{pmatrix}\begin{pmatrix} n\\ u\end{pmatrix}=\begin{pmatrix} t\\ u\end{pmatrix}\begin{pmatrix} n\\ t\end{pmatrix},$$ $$\label{27} \begin{pmatrix} n-v\\ t-v\end{pmatrix}\begin{pmatrix} n\\ v\end{pmatrix}=\begin{pmatrix} t\\ v\end{pmatrix}\begin{pmatrix} n\\ t\end{pmatrix},$$ $$\begin{pmatrix} n-v\\ u-v\end{pmatrix}\begin{pmatrix} n\\ v\end{pmatrix}=\begin{pmatrix} u\\ v\end{pmatrix}\begin{pmatrix} n\\ u\end{pmatrix}.$$ Therefore, $$\left\{\begin{matrix} n\\ 2^{t-1}+2^{u-1}+2^{v-1}\end{matrix}\right\}=\left(\begin{pmatrix} t\\ u\end{pmatrix}\begin{pmatrix} u\\ v\end{pmatrix}-\begin{pmatrix} t\\ u\end{pmatrix}-\begin{pmatrix} t\\ v\end{pmatrix}+1\right)\begin{pmatrix} n\\ t\end{pmatrix}-$$ $$\label{28} -\left(\begin{pmatrix} u\\ v\end{pmatrix}-1\right)\begin{pmatrix} n\\ u\end{pmatrix}+\begin{pmatrix} n\\ v\end{pmatrix}-1.$$ At last, for arbitrary $m$, we obtain the following theorem. \[t6\] If $$\label{29} k=2^{t_1-1}+2^{t_2-1}+\hdots+2^{t_m-1},\;\;t_1>t_2>\hdots>t_m,$$ then $$\left\{\begin{matrix} n\\k\end{matrix}\right\}=(-1)^m\left(1-\sum^m_{i=1}\begin{pmatrix} n\\t_i \end{pmatrix}+\sum_{1\leq i<j\leq m}\begin{pmatrix}n\\t_i\end{pmatrix}\begin{pmatrix}t_i\\t_j\end{pmatrix}-\right.$$ $$\label{30} -\sum_{1\leq i<j\leq l\leq m}\begin{pmatrix}n\\t_i\end{pmatrix}\begin{pmatrix}t_i\\t_j\end{pmatrix}\begin{pmatrix}t_j\\t_l\end{pmatrix} +\hdots\left. +(-1)^m\begin{pmatrix}n\\t_1\end{pmatrix}\prod^{m-1}_{j=1}\begin{pmatrix}t_i\\t_{i+1} \end{pmatrix}\right).$$ Proof.Theorem 6 directly follows from (12), (15), (29) and the following easily proved identity (if to take into account that from (12) and (29) we have $s_i=t+1-t_i,\enskip i=1,....m):$ $$\begin{pmatrix} n\\t_{i_p}\end{pmatrix} \prod^p_{r=2}\begin{pmatrix} n-t_{i_r}\\ t_{i_{r-1}}-t_{i_r}\end{pmatrix}= \begin{pmatrix} n\\ t_{i_1}\end{pmatrix}\begin{pmatrix} t_{i_1}\\ t_{i_2}\end{pmatrix} \begin{pmatrix} t_{i_2}\\ t_{i_3}\end{pmatrix}\hdots\begin{pmatrix} t_{i_{p-1}}\\ t_{i_p}\end{pmatrix}.$$ $\blacksquare$ It is clear that in (\[30\]) $\left\{\begin{matrix} n\\k\end{matrix}\right\}$ is presented as a linear combination of $\begin{pmatrix} n\\i \end{pmatrix},\;\;0\leq i\leq t_1$. More exactly, as (\[30\]) shows, $\left\{\begin{matrix} n\\k\end{matrix}\right\}$ is a linear combination of elements of the last row of the determinant (\[14\]), that does not follow from (\[14\]) directly. Indeed, in case of (\[29\]) we have $$s_1=1,\;\;\; s_i=t_1-t_i+1,\;\;\; i=2,3,\hdots, m.$$ In (\[14\]) $t_1=t$ and, consequently, $$t+1-s_i=t_i,\;\;\;i=1,2,\hdots, m.$$ Thus, $\left\{\begin{matrix} n\\k \end{matrix}\right\}$ is an alternating sum of some elementary symmetric polynomials of binomial coefficients. It is also a polynomial in $n$ of degree $t_1=\lfloor\log_2(2k)\rfloor$. Note that another form of Theorem 6 is \[t7\] In conditions $(\ref{29})$ we have $$\label{31} \left\{\begin{matrix} n\\k \end{matrix}\right\}=(-1)^m+\sum^m_{p=1}c_p\begin{pmatrix} n\\t_p \end{pmatrix},$$ where $$c_p=(-1)^m\left(-1+\sum^m_{j=p+1}\begin{pmatrix} t_p\\t_j \end{pmatrix} -\sum_{p+1\leq j<l\leq m}\begin{pmatrix} t_p\\t_j \end{pmatrix}\begin{pmatrix} t_j\\t_l \end{pmatrix}+ \hdots +\right .$$ $$\label{32} \left. +(-1)^{m-p-1}\prod^m_{j=p+1}\begin{pmatrix} t_{j-1}\\t_j \end{pmatrix}\right).$$ In particular, $$c_m=(-1)^{m+1},$$ $$\label{33} c_{m-1}=(-1)^m\left(-1+\begin{pmatrix} t_{m-1}\\ t_m \end{pmatrix}\right),$$ $$c_{m-2}=(-1)^m\left(-1+\begin{pmatrix} t_{m-2}\\ t_{m-1} \end{pmatrix}+\begin{pmatrix} t_{m-2}\\ t_m \end{pmatrix}-\begin{pmatrix} t_{m-2}\\ t_{m-1} \end{pmatrix}\begin{pmatrix} t_{m-1}\\ t_m \end{pmatrix}\right),$$ etc. \[4\] Let $k=2^m-1$. Then, according to $(\ref{29}),$ we have $$t_m=1,\;\;\; t_{m-1}=2,\;\;\hdots,\;\;\; t_1=m$$ and, by $(\ref{33}),$ we find $$c_m=(-1)^{m+1},$$ $$c_{m-1}=(-1)^m(-1+2)=(-1)^m,$$ $$c_{m-2}=(-1)^m(-1+3+3-3\cdot 2)=(-1)^{m+1},$$ and, by induction, $$c_m=-c_{m-1}= c_{m-2}=\hdots=(-1)^{m-1}c_1=(-1)^{m-1},$$ i.e. $$c_p=(-1)^{p-1}.$$ Thus, by (\[31\]) we have $$\label{34} \left\{\begin{matrix} n\\ 2^m -1\end{matrix}\right\}=(-1)^m+\sum^m_{p=1}(-1)^{p-1}\begin{pmatrix} n\\m-p+1 \end{pmatrix}=\sum^m_{j=0}(-1)^{m-j}\begin{pmatrix} n\\j \end{pmatrix}=\begin{pmatrix} n-1\\m \end{pmatrix}.$$ The latter identity is proved easily by induction over $m$. In particular, putting in (\[34\]) $n=2m$ and $n=2m+1$ we have $$\label{35} \left\{\begin{matrix} 2m\\ 2^m-1\end{matrix}\right\}=\begin{pmatrix} 2m-1\\m\end{pmatrix},\;\;\left\{\begin{matrix} 2m+1\\ 2^m-1\end{matrix}\right\}=\begin{pmatrix} 2m\\m\end{pmatrix}.$$ This proves the analogy for the “central” number $\left\{\begin{matrix} n\\ k\end{matrix}\right\}$ and the central binomial coefficients. Comparing with (\[14\]), we obtain an identity: $$\label{36} \begin{vmatrix}1 & 1& 0& 0&\hdots & 0 \\1&\begin{pmatrix}n-m+1 \\ 1 \end{pmatrix} & 1& 0 & \hdots & 0 \\1&\begin{pmatrix}n-m+2 \\ 2 \end{pmatrix} & \begin{pmatrix}n-m+2 \\ 1 \end{pmatrix}& 1 & \hdots & 0\\\hdots & \hdots & \hdots & \hdots & \hdots & \hdots \\ 1 & \begin{pmatrix}n-1 \\ m-1 \end{pmatrix}& \begin{pmatrix}n-1 \\ m-2 \end{pmatrix} & \begin{pmatrix}n-1 \\ m-3\end{pmatrix} & \hdots & 1\\ 1 & \begin{pmatrix}n \\m\end{pmatrix}& \begin{pmatrix}n \\ m-1 \end{pmatrix} & \begin{pmatrix}n \\ m-2 \end{pmatrix}& \hdots & \begin{pmatrix}n \\ 1 \end{pmatrix}\end{vmatrix}=\begin{pmatrix}n-1 \\ m \end{pmatrix}$$ Formulas (\[31\])-(\[33\]) alow to calculate rather effectively basis polynomials $\left\{\begin{matrix} n\\ k\end{matrix}\right\}$. Nevertheless, there exist a recursions which are more effective for calculations. Recursion relation for basis polynomials ======================================== Consider now $\left\{\begin{matrix} n\\ k\end{matrix}\right\}$ from a more formal point of view as a polynomial (\[30\]) or (\[31\]). If $k$ has larger digits than $n-1$ then $\left\{\begin{matrix} n\\ k\end{matrix}\right\}$ loses its combinatorial sense and could take even negative values. Nevertheless, the formal values of $\left\{\begin{matrix} n\\ k\end{matrix}\right\}$ are useful since the coefficients $c_p$ (\[32\]) could be represented as some values of the basis polynomials. \[t8\] In conditions $(\ref{29})$ in $(\ref{31})$ we have $$c_p=\left\{\begin{matrix} t_p\\ k-2^{t_p-1}\end{matrix}\right\}.$$ Proof.By (\[29\]), $$k_1=k-2^{t_p-1}=2^{t_1-1}+2^{t_2-1}+\hdots +$$ $$+2^{t_{p-1}-1}+2^{t_{p+1}-1}+\hdots + 2^{t_m-1}.$$ Using (\[30\]) for $k_1$ and substituting $n=t_p\enskip,$ we obtain $$\left\{\begin{matrix} t_p\\ k-2^{t_p-1}\end{matrix}\right\}=(-1)^{m-1}\left(1-\sum^m_{i=p+1}\begin{pmatrix} t_p\\ t_i \end{pmatrix}+\sum_{p+1\leq i < j \leq m}\begin{pmatrix} t_p\\ t_i\end{pmatrix}\begin{pmatrix} t_i\\ t_j\end{pmatrix}- \hdots \right)$$ and the comparison with (\[32\]) gives the theorem. $\blacksquare$ From Theorems 7, 8 we obtain a very simple recursion relation. \[t9\] If $k=2^{t_1-1}+2^{t_2-1}+ \hdots + 2^{t_m-1},$ then $$\label{37} \left\{\begin{matrix} n\\ k\end{matrix}\right\}=(-1)^m+\sum^m_{p=1}\left\{\begin{matrix} t_p\\ k-2^{t_p-1}\end{matrix}\right\}\begin{pmatrix}n\\ t_p\end{pmatrix}.$$ Knowing $\left\{\begin{matrix} n\\ j\end{matrix}\right\},\;\; j\leq 20,$ to find $\left\{\begin{matrix} n\\ 21\end{matrix}\right\}$. We have $$21= 2^{5-1}+2^{3-1}+ 2^{1-1},\;\; t_1=5,\;\; t_2=3, \;\;t_3=1.$$ By (\[37\]), we obtain $$\label{38} \left\{\begin{matrix} n\\ 21\end{matrix}\right\}=-1+\left\{\begin{matrix} 5\\ 5\end{matrix}\right\}\begin{pmatrix}n\\ 5\end{pmatrix}+\left\{\begin{matrix} 3\\ 17\end{matrix}\right\}\begin{pmatrix}n\\ 3\end{pmatrix}+\left\{\begin{matrix} 1\\ 20\end{matrix}\right\}\begin{pmatrix}n\\ 1\end{pmatrix}.$$ Using formulas (see Appendix) $$\left\{\begin{matrix} n\\ 5\end{matrix}\right\}=2\begin{pmatrix}n\\ 3\end{pmatrix}-\begin{pmatrix}n\\ 1\end{pmatrix}+1,$$ $$\left\{\begin{matrix} n\\ 17\end{matrix}\right\}=4\begin{pmatrix}n\\ 5\end{pmatrix}-\begin{pmatrix}n\\ 1\end{pmatrix}+1,$$ $$\left\{\begin{matrix} n\\ 20\end{matrix}\right\}=9\begin{pmatrix}n\\ 5\end{pmatrix}-\begin{pmatrix}n\\ 3\end{pmatrix}+1,$$ we conclude that $$\left\{\begin{matrix} n\\ 5\end{matrix}\right\}=2\cdot 10 -5 +1=16,$$ $$\left\{\begin{matrix} 3\\ 17\end{matrix}\right\}=-2,$$ $$\left\{\begin{matrix} 1\\ 20\end{matrix}\right\}=1$$ and, by (\[38\]), we find $$\left\{\begin{matrix} n\\ 21\end{matrix}\right\}=16\begin{pmatrix}n\\ 5\end{pmatrix}-2\begin{pmatrix}n\\ 3\end{pmatrix}+\begin{pmatrix}n\\ 1\end{pmatrix}-1.$$ Another determinant formula for basis polynomials ================================================= In conditions (\[29\]) the determinant (\[14\]) has the form (cf.3,a),b),c)): $$\label{39} \left\{\begin{matrix} n\\ k\end{matrix}\right\}=\begin{vmatrix} 1& 1& 0 & 0& \hdots & 0\\ 1& \begin{pmatrix}n-t_2\\t_1-t_2\end{pmatrix} & 1& 0& \hdots & 0\\ 1& \begin{pmatrix}n-t_3\\t_1-t_3\end{pmatrix}& \begin{pmatrix}n-t_3\\t_2-t_3\end{pmatrix}& 1& \hdots & 0\\ \hdots & \hdots & \hdots & \hdots & \hdots & \hdots \\ 1& \begin{pmatrix}n-t_m\\t_1-t_m\end{pmatrix}& \begin{pmatrix}n-t_m\\t_2-t_m\end{pmatrix}& \begin{pmatrix}n-t_m\\t_3-t_m\end{pmatrix}& \hdots & 1\\ 1 &\begin{pmatrix}n\\ t_1\end{pmatrix}&\begin{pmatrix}n\\ t_2\end{pmatrix}&\begin{pmatrix}n\\ t_3\end{pmatrix}& \hdots &\begin{pmatrix}n\\ t_m\end{pmatrix}\end{vmatrix}.$$ Note that this determinant possesses an astonishing property. If to replace the lower triangular submatrix with the main diagonal of 1’s by the upper one such that the elements $\begin{pmatrix} n-t_j\\ t_i-t_j\end{pmatrix}\;(i< j)$ are mapped to elements $\begin{pmatrix} t_i\\ t_j\end{pmatrix}$ which are symmetric respectively the diagonal of 1’s (and which do not depend on $n(!)$), then the determinant does not change its value. If, in addition, to interchange the places of the first and last rows, then we obtain the following result. \[t10\] $k=2^{t_1-1}+2^{t_2-1}+ \hdots + 2^{t_m-1},\;\;t_1>t_2>\hdots>t_m$, then $$\label{40} \left\{\begin{matrix} n\\ k\end{matrix}\right\}=(-1)^m \begin{vmatrix} 1& \begin{pmatrix}n\\t_1\end{pmatrix}& \begin{pmatrix}n\\t_2\end{pmatrix}& \begin{pmatrix}n\\t_3\end{pmatrix}& \hdots & \begin{pmatrix}n\\t_m\end{pmatrix}\\ 1& 1 &\begin{pmatrix}t_1\\t_2\end{pmatrix}& \begin{pmatrix}t_1\\t_3\end{pmatrix}& \hdots &\begin{pmatrix}t_1\\t_m\end{pmatrix}\\ 1& 0& 1 &\begin{pmatrix}t_2\\t_3\end{pmatrix}&\hdots &\begin{pmatrix}t_2\\t_m\end{pmatrix}\\1 & 0 & 0 & 1 & \hdots & \begin{pmatrix}t_3\\t_m\end{pmatrix}\\\hdots &\hdots &\hdots &\hdots &\hdots &\hdots \\1 & 0 & 0 & 0 & \hdots & 1 \end{vmatrix}.$$ Proof. One can prove this formula using the analysis of the structure of the diagonals and the comparison with (\[30\]). $\blacksquare$ In the case of alternating permutations when $k=k_{n-1}$ is defined by (\[19\]), we obtain an $(m+1)\times (m+1)$ determinant representation of numbers (\[21\]). Thus, for $n=2m,$ we have an identity for the Euler numbers $E_{2m},\;m\geq 1$ (cf.[@1], Table 23.2, [@15], A000364). We drop $(-1)^m$ in order to take account of the sign of $E_{2m}$. $$\label{41} E_{2m}=\begin{vmatrix} 1& \begin{pmatrix}2m\\2m-1\end{pmatrix} & \begin{pmatrix}2m\\2m-3\end{pmatrix}& \begin{pmatrix}2m\\2m-5\end{pmatrix}& \hdots & \begin{pmatrix}2m\\1\end{pmatrix}\\ 1 & 1 & \begin{pmatrix}2m-1\\2m-3\end{pmatrix}& \begin{pmatrix}2m-1\\2m-5\end{pmatrix}& \hdots & \begin{pmatrix}2m-1\\1\end{pmatrix}\\ 1 & 0 & 1 & \begin{pmatrix}2m-3\\2m-5\end{pmatrix}& \hdots &\begin{pmatrix}2m-3\\1\end{pmatrix}\\ 1 & 0 & 0 & 1 & \hdots & \begin{pmatrix}2m-5\\1\end{pmatrix}\\ \hdots &\hdots &\hdots &\hdots &\hdots &\hdots \\ 1 & 0 & 0 & 0 & \hdots & 1 \end{vmatrix}$$ Analogously, putting $n=2m-1,\;\;m\geq 2,$ for the Bernoulli numbers $B_{2m}$ (cf.[@1], Table 23.2) we have the following determinant of an $m\times m$ matrix: $$\label{42} \frac{B_{2m}}{2m}(2^{2m}-1)2^{2m}=\begin{vmatrix} 1& \begin{pmatrix}2m-1\\2m-2\end{pmatrix} & \begin{pmatrix}2m-1\\2m-4\end{pmatrix}&\begin{pmatrix}2m-1\\2m-6\end{pmatrix} & \hdots &\begin{pmatrix}2m-1\\2\end{pmatrix}\\ 1 & 1 & \begin{pmatrix}2m-2\\2m-4\end{pmatrix}& \begin{pmatrix}2m-2\\2m-6\end{pmatrix}& \hdots & \begin{pmatrix}2m-2\\2\end{pmatrix}\\ 1 & 0 & 1 & \begin{pmatrix}2m-4\\2m-6\end{pmatrix}& \hdots &\begin{pmatrix}2m-4\\2\end{pmatrix}\\ 1&0&0&1&\hdots &\begin{pmatrix}2m-6\\2\end{pmatrix}\\ \hdots &\hdots &\hdots &\hdots &\hdots &\hdots \\ 1 & 0 & 0 & 0 & \hdots & 1 \end{vmatrix}$$ The numbers on the left hand side are the tangent numbers ([@15], A 000182). For $m=1,$ we have $$E_2=\begin{vmatrix} 1 & \begin{pmatrix}2\\1\end{pmatrix}\\ 1 & 1\end{vmatrix}=-1,$$ for $m=2,$ we have $$\frac{B_4}{4}16 \cdot 15=\begin{vmatrix}1 & \begin{pmatrix}3\\2\end{pmatrix}\\ 1 & 1\end{vmatrix}=-2,$$ which corresponds to $B_4=-\frac{1}{30}$. For $m=2,$ we have also $$E_4=\begin{vmatrix} 1 & \begin{pmatrix}4\\3\end{pmatrix}& \begin{pmatrix}4\\1\end{pmatrix}\\ 1 & 1& \begin{pmatrix}3\\1\end{pmatrix}\\ 1 & 0 & 1 \end{vmatrix}=5,$$ for $m=3,$ we have $$\frac{B_6}{6}64\cdot 63=\begin{vmatrix} 1 & \begin{pmatrix}5\\4\end{pmatrix}& \begin{pmatrix}5\\2\end{pmatrix}\\ 1 & 1& \begin{pmatrix}4\\2\end{pmatrix}\\ 1 & 0 & 1 \end{vmatrix}=16,$$ which corresponds to $B_6=\frac{1}{42}$. An identity for partial sums of the basis polynomials ===================================================== \[t11\] For $1\leq r\leq n-1,$ we have $$\label{43} \sum^{2^r-1}_{k=0}\left\{\begin{matrix} n\\ k \end{matrix}\right\}=n(n-1) \hdots (n-r+1).$$ Proof. Sum (\[43\]) enumerates the permutations with the $n-r-1$ fixed down points: $1,2,\hdots, n-r-1$. Let us form an arbitrary permutation $\pi$ of such kind. We start with position $n-r+1$. We can choose value of $\pi_{n-r-1}$ by $n$ ways, $\pi_{n-r+2}$ by $n-1$ ways, $\hdots, \pi_n$ by $n-(r-1)$ways. After that $\pi_1>\pi_2>\hdots > \pi_{n-r}$ are defined uniquely. Thus, we obtain (\[43\]).$\blacksquare$ For $r=3,$ we have (cf Appendix) $$\sum^7_{k=0}\left\{\begin{matrix} n\\k \end{matrix}\right\}=1+\left(\begin{pmatrix} n\\ 1 \end{pmatrix}-1\right)+\left(\begin{pmatrix} n\\ 2 \end{pmatrix}-1\right)+\left(\begin{pmatrix} n\\ 2 \end{pmatrix}-\begin{pmatrix} n\\ 1 \end{pmatrix}+1\right)+$$ $$+\left(\begin{pmatrix} n\\ 3 \end{pmatrix}-1\right)+\left(2\begin{pmatrix} n\\ 3 \end{pmatrix}-\begin{pmatrix} n\\ 1 \end{pmatrix}+1\right)+\left(2\begin{pmatrix} n\\ 3 \end{pmatrix}-\begin{pmatrix} n\\ 2 \end{pmatrix}+1\right)+$$ $$+\left(\begin{pmatrix} n\\3 \end{pmatrix}-\begin{pmatrix} n\\ 2\end{pmatrix}+\begin{pmatrix} n\\ 1 \end{pmatrix}-1\right)=6\begin{pmatrix} n\\3 \end{pmatrix}=n(n-1)(n-2).$$ On the positive integer zeros of the basis polynomials ====================================================== \[t12\] If $k=2^{t_1-1}+2^{t_2-1}+\hdots +2^{t_m-1}, \; t_1>t_2>\hdots >t_m\geq 1,$ then the integers $t_1, t_2, \hdots, t_m$ are roots of the basis polynomial $\left\{\begin{matrix} n\\ k\end{matrix}\right\}.$ Proof. Substituting in (\[40\]), $n=t_j,\;j=1,2,\hdots,m,$ we obtain a determinant with two the same rows. $\blacksquare$ Furthermore, the following results are obtained directly from (\[30\]): $$\left\{\begin{matrix} 1\\ k\end{matrix}\right\}=(-1)^m\left(1-\begin{pmatrix} 1\\t_m \end{pmatrix}\right)=\begin{cases} (-1)^m,\;\; if \;\; t_m\geq 2,\\0,\;\;\;\; if \;\;\;\; t_m=1\end{cases};$$ $$\left\{\begin{matrix} 2\\ k\end{matrix}\right\}=(-1)^m\left(1-\begin{pmatrix} 2\\t_{m-1} \end{pmatrix}-\begin{pmatrix} 2\\t_m \end{pmatrix}+\begin{pmatrix} 2\\t_{m-1} \end{pmatrix}\begin{pmatrix} 2\\t_m \end{pmatrix}\right)=$$ $$=\begin{cases} (-1)^m,\;\;if\;\;t_m\geq 3\\ 0,\;\; if\;\;t_m=2\\ 0,\;\; if\;\;t_m=1,\;\;t_{m-1}=2,\\ (-1)^{m-1},\;\; if \;\; t_m=1,\;\; t_{m-1}\geq 3\end{cases};$$ $$\left\{\begin{matrix} 3\\ k\end{matrix}\right\}=(-1)^m\left(1-\begin{pmatrix} 3\\t_{m-2}\end{pmatrix}-\begin{pmatrix} 3\\t_{m-1}\end{pmatrix}-\begin{pmatrix} 3\\t_m\end{pmatrix}+\begin{pmatrix} 3\\t_{m-2}\end{pmatrix}\begin{pmatrix} 3\\t_{m-1}\end{pmatrix}\right.+$$ $$\left.+\begin{pmatrix} 3\\t_{m-2}\end{pmatrix}\begin{pmatrix} 3\\t_m\end{pmatrix}+\begin{pmatrix} 3\\t_{m-1}\end{pmatrix}\begin{pmatrix} 3\\t_m\end{pmatrix}-\begin{pmatrix} 3\\t_{m-2}\end{pmatrix}\begin{pmatrix} 3\\t_{m-1}\end{pmatrix}\begin{pmatrix} 3\\t_m\end{pmatrix}\right)=$$ $$=\begin{cases} (-1)^m,\;\; if \;\; t_m\geq 4\\ 0, \;\; if \;\; t_m=3\\ 0, \;\; if \;\; t_m=2,\;\;t_{m-1}=3\\ 2(-1)^{m-1}, \;\; if \;\; t_m=2,\;\;t_{m-1}> 3\\ 0, \;\; if \;\; t_m=1,\;\;t_{m-1}=2,\;\;t_{m-2}=3\\ 4(-1)^m, \;\; if \;\; t_m=1,\;\;t_{m-1}=2,\;\ t_{m-2}> 3\\ 0, \;\; if \;\; t_m=1,\;\;t_{m-1}=3,\;\;t_{m-2}>3\\ 2(-1)^{m-1}, \;\; if \;\; t_m=1,\;\;t_{m-1}> 3\end{cases} ;$$ etc. Researching the appearance of zero values, we obtain the following result. \[t13\] Let $ a\in\mathbb{N}.$ If $$\label{44} k\equiv 2^{a-1}+j\mod{2^a},$$ where $j\in[0, 2^{a-1}), $ then $\left\{\begin{matrix} a\\ k\end{matrix}\right\}=0$. Proof. By (\[44\]), we have $$k=l 2^a + 2^{a-1} + j,\;\;l \geq 0,\;\; 0 \leq j \leq 2^{a-1}-1.$$ Therefore, if $k=2^{t_1-1}+2^{t_2-1}+\hdots+2^{t_m},$ then there exists $p\in [1,m]$ such that $a=t_p$ and the result follows from Theorem 12. $\blacksquare$ \[t14\] Note that, the conversion of Theorem $13$ implies the conversion of Theorem $12.$ This will be obtained at the end of this article (Theorem $22$). As a corollary from Theorem 13, it follows a more attractive statement. \[t14\] Let $k=2^{t_1-1}+2^{t_2-1}+\hdots+2^{t_m-1}$ and $1\leq i\leq\log_2(2k).$ If $i\neq t_p,\;\;p=1,2,\hdots,m,$ then $\left\{\begin{matrix} i\\ k-2^{i-1}\end{matrix}\right\}=0.$ Proof. Let $t_l< i < t_{l-1}.$ Then $$k-2^{i-1}=2^{t_1-1}+2^{t_2-1}+\hdots+\left(2^{t_{l-1}-1}-2^{i-1}\right) +2^{t_p-1}+\hdots+2^{t_m-1}=$$ $$=2^{t_1-1}+2^{t_2-1}+\hdots+\left(2^{t_{l-1}-2}+2^{t_{l-1}-3}+\hdots+ 2^i+2^{i-1}\right)+2^{t_l-1}+\hdots+2^{t_m-1}\equiv$$ $$\equiv 2^{i-1}+2^{t_l-1}+\hdots+2^{t_m-1}\pmod{2^i}$$ and the theorem directly follows from Theorem 13.$\blacksquare$ Another algorithm of evaluation of basis polynomials ==================================================== \[t15\] If $k=2^{t_1-1}+2^{t_2-1}+\hdots+2^{t_m-1},\; t_1> t_2>\hdots > t_m\geq 1,$ then $$\label{45} \left\{\begin{matrix} n\\ k \end{matrix}\right\}=a_1\begin{pmatrix} n\\ t_m \end{pmatrix}+a_2\begin{pmatrix} n\\ t_{m-1} \end{pmatrix}+\hdots + a_m\begin{pmatrix} n\\ t_1 \end{pmatrix}+(-1)^m,$$ where integers $a_i,\;\; i=1,2,\hdots, m$, are defined by the system of the linear equations $$\label{46} \begin{cases} a_1+(-1)^m=0\\ \begin{pmatrix} t_{m-1}\\ t_m \end{pmatrix} a_1+a_2+(-1)^m=0\\ \begin{pmatrix} t_{m-2}\\ t_m \end{pmatrix} a_1+\begin{pmatrix} t_{m-2}\\ t_{m-1} \end{pmatrix} a_2+a_3+(-1)^m=0\\ \hdots\hdots\hdots\hdots\hdots\hdots\\ \begin{pmatrix} t_1 \\ t_m \end{pmatrix}a_1+\begin{pmatrix} t_1\\ t_{m-1} \end{pmatrix}a_2+\hdots +\begin{pmatrix} t_1\\ t_2 \end{pmatrix}a_{m-1}+a_m+(-1)^m=0 \end{cases}.$$ Proof. From (\[40\]) follows a representation (\[45\]). Substituting in (\[45\]) $n=t_m,t_{m-1},\hdots,t_1$ and using Theorem 12, we obtain system (\[46\]).$\blacksquare$ Let us find $\left\{\begin{matrix} n\\26 \end{matrix}\right\}.$ We have $$26=2^{5-1}+2^{4-1}+2^{2-1}.$$ Thus, $t_1=5,\;\;t_2=4,\;\;t_3=2,\;\;m=3$. By (\[46\]) $$\begin{cases} a_1-1=0\\6a_1+a_2-1=0\\10a_1+5a_2+a_3-1=0\end{cases},$$ whence $a_1=1,\;\; a_2=-5, \;\; a_3=16.$ Consequently, by (\[45\]) $$\left\{\begin{matrix} n\\ 26 \end{matrix}\right\}=\begin{pmatrix} n\\ 2\end{pmatrix}-5\begin{pmatrix} n\\ 4\end{pmatrix}+16\begin{pmatrix} n\\ 5\end{pmatrix}-1.$$ Another recursion relation for basis polynomials ================================================ \[t16\] Let $k=2^{t_1-1}+2^{t_2-1}+\hdots+2^{t_m-1},\; t_1> t_2>\hdots > t_m\geq 1$ . Then, for $l>t_1$ we have $$\label{47} \left\{\begin{matrix} n\\k+2^{l-1} \end{matrix}\right\}=\left\{\begin{matrix} l\\k \end{matrix}\right\}\begin{pmatrix} n\\ l\end{pmatrix}-\left\{\begin{matrix} n\\k \end{matrix}\right\}.$$ Proof. By the latter theorem, $$\label{48} \left\{\begin{matrix} n\\ k+2^{l-1}\end{matrix}\right\}= b_1\begin{pmatrix} n\\ t_m\end{pmatrix}+b_2\begin{pmatrix} n\\ t_{m-1} \end{pmatrix}+\hdots+b_m\begin{pmatrix} n\\ t_1\end{pmatrix}+b_{m+1}\begin{pmatrix} n\\ l\end{pmatrix} +(-1)^{m+1},$$ where $b_i,\;\;i=1,2,\hdots, m+1$, are defined by the following system: $$\label{49} \begin{cases} b_1+(-1)^{m+1}=0\\ \begin{pmatrix}t_{m-1}\\ t_m\end{pmatrix} b_1+b_2+(-1)^{m+1}=0\\ \hdots\hdots\hdots\hdots\hdots\hdots\hdots\\ \begin{pmatrix}t_1\\ t_m\end{pmatrix}b_1+\begin{pmatrix}t_1\\ t_{m-1}\end{pmatrix}b_2+\hdots +\begin{pmatrix}t_1\\ t_2\end{pmatrix}b_{m-1} + b_m+ (-1)^{m+1}=0\\ \begin{pmatrix}l\\ t_m\end{pmatrix}b_1+\begin{pmatrix}l\\ t_{m-1}\end{pmatrix}b_2+\hdots +\begin{pmatrix}l\\ t_1\end{pmatrix}b_m + b_{m+1}+ (-1)^{m+1}=0 \end{cases}.$$ By comparison of the first $m$ equations of (\[49\]) with (\[46\]), we conclude that $$\label{50} b_i=-a_i,\;\; i=1,2,\hdots, m$$ and, by the $(m+1)-th$ equation of (\[49\]) and by (\[45\]), we find that $$\label{51} b_{m+1}=-\left\{\begin{matrix} l\\k \end{matrix}\right\}.$$ Now from (\[45\]) , (\[48\]), (\[50\]) and (\[51\]) we obtain $$\left\{\begin{matrix} n\\k \end{matrix}\right\}+\left\{\begin{matrix} n\\k+2^{l-1} \end{matrix}\right\}=\left\{\begin{matrix} l\\k \end{matrix}\right\}\begin{pmatrix} n\\l \end{pmatrix}$$ and (\[47\]) follows. $\blacksquare$ Starting with $\left\{\begin{matrix} n\\0 \end{matrix}\right\}=1$ and putting $k=0,\;\;l=1,$ we obtain $$\left\{\begin{matrix} n\\1\end{matrix}\right\}= \left\{\begin{matrix} 1\\0\end{matrix}\right\}\begin{pmatrix} n\\1 \end{pmatrix}-\left\{\begin{matrix} n\\0 \end{matrix}\right\}=\begin{pmatrix} n\\1 \end{pmatrix}-1.$$ Furthermore, we consecutively find: $$putting\;\;k=0,\;\; l=2,\;\left\{\begin{matrix} n\\2 \end{matrix}\right\}=\left\{\begin{matrix} 2\\0 \end{matrix}\right\}\begin{pmatrix} n\\2 \end{pmatrix}-\left\{\begin{matrix} n\\0 \end{matrix}\right\}=\begin{pmatrix} n\\2 \end{pmatrix}-1,$$ $$putting\;\;k=1,\;\; l=2,\;\left\{\begin{matrix} n\\3 \end{matrix}\right\}=\left\{\begin{matrix} 2\\1 \end{matrix}\right\}\begin{pmatrix} n\\2 \end{pmatrix}-\left\{\begin{matrix} n\\1 \end{matrix}\right\}=\begin{pmatrix} n\\2 \end{pmatrix}-\begin{pmatrix} n\\1 \end{pmatrix}+1,$$ $$putting\;\;k=0,\;\; l=3,\;\left\{\begin{matrix} n\\4 \end{matrix}\right\}=\left\{\begin{matrix} 3\\0 \end{matrix}\right\}\begin{pmatrix} n\\3 \end{pmatrix}-\left\{\begin{matrix} n\\0 \end{matrix}\right\}=\begin{pmatrix} n\\3 \end{pmatrix}-1,$$ $$putting\;\;k=1,\;\; l=3,\;\left\{\begin{matrix} n\\5 \end{matrix}\right\}=\left\{\begin{matrix} 3\\1 \end{matrix}\right\}\begin{pmatrix} n\\3 \end{pmatrix}-\left\{\begin{matrix} n\\1 \end{matrix}\right\}=2\begin{pmatrix} n\\3 \end{pmatrix}-\begin{pmatrix} n\\1 \end{pmatrix}+1$$ etc. Characteristic conditions for a basis polynomial ================================================ Let $P(n)$ be a polynomial. It is evident that the condition $$P(n)=C\begin{pmatrix} n\\k \end{pmatrix}$$ with a constant $C$ satisfies if and only if $P(r)=0,\;\;r=0,1,\hdots,k$, and $k=deg P(n)$. Concerning $\left\{\begin{matrix} n\\k \end{matrix}\right\}$ we have the following result. Put $$\Delta P(n)=P(n)-P(n-1)$$ and let $\Delta^r P(n)$ be the $r-th$ difference of $P(n)$. \[t17\] For a polynomial $P( n)$ there exists a nonnegative integer $k$ and a constant $C\neq 0$ such that $$\label{52} P(n)=C\left\{\begin{matrix} n\\k \end{matrix}\right\}$$ if and only if the following conditions satisfy: $$\label{53} P(0)\neq 0,\;\;(\Delta^rP(r))P(r)=0,\;\;r=1,2,\hdots,l,$$ where $l=deg P(n)$. Proof. Note that, we have $$\label{54} P(n)=P(0)+\sum^l_{r=1}\Delta^rP(r)\begin{pmatrix} n\\r \end{pmatrix}.$$ Indeed, put $$P(n)=a_0+\sum_{r=1}^la_r\begin{pmatrix} n\\r \end{pmatrix}.$$ Then we consecutively find $$a_0=P(0),$$ $$\Delta P(n)=\sum_{r=1}^la_r\begin{pmatrix} {n-1}\\{r-1} \end{pmatrix} , \enskip a_1=\Delta P(1),...,$$ i.e. $$\Delta^t P(n)=\sum_{r=1}^la_r\begin{pmatrix} {n-t}\\{r-t} \end{pmatrix} , \enskip a_t=\Delta^t P(t),\enskip t=0,...,l.$$ If all $\Delta^rP(r)=0,\;\;r=1,\hdots,l$, then we put $k=0,\;\;C=P(0)$. If $\Delta^rP(r)\neq 0,$ for $r=t_1>t_2>\hdots>t_m\geq 1,$ then by (\[54\]) $$P(n)=P(0)+\sum^m_{i=1}b_{m+1-i}\begin{pmatrix} n\\t_i \end{pmatrix},$$ where $$b_i=\Delta^{t_i}P(t_i),\;\; i=1,2,\hdots,m.$$ Putting $$\frac{b_i}{P(0)}(-1)^m=a_i,\;\;i=1,2,\hdots,m,$$ we have $$\label{55} \frac{(-1)^m}{P(0)}P(n)=(-1)^m+\sum^m_{i=1}a_{m+1-i}\begin{pmatrix} n\\t_i \end{pmatrix}$$ and, according to (\[53\]), $P(t_i)=0,\;\;i=1,2,\hdots, m$. Thus, by Theorem 15, the polynomial (\[55\]) is $\left\{\begin{matrix} n\\k \end{matrix}\right\}$ with $$k=2^{t_1-1}+2^{t_2-1}+\hdots+2^{t_m-1}.$$ The converse statement is evident as well, according to Theorem 12.$\blacksquare$ Consider $P(n)=n^3-3n^2+2n-6$. We have $$\Delta P(n)=3n^2-9n+6, \;\;\Delta^2P(n)=6n-12,\;\; \Delta^3P(n)=6,$$ and see that $$(\Delta P(1))P(1)=(\Delta^2P(2))P(2)=(\Delta^3P(3))P(3)=0.$$ By Theorem 17, we conclude that $P(n)=C\left\{\begin{matrix} n\\k \end{matrix}\right\}$. Now we easily find $C.$ Since the smallest $r$ for which $ \Delta^r P(3)\neq 0, $ then $t_1=3$ and $m=1.$ Therefore, $k=4$ and $C=-P(0)=6$. Thus $P(n)=6\left \{\begin{matrix} n\\4 \end{matrix}\right \}.$ On generating function of the basis polynomials =============================================== Note that, formula (47) gives a possibility to add to $k$ any powers of 2 more than $2^{t_1}.$ Therefore, using some iterations of (47), one can formally to get any $k_1>k.$ Thus we have a natural way to define $\left\{\begin{matrix} n\\k\end{matrix}\right\}$ for $k>2^{n-1}-1.$ Moreover, the following lemma shows that the series $\sum^\infty_{k=0}\left \{\begin{matrix} n\\k\end{matrix}\right \}x^k$ converges in interval $\|x\|<1$ for every $n.$ For a fixed $n,$ the sequence $\{\left\{\begin{matrix} n\\ k\end{matrix}\right\}\}_{k\geq0}$ is bounded. Proof. Directly we consecutively have from (\[30\]): $$\left\{\begin{matrix} 1\\ k\end{matrix}\right\}=(-1)^m\left(1-\begin{pmatrix} 1\\t_m \end{pmatrix}\right)=\begin{cases} (-1)^m,\;\; if \;\; t_m\geq 2,\\0,\;\;\;\; if \;\;\;\; t_m=1\end{cases};$$ $$\left\{\begin{matrix} 2\\ k\end{matrix}\right\}=(-1)^m\left(1-\begin{pmatrix} 2\\t_{m-1} \end{pmatrix}-\begin{pmatrix} 2\\t_m \end{pmatrix}+\begin{pmatrix} 2\\t_{m-1} \end{pmatrix}\begin{pmatrix} 2\\t_m \end{pmatrix}\right)=$$ $$=\begin{cases} (-1)^m,\;\;if\;\;t_m\geq 3\\ 0,\;\; if\;\;t_m=2\\ 0,\;\; if\;\;t_m=1,\;\;t_{m-1}=2,\\ (-1)^{m-1},\;\; if \;\; t_m=1,\;\; t_{m-1}\geq 3\end{cases};$$ $$\left\{\begin{matrix} 3\\ k\end{matrix}\right\}=(-1)^m\left(1-\begin{pmatrix} 3\\t_{m-2}\end{pmatrix}-\begin{pmatrix} 3\\t_{m-1}\end{pmatrix}-\begin{pmatrix} 3\\t_m\end{pmatrix}+\begin{pmatrix} 3\\t_{m-2}\end{pmatrix}\begin{pmatrix} 3\\t_{m-1}\end{pmatrix}\right.+$$ $$\left.+\begin{pmatrix} 3\\t_{m-2}\end{pmatrix}\begin{pmatrix} 3\\t_m\end{pmatrix}+\begin{pmatrix} 3\\t_{m-1}\end{pmatrix}\begin{pmatrix} 3\\t_m\end{pmatrix}-\begin{pmatrix} 3\\t_{m-2}\end{pmatrix}\begin{pmatrix} 3\\t_{m-1}\end{pmatrix}\begin{pmatrix} 3\\t_m\end{pmatrix}\right)=$$ $$=\begin{cases} (-1)^m,\;\; if \;\; t_m\geq 4\\ 0, \;\; if \;\; t_m=3\\ 0, \;\; if \;\; t_m=2,\;\;t_{m-1}=3\\ 2(-1)^{m-1}, \;\; if \;\; t_m=2,\;\;t_{m-1}> 3\\ 0, \;\; if \;\; t_m=1,\;\;t_{m-1}=2,\;\;t_{m-2}=3\\ 4(-1)^m, \;\; if \;\; t_m=1,\;\;t_{m-1}=2,\;\ t_{m-2}> 3\\ 0, \;\; if \;\; t_m=1,\;\;t_{m-1}=3,\;\;t_{m-2}>3\\ 2(-1)^{m-1}, \;\; if \;\; t_m=1,\;\;t_{m-1}> 3\end{cases} ;$$ etc.We see that, for every fixed $n,$ we have a finite number of distinct values of the sequence $\{\left\{\begin{matrix} n\\ k\end{matrix}\right\}\}_{k\geq0}.$ Therefore it is bounded by a constant $C(n).\blacksquare$ Denote, for any $n\in\mathbb{N},$ $$\label{56} F(n,x)=\sum^\infty_{k=0}\left \{\begin{matrix} n\\k \end{matrix}\right \}x^k,\;\; \|x\|<1.$$ Put $$\label{57} \tau(x)=\sum^\infty_{k=0}\tau_k x^k,\;\; \|x\|<1,$$ where ${\tau}$ is the Thue-Morse sequence (\[16\]). \[t18\] For every $n\in\mathbb{N}$ the quotient $\frac{F(n,x)}{\tau(x)}$ is a rational function. Proof. It follows from Theorems \[t9\] and \[t14\] that $$\label{58} \left \{\begin{matrix} n\\k \end{matrix}\right \}=\tau_k+\sum_{1\leq i\leq\log_2(2k)}\begin{pmatrix} n\\ i \end{pmatrix} \left \{\begin{matrix} i\\k-2^{i-1} \end{matrix}\right \},\;\; k\geq 1.$$ Note that, for $k=0,$ we have $$\left \{\begin{matrix} n\\0 \end{matrix}\right \}=1=\tau_0.$$ Therefore, by (\[56\])-(\[58\]), we find $$F(n,x)=\sum^\infty_{k=0}\left \{\begin{matrix} n\\k \end{matrix}\right \}x^k=\tau(x)+\sum_{k\geq 1}\sum_{1\leq i \leq\log_2(2k)} \begin{pmatrix} n\\ i \end{pmatrix}\left \{\begin{matrix} i\\k-2^{i-1} \end{matrix}\right \}x^k=$$ $$=\tau(x)+\sum^n_{i=1}\begin{pmatrix} n\\ i \end{pmatrix}\sum^\infty_{k=2^{i-1}}x^k\left \{\begin{matrix} i\\k-2^{i-1} \end{matrix}\right \}=$$ $$=\tau(x)+\sum^n_{i=1}\begin{pmatrix} n\\ i \end{pmatrix}\sum^\infty_{r=0}\left \{\begin{matrix} i\\r \end{matrix}\right \}x^{r+2^{i-1}}=$$ $$\label{59} =\tau(x)+\sum^n_{i=1}\begin{pmatrix} n\\i \end{pmatrix}x^{2^{i-1}}F(i,x),\;\;\;\|x\|< 1.$$ Thus we obtain us a recursion formula for $F(n,x):$ $$\label{60} \left(1-x^{2^{n-1}}\right) F(n,x)=\tau(x)+\sum^{n-1}_{i=1}\begin{pmatrix} n\\i \end{pmatrix}x^{2^{i-1}}F(i,x),\;\; \|x\|< 1.$$ Put $$\label{61} F(n,x)=\tau(x)\frac{P_n(x)}{(1-x)(1-x^2)\hdots (1-x^{2^{n-2}})}\;\;(\|x\|< 1).$$ Then we obtain a recursion formula for $P_n(x):$ $$P_n(x)=\frac{1}{1-x^{2^{n-1}}}\left((1-x)(1-x^2)\hdots (1-x^{2^{n-1}})+\right.$$ $$\label{62} \left.+\sum^{n-1}_{i=1}\begin{pmatrix} n\\ i \end{pmatrix}(1-x^{2^i})(1-x^{2^{i+1}})\hdots (1-x^{2^{n-1}})\;x^{2^{i-1}}P_i(x)\right).$$ Here we do not cancel $1-x^{2^{n-1}}$ in order to avoid additional conventions. In particular, from (62) we find $$P_1(x)=\frac{1}{1-x}(1-x)=1,$$ $$P_2(x)=\frac{1}{1-x^2}((1-x)(1-x^2)+2(1-x^2)x)=1+x,$$ $$P_3(x)=1+2x+2x^2+x^3,$$ $$P_4(x)=1+3x+5x^2+3x^3+3x^4+5x^5+3x^6+x^7,$$ $$P_5(x)=1+4x+9x^2+6x^3+9x^4+16x^5+11x^6+4x^7+$$ $$4x^8+11x^9+16x^{10}+9x^{11}+6x^{12}+9x^{13}+4x^{14}+x^{15},$$ $$P_6(x)=1+5x+14x^2+10x^3+19x^4+35x^5+26x^6+10x^7+14x^8+40x^9+$$ $$+61x^{10}+35x^{11}+26x^{12}+40x^{13}+19x^{14}+5x^{15}+5x^{16}+19x^{17}+40x^{18}+$$ $$+26x^{19}+35x^{20}+61x^{21}+40x^{22}+14x^{23}+10x^{24}+26x^{25}+35x^{26}+19x^{27}+$$ $$+10x^{28}+14x^{29}+5x^{30}+x^{31},$$ etc. By simple induction we see that $P_n(x)$ is a polynomial in $x$ of degree $2^{n-1}-1$. Thus, the theorem follows from (\[61\]).$\blacksquare$ But (\[61\]) gives us more in view of identity: $$(1-x)(1-x^2)\hdots (1-x^{2^{n-2}})=\sum^{2^{n-1}-1}_{k=0}\tau_k x^k= \tau(x) + o(x^{2^{n-1}-1}).$$ Therefore, from (\[61\]) it follows that $$P_n(x)=F(n,x)(1+o(x^{2^{n-1}-1})).$$ Thus polynomial $P_n(x)$ of degree $2^{n-1}-1$ is a partial sum of (\[56\]). Therefore, we obtain the following result. \[t19\] Polynomial $P_n(x)$ which is defined recursively by $(\ref{62})$ is equal to $$\label{63} P_n(x)=\sum^{2^{n-1}-1}_{k=0}\left\{\begin{matrix} n\\k \end{matrix}\right\} x^k.$$ Moreover, from (\[61\]),for every $n\in \mathbb{N},$ we have an identity $$\label{64} \frac{\sum\limits^\infty\limits_{k=0}\left\{\begin{matrix} n\\k \end{matrix}\right\}x^k}{\sum\limits^{2^{n-1}-1}\limits_{k=0}\left\{\begin{matrix} n\\k \end{matrix}\right\}x^k}=\prod\limits^\infty\limits_{i=n-1}(1-x^{2^i})\;\;(\|x\|< 1).$$ Another type of recursion for basis polynomials =============================================== For any $k\in\mathbb{N},$ let us consider the set $A_k$ of those positive integers $i\leq \log_2(2k)$ for which $\left\lfloor\frac{k}{2^i}-\frac 1 2\right\rfloor=\left\lceil \frac{k+1}{2^i}-1\right\rceil.$ The common values of these expressions denote by $\lambda(k;i):$ $$\label{65} \lambda(k;i)=\left\lfloor\frac{k}{2^i}-\frac 1 2\right\rfloor=\left\lceil \frac{k+1}{2^i}-1\right\rceil.$$ \[t20\] $$\label{66} \left\{\begin{matrix} n\\k \end{matrix}\right\}=\tau_k+\sum\limits_{i\in A_k}\begin{pmatrix}n\\ i \end{pmatrix} \left\{\begin{matrix} i\\k-2^{i-1}-\lambda(k;i)2^i \end{matrix}\right\}\tau_{\lambda(k;,i)}.$$ Proof. Taking into account (\[63\]) and comparing $Coef_{x^k},\enskip k\leq 2^{n-1}-1,$ in both sides of (\[62\]) we find $$\left\{\begin{matrix} n\\k \end{matrix}\right\}=\tau_k+\sum\limits^{n-1}\limits_{i=1}\begin{pmatrix}n\\ i \end{pmatrix} \sum_l \tau_l \left\{\begin{matrix} i\\k-2^{i-1}-2^i l \end{matrix}\right\},$$ where the summing is over those values of $l\geq 0$ for which $$2^{i-1}+2^i l\leq k,\;\;k-2^{i-1}-2^i l \leq deg P_i(x)=2^{i-1}-1.$$ Consequently, $$l\in\left[\frac{k+1}{2^i}-1,\;\;\frac{k}{2^i}-\frac 1 2 \right],$$ Nevertheless, the length of this segment equals to $\frac 1 2 -\frac{1}{2^i}$. This means that $$l=\lambda(k;i)= \left\lfloor\frac{k}{2^i}-\frac 1 2 \right\rfloor=\left\lceil\frac{k+1}{2^i}-1\right\rceil\geq 0.\blacksquare$$ Let $k=2^m$. Then $i\leq \log_2(2k)=m+1$. If $i\leq m$ then $\left\lfloor\frac{k}{2^i}-\frac 1 2\right\rfloor=2^{m-i}-1$ while $\left\lceil\frac{k+1}{2^i}-1\right\rceil=2^{m-i}$. It is left to consider the case $i=m+1$ for which $$\left\lfloor\frac{k}{2^i}-\frac 1 2 \right\rfloor=\left\lceil\frac{k+1}{2^i}-1\right\rceil=0.$$ Thus, by $(\ref{66}),$ we have $$\left\{\begin{matrix} n\\ 2^m\end{matrix}\right\}=-1+\begin{pmatrix} n\\ m+1 \end{pmatrix}\left\{\begin{matrix} m+1\\ 0\end{matrix}\right\}=\begin{pmatrix} n\\ m+1 \end{pmatrix}-1.$$ Conversion of Theorem 12 ======================== We have seen (Theorem 12) that $$\label{67} If\enskip b\equiv 2^{a-1}+c\pmod{2^a},\;\;c\in[0,2^{a-1}),\enskip then\enskip \left\{\begin{matrix} a\\ b\end{matrix}\right\}=0.$$ The conversion of Theorem 12 is based on the following result. \[t21\] Let $a\in\mathbb{N},\;\; r\geq a,\;\;c\in[0,2^{a-1})$. Then for $b=2^rl+c,$ where $l$ is odd, we have $$\label{68} \left\{\begin{matrix} a\\ b\end{matrix}\right\}=\tau_l\left\{\begin{matrix} a\\ c\end{matrix}\right\}.$$ Proof. Let $k=2^{t_1-1}+2^{t_2-1}+\hdots+2^{t_m-1},\;\;t_1>t_2>\hdots >t_m\geq 1$. Comparing Theorems 9 and 20, we conclude that $$\label{69} A_k=\{t_1,t_2,\hdots,t_m\}$$ and, for $i\in A_k,$ $$\label{70} \left\{\begin{matrix} i\\ k-2^{i-1}-\lambda(k;i)2^i\end{matrix}\right\}\tau_{\lambda(k;i)}= \left\{\begin{matrix} i\\ k-2^{i-1}\end{matrix}\right\}.$$ In particular, for $i=t_j$ we have $$\lambda(k;t_j)=\left\lfloor\frac{k}{2^{t_j}}-\frac 1 2 \right\rfloor=\left\lceil\frac{k+1}{2^{t_j}}-1\right\rceil.$$ Thus, $$\lambda(k;t_j)=\left\lfloor\frac{2^{t_1-1}+2^{t_2-1}+\hdots+2^{t_j-1}+\hdots+ 2^{t_m-1}}{2^{t_j}}-\frac 1 2\right\rfloor=$$ $$\label{71} =2^{t_1-t_j-1}+2^{t_2-t_j-1}+\hdots+2^{t_{j-1}-t_j-1}$$ and, consequently, $$\label{72} \tau_{\lambda(k;\enskip t_j)}=(-1)^{j-1}.$$ Therefore, by (\[70\]) (for $i=t_j$) and (\[71\]), we find $$\label{73} (-1)^{j-1}\left\{\begin{matrix} t_j\\ k-2^{t_1-1}-2^{t_2-1}-\hdots-2^{t_{j-1}-1}-2^{t_l-1}\end{matrix}\right\}= \left\{\begin{matrix} t_j\\ k-2^{t_j-1}\end{matrix}\right\}$$ or, taking into account that $k=2^{t_1-1}+\hdots+2^{t_m-1},$ $$(-1)^{j-1}\left\{\begin{matrix} t_j\\ 2^{t_{j+1}-1}+\hdots+2^{t_m-1}\end{matrix}\right\}=$$ $$\label{74} \left\{\begin{matrix} t_j\\ 2^{t_1-1}+\hdots+2^{t_{j-1}-1}+2^{t_{j+1}-1}+\hdots+2^{t_m-1}\end{matrix}\right\}.$$ Put here $$t_j=a,\;\;2^{t_{j+1}-1}+\hdots+2^{t_m-1}=c\in[0, 2^{a-1}),$$ $$t_{j-1}-1=r\geq a,\;\;2^{t_1-1}+\hdots+2^{t_{j-1}-1}=2^rl,$$ where $$l=2^{t_1-1-r}+2^{t_2-1-r}+...+2^{t_{j-1}-1-r}=2^{t_1-1-r}+2^{t_2-1-r}+...+1.$$ In view of $\tau_l=(-1)^{j-1},$ and $$b=2^rl+c,$$ we write (\[74\]) in the form of (\[68\]) $\blacksquare$ Since in Theorem 21 $b-c=2^rl$ then $\tau_l=\tau_{b-c}$. Therefore, Theorem 21, for $b=k,$ one can write in the following form. \[22\] $$\left\{\begin{matrix} a\\ k\end{matrix}\right\}=\tau_{k-i}\left\{\begin{matrix} a\\ i\end{matrix}\right\},\;\; k\equiv i\pmod{2^a},\;\;i=0,1,2,\hdots,2^{a-1}-1.$$ It is worth to add that by (\[15\]) $$\label{75} \left\{\begin{matrix} 0\\ k\end{matrix}\right\}=\tau_k.$$ In particular, taking into account (\[67\]) and (\[75\]), we obtain the following sequences: $$\left\{\begin{matrix} 0\\ k\end{matrix}\right\}: 1,\;-1,\;-1,\;1,\;-1,\;1,\;1,\;-1,\;\hdots$$ $$\left\{\begin{matrix} 1\\ k\end{matrix}\right\}:\mathbf 1,0,-1,0,-1,0,1,0,-1,0,1,0,1,0,-1,\hdots$$ $$\left\{\begin{matrix} 2\\ k\end{matrix}\right\}:\mathbf 1,\mathbf1,0,0,-1,-1,0,0,-1,-1,0,0,1,1,\hdots$$ $$\left\{\begin{matrix} 3\\ k\end{matrix}\right\}:\mathbf 1,\mathbf 2,\mathbf 2,\mathbf 1,0,0,0,0,-1,-2,-2,-1,0,0,0,0,\hdots$$ $$\left\{\begin{matrix} 4\\ k\end{matrix}\right\}:\mathbf {1, 3,5, 3, 3, 5, 3, 1},\underbrace{0,\hdots,0}_8,-1,-3,-5,-3,-3,\hdots$$ $$\left\{\begin{matrix} 5\\ k\end{matrix}\right\}\mathbf {1,4,9,6,9,16,11,4,4,11,16,9,6,9,4,1,}\underbrace{0,\hdots,0}_{16},-1,\hdots$$ $$\left\{\begin{matrix} 6\\ k\end{matrix}\right\}\mathbf{1,5,14,10,19,35,26,10,14,40,61,35,26,40,19,5,5,19,40,26,35,61,}$$ $$\;\;\;\;\;\mathbf{40,14,10,26,35,19,10,14,5,1,}\underbrace{0,\hdots,0}_{32},-1,-5,-14,-10,-19,-35,\hdots$$ etc.We see that, if in sequence $(\left\{\begin{matrix} a\\ k\end{matrix}\right\}) $ to ignore the signs, then it becomes to periodic sequence with period $2^a$ such that the second part of period containing terms of the form $2^{a-1}+c,\enskip c=1,...,2^{a-1},$ by (67), consists of 0’s. It is left to show that all terms $\left\{\begin{matrix} a\\ k\end{matrix}\right\},\enskip k=1,...,2^{a-1},$ of the first part of period are positive. Note that, in order to prove this, it is sufficient to find at least one permutation of elements $1,...,a$ with a given index $k\in[1,2^{a-1}].$ We can consider only positions of 0’s in $a-1$-digit binary expansion of $k.$ Start with initial permutation $\pi=(1,2,...,a).$ Suppose that the first (from the left) run of binary 0’s of $k$ has positions $i,...,i+t-1.$ Let us write in $\pi$ the elements $i,...,i+t,i+t$ in the reverse order. Then we obtain permutation $\pi_1=(1,...,i-1,i+t,i+t-1,...,i+1,i,i+t+1,...,n),$ having the first (from the right) run of descent points exactly in the required positions: $\pi_1(i)=i+t>\pi_1(i+1)=i+t-1>...>\pi_1(i+t-1)=i+1>\pi_(i+t)=i.$ Further, by the same algorithm, we obtain permutation with the first two runs of descent points corresponding to the first two runs of 0’s in the binary expansion of $k,$ etc. At the end of this process we obtain a permutation with index $k.$ So, we proved that $\left\{\begin{matrix} a\\k\end{matrix}\right\}\geq1,\enskip k=1,...,2^{a-1}.$ This gives conversion of Theorem 13 and, consequently, of Theorem 12. Thus we have the following statement. \[t22\] If $k=2^{t_1-1}+2^{t_2-1}+\hdots+2^{t_m-1},\;t_1>t_2>\hdots>t_m\geq 1$, then $t_i,\; i=1,2,\hdots, m$, are only positive integer roots of the polynomials $\left\{\begin{matrix} n\\ k\end{matrix}\right\}$. \[t2\] Let $k=2^{t_1-1}+...+2^{t_m-1},\enskip t_1>...>t_m.$ If $n\neq t_i,\enskip i=1,...,m,$ then $$rank{ \begin{pmatrix} 1& \begin{pmatrix}n\\t_1\end{pmatrix}& \begin{pmatrix}n\\t_2\end{pmatrix}& \begin{pmatrix}n\\t_3\end{pmatrix}& \hdots & \begin{pmatrix}n\\t_m\end{pmatrix}\\ 1& 1 &\begin{pmatrix}t_1\\t_2\end{pmatrix}& \begin{pmatrix}t_1\\t_3\end{pmatrix}& \hdots &\begin{pmatrix}t_1\\t_m\end{pmatrix}\\ 1& 0& 1 &\begin{pmatrix}t_2\\t_3\end{pmatrix}&\hdots &\begin{pmatrix}t_2\\t_m\end{pmatrix}\\1 & 0 & 0 & 1 & \hdots & \begin{pmatrix}t_3\\t_m\end{pmatrix}\\\hdots &\hdots &\hdots &\hdots &\hdots &\hdots \\1 & 0 & 0 & 0 & \hdots & 1 \end{pmatrix} }=m+1.$$ Prooffollows from Theorem \[t22\] and determinant representation (\[40\]) of basis polynomials. $\blacksquare$ In conclusion, note that the bold-faced numbers are coefficients of the corresponding polynomials $P_i(x),\;\;i=1,2,\hdots $ which are defined by recursion (\[62\]). If consider the concatenation sequence of the all bold-faced numbers (cf.[@15],A060351): $$\left\{\begin{matrix} 1\\ 0\end{matrix}\right\};\left\{\begin{matrix} 2\\ 0\end{matrix}\right\},\left\{\begin{matrix} 2\\ 1\end{matrix}\right\};\left\{\begin{matrix} 3\\ 0\end{matrix}\right\},\left\{\begin{matrix} 3\\ 1\end{matrix}\right\},\left\{\begin{matrix} 3\\ 2\end{matrix}\right\},\left\{\begin{matrix} 3\\ 3\end{matrix}\right\};\left\{\begin{matrix} 4\\ 0\end{matrix}\right\},\left\{\begin{matrix} 4\\ 1\end{matrix}\right\},\left\{\begin{matrix} 4\\ 2\end{matrix}\right\},$$ $$\label{76} \left\{\begin{matrix} 4\\ 3\end{matrix}\right\},\left\{\begin{matrix} 4\\ 4\end{matrix}\right\},\left\{\begin{matrix} 4\\ 5\end{matrix}\right\},\left\{\begin{matrix} 4\\ 6\end{matrix}\right\},\left\{\begin{matrix} 4\\ 7\end{matrix}\right\};\hdots$$ then it is easy to see that this sequence one can write in an explicit form: $$\label{77} \{ \left\{\begin{matrix} \lfloor\log_2{k}\rfloor+1\\ k-2^{\lfloor\log_2{k}\rfloor}\end{matrix}\right\} \}^\infty_{k=1}.$$ This sequence is closely connected with asymptotics of $\left\{\begin{matrix} n\\ k\end{matrix}\right\}$. \[t23\] $$\left\{\begin{matrix} n\\ k\end{matrix}\right\}\sim \left\{\begin{matrix} \lfloor\log_2{k}\rfloor+1\\ k-2^{\lfloor\log_2{k}\rfloor}\end{matrix}\right\}\begin{pmatrix} n\\\lfloor\log_2{k}\rfloor+1 \end{pmatrix}\sim$$ $$\label{78} \sim\frac{1}{\lfloor\log_2{k}+1\rfloor!}\{\left\{\begin{matrix} \lfloor\log_2{k}\rfloor+1\\ k-2^{\lfloor\log_2{k}\rfloor}\end{matrix}\right\}n^{\lfloor\log_2{k}\rfloor+1} \;\;(n\rightarrow\infty)$$ Proof.The theorem follows directly from (\[58\])$\blacksquare$ Thus, according to (\[78\]), the first coefficients of $\left\{\begin{matrix} n\\ k\end{matrix}\right\},\;k\geq 1$, as linear combinations of binomial coefficients form sequence (\[76\]) ((cf. Appendix) ). \[4\] Some observations on the basis polynomials have been done by the author as early as 1993 [@13] with the calculation of some first polynomials. But only in the current paper we give a foundation of a more perfect theory of these polynomials. Some open problems ================== 1\. We conjecture that all real roots of the basis polynomials are rational. 2\. We conjecture that a polynomial $\left\{\begin{matrix} n\\ k\end{matrix}\right\},\; k\geq 1$, has only real roots if and only if the number of 0’s in the binary expansion of $k$ less that $2$. In view of Theorem 12 this condition is sufficient (since there is no a place for two conjugate complex roots). Therefore, it is left to prove its necessity. We verified this conjecture up to $k=32$. We have polynomials $\left\{\begin{matrix} n\\ k\end{matrix}\right\}$ with only real roots for $$k=1,2,3,5,6,7,11,13,14,15,23,27,29,30,31,\hdots$$ (cf. sequence A 089633 [@15]). 3\. It is interesting to investigate the sequence $\{k_m\}$ for which the polynomials $\left\{\begin{matrix} n\\ k_m\end{matrix}\right\}$ have a root $n=-1$. The first values of $k_m$ are: $2,5,8,11,23,\hdots$ 4\. Let $D_n^{(a)}$ be the number of alternating permutations without fixed points (i.e.$\pi(i)\neq i,\;\;i=1,2,\hdots,n $). We conjecture that $$\lim_{n\rightarrow \infty}\frac{D_n^{(a)}}{a_n}=e^{-1},$$ where ${a_n}$ is the sequence (\[21\]). 5\. Let $S^{(a)}(n,l)$ be the number of alternating permutations having $l$ cycles (the absolute value of the “alternating” Stirling numbers of the first kind). We conjecture that for a fixed $l$ $$\lim_{n\rightarrow\infty}\frac{nS^{(a)}(n,l)}{a_n(\ln{n})^{l-1}}=\frac{1}{(l-1)!},$$ where $a_n$ is the sequence (\[21\]). The latter means that for each $l$ the events “a permutation is alternative” and “a permutation has $l$ cycles” are asymptotically independent. [18]{} 1. M.Abramowitz and I.A.Stegun (Eds.), Bernoulli and Euler Polynomials and the Euler-Maklaurin Formula in Handbook of Manhematical Functions with Formulas, Graphs, and Manhematical Tables, 9th printing.,New York: Dover,pp. 804-806, 1972. 2\. D.Andre, Sur les Permutation Alternees, J.Math. Pures Appl., 7 (1881),167-184. 3.V.Arnold, Bernoulli-Euler updown numbers associated with function singularities, their combinatorics and arithmetic,Duke Math.J. *63(2) (1990), 537-555. 4. F.C.S.Brown, T.M.A. Fink, K.Willbrand, On arithmetic and asymptotic properties of up-down numbers, Discrete Math.*307 (2007),1722-1736. 5. N.G.de Bruijn, Permutations with given ups and downs, Nieuw Arch.*3 (1970), 61-65. 6. L.Carlitz, Permutations with prescribed pattern, II, Matem. Nachr.*83 (1978), 101-126. 7. H.O.Foulkes, Enumeration of permutations with prescribed up-down and inversion sequences,Discrete Math.*15 (1976), 235-252. 8. S.Goldstein, K.A.Kelly and E.R.Speer,The fractal structure of rarefied sums of the Thue-Morse sequence, J.Number Th.*42 (1992), 1-19. 9. C.L.Mallows, L.A.Shepp, Enumerating pairs of permutations with the same up-down form, Descrete Math. *54(1985), 301-311. . M.Morse, Reccurent geodesics on a surface of negative curvature,Trans. Amer.Math.Soc.*22 (1921), 84-100. . I.Niven, A combinatorial problem of finite sequences, Nieuw Arch.Wisk *3 (1968), 116-123. . B.Shapiro, M.Shapiro, A.Vainshtein, Periodic de Bruijn triangles: exact and asymptotic results, Discrete Math. *298(1-3) (2005), 321-333. . V.S.Shevelev, A classification of permutations by its geometric structure, Deposed in VINITI, no.1457-B93 (1993), 1-20 (in Russian). . V.S.Shevelev, On an arithmetic property of permutation numbers with a given signature associated with the Morse sequence,Izv.Vyssh.Uchebn.Zaved. Sev.-Kavk.Reg.Estestv.Nauki 2 (1996), 20-24 (in Russian; MR99e: 11023).********************************** .N.J.A.Sloane, The On-Line Encyclopedia of Integer Sequences (http://www.research.att.com:/$\sim$njas /sequenses/). . R. Stanley, Alternating permutations and symmetric functions, J. Combin. Theory Series A 114 (2007), 436-460. . G.Szpiro, The number of permutations with a given signature and the expectations of their elements,Discrete Math.226 (2001), 423-430. . G.Viennot. Permutations ayant une forme donnee.Discrete Math.26 (1979), 279-284.**** Appendix. List of the first 32 basis polynomials $$\left\{\begin{array}{c} n \\ 0 \end{array}\right\}=1$$ $$\left\{\begin{array}{c} n \\ 1 \end{array}\right\}=\left(\begin{array}{c} n \\ 1 \end{array}\right)-1$$ $$\left\{\begin{array}{c} n \\ 2 \end{array}\right\}=\left(\begin{array}{c} n \\ 2 \end{array}\right)-1$$ $$\left\{\begin{array}{c} n \\ 3 \end{array}\right\}=\left(\begin{array}{c} n \\ 2 \end{array}\right)-\left(\begin{array}{c} n \\ 1 \end{array}\right)+1$$ $$\left\{\begin{array}{c} n \\ 4 \end{array}\right\}=\left(\begin{array}{c} n \\ 3 \end{array}\right)-1$$ $$\left\{\begin{array}{c} n \\ 5 \end{array}\right\}=2\left(\begin{array}{c} n \\ 3 \end{array}\right)-\left(\begin{array}{c} n \\ 1 \end{array}\right)+1$$** $$\left\{\begin{array}{c} n \\ 6 \end{array}\right\}=2\left(\begin{array}{c} n \\ 3 \end{array}\right)-\left(\begin{array}{c} n \\ 2 \end{array}\right)+1$$ $$\left\{\begin{array}{c} n \\ 7 \end{array}\right\}=\left(\begin{array}{c} n \\ 3 \end{array}\right)-\left(\begin{array}{c} n \\ 2 \end{array}\right)+\left(\begin{array}{c} n \\ 1 \end{array}\right)-1$$ $$\left\{\begin{array}{c} n \\ 8 \end{array}\right\}=\left(\begin{array}{c} n \\ 4 \end{array}\right)-1$$ $$\left\{\begin{array}{c} n \\ 9 \end{array}\right\}=3\left(\begin{array}{c} n \\ 4 \end{array}\right)-\left(\begin{array}{c} n \\ 1 \end{array}\right)+1$$ $$\left\{\begin{array}{c} n \\ 10 \end{array}\right\}=5\left(\begin{array}{c} n \\ 4 \end{array}\right)-\left(\begin{array}{c} n \\ 2 \end{array}\right)+1$$ $$\left\{\begin{array}{c} n \\ 11 \end{array}\right\}=3\left(\begin{array}{c} n \\ 4 \end{array}\right)-\left(\begin{array}{c} n \\ 2 \end{array}\right)+\left(\begin{array}{c} n \\ 1 \end{array}\right)-1$$ $$\left\{\begin{array}{c} n \\ 12 \end{array}\right\}=3\left(\begin{array}{c} n \\ 4 \end{array}\right)-\left(\begin{array}{c} n \\ 3 \end{array}\right)+1$$ $$\left\{\begin{array}{c} n \\ 13 \end{array}\right\}=5\left(\begin{array}{c} n \\ 4 \end{array}\right)-2\left(\begin{array}{c} n \\ 3 \end{array}\right)+\left(\begin{array}{c} n \\ 1 \end{array}\right)-1$$ $$\left\{\begin{array}{c} n \\ 14 \end{array}\right\}=3\left(\begin{array}{c} n \\ 4 \end{array}\right)-2\left(\begin{array}{c} n \\ 3 \end{array}\right)+\left(\begin{array}{c} n \\ 2 \end{array}\right)-1$$ $$\left\{\begin{array}{c} n \\ 15 \end{array}\right\}=\left(\begin{array}{c} n \\ 4 \end{array}\right)-\left(\begin{array}{c} n \\ 3 \end{array}\right)+\left(\begin{array}{c} n \\ 2 \end{array}\right)-\left(\begin{array}{c} n \\ 1 \end{array}\right)+1$$ $$\left\{\begin{array}{c} n \\ 16 \end{array}\right\}=\left(\begin{array}{c} n \\ 5 \end{array}\right)-1$$ $$\left\{\begin{array}{c} n \\ 17 \end{array}\right\}=4\left(\begin{array}{c} n \\ 5 \end{array}\right)-\left(\begin{array}{c} n \\ 1 \end{array}\right)+1$$ $$\left\{\begin{array}{c} n \\ 18 \end{array}\right\}=9\left(\begin{array}{c} n \\ 5 \end{array}\right)-\left(\begin{array}{c} n \\ 2 \end{array}\right)+1$$ $$\left\{\begin{array}{c} n \\ 19 \end{array}\right\}=6\left(\begin{array}{c} n \\ 5 \end{array}\right)-\left(\begin{array}{c} n \\ 2 \end{array}\right)+\left(\begin{array}{c} n \\ 1 \end{array}\right)-1$$ $$\left\{\begin{array}{c} n \\ 20 \end{array}\right\}=9\left(\begin{array}{c} n \\ 5 \end{array}\right)-\left(\begin{array}{c} n \\ 3 \end{array}\right)+1$$ $$\left\{\begin{array}{c} n \\ 21 \end{array}\right\}=16\left(\begin{array}{c} n \\ 5 \end{array}\right)-2\left(\begin{array}{c} n \\ 3 \end{array}\right)+\left(\begin{array}{c} n \\ 1 \end{array}\right)-1$$ $$\left\{\begin{array}{c} n \\ 22 \end{array}\right\}=11\left(\begin{array}{c} n \\ 5 \end{array}\right)-2\left(\begin{array}{c} n \\ 3 \end{array}\right)+\left(\begin{array}{c} n \\ 2 \end{array}\right)-1$$ $$\left\{\begin{array}{c} n \\ 23 \end{array}\right\}=4\left(\begin{array}{c} n \\ 5 \end{array}\right)-\left(\begin{array}{c} n \\ 3 \end{array}\right)+\left(\begin{array}{c} n \\ 2 \end{array}\right)-\left(\begin{array}{c} n \\ 1 \end{array}\right)+1$$ $$\left\{\begin{array}{c} n \\ 24 \end{array}\right\}=4\left(\begin{array}{c} n \\ 5 \end{array}\right)-\left(\begin{array}{c} n \\ 4\end{array}\right) +1$$ $$\left\{\begin{array}{c} n \\ 25 \end{array}\right\}=11\left(\begin{array}{c} n \\ 5 \end{array}\right)-3\left(\begin{array}{c} n \\ 4 \end{array}\right)+\left(\begin{array}{c} n \\ 1 \end{array}\right)-1$$ $$\left\{\begin{array}{c} n \\ 26 \end{array}\right\}=16\left(\begin{array}{c} n \\ 5 \end{array}\right)-5\left(\begin{array}{c} n \\ 4 \end{array}\right)+\left(\begin{array}{c} n \\ 2 \end{array}\right)-1$$ $$\left\{\begin{array}{c} n \\ 27 \end{array}\right\}=9\left(\begin{array}{c} n \\ 5 \end{array}\right)-3\left(\begin{array}{c} n \\ 4 \end{array}\right)+\left(\begin{array}{c} n \\ 2 \end{array}\right)-\left(\begin{array}{c} n \\ 1 \end{array}\right)+1$$ $$\left\{\begin{array}{c} n \\ 28 \end{array}\right\}=6\left(\begin{array}{c} n \\ 5 \end{array}\right)-3\left(\begin{array}{c} n \\ 4 \end{array}\right)+\left(\begin{array}{c} n \\ 3 \end{array}\right)-1$$ $$\left\{\begin{array}{c} n \\ 29 \end{array}\right\}=9\left(\begin{array}{c} n \\ 5 \end{array}\right)-5\left(\begin{array}{c} n \\ 4 \end{array}\right)+2\left(\begin{array}{c} n \\ 3 \end{array}\right)-\left(\begin{array}{c} n \\ 1 \end{array}\right)+1$$ $$\left\{\begin{array}{c} n \\ 30 \end{array}\right\}=4\left(\begin{array}{c} n \\ 5 \end{array}\right)-3\left(\begin{array}{c} n \\ 4 \end{array}\right)+2\left(\begin{array}{c} n \\ 3 \end{array}\right)-\left(\begin{array}{c} n \\ 2 \end{array}\right)+1$$ $$\left\{\begin{array}{c} n \\ 31 \end{array}\right\}=\left(\begin{array}{c} n \\ 5 \end{array}\right)-\left(\begin{array}{c} n \\ 4 \end{array}\right)+\left(\begin{array}{c} n \\ 3 \end{array}\right)-\left(\begin{array}{c} n \\ 2 \end{array}\right)+\left(\begin{array}{c} n \\ 1 \end{array}\right)-1$$	{ "pile_set_name": "ArXiv" }
--- author: - 'Akinori Tanaka,' - 'Hironori Mori,' - and Takeshi Morita title: 'Abelian 3d mirror symmetry on $\mathbb{RP}^2 \times \mathbb{S}^1$ with $N_f=1$' ---	{ "pile_set_name": "ArXiv" }
--- abstract: 'The duality between a $d$-dimensional conformal field theory with relevant deformation and a gravity theory on an asymptotically AdS$_{d+1}$ geometry, has become a suitable tool in the investigation of the emergence of gravity from quantum entanglement in field theory. Recently, we have tested the duality between the mass-deformed ABJM theory and asymptotically AdS$_4$ gravity theory, which is obtained from the KK reduction of the 11-dimensional supergravity on the LLM geometry. In this paper, we extend the KK reduction procedure beyond the linear order and establish non-trivial KK maps between 4-dimensional fields and 11-dimensional fluctuations. We rely on this gauge/gravity duality to calculate the entanglement entropy by using the Ryu-Takayanagi holographic formula and the path integral method developed by Faulkner. We show that the entanglement entropies obtained using these two methods agree when the asymptotically AdS$_4$ metric satisfies the linearized Einstein equation with nonvanishing energy-momentum tensor for two scalar fields. These scalar fields encode the information of the relevant deformation of the ABJM theory. This confirms that the asymptotic limit of LLM geometry is the emergent gravity of the quantum entanglement in the mass-deformed ABJM theory with a small mass parameter. We also comment on the issue of the relative entropy and the Fisher information in our setup.' --- [[[Gravity from Entanglement and RG Flow\ in a Top-down Approach]{} ]{}\ O-Kab Kwon$^{1}$, Dongmin Jang$^{1}$, Yoonbai Kim$^{1}$, D. D. Tolla$^{1,2}$\ [$^{1}$Department of Physics, BK21 Physics Research Division, Institute of Basic Science, Sungkyunkwan University, Suwon 440-746, South Korea\ $^{2}$University College,\ Sungkyunkwan University, Suwon 440-746, South Korea]{}\ [[email protected], [email protected], [email protected], [email protected]]{} ]{} Introduction ============ The holographic approach has become a very efficient technique in computing the entanglement entropy (EE) which is rather challenging in a direct path integral approach in quantum field theories (QFTs) in more than two dimensions. The holographic calculation of the EE, which was proposed by Ryu and Takayanagi (RT) [@Ryu:2006bv; @Ryu:2006ef] and its covariant generalization [@Hubeny:2007xt], drew much attention due to its elegance and implications in condensed matter physics and gravity theories. More recently, the idea of the holographic entanglement entropy (HEE) caught more attention because of its importance in unlocking some mysteries of the gauge/gravity correspondence [@Maldacena:1997re; @Gubser:1998bc] itself. The RT formula relates the EE ($S_A$) of a $(d-1)$-dimensional spatial subregion $A$ in the vacuum state of a $d$-dimensional conformal field theory (CFT), which is living on the boundary of $(d + 1)$-dimensional AdS space, to a surface with minimum area in the bulk of the AdS space with the same boundary as the subregion $A$. This creates a connection between the EE of QFT and the emergent spacetime geometry of the dual gravity theory. The topic of the emergence of gravity from quantum entanglement has shown some significant progresses in recent years [@Ryu:2006bv; @Ryu:2006ef; @Swingle:2009bg; @VanRaamsdonk:2009ar]. Much of the progress in this direction esteemed from the first law of EE ($\delta S_A=\delta\langle H_A\rangle$) [@Bhattacharya:2012mi; @Allahbakhshi:2013rda; @Blanco:2013joa; @Wong:2013gua], which equates the variation of EE due to a transition from a vacuum state of a CFT to some nearby state, and the variation of the vacuum expectation value ($vev$) of some characteristic Hamiltonian associated with the subregion $A$, which is known as the modular Hamiltonian. For a ball-shaped subregion $A$ on the boundary of an asymptotically AdS space, it was shown that the first law of EE is equivalent to a linearized Einstein equation on the AdS background [@Lashkari:2013koa; @Faulkner:2013ica; @Swingle:2014uza]. Alternative approaches and extenstions of this phenomenon were given in [@Lee:2010bg; @Mosk:2016elb; @Czech:2016tqr; @Oh:2017pkr]. See also [@Faulkner:2017tkh] for the emergence of nonlinear gravitational equations from a broad class of CFTs via the EE. In this paper, we pursue this phenomena of the emergent spacetime geometry in the context of the 3-dimensional mass-deformed Aharony-Bergman-Jefferis-Meldacena (mABJM) theory [@Hosomichi:2008jb; @Gomis:2008vc] and the dual gravity theory. The mABJM theory is obtained from the original ${\cal N}=6$ ABJM theory [@Aharony:2008ug] in 3-dimensions by a relevant deformation which preserves the full supersymmetry. The dual gravity theory is the 11-dimensional supergravity on the Lin-Lunin-Maldacena (LLM) geometry [@Lin:2004nb] with ${\mathbb Z}_k$ orbifold and SO(2,1)$\times$SO(4)/${\mathbb Z}_k\times$SO(4)/${\mathbb Z}_k$ isometry, which is asymptotically AdS$_4\times S^7/{\mathbb Z}_k$. In [@Jang:2016tbk], we have presented a compelling evidence for the gauge/gravity duality between the mABJM theory and the 11-dimensional gravity on the LLM geometry. This was achieved by applying the gauge/gravity dictionary [@Gubser:1998bc] to the $vev$s of the chiral primary operator (CPO) with conformal dimension $\Delta=1$ for all supersymmetric vacua of the mABJM theory and the 4-dimensional dual scalar modes obtained from the procedure of the Kaluza-Klein (KK) holography [@Skenderis:2006uy; @Skenderis:2006di; @Skenderis:2007yb] of the 11-dimensional supergravity. We found an exact agreement between the results from the both sides in the large $N$ limit[^1]. In order to show the exact dual relation in [@Jang:2016tbk], we mainly dealt with the matter fields, such as scalar, vector, and tensor fields, without considering the 4-dimensional metric. In this paper, however, we construct the 4-dimensional Einstein-Hilbert action with matter from the KK reduction of the 11-dimensional gravity on the LLM geometries with $\mathbb{Z}_k$ orbifold. The matter content of the 4-dimensional theory is determined by the asymptotic behaviour of the LLM solutions. At quadratic order in the mass parameter $\mu_0$ of the LLM solutions, the matter sector is composed of only one scalar field ($\Psi$) and one pseudoscalar field ($T$). These matter fields as well as the 4-dimensional graviton field $(H_{\mu\nu}$) are related to the 11-dimensional KK modes by some non-trivial field redefinitions, often called the KK maps. Solution of the linearized Einstein equation for $(H_{\mu\nu})$ with the energy-momentum tensor for the two scalar fields is asymptotically AdS$_4$ and encodes the information of all LLM geometries in the small mass limit. The presence of the two matter fields in the 4-dimensional gravity theory corresponds to the deformation of the ABJM theory by two relevant operators, which are a CPO of conformal dimension one (${\cal O}^{(1)}$) dual to $\Psi$ and a gauge invariant operator of conformal dimension two ($\tilde{\cal O}^{(2)}$) dual to $T$. In this setting, where the CFT with relevant deformation and its dual gravity theory are explicitly known, we investigate the emergent gravity phenomena away from the UV fixed point by computing the variation of EE ($\delta S_A$) of the disk-shaped region $A$ in terms of the RT formula [@Ryu:2006bv; @Ryu:2006ef] in the gravity theory and the path integral method developed in [@Faulkner:2014jva] in QFT side. The EE for a CFT with relevant deformation is calculated using the path integral method [@Rosenhaus:2014woa; @Rosenhaus:2014zza; @Faulkner:2014jva; @Faulkner:2015csl], in which the EE is represented in terms of correlation functions. See also [@Haehl:2015rza; @Speranza:2016jwt; @Beach:2016ocq; @Sarosi:2017rsq]. In particular, in [@Faulkner:2014jva] the author showed that the EE of the $d$-dimensional CFT with a relevant deformation can be regarded as a conserved charge in an emergent (auxiliary) $(d+1)$-dimensional gravity theory. The same quantity was also computed in the dual gravity theory [@Liu:2012eea; @Nozaki:2013vta; @Lin:2014hva; @Lashkari:2015hha; @Casini:2016rwj; @Taylor:2016aoi] using the RT formula. The EE from the direct path integral approach is exactly the same as the HEE in RT formula with the metric satisfying the linearized Einstein equation in the presence of a nonvanishing energy-momentum tensor. In our case, the deformations by the two relevant operators, ${\cal O}^{(1)}$ and $\tilde{\cal O}^{(2)}$, belong to the $\Delta< d/2$ and $\Delta>d/2$ cases, which require separate treatments [@Liu:2012eea; @Faulkner:2014jva; @Speranza:2016jwt]. In this setup, we calculate the variation of the EE ($\delta S_A$) of the disk region $A$ up to quadratic order in the deformation parameter in the QFT side and the gravity side separately. In the QFT side, we rely on the exact dual relation established in [@Jang:2016tbk] and use the method developed in [@Faulkner:2014jva], while in the gravity side we use the RT formula for the 4-dimensional metric in the presence of the matter fields[^2]. We show that the $\delta S_A$’s obtained from both sides are equal only when the Einstein equation in the 4-dimensional gravity theory is satisfied. It implies that the asymptotic limit of the LLM geometry is actually the emergent gravity in the small mass expansion. The remaining part of the paper is organized as follows. In section 2, we construct the 4-dimensional gravity from the KK reduction of the 11-dimensional supergravity. We also setup the non-trivial KK maps between the 11-dimensional fields and the 4-dimensional fields. In section 3, we use the RT formula to calculate the HEE from an asymptotically AdS$_4$ metric, which is obtained from the KK reduction of the LLM solution. We then establish the relation between the variation of the HEE and the $vev$ of a conformal dimension one CPO plus the $source$ of a conformal dimension two gauge invariant operator in mABJM theory. In section 4, we review the path integral methods necessary to obtain the EE in a CFT with some relevant deformation. We apply these methods to our setup and obtain the variation of the EE for mABJM theory at quadratic order in the deformation parameter. We use the results in the first law of EE at quadratic order, to show the emergence of the gravity in 4-dimensions in a top-down approach from the quantum entanglement of the 3-dimensional mABJM theory. In section 5, we draw some conclusions and discuss some future directions. Construction of 4-dimensional Gravity Theory {#4dgravity} ============================================ In this section, we construct a 4-dimensional gravity theory on AdS$_4$ background from the KK reduction of the 11-dimensional supergravity theory. In particular, we focus on a 4-dimensional gravity theory whose matter content is determined by the LLM solution with SO(2,1)$\times$SO(4)/${\mathbb Z}_k\times$SO(4)/${\mathbb Z}_k$ isometry. In [@Jang:2016tbk] we performed the KK reduction at linear order and obtained the linearized KK maps among gauge invariant 11-dimensional and 4-dimensional fields. The linearized reduction involves the truncation of the field equations at linear order in the fluctuations on the AdS$_4\times$S$^7$/${\mathbb Z}_k$ solution of the 11-dimensional supergravity. The linearized field equations are solved by the asymptotic expansion of the LLM solutions at linear order in the mass parameter $\mu_0$. However, to solve the field equations at $\mu_0^2$ order or higher, the linearized KK reduction is not sufficient. In this section, we construct the nonlinear KK mapping up to $\mu_0^2$ order by truncating the field equations at quadratic order in the fluctuations. Field equations at quadratic order ---------------------------------- The functional variation of the bosonic part of the 11-dimensional supergravity action gives the following equations for the metric and the 3-form gauge field, $$\begin{aligned} \label{gmunuEoM} &{\bf R}_{pq}-\frac12 {\bf g}_{pq} {\bf R}=\frac{1}{48} \Big(-\frac12{\bf g}_{pq}{\bf F}_{rstu} {\bf F}^{rstu}+4 {\bf F}_{pstu} {\bf F}_q^{~stu}\Big), {\nonumber}\\ &\partial_p(\sqrt{{\bf -g}}{\bf F}^{pqrs})+\frac{1}{2\cdot (4!)^2}\tilde\epsilon^{p_1\cdots p_4q_1\cdots q_4qrs}{\bf F}_{p_1\cdots p_4} {\bf F}_{q_1\cdots q_4}=0,\end{aligned}$$ where we used the index notation $(p,q,r,\cdots =0,1,\cdots,10)$ and $\tilde\epsilon^{012\cdots 10}=-1$ is the Levi-Civita symbol[^3]. The AdS$_4\times S^7$ solution of is given by[^4] $$\begin{aligned} \label{AdS4S7} &ds^2=\frac{L^2}{ 4z^2}\left(-dt^2+dw_1^2+ dw_2^2+dz^2\right)+ L^2 ds^{2}_{S^{7}},{\nonumber}\\ &F_{\mu\nu\rho\sigma}=-\frac6{L}\epsilon_{\mu\nu\rho\sigma}, ~{\rm and~it~is~zero~otherwise},\end{aligned}$$ where we split the 11-dimensional indices into the AdS$_4$ indices $(\mu,\nu,\cdots =0,\cdots,3)$ and the $S^7$ indices $(a,b,\cdots=4,\cdots,10)$. Here $\epsilon_{\mu\nu\rho\sigma} = \sqrt{\|g_{{\rm AdS_4}}\|} \,\tilde \epsilon_{\mu\nu\rho\sigma}$ is the Levi-Civita tensor for the AdS$_4$ space and $L$ is the radius of $S^7$. To obtain nonlinear field equations from , we perturb the fields around AdS$_4\times S^7$ background by expressing the 11-dimensional metric and the 4-form field strength as $$\begin{aligned} \label{fluct} {\bf g}_{pq}=g_{pq}+h_{pq}, \qquad {\bf F}_{pqrs}=F_{pqrs}+f_{pqrs},\end{aligned}$$ and keep all terms up to quadratic order in the fluctuations $h_{pq}$ and $f_{pqrs}$. Applying such perturbation to the metric equation in , we obtain $$\begin{aligned} &\nabla^r\nabla_{p}h_{qr}+\nabla^r\nabla_{q}h_{pr}-\nabla^2h_{pq}-\nabla_q\nabla_ph^r{}_{r}-Rh_{pq}-g_{pq}\left(-R^{rs}h_{rs} +\nabla^r\nabla^sh_{rs}-\nabla^2h^r{}_{r}\right) {\nonumber}\\ &+\frac{1}{48}\Big(F_{rstu}F^{rstu} h_{pq} {-4}g_{pq}h_{rs}F^r{}_{tuv}F^{stuv}\Big)+\frac{1}{24}g_{pq} f_{pqrs}F^{pqrs}-\frac12 h_{rs}F^r{}_{ptu}F_q{}^{stu}{\nonumber}\\ &-\frac1{6}\Big( f_{prst}F_q^{~rst}+F_{prst}f_q^{~rst}\Big)+Q_{pq}=0, \label{lineq2-1}\end{aligned}$$ where the indices are raised (lowered) by the AdS$_4\times S^7$ metric, and the covariant derivatives are also those of the background. Here $Q_{pq}$ denotes terms which are quadratic in the fluctuations, $$\begin{aligned} \label{lineq1-2} Q_{pq}=& -\nabla_r\Big(h^{rs}\big[\nabla_p h_{sq} +\nabla_q h_{sp}-\nabla_s h_{pq}\big]\Big)+\frac12\nabla_qh^{rs}\nabla_p h_{rs}+h^{rs}\nabla_q\nabla_p h_{rs}{\nonumber}\\ &+\frac12\nabla^rh^s{}_s\big[\nabla_p h_{rq}+\nabla_q h_{rp}-\nabla_r h_{pq}\big] +\nabla^r h^s{}_{q}\nabla_r h_{sp}-\nabla^r h^s{}_{q}\nabla_s h_{pr}-g_{pq}R_{rs}h^{rt}h^s{}_t{\nonumber}\\ &+\frac12g_{pq}\nabla_r\Big(h^{rs}\big[2\nabla^t h_{st}-\nabla_s h^t{}_t\big]\Big)-\frac34g_{pq}\nabla^th^{rs}\nabla_t h_{rs}+\frac12g_{pq}\nabla^r h^{st}\nabla_s h_{tr}-\frac12g_{pq}h^{rs}\nabla^2 h_{rs}{\nonumber}\\ &-\frac14g_{pq}\nabla^rh^s{}_s\big[2\nabla^s h_{rs}-\nabla_r h^t{}_t\big] +\frac12g_{pq}h^{rs}\Big(\nabla^t\nabla_rh_{ts}+\nabla^t\nabla_sh_{tr}-\nabla^2h_{rs}-\nabla_r\nabla_sh^t{}_t\Big){\nonumber}\\ &+h_{pq}h^{rs}R_{rs}-h_{pq}\Big(\nabla^r\nabla^sh_{rs}-\nabla^2h^r{}_r\Big)\\ +\frac1{12}&\Big(g_{pq}F_{rstu}F^{rst}~\!\!_{s'}h^{uv}h_v^{s'}+\frac32g_{pq}F_{rstu}F^{rs}~\!\!_{r's'}h^{tr'}h^{us'}-g_{pq}F_{rstu}f^{rst}~\!\!_{s'}h^{us'}{\nonumber}\\ &+\frac12h_{pq}f_{rstu}F^{rstu}-h_{pq}F_{rstu}F^{rst}~\!\!_{s'}h^{us'}+\frac14g_{pq}f_{rstu}f^{rstu}-g_{pq}f_{rstu}F^{rst}~\!\!_{s'}h^{us'}\Big){\nonumber}\\ -\frac1{2}&\Big(F_{pstu}F_q^{~st}~\!\!_{s'}h^{uv}h_v{}^{s'}+F_{pstu}F_q^{~s}~\!\!_{r's'}h^{tr'}h^{us'}-F_{pstu}f_q^{~st}~\!\!_{s'}h^{us'}-f_{pstu}F_q^{~st}~\!\!_{s'}h^{us'}+\frac13f_{pstu}f_q^{~stu}\Big).{\nonumber}\end{aligned}$$ Similarly, from the gauge field equation in we obtain $$\begin{aligned} \label{C3EoM-1} \nabla_p(h^{t}{}_{t}F^{pqrs})+2\nabla_p(4F_{s'}^{~[pqr}h^{s]s'}+f^{pqrs}) &+\frac2{\sqrt{-g}}\frac{1}{(4!)^2}\tilde\epsilon^{p_1\cdots p_4q_1\cdots q_4qrs}f_{p_1\cdots p_4}F_{q_1\cdots q_4}+P^{qrs}=0,\end{aligned}$$ where $$\begin{aligned} \label{C3EoM-2} P^{qrs}=&-\frac12\nabla_p\Big[\Big(h_{tu}h^{tu}-\frac12(h^{t}{}_{t})^2\Big) F^{pqrs}\Big]-8\nabla_p\Big[F_{s'}^{~[pqr}h^{s]t}h_t^{s'}-\frac32F^{r's'[pq}~\!\!h^r{}_{r'}h^{s]}{}_{s'}-f_{s'}^{~[pqr}h^{s]s'}\Big]{\nonumber}\\ &+\nabla_p\Big[h^{t}{}_{t}\big( 4F_{s'}^{~[pqr}h^{s]s'}+f^{pqrs}\big)\Big]+\frac1{\sqrt{-g}}\frac{1}{(4!)^2}\tilde\epsilon^{p_1\cdots p_4q_1\cdots q_4qrs}f_{p_1\cdots p_4}f_{q_1\cdots q_4}.\end{aligned}$$ For later convenience we write and by separating the AdS$_4$ and the $S^7$ indices and also inserting the AdS$_4\times S^7$ background information given in . One can write those quadratic equations in generic gauge, however, having the LLM solutions in mind, we simplify the equations by setting $h_{\mu a}$ and $f_{\mu\nu ab}$ to zero. Then we obtain $$\begin{aligned} \label{lineq2-5} &\nabla^\rho\nabla_{\mu}h_{\nu\rho}+\nabla^\rho\nabla_{\nu}h_{\mu\rho} -(\nabla^\rho\nabla_\rho+\nabla^a\nabla_a) h_{\mu\nu}-\nabla_\mu\nabla_\nu (h^\rho_{~\rho}+h^a_{~a})+\frac{24}{L^2}h_{\mu\nu}{\nonumber}\\ &-g_{\mu\nu}\left[\frac{30}{L^2}h^\rho_{~\rho}-\frac6{L^2}h^a_{~a}+\nabla^\rho\nabla^\sigma h_{\rho\sigma}+\nabla^a\nabla^b h_{ab}-(\nabla^\rho\nabla_\rho+\nabla^a\nabla_a) (h^\sigma_{~\sigma}+h^b_{~b})\right]{\nonumber}\\ &-\frac1{4L}g_{\mu\nu}f_{\rho\sigma\tau\lambda}\epsilon^{\rho\sigma\tau\lambda}+\frac1{L}\Big(f_{\mu \rho\sigma\tau}\epsilon_\nu^{~\rho\sigma\tau}+f_{\nu \rho\sigma\tau}\epsilon_\mu^{~\rho\sigma\tau}\Big)+Q_{\mu\nu}=0,\end{aligned}$$ $$\begin{aligned} \label{lineq2-6} &\nabla^\rho\nabla_{a}h_{\mu\rho} +\nabla^b\nabla_{\mu}h_{a b}-\nabla_\mu\nabla_a (h^\rho_{~\rho}+h^b_{~b})+\frac1{L}f_{a \rho\sigma\tau}\epsilon_\mu^{~\rho\sigma\tau}+Q_{\mu a}=0,\end{aligned}$$ $$\begin{aligned} \label{lineq2-7} &\nabla^c\nabla_{a}h_{b c}+\nabla^c\nabla_{b}h_{a c}-(\nabla^\rho\nabla_\rho+\nabla^c\nabla_c) h_{ab}-\nabla_a\nabla_b (h^\rho_{~\rho}+h^c_{~c})-\frac{12}{L^2}h_{ab}-\frac1{4L}g_{ab}f_{\rho\sigma\tau\lambda}\epsilon^{\rho\sigma\tau\lambda}{\nonumber}\\ &-g_{ab}\left(-\frac{6}{L^2}h^\rho_{~\rho}-\frac6{L^2}h^c_{~c}+\nabla^\rho\nabla^\sigma h_{\rho\sigma}+\nabla^c\nabla^d h_{cd}-(\nabla^\rho\nabla_\rho+\nabla^c\nabla_c) (h^\sigma_{~\sigma}+h^d_{~d})\right)+Q_{ab}=0,\end{aligned}$$ and $$\begin{aligned} &\nabla_\sigma f^{\sigma\mu\nu\rho}+\nabla_a f^{a\mu\nu\rho} -\frac3L(\nabla_\sigma h^\lambda_{~\lambda})\epsilon^{\sigma\mu\nu\rho}-\frac3L(\nabla_\sigma h^a_{~a})F^{\sigma\mu\nu\rho}-\frac{24}L\nabla_\lambda\big(h_\sigma^{[\lambda}\epsilon^{\mu\nu\rho]\sigma}\big)+P^{\mu\nu\rho}=0,{\nonumber}\\ &\nabla_\sigma f^{\sigma\mu\nu a}+P^{\mu\nu a}=0{\nonumber},\qquad \nabla_c f^{c\mu ab}+P^{\mu ab} =0,{\nonumber}\\ &\nabla_\sigma f^{\sigma abc}+\nabla_d f^{dabc} +\frac{1}{(4!)^2}\epsilon^{a_1\cdots a_4\nu_1\cdots \nu_4 abc}f_{a_1\cdots a_4}F_{\nu_1\cdots \nu_4}+P^{abc}=0.\label{C3EoM4-2}\end{aligned}$$ Expansion in spherical harmonics {#Zk} -------------------------------- The KK reduction involves the expansions of the fluctuations $h_{pq}$ and $f_{pqrs}$ in terms of the spherical harmonics on $S^7$. Here we are interested in the asymptotic limit of the LLM geometry with SO(2,1)$\times {\rm SO}(4)/{\mathbb Z}_k \times {\rm SO}(4)/{\mathbb Z}_k$ isometry. In that case, we need to consider expansion in terms of the spherical harmonics on $S^7/\mathbb{Z}_{k}$ with $ {\rm SO}(4)/{\mathbb Z}_k \times {\rm SO}(4)/{\mathbb Z}_k$ symmetry. In the presence of such symmetry, the metric on the $S^7/\mathbb{Z}_{k}$ is written as $$\begin{aligned} ds^2_{S^7/\mathbb{Z}_{k}}=d\tau^2+\frac{d\theta^2 +\sin^2\theta d\phi^2+(d\psi +\cos\theta d\phi)^2}4+\frac{d\tilde\theta^2 +\sin^2\tilde\theta d\tilde\phi^2+(d\tilde\psi +\cos\tilde\theta d\tilde\phi)^2}4,{\nonumber}\end{aligned}$$ with ranges of the angles, $ 0\le \theta,\tilde\theta \le \pi, \,\, 0\le \phi,\tilde\phi \le 2\pi,\,\, 0\le\psi,\tilde\psi \le \frac{4\pi}{k}$. The $\mathbb{Z}_k$ orbifolding acts as $\left(\psi, \, \tilde\psi\right) \to \left( \psi + \frac{4\pi}{k},\, \tilde\psi + \frac{4\pi}{k}\right)$[@Auzzi:2009es; @Cheon:2011gv]. The spherical harmonics with the $ {\rm SO}(4)/{\mathbb Z}_k \times {\rm SO}(4)/{\mathbb Z}_k$ symmetry are dependent only on the $\tau$ coordinate, and they are the same with and without the orbifolding. This implies that expansions of the fluctuations $h_{pq}$ and $f_{pqrs}$ in terms of these spherical harmonics are the same with and without the orbifolding. In [@Jang:2016tbk], we have written a complete form of the expansions in terms of the spherical harmonics on $S^7$. For clarity of presentation, let us recall the expansion, $$\begin{aligned} \label{hpqexp} & h_{\mu\nu}(x,y) = h^{I_1}_{\mu\nu}(x)Y^{I_1}(y), {\nonumber}\\ &h_{\mu a}(x,y) = v_\mu^{I_7}(x)Y_a^{I_7}(y) +s^{I_1}_\mu(x)\nabla_{a}Y^{I_1}(y), {\nonumber}\\ & h_{(ab)}(x,y) = t^{I_{27}}(x)Y_{(ab)}^{I_{27}}(y) +v^{I_7}(x)\nabla_{(a}Y_{b)}^{I_7}(y) +s^{I_1}(x)\nabla_{(a}\nabla_{b)}Y^{I_1}(y), {\nonumber}\\ & h^a_{~a}(x,y) = \phi^{I_1}(x)Y^{I_1}(y), {\nonumber}\\ & f_{\mu\nu\rho\sigma}(x,y) = 4\nabla_{[\mu}s_{\nu\rho\sigma]}^{I_1}(x)Y^{I_1}(y), {\nonumber}\\ & f_{\mu\nu\rho a}(x,y) = 3\nabla_{[\mu}v_{\nu\rho]}^{I_7}(x)Y_a^{I_7}(y) -s^{I_1}_{\mu\nu\rho}(x)\nabla_{a}Y^{I_1}(y), {\nonumber}\\ & f_{\mu\nu ab}(x,y) = 2\nabla_{[\mu}t_{\nu]}^{I_{21}}(x)Y_{[ab]}^{I_{21}}(y) +2v_{\mu\nu}^{I_7}(x)\nabla_{[a}Y_{b]}^{I_7}(y) ,{\nonumber}\\ & f_{\mu abc}(x,y) = \nabla_\mu t^{I_{35}}(x)Y_{[abc]}^{I_{35}}(y) -3 t_\mu^{I_{21}}(x)\nabla_{[a}Y_{bc]}^{I_{21}}(y) ,{\nonumber}\\ & f_{abcd}(x,y) = 4 t^{I_{35}}(x)\nabla_{[a}Y_{bcd]}^{I_{35}}(y),\end{aligned}$$ where $x$ denotes the AdS${}_4$ coordinates and $y$ denotes the $S^7$ coordinates. For the details about the spherical harmonics on $S^7$, see [@Jang:2016tbk]. The parenthesis of two indices $(ab)$ means symmetrized traceless combination, while the square bracket $[ab\cdots]$ denotes complete antisymmetrization of indices. Eventually, we will identify the fluctuations $h_{pq}$ and $f_{pqrs}$ with the deviations of the LLM geometries from the AdS$_4\times S^7$ background. In the gauge choice of the LLM solutions, $h_{\mu a}$ and $f_{\mu\nu ab}$ are zero and as a result most of the KK towers in are absent. Therefore, we use a truncated expansion, $$\begin{aligned} \label{metric-exp1} &h_{\mu\nu}(x,y)=h^{I_1}_{\mu\nu}(x)Y^{I_1}(y),\quad h^\rho{}_{\rho}(x,y)=h^{I_1}(x)Y^{I_1}(y),{\nonumber}\\ &h^a_{~a}(x,y)=\phi^{I_1}(x)Y^{I_1}(y),\quad h_{(ab)}=s^{I_1}(x)\nabla_{(a}\nabla_{b)}Y^{I_1}(y),\end{aligned}$$ and $$\begin{aligned} \label{F4-exp1} &f_{\mu\nu\rho\sigma}(x,y)=4\nabla_{[\mu} s_{\nu\rho\sigma]}^{I_1}(x)Y^{I_1}(y),\quad f_{\mu\nu\rho a}(x,y)=- s^{I_1}_{\mu\nu\rho}(x)\nabla_{a}Y^{I_1}(y),{\nonumber}\\ &f_{\mu abc}(x,y)=\nabla_\mu t^{I_{35}}(x)Y_{abc}^{I_{35}}(y) ,\quad f_{abcd}(x,y)=4 t^{I_{35}}(x)\nabla_{[a}Y_{bcd]}^{I_{35}}(y).\end{aligned}$$ This truncation is not dictated by some symmetry, which would have been a requirement in order to have a consistent truncation if one follows the line of thought of ref. [@Cvetic:2000dm]. However, the equations obtained from the truncated expansion are consistent at quadratic order in the fluctuations. The reason is that those equations are solved order by order in the mass parameter ($\mu_0$) of the LLM solutions, and the modes which are omitted from the expansion are all vanishing, at least up to $\mu_0^2$-order [@Jang:2016tbk]. Plugging and into and then projecting on the scalar harmonics $Y^{I_1}$, we obtain $$\begin{aligned} \label{lineqmn} &-\left(\square+\Lambda^{I_1}-\frac{24}{L^2}\right) h^{I_1}_{\mu\nu}+\nabla^\rho\nabla_{\mu}h^{I_1}_{\nu\rho}+\nabla^\rho\nabla_{\nu}h^{I_1}_{\mu\rho} -\nabla_\mu\nabla_\nu (h^{I_1}+\phi^{I_1}){\nonumber}\\ &+g_{\mu\nu}\left(\square+\Lambda^{I_1}-\frac{30}{L^2}\right)h^{I_1} +g_{\mu\nu}\left(\square+\frac67\Lambda^{I_1}+\frac6{L^2}\right)\phi^{I_1}-g_{\mu\nu}\left(\frac67\Lambda^{I_1}+\frac6{L^2}\right)\Lambda^{I_1}s^{I_1}{\nonumber}\\ &-g_{\mu\nu}\nabla^\rho\nabla^\sigma h^{I_1}_{\rho\sigma}+\frac{1}{L}g_{\mu\nu}\nabla^{\rho} t^{I_1}_\rho+Q^{I_1}_{\mu\nu}=0,\end{aligned}$$ where $\square\equiv\nabla_\mu\nabla^\mu$, $Q^{I_1}_{\mu\nu}=\frac1{\omega_7}\int_{S^7}Q_{\mu\nu}Y^{I_1}$ with the unit volume of the $S^7$, $\omega_7$, and we have set $s^{I_1}_{\mu\nu\rho}=\frac1{3!}\epsilon_{\mu\nu\rho}{}^{\sigma}t^{I_{1}}_\sigma$, and $\Lambda^{I_{1}}=-\frac{I_{1}(I_{1}+6)}{L^{2}}$ is the eigenvalue corresponding to the scalar harmonics $Y^{I_1}$. Taking the trace of the above equation, we obtain $$\begin{aligned} \label{h-eqn} &\Big(2\square+3\Lambda^{I_1}-\frac{96}{L^2}\Big)h^{I_1}-2\nabla^\mu\nabla^\nu h^{I_1}_{\mu\nu}+3\Big(\square+\frac87\Lambda^{I_1}+\frac8{L^2}\Big)\phi^{I_1}+\frac4{L}\nabla^{\rho} t^{I_1}_\rho{\nonumber}\\ &-24\Lambda^{I_1}\Big(\frac17\Lambda^{I_1} +\frac1{L^2}\Big)s^{I_1}+Q^{I_1}=0\end{aligned}$$ with $Q^{I_1}=g^{\mu\nu}Q_{\mu\nu}^{I_1}$. From , we obtain the following equation by projecting on $\nabla^a Y^{I_1}$ with $I_1\ne0$, $$\begin{aligned} \label{leq2-2-2} &-\Big(\frac67 \Lambda^{I_1}+ \frac6{L^2}\Big)\nabla_\mu s^{I_1}+\frac67\nabla_\mu\phi^{I_1} -\nabla^\nu h_{\mu\nu}^{I_1}+\nabla_\mu h^{I_1} -\frac1{L} t^{I_1}_\mu+Q_{\mu}^{I_1} =0,\end{aligned}$$ where $Q_{\mu}^{I_1}=\frac1{\omega_7}\int_{S^7}Q_{\mu a}\nabla^aY^{I_1}$. From , we obtain two scalar equations by projecting on $ g^{ab} Y^{I_1}$ and $\nabla^{(a}\nabla^{b)} Y^{I_1}$, in the latter case $I_1\ne 0$, $$\begin{aligned} & \frac67\Big(\square+\frac{5}7\Lambda^{I_1}+\frac{5}{L^2}\Big)\phi^{I_1} +\Big(\square+\frac67\Lambda^{I_1}+\frac{6}{L^2}\Big)h^{I_1} -\nabla^\mu\nabla^\nu h_{\mu\nu}^{I_1} -\frac1{L}\nabla^{\rho} t^{I_1}_\rho{\nonumber}\\ & -\frac{30}7\Lambda^{I_1}\Big(\frac{\Lambda^{I_1}}7+\frac1{L^2} \Big)s^{I_1}+\tilde Q^{I_1}=0,\label{leq3-4-2}\\ &6\Lambda^{I_1}\Big(\frac{\Lambda^{I_1}}7+\frac1{L^2} \Big)\Big[\Big(\square-\frac57\Lambda^{I_1}\Big)s^{I_1}+h^{I_1}+\frac57\phi^{I_1}\Big]-\hat Q^{I_1}=0 ,\label{leq3-3-2}\end{aligned}$$ where $\tilde Q^{I_1}=\frac1{\omega_7}\int_{S^7}Q_{ab}g^{ab}Y^{I_1}$ and $\hat Q^{I_1}=\frac1{\omega_7}\int_{S^7}Q_{ab}\nabla^{(a}\nabla^{b)} Y^{I_1}$. Similarly, inserting and into , and projecting on the appropriate spherical harmonic elements, we obtain the following set of equations $$\begin{aligned} &4\nabla^{\sigma}\nabla_{[\sigma} s^{I_1}_{\mu\nu\rho]}+\Lambda^{I_1}s^{I_1}_{\mu\nu\rho}-\frac3{L}\epsilon_{\sigma\mu\nu\rho}\nabla^\sigma\big( h^{I_1}+\phi^{I_1}\big)-\frac{24}{L}\nabla^\sigma h^{I_1}_{\lambda[\sigma}\epsilon_{\mu\nu\rho]}~\!\!^{\lambda}+P_{\mu\nu\rho}^{I_1}=0,\label{tSmnr1}\\ &\Lambda^{I_1}\nabla^{\rho} s^{I_1}_{\rho\mu\nu}+P_{\mu\nu}^{I_1}=0,\quad(I_1\ne0),\label{tSmn1} \\ &\left(\square-\frac{12}{L^2}+\Lambda^{I_{35}}\pm\frac{6(I_{35}+3)}{L^2}\right)t_{\pm}^{I_{35}}+P^{I_{35}}=0, \label{tT35}\end{aligned}$$ where $\Lambda^{I_{35}}=-\frac{I_{35}(I_{35}+6)-3}{L^{2}}$ and $$\begin{aligned} &P_{\mu\nu\rho}^{I_1}=\frac1{\omega_7}\int_{S^7}P_{\mu\nu\rho}Y^{I_1},\qquad P_{\mu\nu}^{I_1}=\frac1{\omega_7}\int_{S^7}P_{\mu\nu a}\nabla^aY^{I_1},\qquad P^{I_{35}}=\frac1{\omega_7}\int_{S^7}P^{abc}Y_{abc}^{I_{35}}.\end{aligned}$$ Here we have used the relation, $\epsilon_{abc}~\!\!^{a_1a_2a_3a_4}\nabla_{a_1}Y^{I_{35}}_{a_2a_3a_4} =\pm3!\frac{(I_{35}+3)}{L}Y^{I_{35}}_{abc}$, to obtain the two equations in . For later convenience let us again set $s^{I_1}_{\mu\nu\rho}=\frac1{3!}\epsilon_{\mu\nu\rho}^{~~~~\!\!\lambda}~\!t^{I_1}_\lambda$ and then multiply by $\epsilon_{\mu'}^{\!~~\mu\nu\rho}\nabla_{\nu'}$. Thus we obtain $$\begin{aligned} \label{C3EoM-4b} &\frac{18}L\nabla_{\mu}\nabla_{\nu}(-h^{I_1}+\phi^{I_1})+\Lambda^{I_1}\nabla_\nu t^{I_1}_\mu+\nabla_{\mu}\nabla_{\nu}\nabla^\rho t^{I_1}_\rho+\tilde P^{I_1}_{\mu\nu}=0,\end{aligned}$$ where $\tilde P^{I_1}_{\mu\nu}=\epsilon_{\mu}^{\!~~\rho\sigma\lambda}\nabla_{\nu}P^{I_1}_{\rho\sigma\lambda}$. The trace of the above equation gives $$\begin{aligned} \label{C3EoM-4c} &\frac{18}L\square(-h^{I_1}+\phi^{I_1})+(\square+\Lambda^{I_1})\nabla^\rho t^{I_1}_\rho+\tilde P^{I_1}=0,\qquad \tilde P^{I_1}=g^{\mu\nu}\tilde P^{I_1}_{\mu\nu}.\end{aligned}$$ The 4-dimensional graviton equation at $\mu_0^2$ order ------------------------------------------------------ From the quadratic equations in the previous subsection, one can obtain the quadratic order equations of motions for various 4-dimensional gauge invariant KK modes. In [@Jang:2016tbk] we have obtained the complete 4-dimensional KK spectrum, which in general is composed of three towers of scalar modes, two towers of pseudoscalar modes, two towers of vector modes, one tower of pseudovector mode, and one tower of spin-two mode. We have also found the linear order equations for these modes in generic gauge. In this paper, we have set $h_{\mu a}$ and $f_{\mu\nu ab}$ to zero and as a result some of the KK towers disappear. Furthermore, to construct the equation of motion for 4-dimensional graviton, we focus on the KK zero modes. To that end, we start from the zero modes of the equations , , , , and . Let us rearrange those equations by introducing $u^{I_1}\equiv L\nabla^\rho t^{I_1}_\rho$ and $\hat\psi^{I_1}\equiv 18 h^{I_1}-u^{I_1}$, as follows $$\begin{aligned} \label{lineqmn1} & \square h^0_{\mu\nu}=\frac{24}{L^2}h^0_{\mu\nu}+\nabla^\rho\nabla_{\mu}h^0_{\nu\rho}+\nabla^\rho\nabla_{\nu}h^0_{\mu\rho} -\nabla_\mu\nabla_\nu (h^0+\phi^0)-g_{\mu\nu}\frac{4}{3L^2}\hat\psi^0{\nonumber}\\ & \qquad\quad+Q^{0}_{\mu\nu}-\frac19 g_{\mu\nu}(Q^0+7\tilde Q^0),\\ & \square h^0=-\frac{78}{L^2}h^0+\frac6{L^2}\phi^0+\frac5{L^2}u^0+\nabla^\rho\nabla^\sigma h^0_{\rho\sigma}+\frac13(2Q^{0}-7\tilde Q^0),\label{leq3-4-2a}\\ & \square\phi^{0}=\frac{14}{3L^2}\hat\psi^0-\frac{12}{L^2}\phi^0-\frac79(Q^{0}-2\tilde Q^0),\label{h-eqn1}\\ & \square\hat\psi^0=\frac{84}{L^2}\hat\psi^0-\frac{216}{L^2}\phi^0-14(Q^{0}-2\tilde Q^0)+L\tilde P^{0},\label{C3EoM-4d}\\ & -\nabla_{\mu}\nabla_{\nu}\hat\psi^0+18\nabla_{\mu}\nabla_{\nu}\phi^0+L\tilde P^{0}_{\mu\nu}=0.\label{C3EoM-4e}\end{aligned}$$ We notice that, the linear part of is not the 4-dimension linearized Einstein equation, which is given by $$\begin{aligned} \left(L_{E}+\frac{12}{L^{2}}\right)h^{0}_{\mu\nu}=0\end{aligned}$$ with $L_{E}h^{0}_{\mu\nu}=\frac{1}{2}\left(-\square h^0_{\mu\nu}+\nabla^\rho\nabla_{\mu}h^0_{\nu\rho}+\nabla^\rho\nabla_{\nu}h^0_{\mu\rho}-\nabla_\mu\nabla_\nu h^0\right)$, where $L_{E}$ is the Einstein operator. Therefore, $h^0_{\mu\nu}$ is not the correct 4-dimensional graviton field. Neglecting the quadratic terms in equations -, the combination which satisfies the 4-dimensional linearized Einstein equation is $$\begin{aligned} \hat h^0_{\mu\nu}\equiv h^0_{\mu\nu}-\frac14g_{\mu\nu}\phi^0+\frac1{24}g_{\mu\nu}\hat\psi^0. \end{aligned}$$ However, when we take into account the quadratic terms in -, $\hat h^0_{\mu\nu}$ still does not represent the correct 4-dimensional graviton field. In order to obtain the 4-dimensional graviton field, we need non-trivial field redefinitions to absorb the quadratic terms in the above equations of motion. Such field redefinition will be presented in the next subsection. Now we combine equations - to obtain the following quadratic equation for $\hat h^0_{\mu\nu}$: $$\begin{aligned} \label{lineq1-3-03} \square \hat h^0_{\mu\nu}=&\nabla^\rho\nabla_\mu h^0_{\rho\nu}+\nabla^\rho\nabla_\nu h^0_{\rho\mu}-\nabla_\mu\nabla_\nu h^0+\frac12\nabla_\mu\nabla_\nu\phi^0-\frac1{12}\nabla_{\mu}\nabla_{\nu}\hat\psi^0+\frac{24}{L^2}\hat h^0_{\mu\nu}{\nonumber}\\ &+Q^0_{\mu\nu}-\frac{L}{12}\tilde P^0_{\mu\nu}-\frac12g_{\mu\nu}\left(Q^0-\frac L{12}\tilde P^0\right).\end{aligned}$$ Using the Einstein operator, we write as $$\begin{aligned} \label{lineq1-3-03a} \left(L_E+\frac{12}{L^2}\right)\hat h^0_{\mu\nu}+\frac12Q^0_{\mu\nu}-\frac{L}{24}\tilde P^0_{\mu\nu}-\frac14g_{\mu\nu}\left(Q^0-\frac L{12}\tilde P^0\right)=0.\end{aligned}$$ The explicit forms of the quadratic terms ($Q^0_{\mu\nu}, \tilde P^0_{\mu\nu}$ etc.) are too long to display here, however, we would like to note that they all contain the terms which are quadratic in the fields $h^{I_1}_{\mu\nu}, h^{I_1},\phi^{I_1}, u^{I_1}, s^{I_1},t_\mu^{I_1}, t^{I_{35}}$ and their derivatives. In [@Jang:2016tbk] we have obtained the asymptotic expansion of the values of these fields in the LLM solutions. We have shown that, except for the modes with ${I_1}=2$ and $I_{35}=1$, the asymptotic expansions of the other modes, including the zero modes, are nonlinear in the expansion parameter $\mu_0$. In particular, the leading terms of all the zero modes that appear in the above equations are quadratic in $\mu_0$. Therefore, in order to solve the equations of motion of those zero modes at $\mu_0^2$ order, the quadratic terms in the above equations are built only by the modes with ${I_1}=2$ and $I_{35}=1$. Having said that, we can simplify by using the linearized equations of motion for the ${I_1}=2$ and $I_{35}=1$ modes, which can be read from the list of equations in the previous subsection. $$\begin{aligned} &\square h^2_{\mu\nu}=\frac{40}{L^2}h^2_{\mu\nu}+\nabla^\rho\nabla_\mu h^2_{\rho\nu}+\nabla^\rho\nabla_\nu h^2_{\rho\mu}-\nabla_\mu\nabla_\nu h^2-\nabla_\mu\nabla_\nu \phi^2-\frac4{3L^2}g_{\mu\nu}\hat\psi^2,{\nonumber}\\ &\square\phi^2=\frac{100}{L^2}h^2+\frac{108}{7L^2}\phi^2+\frac{1728}{7L^4}s^2-\frac{14}{3L^2}u^2,\quad \square \hat\psi^2=\frac{100}{L^2}\hat\psi^2+\frac{1944}{7L^2}\phi^2+\frac{31104}{7L^4}s^2,{\nonumber}\\ &\square s^2=-\frac{80}{7L^2}s^2-\frac57\phi^2-h^2,\quad t^2_{\mu}=\frac{54}{7L}\nabla_\mu s^2+\frac{6L}7\nabla_\mu\phi^2-L\nabla^\nu h^2_{\mu\nu}+L\nabla_\mu h^2,{\nonumber}\\ &\square t^1_{+}=-\frac8{L^2} t^1_{+}.\end{aligned}$$ Here we picked only $t^1_+$ from the $t^{I_{35}=1}_\pm$ pair because the leading term in the asymptotic expansion of $t^1_-$ is cubic in $\mu_0$. Then we obtain $$\begin{aligned} \label{lineq1-3-06} &\Big(L_E+\frac{12}{L^2} \Big)\hat h_{\mu\nu}^0+\frac1{40}\Big\{{\frac3{8}}\big(\nabla_{\mu}\nabla_{\nu}h^2h^2-\nabla_{\mu}\nabla_{\nu}h^2\phi^2-h^2\nabla_{\mu}\nabla_{\nu}\phi^2+\frac{19}{21}\nabla_{\mu}\nabla_{\nu}\phi^2\phi^2\big){\nonumber}\\ &+\frac1{48}\Big(-\nabla_{\mu}\nabla_{\nu}h^2u^2-h^2\nabla_{\mu}\nabla_{\nu}u^2{\nonumber}-\nabla_{\mu}h^2\nabla_{\nu}u^2-\nabla_{\nu}h^2\nabla_{\mu}u^2+\nabla_{\mu}\nabla_{\nu}\phi^2u^2{\nonumber}\\ &+\phi^2\nabla_{\mu}\nabla_{\nu}u^2+\nabla_{\mu}\phi^2\nabla_{\nu}u^2+\nabla_{\nu}\phi^2\nabla_{\mu}u^2\Big)+\frac{11}{72}\nabla_\mu h^2\nabla_\nu h^2+\frac{55}{392}\nabla_\mu \phi^2\nabla_\nu \phi^2{\nonumber}\\ &-\frac{3672}{49L^4}\nabla_\mu s^2\nabla_\nu s^2-\frac {95}{168}\big(\nabla_\mu h^2\nabla_\nu\phi^2+\nabla_\nu h^2\nabla_\mu\phi^2\big)-\frac {12}{7L^2}\big(\nabla_\mu h^2\nabla_\nu s^2+\nabla_\nu h^2\nabla_\mu s^2\big){\nonumber}\\ &-\frac {72}{49L^2}\big(\nabla_\mu \phi^2\nabla_\nu s^2+\nabla_\nu \phi^2\nabla_\mu s^2\big)-\frac29\nabla^\rho h^2_{\mu\rho}\nabla^\sigma h^2_{\sigma\nu}+\frac29\big(\nabla_\mu h^2\nabla^\rho h^2_{\rho\nu}+\nabla_\nu h^2\nabla^\rho h^2_{\mu\rho}\big){\nonumber}\\ &+\frac4{21}\big(\nabla_\mu \phi^2\nabla^\rho h^2_{\rho\nu}+\nabla_\nu \phi^2\nabla^\rho h^2_{\mu\rho}\big)+\frac{12}{7L^2}\big(\nabla_\mu s^2\nabla^\rho h^2_{\rho\nu}+\nabla_\nu s^2\nabla^\rho h^2_{\mu\rho}\big)+\nabla_\mu h^{2\rho\sigma}\nabla_\nu h^2_{\rho\sigma}{\nonumber}\\ &-\frac{1}2\nabla_\rho\big( h^{2\rho\sigma}\nabla_\mu h^2_{\sigma\nu} + h^{2\rho\sigma}\nabla_\nu h^2_{\sigma\mu}\big)+\frac12\nabla_\rho\big( h^{2\rho\sigma}\nabla_\sigma h^2_{\mu\nu}\big)-\frac12\nabla^\rho \nabla^\sigma h^2_{\rho\sigma} h^2_{\mu\nu}-\frac{216}{7L^4}s^2\nabla_\mu\nabla_\nu s^2{\nonumber}\\ &+\frac54 h^{2\rho\sigma}\nabla_\nu \nabla_\mu h^2_{\rho\sigma}+\frac14\nabla^\rho(h^2+\phi^2)\big(\nabla_\mu h^2_{\rho\nu}+\nabla_\nu h^2_{\rho\mu}\big)-\frac14\nabla^\rho(h^2+\phi^2)\nabla_\rho h^2_{\mu\nu}+\frac{23}{L^2}h^2h^2_{\mu\nu}{\nonumber}\\ &-\frac12\nabla^\sigma h^2_{\rho\nu}\nabla^\rho h^2_{\sigma\mu}+\frac12\nabla^\sigma h^{2\rho}_{\nu}\nabla_\sigma h^2_{\mu\rho}+\frac{8}{L^2} h^{2\rho}_{\nu}h^2_{\mu\rho}+\frac{17}{2L^2}g_{\mu\nu}h^{2\rho\sigma}h^2_{\rho\sigma}+\frac12g_{\mu\nu}\nabla_\rho\big( h^{2\rho\sigma}\nabla_\sigma \phi^2\big){\nonumber}\\ &+{\frac3{16}}g_{\mu\nu}\nabla^\rho h^2\nabla_\rho h^2{-\frac12}g_{\mu\nu}\nabla^\rho h^2\nabla_\rho \phi^2+{\frac{5}{112}}g_{\mu\nu}\nabla^\rho \phi^2\nabla_\rho \phi^2-\frac{201}{98L^2}g_{\mu\nu}\phi^2\phi^2-\frac{40}{7L^2}g_{\mu\nu}\phi^2h^2{\nonumber}\\ &-\frac{33}{L^2}g_{\mu\nu}h^2h^2-\frac{1}{7L^2}h^2_{\mu\nu}\phi^2+\frac{432}{7L^4}h^2_{\mu\nu} s^2+\frac12h^2_{\mu\nu}\square h^2+\frac34g_{\mu\nu}h^{2\rho\sigma}\nabla^\tau\nabla_\rho h^2_{\tau\sigma}+\frac{26352}{49L^6}g_{\mu\nu}s^2 s^2{\nonumber}\\ &-\frac{108}{49L^4}g_{\mu\nu}\phi^2s^2-\frac{11}{6L^2}h^2_{\mu\nu}u^2-{\frac{1}{18L^2}}g_{\mu\nu}u^2u^2+{\frac{25}{8L^2}g_{\mu\nu}h^2u^2}-\frac1{2}g_{\mu\nu}\Big(\frac34h^{2\rho\sigma}\nabla_\rho \nabla_\sigma h^2{\nonumber}\\ &-\frac12\nabla^\rho \nabla^\sigma h^2_{\rho\sigma} h^2+\frac54 h^{2\rho\sigma} \nabla_\rho\nabla_\sigma \phi^2-\frac34 \nabla^\lambda h^{2\rho\sigma}\nabla_\lambda h^2_{\rho\sigma}+\frac{216}{7L^4}\nabla^\rho s^2\nabla_\rho s^2{+\frac12}h^2\square h^2{\nonumber}\\ &+\frac12\nabla^\rho\phi^2\nabla^\sigma h^2_{\rho\sigma}-{\frac{46}{56L^2}}\phi^2u^2+\frac{216}{7L^4}h^2s^2-\frac{36}{7L^4}u^2s^2\Big)-\frac1{48}g_{\mu\nu}\Big(\frac1{2}\square h^2u^2+\nabla^\lambda h^2\nabla_\lambda u^2{\nonumber}\\ &-\nabla^\lambda \phi^2\nabla_\lambda u^2 \Big)\Big\}+\frac{1}{48}\big(\nabla_{\mu}\nabla_{\nu}t_+^1t_+^1+\frac12\nabla_{\mu} t_+^1\nabla_{\nu} t_+^1\big)+\frac{1}{96}g_{\mu\nu}\big(\nabla_\rho t_+^1\nabla^\rho t_+^1-\frac{16}{L^2}t_+^1t_+^1\big)=0.\end{aligned}$$ In our case, the fluctuation modes $(h^{I_{1}}_{\mu\nu}$, $\phi^{I_{1}}$, etc) represent the deviations of the LLM geometry from the AdS$_{4}\times S^7$ space. In that case, we have shown that the symmetrized-traceless transverse KK graviton mode $\check h^2_{(\mu\nu)}$, which is given by [@Jang:2016tbk] $$\begin{aligned} \check h_{(\mu\nu)}^{2} = \hat\phi^{2}_{(\mu\nu)} +\frac7{30}\hat\psi^{2}_{(\mu\nu)} +\frac{L^2}{8}\nabla_{(\mu}\nabla_{\nu)}\hat\phi^{2} -\frac{7L^2}{720}\nabla_{(\mu}\nabla_{\nu)}\hat\psi^{2},\end{aligned}$$ with $\hat\psi_{\mu\nu}^{2}\equiv 18h^{2}_{\mu\nu}-\frac{L}{2}(\nabla_{\mu}t^{2}_{\nu}+\nabla_{\nu}t^{2}_{\mu})$, $\hat\phi_{\mu\nu}^{2}\equiv -\frac75\left(h_{\mu\nu}^{2}+\nabla_{\mu}\nabla_{\nu}s^{2}\right)$, $\hat\phi^2\equiv \phi^2+\frac{16}{L^2}s^2$, is vanishing at linear order in $\mu_0$. This implies, at linear order in $\mu_0$, $h^2_{\mu\nu}$ is not an independent tensor mode and can be expressed in terms of the scalar modes as $$\begin{aligned} \label{hmn-ans} h_{\mu\nu}^{2} =&-\frac{L^2}{576}\nabla_\mu\nabla_\nu \hat\psi^2+\frac{11L^2}{224}\nabla_\mu\nabla_\nu \phi^2-\frac{3}{14}\nabla_\mu\nabla_\nu s^2-g_{\mu\nu}\Big(\frac1{72}\hat\psi^2+\frac3{28}\hat\phi^2\Big). \end{aligned}$$ We have also shown that, only two gauge invariant combinations of the four scalar fields $h^2,\phi^2, u^2, s^2$ are physical modes. The two gauge invariant physical scalar modes are $\hat \psi^2$ and $\hat\phi^2$ so that we express the four scalars as $$\begin{aligned} \label{GF} h^2=a_1\hat\psi^2+b_1\hat\phi^2,\quad u^2=a_2\hat\psi^2+b_2\hat\phi^2,\quad \phi^2=a_3\hat\psi^2+b_3\hat\phi^2 ,\quad s^2=a_4\hat\psi^2+b_4\hat\phi^2.\end{aligned}$$ In particular, these relations should be valid when we substitute the values of those scalar fields from the asymptotic expansion of the LLM solutions. In that case we find the following relations among the constants $a_i$ and $b_i$, $$\begin{aligned} \label{GF0} b_1=-\frac37(1-6a_1),\quad b_2=-\frac{18}{7}(2-a_2),\quad b_3=\frac37(1+6a_3),\quad b_4=\frac{L^2}{28}\Big(1+\frac{72a_4}{L^2}\Big).\end{aligned}$$ It is more convenient to use the diagonal modes $\check\phi^2=\frac9{70}(7\hat\psi^2+18\hat\phi^2),~\check\psi^2=\frac1{70}(7\hat\psi^2-162\hat\phi^2)$. See [@Jang:2016tbk] for the derivation of these diagonal modes. Then we can write $$\begin{aligned} \label{GF1} &h^2=\frac16\check\psi^2+\frac19\Big(10a_1-\frac16\Big)\check\phi^2,\quad u^2=2\check\psi^2+\frac29\Big(5a_2-1\Big)\check\phi^2,{\nonumber}\\ &\phi^2=-\frac16\check\psi^2+\frac19\Big(10a_3+\frac16\Big)\check\phi^2 ,\quad s^2=-\frac{L^2}{72}\check\psi^2+\frac19\Big(10a_4+\frac{L^2}{72}\Big)\check\phi^2.\end{aligned}$$ In the LLM solution, the asymptotic expansion of $\check\phi^{2}$ is of order $\mu_0^3$ and it does not contribute to the equations of motion at $\mu_0^2$ order. Therefore, we can set $\check\phi^{2}$ to zero and use and in to write the quadratic part of the equation of motion only in terms of $\check\psi^2$ as follows $$\begin{aligned} \label{lineq1-3-06a} &\Big(L_E+\frac{12}{L^2} \Big)\hat h_{\mu\nu}^0+\frac1{34560}\Big\{-\frac{26}3\nabla_\mu\check\psi^2\nabla_\nu\check\psi^2+\frac{28}3\check\psi^2\nabla_\mu\nabla_\nu\check\psi^2+\frac{L^2}3\nabla_\mu\nabla^\rho\check\psi^2\nabla_\nu\nabla_\rho\check\psi^2{\nonumber}\\ &+\frac{L^2}2\nabla^\rho\check\psi^2\nabla_\mu\nabla_\nu\nabla_\rho\check\psi^2+\frac{L^4}{24}\nabla_\mu\nabla^\rho\nabla^\sigma\check\psi^2\nabla_\nu\nabla_\rho\nabla_\sigma\check\psi^2+\frac{L^4}{32}\nabla^\rho\nabla^\sigma\check\psi^2\nabla_\mu\nabla_\nu\nabla_\rho\nabla_\sigma\check\psi^2{\nonumber}\\ &-g_{\mu\nu}\Big(\frac{12}{L^2}\check\psi^2\check\psi^2+\nabla^\rho\check\psi^2\nabla_\rho\check\psi^2+\frac{35L^2}{48}\nabla^\rho\nabla^\sigma\check\psi^2\nabla_\rho\nabla_\sigma\check\psi^2-\frac{L^4}{64}\nabla^\tau\nabla^\rho\nabla^\sigma\check\psi^2\nabla_\tau\nabla_\rho\nabla_\sigma\check\psi^2\Big)\Big\}{\nonumber}\\ &+\frac{1}{48}\big(\nabla_{\mu}\nabla_{\nu} t_+^1t_+^1+\frac12\nabla_{\mu} t_+^1\nabla_{\nu} t_+^1\big) +\frac{1}{96}g_{\mu\nu}\big(\nabla_\rho t_+^{1}\nabla^\rho t_+^1-\frac{16}{L^2} t_+^1 t_+^1\big)=0.\end{aligned}$$ In the next subsection we introduce some non-trivial field redefinitions in order to eliminate the higher derivative terms in the above equation and obtain a linearized equation for the 4-dimensional fields. The KK mapping at quadratic order --------------------------------- Our goal in this section is to apply the KK reduction procedure to the 11-dimensional supergravity on LLM geometry and construct 4-dimensional gravity theory whose solution encodes the information about the asymptotic limit of the LLM geometry. In the previous subsection, using the graviton mode example in , we have shown that the compactification of 11-dimensional supergravity on $S^7$ results in the field equations which contain higher derivative terms. In general, the same is true for all the modes in the KK towers. To absorb the higher derivative terms we need to introduce some field redefinitions. For instance, for some scalar KK mode whose equation of motion contains up to four derivatives, the field redefinition is of the form $$\begin{aligned} \label{field-red1} S^{I} = s^{I}+K_{IJ_1J_2} t^{J_1} t^{J_2} +L_{IJ_1J_2}\nabla_\mu t^{J_1}\nabla^\mu t^{J_2},\end{aligned}$$ where $K_{IJ_1J_2}$, $L_{IJ_1J_2}$ are some numerical coefficients, $s^{I}$ represents a gauge invariant 11-dimensional field and $S^{I}$ is the corresponding 4-dimensional field. See [@Skenderis:2006uy] for a systematic procedure of the KK reduction. The $t^{J_i}$’s represent all the fields that appear in higher derivative part of the equations of motion of $s^I$. If the field equation contains more than four derivatives, then the field redefinition will contain more than two derivatives. For the important scalar modes $\check\psi^2$ and $t^1_+$ discussed in the previous subsection, the equations of motion involve higher derivative terms only if we want to solve them at cubic or higher order in $\mu_0$. Up to quadratic order, their equations of motion are linear and are given by $$\begin{aligned} \label{diagpp} \left(\square-M^2_t\right)t_+^{1}=0,\qquad\left(\square-M^2_\psi \right)\check\psi{^2}=0,\end{aligned}$$ where $M^2_t=M^2_\psi=-\frac{8}{L^2}$. In this case, the field redefinitions are trivial and we can write the corresponding 4-dimensional fields as, $$\begin{aligned} \label{PsiT} \Psi=\check\psi^2 ,\qquad T=t^1_+ .\end{aligned}$$ The 4-dimensional gravity action with matter, which yields the equations of motion in for $\Psi$ and $T$ is given by $$\begin{aligned} \label{4dact} S = \frac{1}{16\pi G_N^{(4)}}\int d^4 x \sqrt{-g}\left(\hat R - 2\Lambda\right) + S_m, \end{aligned}$$ where $\Lambda=-\frac{(d-1)(d-2)}{2L^2_{\rm AdS_4}}=-\frac{12}{L^2}$ and $$\begin{aligned} \label{S_m} S_m=&-\frac{A_{t}}{2}\int d^4x\sqrt{-g}(\nabla_\mu T\nabla^\mu T+M_{t}^2T^{2})-\frac{A_{\psi}}{2}\int d^4x\sqrt{-g}(\nabla_\mu \Psi\nabla^\mu \Psi+M_{\psi}^2\Psi^{2}),\end{aligned}$$ for some over all normalizations $A_{t}$ and $A_{\psi}$ which will be fixed later. The corresponding energy-momentum tensor is $$\begin{aligned} \label{Tmn} \tilde T_{\mu\nu}=-\frac2{\sqrt{-g}}\frac{\delta S_m}{\delta g^{\mu\nu}}&=A_t\Big[\nabla_\mu T\nabla_\nu T-\frac12g_{\mu\nu}\big(\nabla_\rho T\nabla^\rho T+M_t^2 T^{2}\big)\Big]{\nonumber}\\ &+A_{\psi}\Big[\nabla_\mu \Psi\nabla_\nu \Psi-\frac12g_{\mu\nu}\big(\nabla_\rho \Psi\nabla^\rho \Psi+M_{\psi}^2\Psi^{2}\big)\Big].\end{aligned}$$ The next step is to obtain the equation of motion for the 4-dimensional graviton by using the result in and the above energy-momentum tensor. Since the background is AdS$_4$, we consider the 4-dimensional Einstein equation with negative cosmological constant $$\begin{aligned} \hat R_{\mu\nu}-\frac12\hat g_{\mu\nu}\hat R+\Lambda \hat g_{\mu\nu}=8\pi G_N \tilde T_{\mu\nu},\label{4DG}\end{aligned}$$ where $\hat g_{\mu\nu}=g_{\mu\nu}+\delta g_{\mu\nu}$ represents the metric which is deviated from the AdS$_4$ due to the presence of the energy-momentum tensor of the matter fields $\Psi$ and $T$. In order to obtain the equations for the graviton field, we insert the perturbed metric into and keep only up to the terms that are linear in the fluctuations $$\begin{aligned} \delta R_{\mu\nu}-\frac12\delta g_{\mu\nu}R-\frac12g_{\mu\nu}\delta g^{\rho\sigma}R_{\rho\sigma}-\frac12g_{\mu\nu}g^{\rho\sigma}\delta R_{\rho\sigma}-\frac{12}{L^2} \delta g_{\mu\nu}=8\pi G_N \tilde T_{\mu\nu}.\end{aligned}$$ Setting $\delta g_{\mu\nu}= H_{\mu\nu},~\delta g^{\mu\nu}=-H^{\mu\nu}$, we can write $$\begin{aligned} \label{Hmn-eq} &\frac12\Big(-\square H_{\mu\nu}+\nabla^\rho\nabla_\mu H_{\rho\nu}+\nabla^\rho\nabla_\nu H_{\rho\mu}-\nabla_{\mu}\nabla_{\nu} H\Big)+\frac{12}{L^2} H_{\mu\nu}-\frac{6}{L^2}g_{\mu\nu} H {\nonumber}\\ -&\frac12g_{\mu\nu}(\nabla^\rho\nabla^\sigma H_{\rho\sigma}-\square H)=8\pi G_N \tilde T_{\mu\nu}.\end{aligned}$$ Taking the trace, we obtain $$\begin{aligned} -\frac{12}{L^2} H-\nabla^\rho\nabla^\sigma H_{\rho\sigma}+\square H=8\pi G_Ng^{\rho\sigma} \tilde T_{\rho\sigma}.\end{aligned}$$ Plugging this into and using the energy-momentum tensor in , we obtain $$\begin{aligned} \label{Hmn-eq4} \Big(L_E +\frac{12}{L^2}\Big)H_{\mu\nu}&-8\pi G_N A_{t}\Big( \nabla_\mu T\nabla_\nu T+\frac{M_t^2}2g_{\mu\nu}T^{2}\Big){\nonumber}\\ &-8\pi G_N A_{\psi}\Big( \nabla_\mu \Psi\nabla_\nu \Psi+\frac{M_{\psi}^2}2g_{\mu\nu}\Psi^{2}\Big)=0.\end{aligned}$$ The final step is to introduce a field redefinition which eliminates the higher derivative terms in and reduces it to . Since the equation of motion in contains up to six derivative terms in $\check \psi^{2}$ and no higher derivative term in $t^1_+$, the field redefinition should contain up to four derivative terms in $\check \psi^{2}$ and no derivative in $t^1_+$. The required field redefinition is given by $$\begin{aligned} \label{FRD1} H_{\mu\nu}&=\hat h^{0}_{\mu\nu}+g_{\mu\nu}\big(\tilde C_1\check\psi^2\check\psi^2+\tilde C_2\nabla^\rho\check\psi^2\nabla_\rho\check\psi^2\big)+\tilde C_3\nabla_\mu\check\psi^2\nabla_\nu\check\psi^2{\nonumber}\\ &~~+g_{\mu\nu}\tilde C_4\nabla^\rho\nabla^\sigma\check\psi^2\nabla_\rho\nabla_\sigma\check\psi^2+\tilde C_5\nabla_\mu\nabla^\rho\check\psi^2\nabla_\nu\nabla_\rho\check\psi^2+g_{\mu\nu}\tilde C_t t^{1}_+ t^{1}_+.\end{aligned}$$ Inserting into and comparing the result with , we determine the unknown coefficients listed below, $$\begin{aligned} \label{coefs} & \tilde C_1=-\frac{1}{2^6\, 3^3\, 5},\quad \tilde C_2=-\frac{L^2}{2^{11}\,3^3\,5},\quad \tilde C_3=-\frac{7L^2}{2^{11}\,3^4\,5},\quad \tilde C_4=-\frac{L^4}{2^{14}\,3^3\,5},\\ & \tilde C_5=-\frac{L^4}{2^{13}\,3^4\,5},\quad \tilde C_t=-\frac{1}{2^5\,3},\quad 8\pi G_NA_t=\frac{1}{2^5\,3},\quad \quad 8\pi G_NA_{\psi}=\frac{1}{2^8\,3^2}.{\nonumber}\end{aligned}$$ In this paper we focused on obtaining the field redefinition for the graviton mode. A similar procedure determines the field redefinition for higher KK tensor, vector, and scalar modes in 4-dimension. In [@Skenderis:2006uy; @Skenderis:2006di] a similar result was obtained for 5-dimensional KK towers, which are obtained from the dimensional reduction of the 10-dimensional type-IIB supergravity on background which is asymptotically AdS${}_{5}\times S^{5}$. The resulting field redefinition was dubbed the KK map between the 10-dimensional and 5-dimensional fields. In our case, the KK map for graviton field is as in . HEE from *Vev* and *Source* =================================== When the conformal symmetry is broken due to a relevant deformation of a CFT, the CPOs develop non-vanishing one-point functions. In that case, one can infer that the deviation of the EE from it’s value in the CFT is related to the non-vanishing one-point function. From the perspective of the dual gravity, the non-conformal field theory corresponds to a gravity theory on an asymptotically AdS geometry. According to the RT conjecture, the deviation of the geometry from the AdS space induces a variation in HEE. In other words, non-vanishing one-point function of CPO of gauge theory induces the variation of EE and induces change of geometry in dual gravity. The dynamics of the geometry is governed by the linearized Einstein equations with matter fields interactions on the AdS background. Understanding the field theory counterpart of these governing equations is intriguing. Research in this direction is progressing. In this paper, we move further this progress by studying the HEE in mABJM theory. In [@Jang:2016tbk], we presented the calculation of one-point function of the CPO with conformal dimension one as an evidence supporting the duality between the mABJM theory and 11-dimensional supergravity theory on the LLM geometries. We obtained the one-point function, $\langle {\cal O}^{(1)}\rangle$ in the large $N$ limit, using the KK holography method from the 11-dimensional LLM geometry and showed that there is an exact agreement with the results obtained by using the supersymmetric vacua of mABJM theory. We conducted this test for all supersymmetric vacua, which are infinite in number. In order to strengthen the confirmation of the duality of the two theories, those results will be extended to the case of conformal dimension two CPO [@Jang:2018aqr]. In this paper, we exploit this exact correspondence to study the HEE based on the RT conjecture. To that end, we start by reading the asymptotically AdS solutions of the 4-dimensional gravity theory, which are constructed in the previous section, from the 11-dimensional LLM solutions. Applying the RT conjecture to those 4-dimensional gravity solutions, we calculate the leading order deviation of the HEE ($\delta S$) from its value in pure AdS$_4$ space. We compare the result with the known result of the HEE in the LLM geometry [@Kim:2016dzw]. Vacua of mABJM theory and LLM geometries ---------------------------------------- Under the supersymmetry preserving mass deformation of the ABJM theory, the global SU(4) symmetry of the ABJM theory is broken to SU(2)$\times$SU(2)$\times$U(1). To express the vacuum solution which reflects the broken symmetry, we split the scalar fields into $Y^A = (Z^a,\, W^{\dagger a})$, where $A = 1,2,3,4$ and $a = 1,2$. One interesting feature of the mABJM theory is that it has discrete Higgs vacua, which are represented as direct sums of GRVV matrices [@Gomis:2008vc]. The vacua are classified by occupation numbers, $N_n$ and $N_n'$ which are respectively denote the numbers of $n\times (n+1)$ GRVV matrices ${\cal M}_n$ and $(n+1)\times n$ GRVV matrices $\bar {\cal M}_n$, in the direct sums. See [@Cheon:2011gv] for the details. These vacua are supersymmetric if the occupation numbers are in the range, $0\le N_n,\, N_n'\le k$ [@Kim:2010mr]. There is a one-to-one map between the discrete Higgs vacua of the mABJM theory and the LLM geometries with ${\mathbb Z}_k$ orbifold, which have SO(2,1)$\times {\rm SO}(4)/{\mathbb Z}_k \times {\rm SO}(4)/{\mathbb Z}_k$ isometry. The LLM metric and the corresponding 4-form field strength are given by $$\begin{aligned} \label{LLMds2} ds^2 &= -{\bf G}_{tt} ( -dt^2 + dw_1^2 + dw_2^2) + {\bf G}_{xx} (d\tilde x^2 + d\tilde y^2) + {\bf G}_{\theta\theta} ds^2_{S^3/\mathbb{Z}_k} + {\bf G}_{\tilde\theta\tilde\theta} ds^2_{\tilde S^3/\mathbb{Z}_k},{\nonumber}\\ {\bf F}_4 &= -d \left(e^{2\Phi}h^{-2}V \right) \wedge dt\wedge dw_1\wedge dw_2 +\mu_0^{-1} \left[Vd(\tilde y^2e^{2G}) + h^2e^{3G}\star_2 d(\tilde y^2 e^{-2G})\right] \wedge d\Omega_3 {\nonumber}\\ &~~~+\mu_0^{-1} \left[ Vd(\tilde y^2e^{-2G}) -h^2e^{-3G}\star_2 d(\tilde y^2 e^{2G})\right] \wedge d\tilde\Omega_3,\end{aligned}$$ where $\mu_0$ is a mass parameter, $ds^2_{S^3/\mathbb{Z}_k}$ and $ds^2_{\tilde S^3/\mathbb{Z}_k}$ are line elements of three-spheres with ${\mathbb Z}_k$ orbifold, and $d\Omega_3=-(\sin\theta/8)d\theta\wedge d\phi\wedge d\psi$, $d\tilde\Omega_3=-(\sin\tilde\theta/8)d\tilde\theta\wedge d\tilde\phi\wedge d\tilde\psi$ are the volume forms of the two spheres in the Euler coordinate system. The warp factors and the functions defining the 4-form field strength are given by $$\begin{aligned} \label{warpfac} &{\bf G}_{tt}= -\left(\frac{4\mu_0^2 \tilde y\sqrt{\frac14 - Z^2}}{f^2} \right)^{2/3},\quad {\bf G}_{xx}=\left(\frac{f\sqrt{\frac14 - Z^2}}{2 \mu_0\tilde{y}^2}\right)^{2/3}, \quad {\bf G}_{\theta\theta} =\left( \frac{f\tilde y \sqrt{\frac12 + Z}}{2 \mu_0\left(\frac12 -Z\right)} \right)^{2/3},{\nonumber}\\ &{\bf G}_{\tilde\theta\tilde\theta}=\left( \frac{f\tilde y \sqrt{\frac12 - Z}}{2\mu_0 \left(\frac12 + Z\right)}\right)^{2/3},\quad h^2=\frac{\sqrt{\frac14 - Z^2}}{\tilde y}, \quad e^{2\Phi}=\frac{4 \tilde y \mu_0^2 \sqrt{\frac14 - Z^2}}{f^2},\quad e^{2 G}= \frac{\frac12 + Z}{\frac12 - Z}\end{aligned}$$ with $f(\tilde x,\tilde y) = \sqrt{1 - 4 Z^2 - 4\tilde y^2 V^2}.$ Here we notice that the LLM geometry is completely determined by two functions, $$\begin{aligned} \label{ZandV} Z(\tilde x,\tilde y) =\sum_{i=1}^{2N_B\!+\!1}\frac{(-1)^{i\!+\!1} (\tilde x\!-\!\tilde x_i)}{2\sqrt{(\tilde x\!-\!\tilde x_i)^2+\tilde y^2}} \ ,\qquad V(\tilde x,\tilde y) =\sum_{i=1}^{2N_B\!+\!1}\frac{(-1)^{i\!+\!1}}{2\sqrt{(\tilde x\!-\!\tilde x_i)^2+\tilde y^2}},\end{aligned}$$ where $\tilde x_i$’s are numerical parameters. The function $Z(\tilde x,\tilde y)$ at $\tilde y=0$ has a value $\frac12$ for $\tilde x_{2i-1}< \tilde x < \tilde x_{2i}$ and it is $-\frac12$ for $\tilde x_{2i}< \tilde x < \tilde x_{2i+1}$. The geometries are classified by those values of $Z(\tilde x,0)$. That is, the LLM geometries are represented as an infinite strip in the $\tilde x$-direction with $Z(\tilde x,0)=-\frac12$ denoted by black strip and $Z(\tilde x,0)=\frac12$ denoted by white strip. This representation of the LLM geometry is called the droplet picture. So the $\tilde x_i$’s denote the positions of the boundaries between the black and the white strips and $N_B$ is the number of finite-sized black/white strips in the droplet representation. Due to flux quantization condition of the 4-form field strength, the difference between consecutive $\tilde x_i$’s is quantized as [@Cheon:2011gv] $$\begin{aligned} \tilde x_{i+1}-\tilde x_i = 2\pi l_{{\rm P}}^3\mu_0 {\mathbb Z},\end{aligned}$$ where $l_{{\rm P}}$ is the Planck length. Therefore, all LLM geometries are completely determined by these quantized locus $\tilde x_i$’s. In order to consider the gauge/gravity duality near the UV fixed point of the field theory, we need to expand the dual geometry in the asymptotic region. For the asymptotic expansion of the general LLM geometries, it is convenient to introduce new parameters [@Kim:2016dzw], $$\begin{aligned} C_p = \sum_{i=1}^{\infty}(-1)^{i+1} \left( \frac{\tilde x_i}{2\pi l_{{\rm P}}^3 \mu_{0}\sqrt{A}}\right)^p,\end{aligned}$$ where $A$ is defined as[^5] $$\begin{aligned} \label{AA} A = k N -\frac12 \sum_{n=0}^{\infty} \left[l_n (k-l_n) + l_n' (k-l_n')\right]. \end{aligned}$$ Here we introduce a new set of parameters, $\{ l_n, \,l_n'\}$, which are called discrete torsions and used to classify the LLM geometries in the droplet picture. See [@Cheon:2011gv] for the details. The one-to-one correspondence between the vacua of the mABJM theory and the LLM geometries identifies $\{N_n, N_n'\}$ with $\{l_n, l_n'\}$. KK reduction of the LLM geometries {#ast-AdS-metric} ---------------------------------- In the Fefferman-Graham coordinate system, the LLM metric is given by (see [@Jang:2016tbk] for details) $$\begin{aligned} \label{dsFG} ds^2 =&\frac{ L^2}{4 z^2}\left[dz^2 +\frac{4z^2}{L^2} [1+{\tilde g}_1(z,\tau )]\left( -dt^2+dw_1^2+dw_2^2\right) \right]{\nonumber}\\ &+ [1+{\tilde g}_2(z,\tau ) ]d\tau^2+ [1+{\tilde g}_3(z,\tau )] ds_{S^3}^2 + [1+{\tilde g}_4(z,\tau )] ds_{\tilde S^3}^2,\end{aligned}$$ where the $\tilde g_i(z,\tau)$ represents the deviation of the LLM metric from the AdS$_4\times S^7$ background. Similarly, the 4-form field strength can be split into the background and the rest. In the asymptotic region, these deviations become small fluctuations, and the LLM solution can be written as in , where the values of the small fluctuations $h_{pq}$ and $f_{pqrs}$ are read from the asymptotic expansion of $\tilde g_i(z,\tau)$, and the similar terms in the 4-form field strength. Then, these small fluctuations can be expanded in terms of the spherical harmonics on $S^7$, in order to obtain the values of the KK modes ($h^{I_1}_{\mu\nu}, \phi^{I_1}, ~ etc$) of the previous section. In [@Jang:2016tbk], we have listed the full result for all the KK modes up to $\mu_0^2$ order. Here, we are interested only in the graviton mode, which corresponds to a combination of the KK zero modes. We also need ($I_1=2, I_{35}=1$) modes, which appear in the quadratic part of the graviton equation of motion. Therefore, we copy the following results from [@Jang:2016tbk] $$\begin{aligned} \label{hhppp} &h_{ij}^0 = \left[\frac{L^2\mu_0^2}{720}\left(360 + 7 \beta_3^2 \right) + {\cal O}\left(\mu_0^4\right)\right]\eta_{ij},\qquad h_{zz}^0 = 0, {\nonumber}\\ &h^0 = \frac{(\mu_0 z)^2}{60}\left(360 + 7 \beta_3^2\right) + {\cal O}\left(\mu_0^4\right), \qquad \phi^0 = \frac{(\mu_0 z)^2}{15}\left(-80 + \beta_3^2\right) + {\cal O}\left(\mu_0^4\right), {\nonumber}\\ &\hat\psi^0 = \frac{6(\mu_0 z)^2}{5}\left( 80 + \beta_3^2\right)+ {\cal O}(\mu_0^4), \qquad \check\psi^{2} = -24\beta_3\mu_0 z+{\cal O}(\mu_0^3), {\nonumber}\\ &t_+^1 = 16\sqrt{3} \, \mu_0 z + {\cal O}(\mu_0^3),\end{aligned}$$ where $\eta_{ij} = {\rm diag}(-1,1,1)$ and $$\begin{aligned} \label{beta3} \beta_3 = 2 C_1^3 - 3 C_1 C_2 + C_3. \end{aligned}$$ In the previous section, we have established the KK maps which relate the above 11-dimensional KK modes to the corresponding canonical 4-dimensional gravity fields. These maps are given in and . Using these maps, we can write the asymptotically AdS$_4$ solution to the 4-dimensional gravity equations and from the KK reduction of the LLM solution. The results are $$\begin{aligned} \label{Hmn} &H_{ij} = \left[-\frac{(L\mu_0)^2}{180}\left( 30 + \beta_3^2\right)+{\cal O} \left(\mu_0^4\right)\right]\eta_{ij} , \qquad H_{zz} = - \frac{(L\mu_0)^2}{1440}\left(960 + 29\beta_3^2\right) + {\cal O}\left(\mu_0^4\right){\nonumber}\\ &\Psi = -24 \beta_3 \mu_0 z + {\cal O}(\mu_0^3), \qquad T = 16\sqrt{3} \, \mu_0 z + {\cal O}(\mu_0^3).\end{aligned}$$ One-point function for the CPO with $\Delta=1$ ---------------------------------------------- The $vev$ of a CPO with conformal dimension one in mABJM was obtained in [@Jang:2016tbk]. For clarity of presentation, we shortly review that result here. The CPO in ABJM theory with conformal dimension one, which preserves the SU(2)$\times$SU(2)$\times$U(1) global symmetry and has non-vanishing $vev$, is given by $$\begin{aligned} \label{CPO} {\cal O}^{(1)} = \frac1{2\sqrt{2}} {\rm Tr}\left( Z^a Z_a^\dagger - W^{\dagger a} W_a\right).\end{aligned}$$ Since the supersymmetry of the ABJM theory protects the scalar fields from quantum corrections and the contributions from the multi-trace terms are suppressed by $1/N$ as compared to single-trace terms, the $vevs$ of the CPOs are exactly determined by the classical values for scalar fields in the large $N$ limit [@Jang:2016tbk]. Based on this argument, we obtained the $vev$ of the CPO with conformal dimension one for all supersymmetric vacua in large $N$ limit, $$\begin{aligned} \label{vevCPO2} \langle {\cal O}^{(1)}\rangle_m = \frac{k \mu }{4\sqrt{2}\, \pi } \sum_{n=0}^{\infty}n(n+1)(N_n - N_n'),\end{aligned}$$ where $\langle\cdots\rangle_m$ represents the $vev$ in the mABJM theory and $\mu$ is the mass parameter related to the LLM geometry mass parameter as $\mu = 4\mu_0$. In the gauge/gravity duality, the relation between the conformal dimension of gauge invariant operators and the mass of the dual scalar modes in the 4-dimensional gravity theory is given by $$\begin{aligned} \label{ScalarMass} \frac{m_\phi^2 L^2}{4} = \Delta (\Delta -3).\end{aligned}$$ In our case the gravity mode dual to the CPO with conformal dimension one is the mode $\Psi$ in . The gauge/gravity duality dictionary states that, the $vev$ of a CPO with conformal dimension $\Delta$ is determined by the coefficient of $z^\Delta$ in the asymptotic expansion of the dual scalar field. According to this rule, the $vev$ of the CPO of conformal dimension one is determined by the asymptotic expansion of $\Psi$ in as $$\begin{aligned} \label{dual_rel} \langle {\cal O}^{(1)}\rangle_m = - \frac{24 N^2}{\sqrt{\lambda}}\, {\mathbb N} \, \beta_3 \mu_0,\end{aligned}$$ where ${\mathbb N}$ depends on the normaliation of the scalar field $\Psi$ and $\lambda= N/k$ is the ’t Hooft coupling constant in the ABJM theory. The overall factor $N^2/\sqrt{\lambda}$ is originated from the gauge/gravity dual relation in 4-dimensional gravity, $$\begin{aligned} \frac1{16\pi G_N^{(4)}}\sim \frac{N^2}{\sqrt{\lambda} L^2}.\end{aligned}$$ In order to fix the normalization ${\mathbb N}$, we use the identity (see [@Jang:2016tbk] for the proof) $$\begin{aligned} \label{beta3_2} \beta_3= \frac{3}{A^{\frac32}}\sum_{n=0}^{N_B}n(n+1)(l_n - l_n'). \end{aligned}$$ For $k=1$ and the general $N_B$ or the general $k$ and $N_B=1$ cases, this identity is valid for any $N\ge 2$. However, for both $k$ and $N_B$ greater than one, the right-hand side of is only the leading order term in the $\frac1N$-expansion of $\beta_3$. In the large $N$ limit, we note that the leading contribution of $A$ is $A = kN$. Therefore, recalling that the field theory result in is obtained in the large $N$ limit, we can fix the normalization as ${\mathbb N} = -\frac{\sqrt{2}}{144\pi}$ by comparing the field theory result and the gravity result in the large $N$ limit. The one-to-one map $\{l_n,\, l_n'\} \Longleftrightarrow \{N_n,\, N_n'\}$ is also used. Then the $vev$ can be written as $$\begin{aligned} \label{vevCPO3} \langle {\cal O}^{(1)}\rangle_m =\frac{N^2\mu_0}{3\sqrt{2}\,\pi\sqrt{\lambda}}\, \beta_3.\end{aligned}$$ In [@Jang:2016tbk], we have verified that $\beta_3$ is independent of $N$, so that the overall normalization factor in is proportional to $N^{\frac32}$ in the case $k=1$, which is the well-known relation in M2-brane theory. The above relation gives an exact dual relation in large $N$ limit for all supersymmetric vacua in mABJM theory and the corresponding LLM geometries with ${\mathbb Z}_k$ orbifold. HEE from LLM geometries in 4-dimensions --------------------------------------- According to the RT conjecture, the HEE with a subspace $A$ at a fixed time on the boundary of ($d$+1)-dimensional AdS geometry is given by $$\begin{aligned} \label{S_A} S_{A} = \frac{{\rm Min}(\gamma_A)}{4 G_N},\end{aligned}$$ where $G_N$ is the Newton constant in the ($d$+1)-dimensional gravity theory and $\gamma_A$ is an area of the surface stretched to the bulk direction, which has the same boundary with the subsystem $A$. The surface is expressed by the induced metric, $$\begin{aligned} \label{indmet} g^{(0)}_{ij} = \frac{\partial w^\mu}{\partial \sigma^i}\frac{\partial w^\nu}{\partial \sigma^j} g_{\mu\nu}, \end{aligned}$$ where $\sigma^1,\cdots, \sigma^{d-1}$ are coordinates on the surface, $g_{\mu\nu}$ is the asymptotically AdS$_{d+1}$ bulk metric, and $w^0, w^1,\cdots, w^{d}$ are the bulk coordinates. The area $\gamma_A$ is given by $$\begin{aligned} \label{gamA} \gamma_A = \int d^{d-1}\sigma \sqrt{\det \tilde g_{ij}}\,. \end{aligned}$$ In section \[4dgravity\], we have constructed a 4-dimensional gravity theory using KK reduction from the 11-dimensional gravity. We showed that, up to the quadratic order in the mass parameter $\mu_0$, there are only two scalar fields that are coupled to the 4-dimensional metric. These two scalars carry the information of the asymptotic expansion of the LLM geometry. In the absence of these scalar fields, the geometry is pure AdS$_4$, and the scalar fields induce the deviation from the AdS$_4$ space. In this section, we use the asymptotically AdS$_4$ metric in subsection \[ast-AdS-metric\] to compute the variation of the HEE $\delta S_A$ from its original value $S^0_A$ in pure AdS$_ 4$ geometry. ### HEE from pure AdS geometry in 4-dimensions Before we proceed to the calculation of $\delta S_A$, let us summarize the calculation of the HEE for the pure AdS$_4$ [@Ryu:2006ef]. The 4-dimensional AdS metric is given by $$\begin{aligned} ds^2_{{\rm AdS}} = \frac{L_{{\rm AdS}}^2}{z^2}\left( -dt^2 + dw_1^2 + dw_2^2 + dz^2 \right),\end{aligned}$$ where $L_{{\rm AdS}}$ denotes the radius of the AdS$_4$ geometry and $z$ represents the holographic direction. To obtain the HEE for a subspace $A$ which is a disk of radius $l$, we consider a mapping for the codimension 2 coordinates $\sigma^{1,2}$, $$\begin{aligned} \label{map1} t = {\rm constant}, \quad w_1 = \sigma^2\cos\sigma^1, \quad w_2 = \sigma^2\sin\sigma^1, \quad z = z(\sigma^2). \end{aligned}$$ The components of the induced metric are given by $$\begin{aligned} \label{tilgij2} \tilde g_{11}^{(0)} = \frac{L_{{\rm AdS}}^2 \,\rho^2}{z^2}, \qquad \tilde g_{12}^{(0)} = 0, \qquad \tilde g_{22}^{(0)} = \frac{L_{{\rm AdS}}^2 }{z^2}\left(1 + z'^2\right), \end{aligned}$$ where we set $\sigma^2 = \rho$ and $z'\equiv \left(\partial z/\partial\rho\right)$. Then the area of the surface $\gamma_A$ is given by $$\begin{aligned} \label{gamA2} \gamma_A = \int_0^{2\pi}d\sigma^1\int_0^{l}d\rho \sqrt{\det\tilde g_{ij}} = 2\pi L_{{\rm AdS}}^2 \int_0^{l}d\rho \frac{\rho}{z^2}\sqrt{1+ z'^2}.\end{aligned}$$ The solution of $z(\rho)$ which minimizes $\gamma_A$ is $$\begin{aligned} \label{solz} z(\rho) = \sqrt{l^2 - \rho^2}\end{aligned}$$ with boundary conditions $$\begin{aligned} z(l) =0,\qquad z'(0) = 0. \end{aligned}$$ Computing the minimum area of $\gamma_A$, we obtain the HEE, $$\begin{aligned} S_{A}^{(0)} = \frac{\pi L^2}{8 G_N^{(4)}}\left(\frac{l}{\epsilon} - 1\right),\end{aligned}$$ where $\epsilon$ is the UV cut-off in the $z$-direction. Variation of HEE from LLM geometries in 4-dimensions ---------------------------------------------------- The asymptotically AdS$_4$ metric can be split into the pure AdS part and fluctuations as $$\begin{aligned} \label{4Dgmn} \hat g_{\mu\nu} = g_{\mu\nu} + H_{\mu\nu},\end{aligned}$$ where $g_{\mu\nu}$ is metric of the pure AdS$_4$. Then the induced metric is given by $$\begin{aligned} \label{defmet} \tilde g_{ij} &= \frac{\partial w^\mu}{\partial \sigma^i}\frac{\partial w^\nu}{\partial \sigma^j} \hat g_{\mu\nu} = \frac{\partial w^\mu}{\partial \sigma^i}\frac{\partial w^\nu}{\partial \sigma^j}\left(g_{\mu\nu} + H_{\mu\nu}\right) = \tilde g_{ij}^{(0)} + \tilde H_{ij},\end{aligned}$$ where $\tilde g_{ij}^{(0)}$ is the induced AdS$_4$ metric and $ \tilde H_{ij}$ is determined from the asymptotic AdS$_4$ solution given in . Specifically, $$\begin{aligned} \label{indH} \tilde H_{11} &= \rho^2 H_{11} = -\frac{( L\mu_0)^2\rho^2}{180}\left( 30 + \beta_3^2\right), {\nonumber}\\ \tilde H_{22} &= H_{11} + z'^2 H_{zz} = -\frac{(L\mu_0)^2}{1440}\left[ 240 + 8\beta_3^2 + \left(960 + 29 \beta_3^2\right) z'^2\right]. \end{aligned}$$ Apply the mapping to the induced metric , we obtain the variation of the area, $$\begin{aligned} \label{delEE} \delta \gamma_A &=\frac{1}{2}\int d^2\sigma \sqrt{\det\tilde g^{(0)}}\, \tilde g^{(0)ij}\tilde H_{ij}=\pi\int_0^l d\rho \sqrt{\det\tilde g^{(0)}}\, \tilde g^{(0)ij}\tilde H_{ij}.\end{aligned}$$ Inserting into , we obtain, $$\begin{aligned} \label{gammaA1} \delta\gamma_A &= -\frac{\pi L^2 \mu_0^2}{1440}\int_0^l d\rho \frac{\rho}{\sqrt{1 + z'^2}}\left[\left(1200 + 37\beta_3^2\right) z'^2 + 16 \left(30 + \beta_3^2\right)\right] {\nonumber}\\ &= -\frac{\pi L^2 (\mu_0 l)^2}{48}\left(32 + \beta_3^2\right).\end{aligned}$$ In the last step of , we have used the solution of $z(\rho)$ given in . Therefore, the HEE up to $\mu_0^2$-order is given by $$\begin{aligned} \label{4dHEE} S_A = S_A^{(0)} + \delta S_A &= \frac{\pi L^2}{8 G_N^{(4)}}\left[\frac{l}{\epsilon} - 1 - \frac43\left(1+ \frac{\beta_3^2}{32}\right)\left(\mu_0 l\right)^2\right] {\nonumber}\\ &= \frac{\sqrt{2}\,\pi N^2}{3\sqrt{\lambda}}\left[\frac{l}{\epsilon} - 1 - \frac43\left(1+ \frac{\beta_3^2}{32}\right)\left(\mu_0 l\right)^2\right],\end{aligned}$$ where the KK reduction relates the 4-dimensional and the 11-dimensional Newton’s constant as $$\begin{aligned} \frac1{G_N^{(4)}} = \frac{{\rm vol}(S^7/{\mathbb Z}_k)}{G_N^{(11)}} = \frac{\pi^4 L^7}{3k\, G_N^{(11)}}= \frac{8\sqrt{2}\, N^2}{3\sqrt{\lambda}\, L^2}. \end{aligned}$$ Here we used the gauge/gravity dual relation in the ABJM theory and the 11-dimensional supergravity, $16\pi G_N^{(11)} = (2\pi)^8 l_{{\rm P}}^9$ and $L^6 = 32\pi^2 k N l_{{\rm P}}^6$. Inserting this into , we reproduce the HEE from the LLM geometries in 11-dimensions [@Kim:2016dzw]. As shown in section 2, the matter fields in the 4-dimensional gravity theory are determined by the 11-dimensional geometry. The equality of the HEE obtained from pure geometrical 11-dimensional gravity theory and the one obtained from the 4-dimensional matter-gravity theory shows that the information of the LLM geometry in the asymptotic limit is exactly encoded in the solutions of the matter fields in 4-dimensions. Reversing this statement, the RT formula may play some role in the construction of geometrical solutions in higher dimensional theories from the solutions of matter fields in lower dimensional theories. It is intriguing to examine this possibility in our setup. However, it is highly non-trivial to achieve this goal because one has to extract the information of metric, which is local, from the HEE, which is non-local. HEE from holographic mapping of *vev* and *source* ---------------------------------------------------------- In [@Jang:2016tbk], we have discussed the gauge/gravity maps between gauge invariant operators in ABJM theory and scalar modes in 11-dimensional supergravity, which encode the information of LLM geometries with $S^7/{\mathbb Z}_k$ orbifold. Among the five infinite KK towers of scalar modes listed in [@Jang:2016tbk], only the two scalar modes in are nonvanishing in the leading order of the asymptotic expansion of the LLM geometries. Using the relations between mass and conformal dimension listed in [@Jang:2016tbk], we observe that the scalar field $\Psi$ with $M^2_{\Psi}=\frac{I(I-6)}{L^2}\big\|_{I=2}$ is dual to CPO of conformal dimension $\Delta=\frac I2\big\|_{I=2}=1$ while the pseudoscalar field $T$ with $M^2_{T}=\frac{(I-3)(I+3)}{L^2}\big\|_{I=1}$ is dual to gauge invariant operator with conformal dimension $\Delta= \frac{I+3}2\big\|_{I=1}=2$. Then, the gauge/gravity dictionary implies[^6] the expansion of these two fields in powers of the holographic coordinate $z$ should read $$\begin{aligned} \label{asypExp} \Psi(z)=V_{\Psi}z+S_{\Psi}z^2+\cdots,\quad T(z)=S_{T}z+V_{T}z^2+\cdots,\end{aligned}$$ where the coefficients $V_{\Psi,T}$ are determined by the $vevs$ of the dual operators and the coefficients $S_{\Psi,T}$ are determined by the values of the external sources which are coupled to those operators. Comparing with our results in , we see that $S_{\Psi}=V_{T}=0$, while $V_{\Psi}=-24\beta_3\mu_0$ and $S_{T}=16\sqrt3\mu_0$. Therefore, one can identity the coefficients of $z$ in $\Psi(z)$ as [$vevs$]{} of the CPO with $\Delta = 1$ and the coefficients of $z$ in $T(z)$ as the [source]{} of a gauge invariant operator with $\Delta = 2$ in 3-dimensions. The coupling of these two scalars to gravity in 4-dimensions causes the deformation of the induced metric in . More precisely, the $\beta_3$ terms in are due to the coupling with $\Psi$, while the numerical terms are due to coupling with $T$. As we discussed in the previous subsection, the matric deformation determines the variation of the HEE ($\delta S_A$), which means that the $\beta_3$-term in $\delta S_A$ in is originated from the [$vevs$]{} of the CPO, $\langle{\cal O}^{(1)}\rangle_m$, while the numerical term in $\delta S_A$ is originated from the [source]{} of the gauge invariant operator $J_{\tilde {\cal O}^{(2)}}$. The variation of HEE is actually related to the squares of the [$vevs$]{} and the [source]{} term. Therefore, up to $\mu_0^2$-order in the large $N$ limit, the HEE can be written as $$\begin{aligned} \label{deltaSA} \delta S_A = -\frac{4\sqrt{2} \pi\, N^2 l^2}{\sqrt{\lambda}}\left[\frac1{9}\left(J_{\tilde {\cal O}^{(2)}}\right)^2 + \frac1{16} \left(\frac{\pi\sqrt{\lambda}\langle {\cal O}^{(1)}\rangle_m}{N^2}\right)^2\right],\end{aligned}$$ where we set the source of the operator $\tilde {\cal O}^{(2)}$ as $J_{\tilde {\cal O}^{(2)}} =\mu_0$. The CPO of conformal dimension one is as in , and the gauge invariant operator $\tilde {\cal O}^{(2)}$ is built from the fermionic fields of the ABJM theory, $(\psi_{A},$ $A=1,2,3,4)$, and is given by [@Bak:2010ry] $$\begin{aligned} \label{CPOt} \tilde{\cal O}^{(2)}=\tilde C_A^B{\rm Tr}\big(\psi^{\dagger A}\psi_{B}\big),\end{aligned}$$ where $\tilde C_A^B$ are traceless. See [@Bena:2000zb] for the source of the mass deformation by adding fermion mass terms in the dual field theory. In general the CPOs of conformal dimension $\Delta$ are given by $$\begin{aligned} \label{CPOD} {\cal O}^{(\Delta)}=C_{A_1,\cdots,A_{\Delta}}^{B_1,\cdots,B_{\Delta}}{\rm Tr}\big(Y^{A_1}Y^\dagger_{B_1}\cdots Y^{A_{\Delta}}Y^\dagger_{B_{\Delta}}\big).\end{aligned}$$ The gauge invariant operators $\tilde{\cal O}^{(\Delta)}$ are descendents of the CPOs and they are given by $$\begin{aligned} \label{CPOtD} \tilde{\cal O}^{(\Delta)}=\tilde C_{a,A_{1},\cdots,A_{\Delta-2}}^{b,B_1,\cdots,B_{\Delta-2}}{\rm Tr}\big(\psi^{\dagger a}\psi_{b}Y^{A_1}Y^\dagger_{B_1}\cdots Y^{A_{\Delta-2}}Y^\dagger_{B_{\Delta-2}}\big).\end{aligned}$$ The CPOs ${\cal O}^{(\Delta)}$ are dual to the scalar KK modes $\Psi^{I}$ $(I=2,4,6,\cdots)$ with $\Delta=\frac I2$, and the gauge invariant operators $\tilde{\cal O}^{(\Delta)}$ are dual to the pseudoscalar KK modes $T^I$ $(I=1,3,5,\cdots)$ with $\Delta=\frac {I+3}2$. We obtain the expansions in holographic coordinate, for those two KK towers of dual scalars from the asymptotic expansion of the LLM solutions, which are given by $$\begin{aligned} \label{astExp-2} &\Psi^{2+4i}(z)=\tilde\psi_1(\mu_0z)+\tilde\psi_3 (\mu_0z)^3+\cdots+\tilde\psi_{\Delta} (\mu_0z)^{\Delta}+\tilde\psi_{\Delta+2} (\mu_0z)^{\Delta+2}+\cdots,\quad(\Delta=1+2i),{\nonumber}\\ & \Psi^{4+4i}(z)=\tilde\psi_2(\mu_0z)^2+\tilde\psi_4 (\mu_0z)^4+\cdots+\tilde\psi_{\Delta} (\mu_0z)^{\Delta}+\tilde\psi_{\Delta+2} (\mu_0z)^{\Delta+2}+\cdots,\quad(\Delta=2+2i),{\nonumber}\\ &T^{1+4i}(z)=t_1(\mu_0z)+t_3 (\mu_0z)^3+\cdots+t_{\Delta-1} (\mu_0z)^{\Delta-1}+t_{\Delta+1} (\mu_0z)^{\Delta+1}+\cdots,\quad(\Delta=2+2i),{\nonumber}\\ &T^{3+4i}(z)=t_2(\mu_0z)^2+t_4 (\mu_0z)^4+\cdots+t_{\Delta-1} (\mu_0z)^{\Delta-1}+t_{\Delta+1} (\mu_0z)^{\Delta+1}+\cdots,\quad(\Delta=3+2i),\end{aligned}$$ where $i=0,1,2,\cdots$. As mentioned before, in these expansions the coefficient of $z^{\Delta}$ is determined by the $vev$ of the dual operator while the coefficient of $z^{d-\Delta}\|_{d=3}=z^{3-\Delta}$ is determined by the value of the external source which is coupled to the operator. Among the towers of the KK modes in , the only mode which has nonvanishing coefficient of $z^{3-\Delta}$ is $T=T^{1+4i}\big\|_{i=0}$. Therefore, the dual operator $\tilde{\cal O}^{(2)}$ is the only operator which is coupled to an external source, while the rest have no external source term. In addition, from , we see that all the gauge invariant operators dual to the pseudo-scalar fields $T^{I}$ have vanishing $vev$s whereas all the CPOs which are dual to the scalar fields $\Psi^{I}$ have non-vanishing $vev$s. This suggests that, the variation of the HEE ($\delta S_A$) can depend on the source term of the gauge invariant operator $\tilde{\cal O}^{(2)}$ and the $vevs$ of the CPOs ${\cal O}^{(\Delta)}(\Delta=1,2,3\cdots)$. However, as can be seen from , for $\Delta\ge2$, the $vevs$ are at least quadratic in $\mu_0$, and cannot contribute to $\delta S_A$ at quadratic order in $\mu_0$. Therefore, at quadratic order in $\mu_0$, the variation of HEE is fully determined by the source term of the operator $\tilde{\cal O}^{(2)}$ and the $vev$ of the CPO ${\cal O}^{(1)}$, which is the result in . Comments on relative entropy and Fisher information --------------------------------------------------- The variation of the EE, $\delta S_A$ in , is connected with the relative entropy, which is defined as $$\begin{aligned} S(\rho \|\| \sigma) \equiv {\rm tr}\left(\rho \log\rho\right) - {\rm tr}\left(\rho \log\sigma\right),\end{aligned}$$ where $\rho$ is a deformed density matrix from some reference density matrix $\sigma$. For the ball-shaped region $A$, the relative entropy is represented as $$\begin{aligned} \label{RelE} S(\rho_A \|\| \sigma_A) = \Delta \langle H_A\rangle - \Delta S_A, \end{aligned}$$ where $H_A$ is the modular Hamiltonian associated with the region $A$. See below for details. At first order in the deformation parameter, the relative entropy vanishes and becomes the first law of the EE. Therefore, the leading nonvanishing contribution to relative entropy is quadratic in the deformation parameter. In our case the deformation parameter is $\mu_0$, and due to the supersymmetry of the mABJM theory, $\Delta \langle H_A\rangle_{\mu_0^2} =0$ (see the discussion below the equation ). Then using the result in , we have $$\begin{aligned} \label{relSA} S(\rho_A \|\| \sigma_A)\|_{\mu_0^2} = \frac{\pi L^2}{6 G_N^{(4)}}\left( 1 + \frac{\beta_3^2}{32}\right)\left(\mu_0 l\right)^2. \end{aligned}$$ The relative entropy is positive definite as a measure of [distance]{} between two quantum states and monotonically increasing with the size of the subsystem [@Blanco:2013joa; @Lin:2014hva; @Lashkari:2014kda; @Lashkari:2015hha; @Jafferis:2015del; @Lin:2016fua]. The relation in represents those properties clearly. The positivity of the relative entropy states the positivity of the Fisher information metric with one deformation parameter, $$\begin{aligned} F_{\mu_0\mu_0} = \frac{d^2 }{d\mu_0^2}S(\rho_A \|\| \sigma_A) = \frac{\pi L^2 l^2}{3 G_N^{(4)}}\left( 1 + \frac{\beta_3^2}{32}\right).\end{aligned}$$ This quantity is also known as the fidelity susceptibility. In [@Lashkari:2015hha], it was shown that the Fisher information of the reduced density matrix of a ball-shaped subregion at the CFT vacuum is connected to the canonical energy [@Hollands:2012sf] for perturbations in the corresponding Rindler wedge of the AdS geometry. See also [@Hyun:2016dvt] for the canonical energy by using the Euler-Lagrange expression and its connection to the Fisher information. In most cases, for instance [@Nozaki:2013vta; @Lin:2014hva; @Lashkari:2015hha; @Beach:2016ocq; @Banerjee:2017qti], the second order deformation in the gravity is connected to nonvanishing $vev$ of gauge invariant operator in QFT. Though the field fluctuations in gravity side in our case are originated from the $vev$ and $source$ of gauge invariant operators, we expect that the interpretation of the Fisher information in QFT as the canonical energy in the gravity side is correct since the roles of the dual fields $\Psi$ and $T$ in gravity side are indistinguishable. Gravity from Entanglement and RG Flow ===================================== In the previous section, we have determined the HEE by applying the RT formula in 4-dimensional gravity, which is obtained from the KK reduction of the 11-dimensional gravity on the LLM geometry. At the quadratic order in $\mu_0$, our result is in complete agreement with the one obtained by applying the RT formula in the 11-dimensional supergravity before the KK reduction [@Kim:2016dzw]. This indirectly proves that the solution of the 4-dimensional gravity theory we have built contains all the information of the 11-dimensional LLM geometry near the UV fixed point. In the dual gauge theory, the asymptotic limit of the LLM geometry describes the RG flow from the UV fixed point where the ABJM theory lives. In this section we discuss the first law-like relation for the EE when there is RG flow due to relevant perturbations from the UV fixed point. Emergent gravity from relevant perturbations in CFT {#relpert} --------------------------------------------------- We consider an Euclidean CFT action with a relevant deformation, $$\begin{aligned} \label{defAct} I = I^{(0)} + \tilde\lambda \int d^d w \,{\cal O}^{(\Delta)},\end{aligned}$$ where $I^{(0)}$ is the $d$-dimensional CFT action, ${\cal O}^{(\Delta)}$ is a gauge invariant operator with conformal dimension $\Delta$, and $\tilde\lambda$ is the deformation parameter with mass dimension $d-\Delta>0$. In QFT, the EE is defined as $$\begin{aligned} \label{SEE} S_{A} = - {\rm tr} \left(\rho_A \ln \rho_A\right),\end{aligned}$$ where the total space of states is divided into two subregions $A$, $B$ and $\rho_A$ is the reduced density matrix of the subregion $A$ at a given time. Here we consider the subregion $A$ is in a shape of $(d-1)$-dimensional ball of radius $l$. Under the relevant deformation , the density matrix is deformed as $\rho_A = \rho_A^{(0)} + \delta\rho_A$ with the matrix $\rho_A^{(0)}$ is for the undeformed CFT. The EE is calculated in the path integral formalism using the perturbative expansion in the deformation parameter $\tilde\lambda$. At the linear order in $\tilde\lambda$, the perturbative evaluation of the EE produces the relation $ \delta S = \delta \langle H_A\rangle_{\tilde\lambda}$, where $H_A$ is the modular Hamiltonian defined as $H_A\equiv -\ln \rho_A $. This relation is known as the first law of the EE [@Blanco:2013joa; @Bhattacharya:2012mi]. For a ball-shaped subregion $A$, the modular Hamiltonian is expressed in terms of energy-momentum tensor (see below). At the quadratic order of the perturbative expansion, it was pointed out that the variation of EE gets contributions from the two point function $\langle {\cal O}^{(\Delta)} {\cal O}^{(\Delta)}\rangle$ and the three point function $\langle H_A {\cal O}^{(\Delta)} {\cal O}^{(\Delta)}\rangle$ [@Rosenhaus:2014woa; @Rosenhaus:2014zza]. Later, it was found that in the evaluation of the two point function $\langle {\cal O}^{(\Delta)} {\cal O}^{(\Delta)}\rangle$ there exists an additional finite contribution at the $\tilde\lambda^2$-order, which comes from the non-commutative property between the matrix representations of $\rho_A^{(0)}$ and $\delta\rho_A$ [@Faulkner:2014jva]. As we will see later, this finite term reflects the deviation of EE away from the UV fixed point under the relevant deformation and was identified quantitatively in the holographic picture via the RT conjecture. Next, we follow [@Faulkner:2014jva] and briefly discuss the two terms contributing to the variation of EE at $\tilde\lambda^2$-order. The first one is given by $$\begin{aligned} \label{delS1} \delta S^{(1)} = \frac{\tilde\lambda^2}2 \int d^dw\int d^dw' \langle H_A {\cal O}^{(\Delta)}(w){\cal O}^{(\Delta)}(w')\rangle_0 = \delta\langle H_A\rangle_{\tilde\lambda^2},\end{aligned}$$ where $\langle\cdots\rangle_0$ denotes the $n$-point functions of the undeformed theory. The result in has also been obtained in [@Rosenhaus:2014woa; @Rosenhaus:2014zza]. For the ball-shaped region $A$, the explicit form of the modular Hamiltonian $H_A$ for the $d$-dimensional CFT is obtained in [@Casini:2011kv], $$\begin{aligned} \label{H_A} H_A = 2\pi \int_{B(l,\vec w_0)} d\Sigma^i \zeta^j T_{ij}= 2\pi \int_{B(l,\vec w_0)} d^{d-1} w \frac{l^2 - \|\vec w - \vec w_0\|^2}{ 2 l}\, T_{tt}(t_0,\vec w),\end{aligned}$$ where $i =0,1,\cdots,d-1$, the ball $B(l,\vec w_0)$ is on a time slice $t=t_0$, it is of radius $l$ and centered at $\vec w = \vec w_0$. The $d\Sigma^i$ is the volume form on the $(d-1)$-dimensional surface perpendicular to a unit vector in $i$-th direction and $\zeta$ is the conformal Killing vector defined as $$\begin{aligned} \label{Kzeta} \zeta = \left(\frac{l^2 -\rho^2 - t^2}{2l}\right)\partial_t - \frac{\rho t}{l}\partial_\rho\end{aligned}$$ with radius $\rho=\sqrt{w_1^2 + \cdots + w_{d-1}^2}$. Here $T_{tt}$ denotes the $(tt)$-component of the energy-momentum tensor in $d$-dimensional CFT. Inserting into , the calculation of $\delta S^{(1)}$ is reduced to the evaluation of the three point function $\langle T_{tt}(w) {\cal O}^{(\Delta)} (w') {\cal O}^{(\Delta)}(w'')\rangle_0$. In the mABJM theory with supersymmetric discrete Higgs vacua, the three point function is vanishing in the large $N$ limit. In order to see this, we expand the field near the vacua as $Y^{A=1,2,3,4} = Y_0^A + \tilde Y^A$ with the vacuum configuration $Y_0^A$. Then the gauge invariant operators are written as $$\begin{aligned} {\cal O}^{(\Delta)}(w) = {\cal O}^{(\Delta)}_0 + \sum_i\tilde {\cal O}_i^{(\Delta)}(w), \label{opDel}\end{aligned}$$ where ${\cal O}^{(\Delta)}_0\equiv {\cal O}^{(\Delta)}(w)\|_{Y^A = Y_0^A}$ and $\tilde {\cal O}^{(\Delta)}_i$’s are operators expanded around the vacuum $Y_0^A$. Then the three point function can formally be written as $$\begin{aligned} \langle T_{tt}(w) {\cal O}^{(\Delta)} (w') {\cal O}^{(\Delta)}(w'')\rangle_0 &= \sum_{i,j} \langle T_{tt}(w) \tilde{\cal O}_i^{(\Delta)} (w') \tilde{\cal O}_j^{(\Delta)}(w'')\rangle_0+\left({\cal O}_0^{(\Delta)}\right)^2\langle T_{tt}(w)\rangle_0{\nonumber}\\ &+{\cal O}_0^{(\Delta)}\sum_i\left[\langle T_{tt}(w)\tilde {\cal O}_i^{(\Delta)}(w')\rangle_0+\langle T_{tt}(w)\tilde {\cal O}_i^{(\Delta)}(w'')\rangle_0\right]. \end{aligned}$$ In CFT, the conformal invariance dictates that the two point function $\langle {\cal O} T_{\mu\nu}\rangle_0$ is vanishing. In addition, the first term on the right-hand side in the above equation is a multi-trace term and is suppressed by $1/N$ as compared with the single trace terms. Therefore the three point function in the large $N$ limit is given by $$\begin{aligned} \langle T_{tt}(w) {\cal O}^{(\Delta)} (w') {\cal O}^{(\Delta)}(w'')\rangle_0 &= \left({\cal O}_0^{(\Delta)}\right)^2\langle T_{tt}(w)\rangle_0 + \frac1{N}-{\rm corrections}. \end{aligned}$$ Since the mABJM theory is a supersymmetric gauge theory, the $vev$ of the energy-momentum tensor $\langle T_{tt}(w)\rangle_0$ is vanishing. Therefore, we expect that $\delta S^{(1)}$ for the mABJM theory is vanishing in the large $N$-limit. The second contribution to the variation of the EE is obtained in [@Faulkner:2014jva] and it is given by $$\begin{aligned} \label{delS2} \delta S^{(2)} = -2\pi \int_{{\cal H}^+} d\Sigma^\mu\xi^\nu \tilde{T}_{\mu\nu},\end{aligned}$$ where $\mu,\,\nu = 0,1,\cdots d$ including one additional direction, which will be identified in dual gravity theory as the holographic direction. Here the integration is performed over the region ${\cal H}^+$, which is the future part of the Rindler horizon in the emergent AdS$_{d+1}$ space and $d\Sigma^\mu$ is the surface element on the horizon. The energy-momentum tensor $\tilde{T}_{\mu\nu}$ of an auxiliary scalar field $\tilde\phi$, is $$\begin{aligned} \label{TAab} \tilde{T}_{\mu\nu} = \nabla_\mu\tilde\phi \nabla_\nu\tilde\phi - \frac12 g_{\mu\nu}^{(0)}\left( \nabla_\lambda\phi\nabla^\lambda\tilde\phi + m^2\tilde\phi^2\right),\end{aligned}$$ which satisfies the conservation law $\nabla_\mu\tilde{T}^{\mu\nu} = 0$, and this means the $\delta S^{(2)}$ in is a conserved charge. The field $\tilde\phi$ satisfies the Klein-Gordon equation, $$\begin{aligned} \label{phieq} \nabla^2\tilde\phi - \frac{\Delta (\Delta- d)}{L_{{\rm AdS}}^2}\tilde\phi = 0\end{aligned}$$ with some boundary conditions at $z=0$. For calculation of $\delta S^{(2)}$, we consider an explicit example of the emergent space. To do that, we introduce the AdS$_{d+1}$ geometry in the Poincare patch, $$\begin{aligned} \label{ds2} ds^2 = \frac1{z^2}\left(-dt^2 + dz^2 + d\rho^2 + \rho^2 d\Omega_{d-2}^2\right), \end{aligned}$$ where $d\Omega_{d-2}^2$ is the line element on $S^{d-2}$ with unit radius. The Killing vector $\zeta$ in of the $d$-dimensional CFT extends to a Killing vector in the emergent bulk geometry, $$\begin{aligned} \label{Kvecxi} \xi = \frac{l^2 -z^2-\rho^2-t^2}{2l}\partial_t - \frac{t}{l}\left(z\partial_z + \rho\partial_\rho\right).\end{aligned}$$ When we choose a spatial slice at $t=0$, the AdS Rindler horizon is reduced to the minimal surface satisfying the relation $z(\rho) = \sqrt{l^2-\rho^2}$ in . Then the variation $\delta S^{(2)}$ in is given by $$\begin{aligned} \label{delS2-3} \delta S^{(2)} = -\frac{\pi \Omega_{d-2} L_{{\rm AdS}}^{d-1}}{l}\int_{z_\Lambda}^l dz z^{1-d}\int_0^{\sqrt{l^2-z^2}}d\rho \rho^{d-2}\left(l^2-z^2-\rho^2\right)\tilde T_{tt},\end{aligned}$$ where $\Omega_{d-2}= \frac{2\pi^{(d-1)/2}}{\Gamma\left((d-1)/2\right)}$. We used the volume form $d\Sigma^t = L_{{\rm AdS}}^{d-1}z^{1-d}dz \rho^{d-2} d\rho d\Omega_{d-2}$ and the Killing vector $\xi^t = \frac{l^2-z^2-\rho^2}{2l}$, and also introduced the cutoff scale $z_\Lambda$ to regularize the divergence at $z=0$. In the $z\to 0$ limit, the $\tilde\phi$ configuration satisfying the equation is given by $$\begin{aligned} \label{phisol} \tilde\phi(z, w)\longrightarrow V_{\tilde\lambda}( w) z^\Delta + S_{\tilde\lambda}( w) z^{d-\Delta}.\end{aligned}$$ Since this auxiliary field is absent in the undeformed theory, the coefficients $V_{\tilde\lambda}$ and $S_{\tilde\lambda}$ also depend on the deformation parameter[^7]. Inserting the solution into , one can obtain $\delta S^{(2)}$ in the path integral method. For a very small ball-shaped region, $l\ll L_{{\rm AdS}}$, the coefficients $V_{\tilde\lambda}$ and $S_{\tilde\lambda}$ can be regarded as constants, which are consistent with our case. In the literature the $\Delta<\frac d2$ and $\Delta>\frac d2$ cases are treated separately, whereas the $\Delta=\frac d2$ case needs a special treatment [@Casini:2016rwj; @Speranza:2016jwt]. [$\bullet$ $\Delta < \frac{d}{2}$ case:]{} To obtain the leading contribution for the small value of $\tilde\lambda$ in $\delta S^{(2)}$, it is enough to consider the asymptotic behavior of $\tilde\phi(z,w)$ as $$\begin{aligned} \label{glam=0} \tilde\phi(z,w) = V_{\tilde\lambda} z^\Delta, \end{aligned}$$ where $V_{\tilde\lambda} = \mathbb{N}_V\tilde\lambda$ with a numerical factor $\mathbb{N}_V$. Then we obtain the $(tt)$-component of the energy-momentum tensor, $T_{tt}^A = V_{\tilde\lambda}^2 z^{2\Delta -2} \Delta \left(\Delta - \frac{d}{2}\right)$. Inserting this into , we obtain [@Blanco:2013joa; @Speranza:2016jwt] $$\begin{aligned} \label{delS2-1} \delta S^{(2)} &= \frac{\pi \Omega_{d-2} \mathbb{N}_V^2\tilde\lambda^2L_{{\rm AdS}}^{d-1}}{l}\Delta\left(\frac{d}{2} - \Delta\right)\int_{z_\Lambda}^{l} z^{2\Delta - d -1} \int_0^{\sqrt{l^2- z^2}} d\rho \rho^{d-2}\left(l^2 - z^2- \rho^2\right) {\nonumber}\\ &= -\frac{\pi \Omega_{d-2} \mathbb{N}_V^2\tilde\lambda^2L_{{\rm AdS}}^{d-1} l^{2\Delta}\Delta \Gamma\left(\frac{d+3}{2}\right)\Gamma\left(\Delta-\frac{d}{2}+1\right)}{(d^2-1) \Gamma\left(\frac32 + \Delta\right)}+ \delta S^{(2)}_{{\rm div}},\end{aligned}$$ where $\delta S^{(2)}_{{\rm div}}$ includes the divergent pieces depending the UV-cutoff $z_\Lambda$, $$\begin{aligned} \delta S^{(2)}_{{\rm div}}= -\frac{\pi \Omega_{d-2} \mathbb{N}_V^2\tilde\lambda^2L_{{\rm AdS}}^{d-1} l^{d}\Delta}{(d^2-1)z_\Lambda^{d-2\Delta}}+ {\cal O}\left(z_\Lambda^{-d + 2\Lambda + 2}\right).\end{aligned}$$ [$\bullet$ $\Delta > \frac{d}{2}$ case:]{} In order to extract the leading contribution for the small value of $\tilde\lambda$, we consider the asymptotic behavior of $\tilde\phi(z,w)$ as $$\begin{aligned} \label{flam=0} \tilde\phi(z,w) = S_{\tilde\lambda} z^{d-\Delta}, \end{aligned}$$ where $S_{\tilde\lambda} = \mathbb{N}_S \tilde\lambda$ with a numerical factor $\mathbb{N}_S$. Then we obtain the $(tt)$-component of the energy-momentum tensor, $T_{tt}^A = S_{\tilde\lambda}^2 z^{2d -2\Delta -2} \left(\Delta - \frac{d}{2}\right) \left(\Delta -d\right)$. Inserting this into , we obtain the result in [@Faulkner:2014jva], $$\begin{aligned} \label{delS2-2} \delta S^{(2)} &= \frac{\pi \Omega_{d-2} \mathbb{N}_S^2\tilde\lambda^2L_{{\rm AdS}}^{d-1}}{l}(d-\Delta)\left(\Delta-\frac{d}{2} \right)\int_{z_\Lambda}^{l} z^{d-2\Delta -1} \int_0^{\sqrt{l^2- z^2}} d\rho \rho^{d-2}\left(l^2 - z^2- \rho^2\right) {\nonumber}\\ &= -\frac{\pi \Omega_{d-2} \mathbb{N}_S^2\tilde\lambda^2L_{{\rm AdS}}^{d-1} l^{2(d-\Delta)}(d-\Delta) \Gamma\left(\frac{d+3}{2}\right)\Gamma\left(\frac{d}{2}-\Delta +1\right)}{(d^2-1) \Gamma\left(\frac32 +d- \Delta\right)}+ \delta S^{(2)}_{{\rm div}},\end{aligned}$$ where $$\begin{aligned} \delta S^{(2)}_{{\rm div}}= -\frac{\pi \Omega_{d-2} \mathbb{N}_S^2\tilde\lambda^2L_{{\rm AdS}}^{d-1} l^{d}(d-\Delta)}{(d^2-1) z_\Lambda^{2\Delta-d}}+ {\cal O}\left(z_\Lambda^{d - 2\Lambda + 2}\right).\end{aligned}$$ The leading finite contribution in was also obtained in [@Liu:2012eea] using the RT formula. In the calculations of $\delta S^{(1)}$ in and $\delta S^{(2)}$ in , there exist divergent terms. One needs to subtract those terms by adding an appropriate count term in $\tilde\lambda^2$-order $$\begin{aligned} \label{deltaS_A} \delta S_A = \langle H_A\rangle_{\tilde\lambda} -2\pi \int_{{\cal H}^+} d\Sigma^\mu\xi^\nu \tilde T_{\mu\nu} + S_{ct}.\end{aligned}$$ Gravity from entanglement in the mABJM theory --------------------------------------------- In the previous subsection, using the path integral method in $d$-dimensional CFT in the presence of the relevant deformation , we have summarized that some part of $\delta S$ in $\tilde\lambda^2$-order becomes a conserved charge in $(d+1)$-dimensional theory. The conserved charge is defined by introducing one additional coordinate $z$ and one auxiliary field $\tilde \phi$. The appearance of the additional coordinate indicates the emergence of gravity from the entanglement in QFT. We expect the emergent gravity identifies with the gravity theory which is dual to the QFT. In particular, the energy-momentum tensor of the auxiliary field is expected to be identified with the energy-momentum tensor of a dynamical scalar field in the dual gravity theory. In this paper, we have constructed a 4-dimensional gravity theory with the matter sector composed of two scalar fields $\Psi$ and $T$. The result we summarized in the previous subsection suggests that these two scalar fields should emerge from the calculation of the variation of the EE in the 3-dimensional dual mABJM theory. In this subsection, we discuss this phenomena in detail and show the emergence of 4-dimensional Einstein equation from the EE analysis. In order to calculate the quantity $\delta S^{(2)}$ in from the energy-momentum tensor defined in , we treat the contributions from $\Psi$ and $T$ separately. The operator which is dual to the field $\Psi$ is of conformal dimension one and it corresponds to the case $\Delta <\frac{d}{2}$, while that of the field $T$ is of conformal dimension two and corresponds to the case $\Delta >\frac{d}{2}$. Comparing the energy-momentum tensors in and , we need to rescale the scalar fields as $\tilde \Psi = \sqrt{A_\psi} \Psi$ and $\tilde T = \sqrt{A_t} T$ and then asymptotic behaviours of the scalar fields in are rescaled as $$\begin{aligned} \label{tilPsi} \tilde \Psi = -24\sqrt{A_\psi}\,\beta_3\mu_0 z + {\cal O}(\mu_0^3), \qquad \tilde T = 16\sqrt{3}\sqrt{A_t}\,\mu_0 z + {\cal O}(\mu_0^3).\end{aligned}$$ The energy-momentum tensors for $\tilde \Psi$ and $\tilde T$ up to $\mu_0^2$-order are read from , $$\begin{aligned} \label{tilTij} \tilde T_{ij}^{(\tilde \Psi)} &= 288 A_\psi\beta_3^2\mu_0^2\eta_{ij},\qquad \tilde T_{zz}^{(\tilde \Psi)} = 864 A_\psi\beta_3^2\mu_0^2, {\nonumber}\\ \tilde T_{ij}^{(\tilde T)} &= 384 A_t\mu_0^2,\qquad \tilde T_{zz}^{(\tilde T)} = 1152 A_t\mu_0^2.\end{aligned}$$ For the field $\tilde\Psi$ in , we obtain the variation of the EE from after the cancellation of the divergent term, $$\begin{aligned} \delta S^{(2)}_\Psi = -96 \pi^2 L^2 A_\psi \beta_3^2 (\mu_0 l)^2.\end{aligned}$$ Similarly, for the scalar field $\tilde T$, the result is obtained from , $$\begin{aligned} \delta S^{(2)}_T = -128 \pi^2 L^2 A_t (\mu_0 l)^2.\end{aligned}$$ Therefore, the total variation of the EE is given by $$\begin{aligned} \label{delS2-4} \delta S^{(2)} = \delta S^{(2)}_\Psi+\delta S^{(2)}_T = -128\pi^2 L^2 \left(A_t+ \frac{3 A_\psi \beta_3^2}{4} \right)(\mu_0 l)^2.\end{aligned}$$ In the subsection \[relpert\], we argued that the quantity $\delta S^{(1)}$ for the mABJM theory is vanishing in the large $N$ limit. This is because of vanishing vacuum energy density $\langle T_{ij}(w)\rangle_0 = 0$ in supersymmetric theories. Now we examine again the vanishing of $\delta S^{(1)}$ for the mABJM theory using the gauge/gravity duality dictionary. To that end, we start from the deformed 4-dimensional metric in and , $$\begin{aligned} \label{oldFG} ds^2 = \frac{L^2}{4z^2}\left[\left(1- {\cal B}(\mu_0z)^2+{\cal O}( z^4)\right)dz^2+\left(1- {\cal A}(\mu_0z)^2+{\cal O}( z^4)\right)\eta_{ij} dx^i dx^j \right],\end{aligned}$$ where $$\begin{aligned} {\cal A} =\frac1{45}\left(30 + \beta_3^2\right), \qquad {\cal B} = \frac1{360}\left(960 + 29\beta_3^2\right). \end{aligned}$$ The FG coordinate system is convenient to read the $vev$ of the energy-momentum tensor from the asymptotic expansion of the metric. Therefore, we apply the coordinate transformation, $$\begin{aligned} z\,\longrightarrow\, \tilde z = z - \frac{\mu_0^2 {\cal B}}{4} z^3, \end{aligned}$$ to the metric in order to write it in FG-coordinate system, $$\begin{aligned} \label{newFG} ds^2 = \frac{L^2}{4\tilde z^2}\left[ d\tilde z^2 + \left(1- \frac12\left(2 {\cal A} + {\cal B}\right)(\mu_0 \tilde z)^2 + {\cal O}(\tilde z^4)\right)\eta_{ij}dx^i dx^j\right].\end{aligned}$$ This asymptotic behavour of the metric tells us the fact that $\langle T_{ij}\rangle_0$ is vanishing[^8]. This confirms the claim that $\delta S^{(1)} =0$ for the mABJM theory. In conclusion, the quantity $\delta S^{(2)}$ represents the full variation of the EE in the mABJM theory, which coincides with the variation of area obtained in . On the other hand, for a general metric perturbation, the variation of the area is also given [@Wald:1993nt; @Faulkner:2013ica], $$\begin{aligned} \label{gammaA3} \delta \gamma_A = -\int_{{\cal H}_0} d\Sigma^\mu\xi^\nu \delta G_{\mu\nu} + \delta \gamma_A^{(ct)}, \end{aligned}$$ where ${\cal H}_0$ is a region on the space-like surface $t=0$ squeezed between the minimum area and the disk $A$ with the UV cut-off at $z=z_\Lambda$. The term $\delta \gamma_A^{(ct)}$ is introduced to cancel out the divergences which arise from $z_\Lambda\to 0$ limit. Here the variation of the Einstein tensor $\delta G_{\mu\nu}$ is read from the left-hand side of , $$\begin{aligned} \label{deltaG} \delta G_{\mu\nu} = &\frac12\Big(-\square H_{\mu\nu}+\nabla^\rho\nabla_\mu H_{\rho\nu}+\nabla^\rho\nabla_\nu H_{\rho\mu}-\nabla_{\mu}\nabla_{\nu} H\Big)+\frac{12}{L^2} H_{\mu\nu}-\frac{6}{L^2}g_{\mu\nu} H {\nonumber}\\ -&\frac12g_{\mu\nu}(\nabla^\rho\nabla^\sigma H_{\rho\sigma}-\square H).\end{aligned}$$ Inserting into , we obtain $$\begin{aligned} \label{deltaG2} \delta G_{ij} = \frac{\mu_0^2}{8}\left(32 + \beta_3^2\right)\eta_{ij},\qquad \delta G_{zz} = \frac{3\mu_0^2}{8} \left(32 + \beta_3^2\right). \end{aligned}$$ Plugging into , we obtain $$\begin{aligned} \label{gammaA4} -\int_{{\cal H}_0} d\Sigma^t \xi^t \delta G_{tt} = -\frac{\pi L^2}{48}\left(32 + \beta_3^2\right) (\mu_0 l)^2 + \frac{\pi L^2}{128}\frac{l}{z_\Lambda}\left(32 + \beta_3^2\right) (\mu_0 l)^2.\end{aligned}$$ The divergent term in $z_\Lambda \to 0$ limit in is cancelled by $\delta \gamma_A^{(ct)}$ in and the finite variation of the area is equivalent to the $\delta\gamma_A$ in . Finally, we identify the variation of the EE obtained from field theory calculations with the one obtained from the RT formula, $$\begin{aligned} \label{tilTG} -2\pi \int_{{\cal H}^+} d\Sigma^\mu\xi^\nu \tilde T_{\mu\nu} = -\frac{1}{4G_N^{(4)}}\int_{{\cal H}_0} d\Sigma^\mu\xi^\nu \delta G_{\mu\nu}. \end{aligned}$$ For the Killing vector $\xi^\mu$, the term in the left-hand side of defines a conserved quantity and thus one can choose any surface homologous to ${\cal H}^+$ as the integration surface. For the choice of the surface at $t=0$, the integration surfaces of both sides in are identified. Then, from this relation, we obtain $$\begin{aligned} \label{EinEq} \delta G_{\mu\nu} = 8\pi G_N^{(4)} \tilde T_{\mu\nu},\end{aligned}$$ which is the linearized Einstein equation with matter fields. Inserting and into , one can determine the numerical factors $A_\psi$ and $A_t$, which exactly match the values given in . Therefore, one can see that the RT formula satisfying the Einstein equation reproduces the variation of the EE calculated in the path integral method in the field theory side. Conclusion ========== In this paper, we investigated the phenomena of the emergent gravity in 4-dimensions from the EE of the 3-dimensional mABJM theory. Using the path integral method developed in [@Faulkner:2014jva] and the RT formula, we clarified the relation between the emergent (auxiliary) gravity and the Einstein-Hilbert action with two scalar fields in 4-dimensions, which is obtained from the KK reduction of the 11-dimensional supergravity on the LLM geometries with ${\mathbb Z}_k$ orbifold. Our analysis relies heavily on the gauge/gravity duality between the mABJM theory and the 11-dimensional supergravity on the LLM geometries. In order to setup the gauge/gravity dictionary, we need to construct the 4-dimensional gravity using the KK reduction on the compact manifold $S^7/{\mathbb Z}_k$. In our previous work [@Jang:2016tbk], we showed an exact dual relation for the $vev$ of a CPO of conformal dimension one (${\cal O}^{(1)}$) in mABJM theory and a scalar field in an asymptotically AdS$_4$ gravity theory in the large $N$ limit. However, the connection between the 4-dimensional graviton mode and the 11-dimensional fluctuations was missing. In this paper, for the minimal ingredients that encode all the information of the LLM geometries with ${\mathbb Z}_k$ orbifold in the asymptotic limit up to $\mu_0^2$-order, we completed the non-trivial KK maps between the 4-dimensional fields and the 11-dimensional fluctuations on AdS$_4\times S^7/{\mathbb Z}_k$. The resulting 4-dimensional fields are composed of the graviton mode $H_{\mu\nu}$, one scalar field $\Psi$, and one pseudoscalar field $T$. In the matter sector, the asymptotic behaviour of the scalar field $\Psi$ determines the $vev$ of ${\cal O}^{(1)}$ whereas the asymptotic behaviour of the pseudoscalar field $T$ determines the $source$ which couples to a gauge invariant operator of conformal dimension two ($\tilde{\cal O}^{(2)}$). The presence of these scalar fields in the 4-dimensional gravity theory implies the existence of some relevant deformation in the dual CFT and as a result the EE is expected to show a variation from that of the CFT. Employing the holographic RT formula for subregion $A$, which is a disk of radius $l$, we calculated the leading order contribution to the variation of the EE $\delta S_A$ in mABJM theory by using the 4-dimensional metric ($g_{\mu\nu} +H_{\mu\nu}$), which encodes the information of the asymptotic LLM geometries in small mass limit. We showed that, the leading order contribution to the $\delta S_A$ is quadratic in the deformation parameter $\mu_0$. At such leading order, $\delta S_A$ is completely fixed by the $vev$ of ${\cal O}^{(1)}$ and the $source$ which couples to $\tilde{\cal O}^{(2)}$. The obtained $\delta S_A$ is negative, which is consistent with the $F$-theorem in 3-dimensional gauge theory and describes the RG flow from UV fixed point of the conformal invariant ABJM theory. Based on a recent progress in the computation of the EE by using the path integral method in CFT with some relevant deformations, we reproduced the HEE for the mABJM theory. In order to calculate the variation of EE using the path integral method, it is necessary to introduce an additional coordinate $z$ and one auxiliary scalar field for every relevant operator added to the CFT action. In the quadratic order of approximation, the mABJM theory is regarded as a deformation of the ABJM theory by the relevant operators ${\cal O}^{(1)}$ and $\tilde{\cal O}^{(2)}$. Therefore, we need to introduce two auxiliary scalar fields. We identified these auxiliary fields with the scalar fields $\Psi$ and $T$, in the calculation of the EE using the path integral approach. Consistent with the holographic method, the leading order contribution to $\delta S_A$ from the path integral approach is quadratic in the deformation parameter. Furthermore, the variation of HEE obtained from the RT formula and the $\delta S_A$ from the path integral methods are equal, when the linearized Einstein equation with the energy-momentum tensor of the two scalar fields $\Psi$ and $T$, is satisfied. This Einstein equation agrees with the one we obtained from the KK reduction of the 11-dimensional supergravity on LLM geometries. This agreement and the appearance of the additional coordinate $z$ are the indications of the emergence of an asymptotically AdS$_4$ gravity from the quantum entanglement of the 3-dimensional mABJM theory. In the calculation of $\delta S_A$ in terms of the path integral method developed in [@Faulkner:2014jva], we used the energy-momentum tensor for the two scalar fields $\Psi$ and $T$ in the dual 4-dimensional gravity theory, relying on the exact dual relation in our previous work in [@Jang:2016tbk]. However, in general, one can compute the $\delta S_A$ up to $\mu_0^2$-order in the path integral method by using the mABJM theory directly without using the dual theory. We leave this for future work. Acknowledgements {#acknowledgements .unnumbered} ================ OK would like to thank the participants of the 5th IBS Brainstrom workshop for stimulating discussions and appreciates APCTP for its hospitality during completion of this work. DT would like to thank the physics department of Addis Ababa University for hospitality, during the visit to present part of this work. This work was supported by the National Research Foundation of Korea(NRF) grant with grant number NRF-2016R1D1A1B03931090 (Y.K.), NRF-2017R1D1A1A09000951 (O.K.), and NRF-2017R1D1A1B03032523 (D.T.). [99]{} S. Ryu and T. Takayanagi, “Holographic derivation of entanglement entropy from AdS/CFT,” Phys. Rev. Lett. [96]{}, 181602 (2006) \[hep-th/0603001\]. S. Ryu and T. Takayanagi, “Aspects of Holographic Entanglement Entropy,” JHEP [0608]{}, 045 (2006) \[hep-th/0605073\]. V. E. Hubeny, M. Rangamani and T. Takayanagi, “A Covariant holographic entanglement entropy proposal,” JHEP [0707]{}, 062 (2007) \[arXiv:0705.0016 \[hep-th\]\]. J. M. Maldacena, “The Large N limit of superconformal field theories and Int. J. Theor. Phys. [38]{}, 1113 (1999) \[Adv. Theor. Math. Phys. [2]{}, 231 (1998)\] \[hep-th/9711200\]. S. S. Gubser, I. R. Klebanov and A. M. Polyakov, “Gauge theory correlators from noncritical string theory,” Phys. Lett. B [428]{}, 105 (1998) \[hep-th/9802109\]; E. Witten, “Anti-de Sitter space and holography,” Adv. Theor. Math. Phys. [2]{}, 253 (1998) \[hep-th/9802150\]. B. Swingle, “Entanglement Renormalization and Holography,” Phys. Rev. D [86]{}, 065007 (2012) \[arXiv:0905.1317 \[cond-mat.str-el\]\]. M. Van Raamsdonk, “Comments on quantum gravity and entanglement,” arXiv:0907.2939 \[hep-th\]; “Building up spacetime with quantum entanglement,” Gen. Rel. Grav. [42]{}, 2323 (2010) \[Int. J. Mod. Phys. D [19]{}, 2429 (2010)\] \[arXiv:1005.3035 \[hep-th\]\]. J. Bhattacharya, M. Nozaki, T. Takayanagi and T. Ugajin, “Thermodynamical Property of Entanglement Entropy for Excited States,” Phys. Rev. Lett. [110]{}, no. 9, 091602 (2013) \[arXiv:1212.1164 \[hep-th\]\]. D. Allahbakhshi, M. Alishahiha and A. Naseh, “Entanglement Thermodynamics,” JHEP [1308]{}, 102 (2013) \[arXiv:1305.2728 \[hep-th\]\]. D. D. Blanco, H. Casini, L. Y. Hung and R. C. Myers, “Relative Entropy and Holography,” JHEP [1308]{}, 060 (2013) \[arXiv:1305.3182 \[hep-th\]\]. G. Wong, I. Klich, L. A. Pando Zayas and D. Vaman, “Entanglement Temperature and Entanglement Entropy of Excited States,” JHEP [1312]{}, 020 (2013) \[arXiv:1305.3291 \[hep-th\]\]. N. Lashkari, M. B. McDermott and M. Van Raamsdonk, “Gravitational dynamics from entanglement ’thermodynamics’,” JHEP [1404]{}, 195 (2014) \[arXiv:1308.3716 \[hep-th\]\]. T. Faulkner, M. Guica, T. Hartman, R. C. Myers and M. Van Raamsdonk, “Gravitation from Entanglement in Holographic CFTs,” JHEP [1403]{}, 051 (2014) \[arXiv:1312.7856 \[hep-th\]\]. B. Swingle and M. Van Raamsdonk, “Universality of Gravity from Entanglement,” arXiv:1405.2933 \[hep-th\]. J. W. Lee, H. C. Kim and J. Lee, “Gravity from Quantum Information,” J. Korean Phys. Soc. [63]{}, 1094 (2013) \[arXiv:1001.5445 \[hep-th\]\]. B. Mosk, “Holographic equivalence between the first law of entanglement entropy and the linearized gravitational equations,” Phys. Rev. D [94]{}, no. 12, 126001 (2016) \[arXiv:1608.06292 \[hep-th\]\]. B. Czech, L. Lamprou, S. McCandlish, B. Mosk and J. Sully, “Equivalent Equations of Motion for Gravity and Entropy,” JHEP [1702]{}, 004 (2017) \[arXiv:1608.06282 \[hep-th\]\]. E. Oh, I. Y. Park and S. J. Sin, “Complete Einstein equation from the genenalized First Law of Entanglement,” arXiv:1709.05752 \[hep-th\]. T. Faulkner, F. M. Haehl, E. Hijano, O. Parrikar, C. Rabideau and M. Van Raamsdonk, “Nonlinear Gravity from Entanglement in Conformal Field Theories,” JHEP [1708]{}, 057 (2017) \[arXiv:1705.03026 \[hep-th\]\]. K. Hosomichi, K. M. Lee, S. Lee, S. Lee and J. Park, “N=5,6 Superconformal Chern-Simons Theories and M2-branes on Orbifolds,” JHEP [0809]{}, 002 (2008) \[arXiv:0806.4977 \[hep-th\]\]. J. Gomis, D. Rodriguez-Gomez, M. Van Raamsdonk and H. Verlinde, “A Massive Study of M2-brane Proposals,” JHEP [0809]{}, 113 (2008) \[arXiv:0807.1074 \[hep-th\]\]. O. Aharony, O. Bergman, D. L. Jafferis and J. Maldacena, “N=6 superconformal Chern-Simons-matter theories, M2-branes and their gravity duals,” JHEP [0810]{}, 091 (2008) \[arXiv:0806.1218 \[hep-th\]\]. H. Lin, O. Lunin and J. M. Maldacena, “Bubbling AdS space and 1/2 BPS geometries,” JHEP [0410]{}, 025 (2004) \[hep-th/0409174\]. D. Jang, Y. Kim, O. K. Kwon and D. D. Tolla, “Exact Holography of the Mass-deformed M2-brane Theory,” Eur. Phys. J. C [77]{}, no. 5, 342 (2017) \[arXiv:1610.01490 \[hep-th\]\]; “Mass-deformed ABJM Theory and LLM Geometries: Exact Holography,” JHEP [1704]{}, 104 (2017) \[arXiv:1612.05066 \[hep-th\]\]. K. Skenderis and M. Taylor, “Kaluza-Klein holography,” JHEP [0605]{}, 057 (2006) \[hep-th/0603016\]. K. Skenderis and M. Taylor, “Holographic Coulomb branch vevs,” JHEP [0608]{}, 001 (2006) \[hep-th/0604169\]. K. Skenderis and M. Taylor, “Anatomy of bubbling solutions,” JHEP [0709]{}, 019 (2007) \[arXiv:0706.0216 \[hep-th\]\]. D. Jang, Y. Kim, O. K. Kwon and D. D. Tolla, “Holography of Massive M2-brane Theory: Non-linear Extension,” arXiv:1803.10660 \[hep-th\]. T. Faulkner, “Bulk Emergence and the RG Flow of Entanglement Entropy,” JHEP [1505]{}, 033 (2015) \[arXiv:1412.5648 \[hep-th\]\]. V. Rosenhaus and M. Smolkin, “Entanglement Entropy: A Perturbative Calculation,” JHEP [1412]{}, 179 (2014) \[arXiv:1403.3733 \[hep-th\]\]. V. Rosenhaus and M. Smolkin, “Entanglement Entropy for Relevant and Geometric Perturbations,” JHEP [1502]{}, 015 (2015) \[arXiv:1410.6530 \[hep-th\]\]. T. Faulkner, R. G. Leigh and O. Parrikar, “Shape Dependence of Entanglement Entropy in Conformal Field Theories,” JHEP [1604]{}, 088 (2016) \[arXiv:1511.05179 \[hep-th\]\]. F. M. Haehl, “Comments on universal properties of entanglement entropy and bulk reconstruction,” JHEP [1510]{}, 159 (2015) \[arXiv:1508.00766 \[hep-th\]\]. A. J. Speranza, “Entanglement entropy of excited states in conformal perturbation theory and the Einstein equation,” JHEP [1604]{}, 105 (2016) \[arXiv:1602.01380 \[hep-th\]\]. M. J. S. Beach, J. Lee, C. Rabideau and M. Van Raamsdonk, “Entanglement entropy from one-point functions in holographic states,” JHEP [1606]{}, 085 (2016) \[arXiv:1604.05308 \[hep-th\]\]. G. Sárosi and T. Ugajin, “Modular Hamiltonians of excited states, OPE blocks and emergent bulk fields,” arXiv:1705.01486 \[hep-th\]. H. Liu and M. Mezei, “A Refinement of entanglement entropy and the number of degrees of freedom,” JHEP [1304]{}, 162 (2013) \[arXiv:1202.2070 \[hep-th\]\]. M. Nozaki, T. Numasawa, A. Prudenziati and T. Takayanagi, “Dynamics of Entanglement Entropy from Einstein Equation,” Phys. Rev. D [88]{}, no. 2, 026012 (2013) \[arXiv:1304.7100 \[hep-th\]\]. J. Lin, M. Marcolli, H. Ooguri and B. Stoica, “Locality of Gravitational Systems from Entanglement of Conformal Field Theories,” Phys. Rev. Lett. [114]{}, 221601 (2015) \[arXiv:1412.1879 \[hep-th\]\]. N. Lashkari and M. Van Raamsdonk, “Canonical Energy is Quantum Fisher Information,” JHEP [1604]{}, 153 (2016) \[arXiv:1508.00897 \[hep-th\]\]. H. Casini, D. A. Galante and R. C. Myers, “Comments on Jacobson’s “entanglement equilibrium and the Einstein equation”,” JHEP [1603]{}, 194 (2016) \[arXiv:1601.00528 \[hep-th\]\]. M. Taylor and W. Woodhead, “Renormalized entanglement entropy,” JHEP [1608]{}, 165 (2016) \[arXiv:1604.06808 \[hep-th\]\]. C. Kim, K. K. Kim and O. K. Kwon, “Holographic Entanglement Entropy of Anisotropic Minimal Surfaces in LLM Geometries,” Phys. Lett. B [759]{}, 395 (2016) \[arXiv:1605.00849 \[hep-th\]\]. K. K. Kim, O. K. Kwon, C. Park and H. Shin, “Renormalized Entanglement Entropy Flow in Mass-deformed ABJM Theory,” Phys. Rev. D [90]{}, no. 4, 046006 (2014) \[arXiv:1404.1044 \[hep-th\]\]; “Holographic entanglement entropy of mass-deformed Aharony-Bergman-Jafferis-Maldacena theory,” Phys. Rev. D [90]{}, no. 12, 126003 (2014) \[arXiv:1407.6511 \[hep-th\]\]. R. Auzzi and S. P. Kumar, “Non-Abelian Vortices at Weak and Strong Coupling in mass-deformed ABJM Theory,” JHEP [0910]{}, 071 (2009) \[arXiv:0906.2366 \[hep-th\]\]. S. Cheon, H. C. Kim and S. Kim, “Holography of mass-deformed M2-branes,” arXiv:1101.1101 \[hep-th\]. M. Cvetic, H. Lu and C. N. Pope, “Consistent Kaluza-Klein sphere reductions,” Phys. Rev. D [62]{}, 064028 (2000) \[hep-th/0003286\]. H. C. Kim and S. Kim, “Supersymmetric vacua of mass-deformed M2-brane theory,” Nucl. Phys. B [839]{}, 96 (2010) \[arXiv:1001.3153 \[hep-th\]\]. D. Bak, K. K. Kim and S. Yun, “Symmetry Breaking Phase Transitions in ABJM Theory with a Finite U(1) Chemical Potential,” Phys. Rev. D [84]{}, 086010 (2011) \[arXiv:1011.5749 \[hep-th\]\]. I. Bena, “The M theory dual of a three-dimensional theory with reduced supersymmetry,” Phys. Rev. D [62]{}, 126006 (2000) \[hep-th/0004142\]. N. Lashkari, C. Rabideau, P. Sabella-Garnier and M. Van Raamsdonk, “Inviolable energy conditions from entanglement inequalities,” JHEP [1506]{}, 067 (2015) \[arXiv:1412.3514 \[hep-th\]\]. D. L. Jafferis, A. Lewkowycz, J. Maldacena and S. J. Suh, “Relative entropy equals bulk relative entropy,” JHEP [1606]{}, 004 (2016) \[arXiv:1512.06431 \[hep-th\]\]. B. Ning and F. L. Lin, “Relative Entropy and Torsion Coupling,” Phys. Rev. D [94]{}, no. 12, 126007 (2016) \[arXiv:1607.00263 \[hep-th\]\]. S. Hollands and R. M. Wald, “Stability of Black Holes and Black Branes,” Commun. Math. Phys. [321]{}, 629 (2013) \[arXiv:1201.0463 \[gr-qc\]\]. S. Hyun, S. A. Park and S. H. Yi, “Canonical energy and hairy AdS black holes,” Phys. Rev. D [94]{}, no. 4, 044014 (2016) \[arXiv:1603.02530 \[hep-th\]\]. S. Banerjee, J. Erdmenger and D. Sarkar, “Connecting Fisher information to bulk entanglement in holography,” arXiv:1701.02319 \[hep-th\]. H. Casini, M. Huerta and R. C. Myers, “Towards a derivation of holographic entanglement entropy,” JHEP [1105]{}, 036 (2011) \[arXiv:1102.0440 \[hep-th\]\]. V. Balasubramanian and P. Kraus, “A Stress tensor for Anti-de Sitter gravity,” Commun. Math. Phys. [208]{}, 413 (1999) \[hep-th/9902121\]. S. de Haro, S. N. Solodukhin and K. Skenderis, “Holographic reconstruction of space-time and renormalization in the AdS / CFT correspondence,” Commun. Math. Phys. [217]{}, 595 (2001) \[hep-th/0002230\]. K. Skenderis, “Asymptotically Anti-de Sitter space-times and their stress energy tensor,” Int. J. Mod. Phys. A [16]{}, 740 (2001) \[hep-th/0010138\]. M. Bianchi, D. Z. Freedman and K. Skenderis, “Holographic renormalization,” Nucl. Phys. B [631]{}, 159 (2002) \[hep-th/0112119\]. R. M. Wald, “Black hole entropy is the Noether charge,” Phys. Rev. D [48]{}, no. 8, R3427 (1993) \[gr-qc/9307038\]; V. Iyer and R. M. Wald, “Some properties of Noether charge and a proposal for dynamical black hole entropy,” Phys. Rev. D [50]{}, 846 (1994) \[gr-qc/9403028\]. [^1]: See also [@Jang:2018aqr] for the extension to the $vevs$ of the CPO with conformal dimension $\Delta=2$. [^2]: The HEE obtained from the 4-dimensional metric agrees with the HEE which was obtained from the 11-dimensional LLM geometries before the KK reduction [@Kim:2016dzw]. See also [@Kim:2014yca]. [^3]: We use a notation in which the quantities in 11-dimensional supergravity are denoted by bold font symbols whereas the normal font symbols are reserved for AdS$_4\times S^7$ values of those quantities. [^4]: For the ${\mathbb Z}_k$ orbifold, we discuss in the subsection \[Zk\]. [^5]: In [@Lin:2004nb], there is also an alternative representation of the LLM solutions in terms of the Young diagrams. In that case, $A$ denotes the area of the Young diagram. [^6]: The gauge/gravity duality dictionary for the scalar mode is following: Under the asymptotic expansion of a scalar field, $$\begin{aligned} \phi(z,x_{i}) = \phi_{1}(x_{i})z +\cdots +\phi_{d-\Delta}(x_{i})z^{d-\Delta} +\cdots +\phi_{\Delta}(x_{i})z^{\Delta} +\cdots,\end{aligned}$$ the $vev$ of the dual gauge invariant operator with conformal dimension $\Delta$ is proportional to $\phi_{\Delta}(x_{i})$ and the $source$ of that operator is proportional to $\phi_{d-\Delta}(x_{i})$. [^7]: If one identifies this auxiliary field with a scalar field in a gravity theory, the gauge/gravity dictionary dictates that $V_{\tilde\lambda}$ corresponds to the $vev$ of a gauge invariant operator with conformal dimension $\Delta$, while $S_{\tilde\lambda}$ corresponds to the source coupled to the operator. [^8]: For the metric in the FG coordinate system in ($d+1$)-dimensional gravity theory, $$\begin{aligned} ds^2 = \frac{L_{\rm AdS}^2}{z^2}\left(dz^2 + g_{ij}(x,z) dx^idx^j\right),\end{aligned}$$ where $L_{\rm AdS}$ is the radius of the AdS$_{d+1}$ geometry and the metric is expanded as $g_{ij}(x,z) = g_{(0)ij}(x) + z^2 g_{(2)ij}(x) +\cdots + z^d g_{(d)ij}(x)+\cdots$ in the asymptotic limit, the $vev$ of the energy-momentum tensor operator is given by [@Balasubramanian:1999re; @deHaro:2000vlm; @Skenderis:2000in; @Bianchi:2001kw] $$\begin{aligned} \label{vevTij} \langle T_{ij}\rangle = \frac{d L_{{\rm AdS}}^{d-1}}{16\pi G_N}\, g_{(d)ij}. \end{aligned}$$	{ "pile_set_name": "ArXiv" }
--- abstract: \| We present a code implementing the linearized self-consistent quasiparticle GW method (scQPGW) in the LAPW basis. Our approach is based on the linearization of the self-energy around zero frequency which differs it from the existing implementations of the scQPGW method. The linearization allows us to use Matsubara’s frequencies instead of real ones. As a result it gives us an advantage in terms of efficiency, allowing us easily switch to the imaginary time representation the same way as in the space time method. The all electron LAPW basis set eliminates the need for pseudopotentials. We discuss the advantages of our approach, such as its $N^{3}$ scaling with the system size, as well as its shortcomings. We apply our approach to study electronic properties of selected semiconductors, insulators, and simple metals and show that our code produces results very close to the previously published scQPGW data. Our implementation is a good platform for further many body diagrammatic resummations such as GW+DMFT. address: - '$^{a}$Department of Physics and Astronomy, Rutgers University, Piscataway, NJ 08856' - '$^{b}$Brookhaven National Laboratory, NY' author: - 'A. L. Kutepov$^{a,}$' - 'V. S. Oudovenko$^{a}$' - 'G. Kotliar$^{a,b}$' title: 'Linearized self-consistent quasiparticle GW method: Application to semiconductors and simple metals' --- [PROGRAM SUMMARY/NEW VERSION PROGRAM SUMMARY]{} [Program Title: LqsgwFlapw]{}\ [Journal Reference:]{}\ [Catalogue identifier:]{}\ [Licensing provisions: GNU General Public License]{}\ [Programming language:]{} Fortran 90\ [Computer:]{} Windows workstations, Linux/UNIX servers/workstations or clusters\ [Operating system: LINUX, MAC OS X, Windows]{}\ [RAM: 2-10]{} gigabytes per CPU (depending on system size).\ [Number of processors used: 1-192]{}\ [Keywords:]{} GW; quasiparticle approximation\ [Classification: 7.3]{}\ [External routines/libraries:]{} BLAS, LAPACK, MPI(optional)\ [Nature of problem:]{}\ Direct implementation of GW method scales as $N^4$ with the system size, which quickly becomes prohibitively time consuming even for the modern computers.\ [Solution method:]{}\ We implemented GW method using the fact that some operations are better to perform in real space, whereas other are more computationally efficient in the reciprocal space. This makes our approach scale as $N^{3}$.\ [Restrictions:]{}\ Limiting factor usually is memory available in a computer. Memory 10 GB/core allows us to study systems up to 15 atoms per unit cell.\ [Running time:]{}\ From 10 minutes up to a few days. (Depending on the system size.)\ Introduction {#intr} ============ The so called GW method was originally proposed by Hedin[@pr_139_A796] and was first applied to real materials by Hybertsen and Louie[@prb_34_5390] and by Godby et al[@prb_37_10159]. There are many successful implementations of this method in open source codes using plane waves basis sets such as BerkeleyGW [@cpc_183_1269], ABINIT [@zk_220_558], and West [@jctc_11_2680]. There are also codes implementating GW in all electron basis sets such as exciting [@jcm_26_363202] and SPEX [@prb_81_125102]. There are numerous computational developments in this area (see for example [@prb_93_115203; @prb_93_125210; @arx_1501_03141] and references therein). For our present study we found particularly useful publication by Rieger et al on the space-time method[@cpc_117_211] and work by Ku and Eguiluz on the application of Matsubara time in GW calculations[@prl_89_126401]. Our main goal in the present work is implementation and testing the self-consistent quasi-particle GW method (scQPGW) which is a promising tool for studying electronic structure of moderately correlated materials, atoms and molecules[@tcc_347_99]. Whereas the current method usually overestimates the widths of spectral features in materials (such as band gaps, band widths, and exchange splitting) it is generally more accurate than the local density approximation (LDA). scQPGW method also has clear advantages in comparison with another popular approximation of GW - one shot GW method (implemented after self-consistent LDA calculation), as not being dependent on the starting point. All previous implementations of scQPGW method[@prb_76_165106; @prl_99_246403; @prb_92_041115] are based on real frequencies. We have found however, that similar results can be obtained with an approach based on imaginary frequency representation. We exploit the fact that we can easily transform functions from imaginary frequency to imaginary time (and back) to enhance the computational efficiency. We use all-electron approximation (Linear Augmented Plane Wave method, LAPW) as a basis of our approach, thus avoiding the need in pseudopotentials. Our implementation of the GW method was outlined in our previous work[@prb_85_155129], where we presented the general scheme of the approach with detailed description of scGW method in its fully relativistic form and the application of the scheme to Am and Pu. Also, we provided there the total energy evaluation which was used earlier in Ref.[@prb_80_041103]. In this work we present the non-relativistic version of scGW method with special emphasis on the scaling of the most time consuming steps of the algorithm. Particularly, we stress on the overall scaling $N^{3}$ with the system size which is similar to the scaling of LDA. In the end we present a few numerical examples obtained using quasiparticle approximation scQPGW. Our implementation of scQPGW approach was used recently in Ref.[@arx_1504_07569] in the combination of scQPGW and one-shot DMFT. The basics of GW method {#GW} ======================= Below we outline the basic formulae of the method introduced earlier in Ref.[@prb_85_155129]. The input for every iteration is the Green function $G$ which is renewed until the self consistency reached. Then we perform a few steps, calculating the intermediate functions such as: polarizability $$\begin{aligned} \label{GW_1} P(12;\tau)=-G(12;\tau)G(21;\beta-\tau),\end{aligned}$$ screened interaction $$\begin{aligned} \label{GW_2} W(12;\nu)=V(12)+\int d(34) V(13)P(34;\nu)W(42;\nu),\end{aligned}$$ self energy $$\begin{aligned} \label{GW_3} \Sigma(12;\tau)=-G(12;\tau)W(21;\tau),\end{aligned}$$ new Green’s function $$\begin{aligned} \label{GW_4} G(12;\omega)&=G_{0}(12;\omega)\nonumber\\&+\int d(34) G_{0}(13;\omega)\Sigma(34;\omega)G(42;\omega).\end{aligned}$$ In scQPGW approach the last step is replaced with a special construction of quasiparticle Green’s function, which was introduced in [@prb_85_155129] $$\begin{aligned} \label{GW_5} G\leftarrow G[G_{0};\Sigma].\end{aligned}$$ The details of this step in our implementation of QPGW approach are given in section \[QP\]. Representation of band states in LAPW method {#LAPW} ============================================ In the LAPW method [@prb_12_3060] one represents the band states in the interstitial region as a linear combination of plane waves $$\begin{aligned} \label{LAPW_1} \Psi^{\alpha\mathbf{k}}_{\lambda}(\mathbf{r})=\frac{1}{\Omega_{0}}\sum_{\mathbf{G}}A^{\alpha\mathbf{k}\lambda}_{\mathbf{G}} e^{i(\mathbf{k}+\mathbf{G})\mathbf{r}},\end{aligned}$$ where $\alpha$ is the spin index, $\mathbf{k}$ is the point in the Brillouin zone, $\mathbf{G}$ labels plane waves, and $\Omega_{0}$ being the volume of the unit cell. Index $\lambda$ stands for the band states, which in this work are the eigen states of an effective Hartree-Fock Hamiltonian constructed with the quasi-particle Greens function [@prb_85_155129]. Inside the MT-sphere at atom $\mathbf{t}$ it is convenient to represent band states as linear combinations of orbitals belonging to that MT-sphere $$\begin{aligned} \label{LAPW_2} \Psi^{\alpha\mathbf{k}}_{\lambda}(\mathbf{r})\|_{\mathbf{t}}=\sum_{L}Z^{\alpha\mathbf{k}\lambda}_{\mathbf{t}L} \varphi^{\alpha\mathbf{t}}_{L}(\mathbf{r}),\end{aligned}$$ where index $L$ combines angular momentum indexes $l,m$ and any additional indexes to distinguish the orbitals inside the sphere (for example, it distinguishes between the solutions of radial equations and their energy derivatives). Representations (\[LAPW\_1\]) and (\[LAPW\_2\]) will be used throughout the paper. Product basis conventions {#PB_def} ========================= We define product basis functions $M^{\mathbf{q}}_{K}(\mathbf{r})$ as plane waves in the interstitial region and optimized basis functions inside MT spheres. We use the index $K$ to label the product basis functions in all MT-spheres and in the interstitial region. When index $K$ runs over the functions inside MT-sphere $\mathbf{t}$ then $$\begin{aligned} \label{PB_1} M^{\mathbf{q}}_{K}(\mathbf{r}+\mathbf{R})=\Big\{\begin{array}{cc} 0 & \mathbf{r}\notin \mathbf{t} \\ e^{i\mathbf{q}\mathbf{R}}M^{\mathbf{t}}_{K}(\mathbf{r}) & \mathbf{r}\in \mathbf{t} \end{array} .\end{aligned}$$ When it runs over the functions in the interstitial region, we associate the index with plane waves $\mathbf{G}_{K}$: $$\begin{aligned} \label{PB_1a} M^{\mathbf{q}}_{K}(\mathbf{r}+\mathbf{R})=\Big\{\begin{array}{cc} e^{i\mathbf{q}\mathbf{R}}e^{i(\mathbf{q}+\mathbf{G}_{K})\mathbf{r}} & \mathbf{r}\in Int \\ 0 & \mathbf{r}\notin Int \end{array} .\end{aligned}$$ The word “optimized” means that we build the space of all products of orbitals $\varphi^{\alpha\mathbf{t}}_{L}$ in each MT-sphere and construct linear independent ortho-normal basis in this space, following the Ref.[@prb_81_125102]. The defined above product basis is not ortho-normal in the interstitial region. So we also define the dual product basis $$\begin{aligned} \label{PB_2} \widetilde{M}^{\mathbf{q}}_{K}(\mathbf{r})=\Big\{\begin{array}{cc} \sum_{K'\in Int}\langle M^{\mathbf{q}}_{K'}\|M^{\mathbf{q}}_{K}\rangle^{-1}M^{\mathbf{q}}_{K'} & \mathbf{r}\in Int \\ M^{\mathbf{q}}_{\mathbf{t}K}(\mathbf{r}) & \mathbf{r}\in \mathbf{t} \end{array} ,\end{aligned}$$ which is ortho-normal to the basis (\[PB\_1\]): $\langle \widetilde{M}^{\mathbf{q}}_{K'}\|M^{\mathbf{q}}_{K}\rangle=\delta_{KK'}$. We expand the polarizability in dual basis $$\begin{aligned} \label{PB_3} P(\mathbf{r};\mathbf{r}';\tau)=\frac{1}{N_{\mathbf{k}}}\sum_{\mathbf{q}}\sum_{KK'}\widetilde{M}^{\mathbf{q}}_{K}(\mathbf{r})P^{\mathbf{q}}_{KK'}(\tau) \widetilde{M}^{^{}\mathbf{q}}_{K'}(\mathbf{r}'),\end{aligned}$$ with $N_{\mathbf{k}}$ being the number of points in the Brillouin zone. Correspondingly to express in product basis the interaction we have to calculate the integral $$\begin{aligned} \label{PB_4} W^{\mathbf{q}}_{KK'}(\tau)=\int\int d\mathbf{r}d\mathbf{r}' \widetilde{M}^{^{}\mathbf{q}}_{K}(\mathbf{r})W(\mathbf{r};\mathbf{r}';\tau) \widetilde{M}^{\mathbf{q}}_{K'}(\mathbf{r}').\end{aligned}$$ Because of the orthogonality of the original and dual product basis sets it is convenient to think of the interaction as an expansion in original product basis set: $$\begin{aligned} \label{PB_5} W(\mathbf{r};\mathbf{r}';\tau)=\frac{1}{N_{\mathbf{k}}}\sum_{\mathbf{q}}\sum_{KK'}M^{\mathbf{q}}_{K}(\mathbf{r})W^{\mathbf{q}}_{KK'}(\tau) M^{^{}\mathbf{q}}_{K'}(\mathbf{r}').\end{aligned}$$ Quasiparticle approximation {#QP} =========================== Different from the QPscGW method by Kotani et al.[prb\_76\_165106]{}, our method is based exclusively on imaginary axis data. We proceed as follows. In Dyson’s equation for the Green function $$\begin{aligned} \label{qp_1} G^{-1}_{\lambda\lambda^{\prime }}(\mathbf{k};\omega)=(i\omega+\mu-\varepsilon^{\mathbf{k}}_{\lambda})\delta_{\lambda\lambda^{\prime }} -\Sigma^{c}_{\lambda\lambda^{\prime }}(\mathbf{k};\omega),\end{aligned}$$ where band indices $(\lambda ,\lambda ^{\prime })$ correspond to the effective exchange Hamiltonian[@prb_85_155129], we approximate frequency dependence of the self energy by a linear function $$\begin{aligned} \label{qp_2} \Sigma^{c}_{\lambda\lambda^{\prime }}(\mathbf{k};\omega)=\Sigma^{c}_{\lambda\lambda^{\prime }}(\mathbf{k};\omega=0) +\frac{\partial\Sigma^{c}_{\lambda\lambda^{\prime }}(\mathbf{k};\omega)}{\partial (i\omega)}\|_{\omega=0}(i\omega).\end{aligned}$$ With this approximation the Dyson equation is simplified $$\begin{aligned} \label{qp_3} G^{-1}_{\lambda\lambda^{\prime }}(\mathbf{k};\omega)=Z^{-1}_{\lambda\lambda^{\prime }}(\mathbf{k})(i\omega)+ (\mu-\varepsilon^{\mathbf{k}}_{\lambda})\delta_{\lambda\lambda^{\prime }} -\Sigma^{c}_{\lambda\lambda^{\prime }}(\mathbf{k};0),\end{aligned}$$ where we have introduced a renormalization factor $Z$ matrix (not to be confused with the expansion coefficients in Eq. (\[LAPW\_2\])): $$\begin{aligned} \label{qp_4} Z^{-1}_{\lambda\lambda^{\prime }}(\mathbf{k})=\delta_{\lambda\lambda^{\prime }}-\frac{\partial\Sigma^{c}_{\lambda\lambda^{\prime }}(\mathbf{k};\omega)}{\partial (i\omega)}\|_{\omega=0}.\end{aligned}$$ Representing $Z$-factor as a symmetrical product $$\begin{aligned} \label{qp_5} Z^{-1}_{\lambda\lambda^{\prime }}(\mathbf{k})=\sum_{\lambda^{\prime \prime }}Z^{-1/2}_{\lambda\lambda^{\prime \prime }}(\mathbf{k})Z^{-1/2}_{\lambda^{\prime \prime }\lambda^{\prime }}(\mathbf{k}),\end{aligned}$$ we reduce the Dyson equation to the following form $$\begin{aligned} \label{qp_6} &\sum_{\lambda^{\prime \prime }\lambda^{\prime \prime \prime }}Z^{1/2}_{\lambda\lambda^{\prime \prime }}(\mathbf{k}) G^{-1}_{\lambda^{\prime \prime }\lambda^{\prime \prime \prime }}(\mathbf{k};\omega)Z^{1/2}_{\lambda^{\prime \prime \prime }\lambda^{\prime }}(\mathbf{k})= i\omega\delta_{\lambda\lambda^{\prime }}+ \notag \\ &\sum_{\lambda^{\prime \prime }\lambda^{\prime \prime \prime }}Z^{1/2}_{\lambda\lambda^{\prime \prime }}(\mathbf{k}) [(\mu-\varepsilon^{\mathbf{k}}_{\lambda^{\prime \prime }})\delta_{\lambda^{\prime \prime }\lambda^{\prime \prime \prime }} -\Sigma^{c}_{\lambda^{\prime \prime }\lambda^{\prime \prime \prime }}(\mathbf{k};0)]Z^{1/2}_{\lambda^{\prime \prime \prime }\lambda^{\prime }}(\mathbf{k}).\end{aligned}$$ The second term on the right hand side of this equation is a Hermitian matrix, the quasiparticle Hamiltonian matrix. It is diagonalized in subroutine `BANDS_QP`. $$\begin{aligned} \label{qp_7} &\mu\delta_{\lambda\lambda^{\prime }}-H^{\mathbf{k}}_{\lambda\lambda^{\prime }} \notag \\ &=\sum_{\lambda^{\prime \prime }\lambda^{\prime \prime \prime }}Z^{1/2}_{\lambda\lambda^{\prime \prime }}(\mathbf{k}) [(\mu-\varepsilon^{\mathbf{k}}_{\lambda^{\prime \prime }})\delta_{\lambda^{\prime \prime }\lambda^{\prime \prime \prime }} -\Sigma^{c}_{\lambda^{\prime \prime }\lambda^{\prime \prime \prime }}(\mathbf{k};0)]Z^{1/2}_{\lambda^{\prime \prime \prime }\lambda^{\prime }}(\mathbf{k}) \notag \\ &=\sum_{i}Q^{\mathbf{k}}_{\lambda i}E^{\mathbf{k}}_{i}Q^{^{\dagger}\mathbf{k}}_{i\lambda^{\prime }},\end{aligned}$$ where $E_{i}^{\mathbf{k}}$ are the effective eigenvalues. After that, we can rewrite (\[qp\_6\]) as follows $$\begin{aligned} \label{qp_8} \sum_{\lambda^{\prime \prime }\lambda^{\prime \prime \prime }}Z^{1/2}_{\lambda\lambda^{\prime \prime }}(\mathbf{k}) &G^{-1}_{\lambda^{\prime \prime }\lambda^{\prime \prime \prime }}(\mathbf{k};\omega)Z^{1/2}_{\lambda^{\prime \prime \prime }\lambda^{\prime }}(\mathbf{k}) \notag \\ &= \sum_{i}Q^{\mathbf{k}}_{\lambda i}\big[i\omega+\mu-E^{\mathbf{k}}_{i}\big]Q^{^{\dagger}\mathbf{k}}_{i\lambda^{\prime }},\end{aligned}$$ or, for the Green function $$\begin{aligned} \label{qp_9} G^{\mathbf{k}}_{\lambda\lambda^{\prime }}(\omega)= \sum_{i}\frac{(Z^{1/2}Q)^{\mathbf{k}}_{\lambda i}(Q^{\dagger}Z^{1/2})^{\mathbf{k}}_{i\lambda^{\prime }}}{i\omega+\mu-E^{\mathbf{k}}_{i}}.\end{aligned}$$ This expression differs from the full GW Greens function by a linear approximation for the frequency dependent self energy. At this point, we construct the quasiparticle Greens funcion (step (\[GW\_5\]) in section \[GW\]) by setting $Z_{\lambda \lambda ^{\prime }}^{\mathbf{k}}=\delta _{\lambda \lambda ^{\prime }}$ in the above equation to obtain $$\begin{aligned} \label{qp_10} G^{\mathbf{k}}_{\lambda\lambda^{\prime }}(\omega)= \sum_{i}\frac{Q^{\mathbf{k}}_{\lambda i}Q^{^{\dagger}\mathbf{k}}_{i\lambda^{\prime }}}{i\omega+\mu-E^{\mathbf{k}}_{i}}.\end{aligned}$$ Polarizability calculation and scaling {#P_R} ====================================== In accordance with the MT-geometry there are three essentially different contributions to the polarizability corresponding to i)when both space arguments of $P$ belong to MT spheres (Mt-Mt); ii) one of them belongs to a MT-sphere and another belongs to the Interstitial region (Mt-Int); iii) both arguments belong to the interstitial region (Int-Int). Below we consider three cases separately, describing how the Green function is transformed from the band representation to the real space, how we calculate the polarizability, and how we transform it from the real space to the reciprocal space. Our parallelization strategy here is to use two-dimensional grid of MPI-processes. The first MPI-dimension in polarizability calculations is associated with $\tau$-variable with each process doing calculation only on its own set of $\tau$-indexes. It is most efficient because all formulae in this section are totally independent for different $\tau$’s. The second dimension of MPI grid is used whenever it is appropriate as described briefly below. Namely, every process associated with the second dimension of the MPI grid is carrying out calculations on its own set of $\mathbf{k}$ points, or on its own set of triplets ($\mathbf{Rtt}'$). In the following sections we will present the scalings associated with the principle steps of the algorithm. For convenience we summarize main notations here: $N_{at}$ is the number of atoms in the unit cell; $N_{orb}$ is the number of orbitals per atom in the LAPW+LO representation (for typical numbers see Table \[Time\] below). The number of bands is approximately equal to $N_{at}N_{orb}$, so we will not use the number of bands below. Further, the number of plane waves in the interstitial region used to represent the fermionic functions approximately equals the number of bands, so we do not use it below as well. Bosonic functions make the major impact on the calculation time. So, it is practical to take into account their numbers more carefully. $N^{Mt}_{pb}$ is the number of product basis orbitals inside MT-sphere (per atom); $N^{Int}_{pb}$ is the number of plane waves associated with product basis in the interstitial region (per atom); $N_{r}$ is the number of points in the regular real space mesh in the unit cell (per atom); $N_{k}$ is the number of points in the whole Brillouin zone; $N_{\tau}$ is the number of points in $\tau$-mesh. The number of points in the fermionic and bosonic frequency meshes is about the same as the number of $\tau$-points, so we use the latter in all cases. Mt-Mt part of polarizability {#P_MM} ---------------------------- When both space arguments belong to MT-spheres, real space representation means that we represent $G$ as an expansion in local orbitals inside the spheres at $\mathbf{t}$ and $\mathbf{t}'$ in the unit cells separated by translation vector $\mathbf{R}$ $$\begin{aligned} \label{GR_1} G^{\alpha}(\mathbf{r};\mathbf{r}';\tau)\|_{\mathbf{t}+\mathbf{R};\mathbf{t}'}=\sum_{LL'}\varphi^{\alpha\mathbf{t}}_{L}(\mathbf{r}) G_{\mathbf{t}L;\mathbf{t}'L'}^{\alpha\mathbf{R}}(\tau )\varphi^{\alpha\mathbf{t}'}_{L'}(\mathbf{r}'),\end{aligned}$$ with the coefficients found with (\[LAPW\_2\]): $$\begin{aligned} \label{GR_2} G_{\mathbf{t}L;\mathbf{t}'L'}^{\alpha\mathbf{R}}(\tau )=\frac{1}{N_{ \mathbf{k}}}\sum_{\mathbf{k}}e^{i\mathbf{k}\mathbf{R}} \sum_{\lambda \lambda'}Z_{\mathbf{t}L}^{\alpha\mathbf{k}\lambda}G_{\lambda \lambda'}^{\alpha\mathbf{k}}(\tau)Z_{\mathbf{t}'L'}^{^{}\alpha\mathbf{k}\lambda'}.\end{aligned}$$ So the first step in Mt-Mt case is to transform Green’s function from band representation to the representation (\[GR\_1\]) using (\[GR\_2\]). The scaling associated with the evaluation of (\[GR\_2\]) is $\left[(N_{at}N_{orb})^{3}N_{k}+(N_{at}N_{orb})^{2}N_{k}\ln N_{k}\right]N_{\tau}$. The first term corresponds to the sum over ($\lambda,\lambda'$) indexes which scales as $(N_{at}N_{orb})^{3}$ for every $\mathbf{k}$-point and $\tau$. The second term is related to the fast Fourier transform from $\mathbf{k}$ space to the $\mathbf{R}$ space which scales as $N_{k}\ln N_{k}$ for each matrix element and $\tau$. We use second dimension of MPI-grid to calculate matrix products (sums over band indexes) spreading different $\mathbf{k}$’s over the MPI processes. Then we switch MPI parallelization to perform FFT for different indexes ($\mathbf{t}'L'$). In the code the above Green function transformation is performed in `G_RS_FROM_KS_MM` subroutine. The expression for the polarizability then follows from (\[GR\_1\]) and (\[GW\_1\]) $$\begin{aligned} \label{P_MM_1} P_{\mathbf{t}K;\mathbf{t}'K'}^{\mathbf{R}}(\tau)=&-\sum_{\alpha} \sum_{LL''}\langle M_{K}^{\mathbf{t}}\|\varphi_{L}^{\alpha\mathbf{t}}\varphi _{L''}^{\alpha\mathbf{t}}\rangle \nonumber \\ & \times \sum_{L'}G_{\mathbf{t}L;\mathbf{t}'L'}^{\alpha\mathbf{R}}(\tau )\sum_{L'''} G_{\mathbf{t}L'';\mathbf{t}'L'''}^{\alpha;\mathbf{R}}(\beta -\tau) \nonumber \\ &\times \langle \varphi_{L'}^{\alpha\mathbf{t}'}\varphi_{L'''}^{\alpha\mathbf{t}'}\|M_{K'}^{\mathbf{t}'}\rangle ,\end{aligned}$$ The scaling of (\[P\_MM\_1\]) is $\left[2N_{orb}^{3}N^{Mt}_{pb}+(N_{orb}N_{pb}^{Mt})^{2}\right]N_{at}^{2}N_{k}N_{\tau}$. To evaluate (\[P\_MM\_1\]) we use second MPI dimension to parallelize the triplets ($\mathbf{R};\mathbf{t};\mathbf{t}'$). In the code (\[P\_MM\_1\]) is implemented in `P_MM_R` subroutine. Transform to the reciprocal space consists in one FFT transform. $$\begin{aligned} \label{P_TR_1} P_{\mathbf{t}K;\mathbf{t}'K'}^{\mathbf{q}}(\tau)=\sum_{\mathbf{R}}e^{-i\mathbf{q}\mathbf{R}}P_{\mathbf{t}K;\mathbf{t}'K'}^{\mathbf{R}}(\tau),\end{aligned}$$ which is implemented in the subroutine `P_MM_Q_FROM_R`. The scaling of (\[P\_TR\_1\]) is $(N_{pb}^{Mt}N_{at})^{2}N_{k}\ln N_{k}N_{\tau}$. Mt-Int part of polarizability {#P_MI} ----------------------------- In this case the second space argument in Eq.(\[GR\_3\]) runs over the regular $\mathbf{r}$-mesh in the whole unit cell, whereas for the first space argument we use an expansion in local orbitals: $$\begin{aligned} \label{GR_3} G^{\alpha}(\mathbf{r};\mathbf{r}';\tau)\|_{\mathbf{r}\in\mathbf{t}+\mathbf{R}}=\sum_{L}\varphi^{\alpha\mathbf{t}}_{L}(\mathbf{r}) G_{\mathbf{t}L;\mathbf{r}'}^{\alpha\mathbf{R}}(\tau ).\end{aligned}$$ The corresponding coefficients $G_{\mathbf{t}L;\mathbf{r}'}^{\alpha\mathbf{R}}(\tau)$ are obtained in two steps (subroutine `G_RS_FROM_KS_MI`): $$\begin{aligned} \label{GR_5} G_{\mathbf{t}L;\mathbf{G}'}^{\alpha\mathbf{k}}(\tau)=\frac{1}{\sqrt{\Omega_{0}}}\sum_{\lambda \lambda'} Z_{\mathbf{t}L}^{\alpha\mathbf{k}\lambda}G_{\lambda \lambda'}^{\alpha\mathbf{k}}(\tau) A_{\mathbf{G}'}^{^{}\alpha\mathbf{k}\lambda'},\end{aligned}$$ and $$\begin{aligned} \label{GR_6} G^{\alpha\mathbf{R}}_{\mathbf{t}L;\mathbf{r}'}(\tau )=\frac{1}{N_{\mathbf{k}}}\sum_{\mathbf{k}}e^{i\mathbf{k}\mathbf{R}} \sum_{\mathbf{G}'}e^{-i(\mathbf{k}+\mathbf{G}')\mathbf{r}'} G_{\mathbf{t}L;\mathbf{G}'}^{\alpha\mathbf{k}}(\tau ).\end{aligned}$$ The scaling of (\[GR\_5\]) is $N_{at}^{3}N_{orb}^{3}N_{k}N_{\tau}$. The scaling associated with the evaluation of (\[GR\_6\]) is $N_{at}^{2}N_{orb}N_{r}N_{k}N_{\tau}\left[\ln(N_{at}N_{r})+\ln N_{k}\right]$. MPI-parallelization is used in (\[GR\_5\]) and (\[GR\_6\]) to perform calculations for different $\mathbf{k}$’s independently. The expression for the polarizability follows from (\[GR\_3\]) and (\[GW\_1\]) $$\begin{aligned} \label{PR_2} P^{\mathbf{R}}_{\mathbf{t}K;\mathbf{r}'}(\tau)=-\sum_{\alpha}\sum_{LL'} \langle M^{\mathbf{t}}_{K}\|\varphi^{\alpha\mathbf{t}}_{L}\varphi^{\mathbf{t}}_{L'}\rangle G^{\alpha\mathbf{R}}_{\mathbf{t}L;\mathbf{r}'}(\tau)G^{\alpha\mathbf{R}}_{\mathbf{t}L';\mathbf{r}'}(\beta-\tau).\end{aligned}$$ The scaling of (\[PR\_2\]) is $(N_{at}N_{orb})^{2}N^{Mt}_{pb}N_{r}N_{k}N_{\tau}$. MPI-parallelization is used in (\[PR\_2\]) to perform the calculations for different $\mathbf{R}$’s independently (subroutine `P_IM_R`). The reciprocal space representation in original product basis is obtained with two FFTs: $$\begin{aligned} \label{P_TR_2} \widetilde{P}^{\mathbf{q}}_{\mathbf{t}K;\mathbf{G}'}(\tau)=\frac{1}{N_{\mathbf{r}}}\sum_{\mathbf{r}'}e^{i(\mathbf{q}+\mathbf{G}')\mathbf{r}'} \sum_{\mathbf{R}}e^{-i\mathbf{q}\mathbf{R}}P_{\mathbf{t}K;\mathbf{r}'}^{\mathbf{R}}(\tau).\end{aligned}$$ The scaling of (\[P\_TR\_2\]) is $N_{at}^{2}N_{k}N_{\tau}N^{Mt}_{pb}N_{r}\left[\ln N_{k}+\ln (N_{at}N_{r})\right]$. Representation in the dual basis is obtained after an additional step $$\begin{aligned} \label{P_TR_2A} P^{\mathbf{q}}_{\mathbf{t}K;K'}(\tau)=\sum_{\mathbf{G}'}\widetilde{P}^{\mathbf{q}}_{\mathbf{t}K;\mathbf{G}'}(\tau)\langle e^{i(\mathbf{q}+\mathbf{G}')\mathbf{r}'}\|M^{\mathbf{q}}_{K'}\rangle_{Int},\end{aligned}$$ where $\langle e^{i(\mathbf{q}+\mathbf{G}')\mathbf{r}'}\|M^{\mathbf{q}}_{K'}\rangle_{Int}$ represents the integral of the product of two plane waves over the interstitial region and it is done analytically. The scaling of (\[P\_TR\_2A\]) is $N_{at}^{3}N_{k}N_{\tau}N^{Mt}_{pb}N_{pb}^{^{2}Int}$. MPI-parallelization is used in (\[P\_TR\_2\]) and (\[P\_TR\_2A\]) to perform calculations for different $\mathbf{q}$’s independently (subroutine `P_IM_Q_FROM_R`). Int-Int part of polarizability {#P_II} ------------------------------ In this case both space arguments run over the regular mesh in the unit cell. The real space representation for $G$ is obtained in two steps (subroutines `G_K_G_R1` and `G_RR_R_R1_STAR`): $$\begin{aligned} \label{GR_3A} G_{\mathbf{G};\mathbf{G}'}^{\alpha\mathbf{k}}(\tau )=\frac{1}{\Omega_{0}}\sum_{\lambda \lambda'} A_{\mathbf{G}}^{\alpha\mathbf{k}\lambda}G_{\lambda \lambda'}^{\alpha\mathbf{k}}(\tau)A_{\mathbf{G}'}^{^{}\alpha\mathbf{k}\lambda'},\end{aligned}$$ and $$\begin{aligned} \label{GR_4} G^{\alpha\mathbf{R}}_{\mathbf{r};\mathbf{r}'}(\tau )= \frac{1}{N_{\mathbf{k}}}\sum_{\mathbf{k}}e^{i\mathbf{k}\mathbf{R}}\sum_{\mathbf{G};\mathbf{G}'}e^{i(\mathbf{k}+\mathbf{G})\mathbf{r}} G_{\mathbf{G};\mathbf{G}'}^{\alpha\mathbf{k}}(\tau )e^{-i(\mathbf{k}+\mathbf{G}')\mathbf{r}'}.\end{aligned}$$ The scalings of (\[GR\_3A\]) and (\[GR\_4\]) are $(N_{at}N_{orb})^{3}N_{k}N_{\tau}$ and $(N_{at}N_{r})^{2}N_{k}N_{\tau}\left[2\ln (N_{at}N_{r})+\ln N_{k}\right]$ correspondingly. MPI-parallelization is used in (\[GR\_3A\]) and (\[GR\_4\]) to perform calculations for different $\mathbf{k}$’s independently. The formula for the polarizability is very simple in this case $$\begin{aligned} \label{PR_3} P^{\mathbf{R}}_{\mathbf{r}\mathbf{r}'}(\tau)=-\sum_{\alpha}G^{\alpha\mathbf{R}}_{\mathbf{r}\mathbf{r}'}(\tau) G^{\alpha\mathbf{R}}_{\mathbf{r}\mathbf{r}'}(\beta-\tau).\end{aligned}$$ The scaling of (\[PR\_3\]) is $(N_{at}N_{r})^{2}N_{k}N_{\tau}$. In (\[PR\_3\]) we use MPI processes associated with index $\mathbf{r}'$ and $\tau$. The reciprocal space representation in the original product basis is obtained with three FFTs: $$\begin{aligned} \label{P_TR_3} \widetilde{P}^{\mathbf{q}}_{\mathbf{G}\mathbf{G}'}(\tau)&=\frac{1}{N_{\mathbf{r}}}\sum_{\mathbf{r}}e^{i(\mathbf{q}+\mathbf{G})\mathbf{r}} \frac{1}{N_{\mathbf{r}}}\sum_{\mathbf{r}'}e^{-i(\mathbf{q}+\mathbf{G}')\mathbf{r}'} \nonumber \\ &\times\sum_{\mathbf{R}}e^{-i\mathbf{q}\mathbf{R}}P^{\mathbf{R}}_{\mathbf{r};\mathbf{r}'}(\tau).\end{aligned}$$ The scalings of Eq. (\[P\_TR\_3\]) is $N_{at}^{2}N_{r}N_{k}N_{\tau}\left[(N_{r}+N_{pb}^{Int})\ln (N_{at}N_{r})+N_{r}\ln N_{k}\right]$. Representation in the dual basis follows as additional matrix multiplications $$\begin{aligned} \label{P_TR_3A} P^{\mathbf{q}}_{K;K'}(\tau)=\sum_{\mathbf{G}\mathbf{G}'}\langle e^{i(\mathbf{q}+\mathbf{G})\mathbf{r}}\|M^{\mathbf{q}}_{K}\rangle^{}_{Int}\widetilde{P}^{\mathbf{q}}_{\mathbf{G}\mathbf{G}'}(\tau)\langle e^{i(\mathbf{q}+\mathbf{G}')\mathbf{r}'}\|M^{\mathbf{q}}_{K'}\rangle_{Int}.\end{aligned}$$ The scaling of (\[P\_TR\_3A\]) is $(N_{at}N_{pb}^{Int})^{3}N_{k}N_{\tau}$. MPI-parallelization is used in (\[P\_TR\_3\]) and (\[P\_TR\_3A\]) to perform calculations for different $\mathbf{q}$’s independently. Formulae (\[PR\_3\])-(\[P\_TR\_3A\]) are implemented in the subroutine `P_II_SOLID`. Screened interaction {#W_bas} ==================== Equation (\[GW\_2\]) in reciprocal space reads as the following $$\begin{aligned} \label{W_1} W^{\mathbf{q}}_{KK'}(\nu)=V^{\mathbf{q}}_{KK'}+\sum_{K''K'''} V^{\mathbf{q}}_{KK''}P^{\mathbf{q}}_{K''K'''}(\nu)W^{\mathbf{q}}_{K'''K'}(\nu).\end{aligned}$$ The scaling of (\[W\_1\]) is $(N_{at}[N_{pb}^{Mt}+N_{pb}^{Int}])^{3}N_{k}N_{\tau}$. We associate two-dimensional mesh of MPI-processes with variables $\mathbf{q}$ and $\nu$. Formula (\[W\_1\]) is implemented in the subroutine `WS_K_NU_SOLID_0`. Dynamic self energy {#sigma} =================== According to the division of the screened interaction into bare Coulomb V and dynamic part $\widetilde{W}$ ($W=V+\widetilde{W}$) the self energy is also divided into static and dynamic. Here we consider the evaluation of the dynamic part only. Static part is evaluated similarly with obvious simplifications in the formulae. In accordance with MT-geometry there are three essentially different contributions to the self energy corresponding to i)when both space arguments of belong to MT spheres (Mt-Mt); ii) one of them belongs to a MT-sphere and another belongs to the Interstitial (Mt-Int); iii) both arguments belong to the interstitial region (Int-Int). Below we consider three cases separately, describing how the screened interaction is transformed from reciprocal space to the real space, how we calculate the dynamic self energy, and how we transform it from real space back to reciprocal space and band representation. Our parallelization strategy here is similar to the strategy in polarizability calculations. Mt-Mt part of self energy {#S_MM} ------------------------- When both space arguments belong to MT-spheres, real space representation means that we represent $\widetilde{W}$ as an expansion in product basis functions inside the spheres at $\mathbf{t}$ and $\mathbf{t}'$ in the unit cells separated by translation vector $\mathbf{R}$ $$\begin{aligned} \label{S_MM_1} \widetilde{W}_{\mathbf{t}K;\mathbf{t}'K'}^{\mathbf{R}}(\tau)=\frac{1}{N_{\mathbf{k}}}\sum_{\mathbf{q}}e^{i\mathbf{q}\mathbf{R}} \widetilde{W}_{\mathbf{t}K;\mathbf{t}'K'}^{\mathbf{q}}(\tau).\end{aligned}$$ We use the second dimension of MPI-grid to calculate matrix products (sums over band indexes) spreading different $\mathbf{q}$’s over the MPI processes. Formula (\[S\_MM\_1\]) is implemented in the subroutine `W_MM`. The scaling of (\[S\_MM\_1\]) is $(N_{at}N_{pb}^{Mt})^{2}N_{\tau}N_{k}\ln N_{k}$. The expression for the self energy follows from (\[GR\_1\]) and (\[GW\_3\]) (subroutine `SIGC_MM_R`) $$\begin{aligned} \label{S_MM_2} \Sigma_{\mathbf{t}L;\mathbf{t}'L'}^{\alpha\mathbf{R}}(\tau)=&-\sum_{L''L'''}\sum_{KK'}\langle \varphi_{L}^{\alpha\mathbf{t}}\|\varphi _{L''}^{\alpha\mathbf{t}}M_{K}^{\mathbf{t}}\rangle \nonumber \\ & \times G_{\mathbf{t}L;\mathbf{t}'L'}^{\alpha\mathbf{R}}(\tau ) \widetilde{W}_{\mathbf{t}K;\mathbf{t}'K'}^{\mathbf{R}}(\beta-\tau) \nonumber \\ &\times \langle \varphi_{L'}^{\alpha\mathbf{t}'}\|\varphi_{L'''}^{\alpha\mathbf{t}'}M_{K'}^{\mathbf{t}'}\rangle ,\end{aligned}$$ The scaling of (\[S\_MM\_2\]) is $\left[2N_{orb}^{3}N^{Mt}_{pb}+(N_{orb}N_{pb}^{Mt})^{2}\right]N_{at}^{2}N_{k}N_{\tau}$. To evaluate (\[S\_MM\_2\]) we use second MPI dimension to parallelize the triplets ($\mathbf{R};\mathbf{t};\mathbf{t}'$). Transform to the reciprocal space consists in one FFT transform (subroutine `SIGC_MM_K_FROM_R`). $$\begin{aligned} \label{S_TR_1} \Sigma_{\mathbf{t}L;\mathbf{t}'L'}^{\mathbf{k}}(\tau)=\sum_{\mathbf{R}}e^{-i\mathbf{k}\mathbf{R}}\Sigma_{\mathbf{t}L;\mathbf{t}'L'}^{\mathbf{R}}(\tau),\end{aligned}$$ which scales as $(N_{at}N_{orb})^{2}N_{\tau}N_{k}\ln N_{k}$. Mt-Int part of self energy {#S_MI} -------------------------- In this case the transform of $W$ to real space involves two FFT’s (subroutine `V_IM_R_FROM_K`): $$\begin{aligned} \label{W_RT_2} \widetilde{W}^{\mathbf{R}}_{\mathbf{t}K;\mathbf{r}}(\tau)= \frac{1}{N_{\mathbf{k}}}\sum_{\mathbf{q}}e^{i\mathbf{q}\mathbf{R}} \sum_{\mathbf{G}}e^{-i(\mathbf{q}+\mathbf{G})\mathbf{r}}\widetilde{W}_{\mathbf{t}K;\mathbf{G}}^{\mathbf{q}}(\tau).\end{aligned}$$ The scaling of (\[W\_RT\_2\]) is $N_{at}^{2}\ln (N_{at}N_{r}) N_{pb}^{Mt}N_{r}N_{k}N_{\tau}$. MPI-parallelization is used in (\[W\_RT\_2\]) to perform calculations for different $\mathbf{q}$’s independently. The expression for the self energy follows from (\[GR\_3\]) and (\[GW\_3\]) $$\begin{aligned} \label{S_MI_2} \Sigma^{\alpha\mathbf{R}}_{\mathbf{t}L;\mathbf{r}'}(\tau)=-\sum_{L'K} \langle \varphi^{\alpha\mathbf{t}}_{L}\|\varphi^{\mathbf{t}}_{L'}M^{\mathbf{t}}_{K}\rangle G^{\alpha\mathbf{R}}_{\mathbf{t}L;\mathbf{r}'}(\tau)\widetilde{W}^{\mathbf{R}}_{\mathbf{t}K;\mathbf{r}'}(\beta-\tau),\end{aligned}$$ which scales as $N_{at}^{2}N_{orb}^{2}N_{pb}^{Mt}N_{r}N_{k}N_{\tau}$. MPI-parallelization is used in (\[S\_MI\_2\]) to perform calculations for different $\mathbf{R}$’s independently. Formula (\[S\_MI\_2\]) is implemented in the subroutine `SIGC_IM_R`. Transformation to the band states representation is achieved in a few steps. They are implemented in the subroutine `SIGC_IM_K_FROM_R`. First we apply FFT $$\label{S_MI_3} \Sigma^{\alpha\mathbf{k}}_{\mathbf{t}L;\mathbf{r}'}(\tau)=\sum_{\mathbf{R}}e^{-i\mathbf{k}\mathbf{R}} \Sigma^{\alpha\mathbf{R}}_{\mathbf{t}L;\mathbf{r}'}(\tau),$$ with scaling $N_{at}^{2}N_{orb}N_{r}N_{\tau}N_{k}\ln N_{k}$. At this point the function is represented by its values at the homogeneous $\mathbf{r}'$-mesh in the whole unit cell. In order to perform integration over the interstitial region we again apply FFT to transform it into equivalent linear combination of plane waves $$\begin{aligned} \label{S_MI_4} \Sigma^{\alpha\mathbf{k}}_{\mathbf{t}L;\mathbf{r}'}(\tau)=\sum_{\mathbf{G}'} \tilde{\Sigma}^{\alpha\mathbf{k}}_{\mathbf{t}L;\mathbf{G}'} e^{-i(\mathbf{k}+\mathbf{G}')\mathbf{r}'},\end{aligned}$$ with the coefficients $$\begin{aligned} \label{S_MI_5} \tilde{\Sigma}^{\alpha\mathbf{k}}_{\mathbf{t}L;\mathbf{G}'}(\tau)=\frac{1}{N_{\mathbf{r}}}\sum_{\mathbf{r}'} e^{i(\mathbf{k}+\mathbf{G}')\mathbf{r}'}\Sigma^{\alpha\mathbf{k}}_{\mathbf{t}L;\mathbf{r}'}(\tau).\end{aligned}$$ The scaling of (\[S\_MI\_5\]) is $N_{at}^{2}N_{orb}N_{r}N_{k}N_{\tau}\ln (N_{at}N_{r})$. The form (\[S\_MI\_4\]) allows us to integrate over the interstitial region analytically and we obtain $$\begin{aligned} \label{S_MI_6} \Sigma^{\alpha\mathbf{k}}_{\mathbf{t}L;\mathbf{G}'}(\tau)=\frac{1}{\sqrt{\Omega_{0}}} \sum_{\mathbf{G}''}\tilde{\Sigma}^{\alpha\mathbf{k}}_{\mathbf{t}L;\mathbf{G}''}(\tau)S^{\mathbf{k}}_{\mathbf{G}''\mathbf{G}'},\end{aligned}$$ which scales as $N_{at}^{3}N_{orb}^{2}N_{pb}^{Int}N_{k}N_{\tau}$. $S^{\mathbf{k}}_{\mathbf{G}''\mathbf{G}'}$ in (\[S\_MI\_6\]) is the integral of the product of two plane waves ($e^{-i(\mathbf{k}+\mathbf{G}'')\mathbf{r}}$ and $e^{i(\mathbf{k}+\mathbf{G}''')\mathbf{r}}$) taken over the interstitial region. Finally, the contribution to the band state representation follows $$\label{S_MI_7} \Sigma^{\alpha\mathbf{k}}_{\lambda\lambda'}(\tau)\|^{Mt}_{Int}=\sum_{\mathbf{t}L}\sum_{\mathbf{G}'} Z^{^{}\alpha\mathbf{k}\lambda}_{\mathbf{t}L} \Sigma^{\alpha\mathbf{k}}_{\mathbf{t}L;\mathbf{G}'}(\tau)A^{\alpha\mathbf{k}\lambda'}_{\mathbf{G}'}+H.C..$$ The scaling of (\[S\_MI\_7\]) is $(N_{a}N_{orb})^{3}N_{k}N_{\tau}$. MPI-parallelization is used in (\[S\_MI\_3\]-\[S\_MI\_7\]) to perform calculations for different $\mathbf{k}$’s independently. Int-Int part of self energy {#S_II} --------------------------- In this case both space arguments run over the regular mesh in the unit cell. Real space representation for $W$ is obtained with three FFT’s (subroutines `W_Q_G_R1` and `W_RR_R_R1_STAR`): $$\begin{aligned} \label{S_II_1} \widetilde{W}^{\mathbf{R}}_{\mathbf{r};\mathbf{r}'}(\tau)&=\frac{1}{N_{\mathbf{k}}}\sum_{\mathbf{q}}e^{i\mathbf{q}\mathbf{R}} \sum_{\mathbf{G}} e^{-i(\mathbf{q}+\mathbf{G})\mathbf{r}} \sum_{\mathbf{G}'}e^{i(\mathbf{q}+\mathbf{G}')\mathbf{r}'} \nonumber \\ &\times \widetilde{W}^{\mathbf{q}}_{\mathbf{G}\mathbf{G}'}(\tau),\end{aligned}$$ with scaling $N_{at}^{2}N_{r}N_{k}N_{\tau}\nonumber\\ \times\left[N_{pb}^{Int}\ln (N_{at}N_{r})+N_{r}\ln (N_{at}N_{r}) + N_{r}\ln N_{k}\right]$. MPI-parallelization is used in (\[S\_II\_1\]) to perform calculations for different $\mathbf{k}$’s independently. The formula for the self energy is very simple in this case $$\label{S_II_2} \Sigma^{\alpha\mathbf{R}}_{\mathbf{r}\mathbf{r}'}=-G^{\alpha\mathbf{R}}_{\mathbf{r}\mathbf{r}'}(\tau ) \widetilde{W}^{\mathbf{R}}_{\mathbf{r}\mathbf{r}'}(\beta -\tau ).$$ The scaling of (\[S\_II\_2\]) is $(N_{at}N_{r})^{2}N_{\tau}N_{k}$. In (\[S\_II\_2\]) we use MPI processes associated with index $\mathbf{r}'$. Then we apply FFT $$\label{S_II_3} \Sigma^{\alpha\mathbf{k}}_{\mathbf{r}\mathbf{r}'}(\tau)=\sum_{\mathbf{R}}e^{-i\mathbf{k}\mathbf{R}} \Sigma^{\alpha\mathbf{R}}_{\mathbf{r}\mathbf{r}'}(\tau).$$ The scaling of (\[S\_II\_3\]) is $(N_{at}N_{r})^{2}N_{\tau}N_{k}\ln N_{k}$. Similar to the MT-Int case, we use FFT to transform it into equivalent linear combination of plane waves $$\begin{aligned} \label{S_II_4} \Sigma^{\alpha\mathbf{k}}_{\mathbf{r}\mathbf{r}'}(\tau)=\sum_{\mathbf{G}}\sum_{\mathbf{G}'} e^{i(\mathbf{k}+\mathbf{G})\mathbf{r}}\tilde{\Sigma}^{\alpha\mathbf{k}}_{\mathbf{G};\mathbf{G}'}(\tau)e^{-i(\mathbf{k}+\mathbf{G}')\mathbf{r}'},\end{aligned}$$ with the coefficients $$\begin{aligned} \label{S_II_5} \tilde{\Sigma}^{\alpha\mathbf{k}}_{\mathbf{G};\mathbf{G}'}(\tau)=\frac{1}{N^{2}_{\mathbf{r}}}\sum_{\mathbf{r}\mathbf{r}'} e^{-i(\mathbf{k}+\mathbf{G})\mathbf{r}}e^{i(\mathbf{k}+\mathbf{G}')\mathbf{r}'}\Sigma^{\alpha\mathbf{k}}_{\mathbf{r}\mathbf{r}'}(\tau).\end{aligned}$$ The scaling of (\[S\_II\_5\]) is $N_{at}^{2}N_{r}\ln(N_{at}N_{r})N_{k}N_{\tau}\left[N_{r}+N_{pb}^{Int}\right]$. The form (\[S\_II\_4\]) allows us to integrate over the interstitial region analytically and as a result we obtain $$\begin{aligned} \label{S_II_6} \Sigma^{\alpha\mathbf{k}}_{\mathbf{G};\mathbf{G}'}(\tau)=\frac{1}{\Omega_{0}} \sum_{\mathbf{G}''\mathbf{G}'''}S^{\mathbf{k}}_{\mathbf{G}\mathbf{G}''} \tilde{\Sigma}^{\alpha\mathbf{k}}_{\mathbf{G}'';\mathbf{G}'''}(\tau) S^{\mathbf{k}}_{\mathbf{G}'''\mathbf{G}'},\end{aligned}$$ with scaling $N_{at}^{3}N_{pb}^{Int}N_{orb}N_{k}N_{\tau}\left[N_{pb}^{Int}+N_{orb}\right]$. Finally, the contribution to the band state representation from the interstitial is given by $$\label{S_II_7} \Sigma^{\alpha\mathbf{k}}_{\lambda\lambda'}(\tau)\|^{Int}_{Int}=\sum_{\mathbf{G}\mathbf{G}'} A^{^{*}\alpha\mathbf{k}\lambda}_{\mathbf{G}} \Sigma^{\alpha\mathbf{k}}_{\mathbf{G};\mathbf{G}'}(\tau)A^{\alpha\mathbf{k}\lambda'}_{\mathbf{G}'}.$$ The scaling of (\[S\_II\_7\]) is $(N_{at}N_{orb})^{3}N_{k}N_{\tau}$. MPI-parallelization is used in (\[S\_II\_3\]-\[S\_II\_7\]) to perform calculations for different $\mathbf{k}$’s independently. Formulae (\[S\_II\_2\])-(\[S\_II\_7\]) have been implemented in the subroutine `SIGC_II_SOLID`. Results {#res} ======= In this section we show how our linearized version of scQPGW performs and compare the results to other (non-linearized) implementations of the scQPGW method in other basis sets and to experimental data. In the Table\[B\_Gaps\] we present our calculated band gaps for selected semiconductors and insulators obtained with linearized scQPGW and compare them with previous scQPGW calculations and experiment. As one can see, our results are pretty close to the non-linearized scQPGW results and systematically overestimate the band gaps. The overestimation generally is in 10-24% range for all studied materials, excluding antiferromagnetically ordered NiO (error is only 3.9%) and the f-eletron compound CeO$_{2}$ where the error is large (70%). Let us also mention that other QPscGW methods produce large error as well for the current material. Table \[B\_Widths\] presents the band widths of alkali metals Na and K. Formally, alkali metals belong to the s-materials. But as one can see from the table the error in calculated band width (20-30%) is a bit larger than the error in the calculated band gaps for sp-semiconductors. Also, the error increases when the density of valence electrons is reduced (when going from Sodium to Potassium). That fact was expected because the electron gas of lower density corresponds to more correlated situation. ------------- ---------------------------------- ------------------ ------------------ ------------- ------- Present [@prb_76_165106; @prl_96_226402] [@prl_99_246403] [@prb_92_041115] work Exp. Si 1.23 1.41 1.47 1.40(14.8%) 1.22 SiC 2.14 2.88 2.90 3.08(22.7%) 2.51 C 6.52 6.18 6.40 6.71(14.1%) 5.88 GaAs 1.93 1.85 1.75 2.08(23.1%) 1.69 ZnO 3.87 3.8 4.61 4.47(24.2%) 3.60 NiO 4.8 4.97 4.47(3.9%) 4.3 Cu$_{2}$O 2.36 2.65 2.42(10.0%) 2.20 TiO$_{2}$ 3.78 4.22 3.80(22.6%) 3.1 SrTiO$_{3}$ 4.19 4.01(21.5%) 3.3 CeO$_{2}$ $\sim$5 5.83(70.1%) 3-3.5 ------------- ---------------------------------- ------------------ ------------------ ------------- ------- : Band gaps (eV) of selected semiconductors and insulators. Experimental data have been cited from Ref.[@prb_92_041115] and [@prl_96_226402]. For the present work results we also include the error (%) relative to the experiment.[]{data-label="B_Gaps"} ---- ------------------ ------------- ------ Present [@prl_96_226402] work Exp. Na 3.0 3.16(19.2%) 2.65 K 2.07(29.4%) 1.60 ---- ------------------ ------------- ------ : Band widths (eV) of alkali metals. Experimental data have been taken from Ref.[@prl_60_1558] and Ref.[@prb_54_7758]. For the present work results we also include the error (%) relative to the experiment.[]{data-label="B_Widths"} $N_{PB}$ $N_{LAPW+LO}$ $N_{k}$ $n_{\tau}/n_{k}$ P W $\Sigma$ G ------------- ---------- --------------- ---------- ------------------ ----- ------ ---------- ----- Si 910 129 $12^{3}$ 24/3 262 167 705 4 SiC 903 118 $8^{3}$ 24/3 89 140 345 3 C 973 140 $12^{3}$ 24/3 167 235 582 3 GaAs 1116 158 $4^{3}$ 24/3 23 46 121 0 NiO 2072 420 $4^{3}$ 24/6 427 334 1320 57 Cu$_{2}$O 3390 794 $4^{3}$ 24/2 121 1067 451 178 TiO$_{2}$ 2552 496 $4^{3}$ 24/6 289 398 922 26 SrTiO$_{3}$ 2598 373 $4^{3}$ 24/5 85 223 304 3 CeO$_{2}$ 1361 225 $6^{3}$ 24/4 99 89 240 2 K 628 76 $8^{3}$ 24/1 22 31 56 2 Ni 475 68 $16^{3}$ 16/3 486 170 707 10 Table \[Time\] shows the time which was needed to evaluate the main quantities (P, W, $\Sigma$, and G) during one iteration. As it can be seen for the materials studied, the calculation of the self energy is the most time consuming. However, increasing the size of the product basis (which is proportional to the number of atoms in the unit cell) will eventually make the evaluation of W the most time consuming. This is clear from the scaling considerations: whereas many parts of the algorithm scale as $N_{at}^{3}$ the evaluation of W has the biggest prefactor. Conclusions {#concl .unnumbered} =========== We presented an implementation of the scQPGW method in LAPW basis set which scales as $N^{3}$ with the number of atoms. Further improvements of the algorithm for large systems would require a removal of the computational bottleneck which is the matrix inversion in Eq.(\[W\_1\]). In its current form, this code can serve as a starting point for further diagrammatic many body studies on the Matsubara axis in an all electron basis as was done for example in Ref.[@arx_1504_07569]. Acknowledgments {#acknow .unnumbered} =============== This work was supported by the U.S. Department of energy, Office of Science, Basic Energy Sciences as apart of the Computational Materials Science Program. We thank Sangkook Choi for many discussions. [10]{} url \#1[`#1`]{}urlprefixhref \#1\#2[\#2]{} \#1[\#1]{} , [New Method for Calculating the One-Particle Green’s Function with Application to the Electron-Gas Problem]{}, Phys. Rev. 139 (1965) A796. , [Electron correlation in semiconductors and insulators: Band gaps and quasiparticle energies]{}, Phys. Rev.B 34 (1986) 5390. , [Self-energy operators and exchange-correlation potentials in semiconductors]{}, Phys. Rev. B 37 (1988) 10159. , Berkeleygw: A massively parallel computer package for the calculation of the quasiparticle and optical properties of materials and nanostructures, Comp. Phys. Comm. 183 (2012) 1269. , A brief introduction to the abinit software package, Z. Kristallogr. 220 (2005) 558. , [Large Scale GW Calculations]{}, J. Chem. Theor. Comp. 11 (2015) 2680. , [exciting: a full-potential all-electron package implementing density-functional theory and many-body perturbation theory]{}, J. Phys.: Condens. Matter 26 (2014) 363202. , [Efficient implementation of the GW approximation within the all-electron FLAPW method]{}, Phys. Rev. B 81 (2010) 125102. , [GW with linearized augmented plane waves extended by high-energy local orbitals]{}, Phys. Rev. B 93 (2016) 115203. , [All-electron self-consistent GW in the Matsubara-time domain: Implementation and benchmarks of semiconductors and insulators]{}, Phys. Rev. B 93 (2016) 125210. , [Large scale GW calculations]{}, arXiv.cond.mat.:1501.03141. , [The GW space-time method for the self-energy of large systems]{}, Comp. Phys. Comm. 117 (1999) 211. , [Band-gap problem in semiconductors revisited: effects of core states and many-body self-consistency]{}, Phys. Rev. Lett. 89 (2002) 126401. , [Quasiparticle Self-Consistent GW Method for the Spectral Properties of Complex Materials]{}, Top Curr Chem. 347 (2014) 99. , [Quasiparticle self-consistent GW method: A basis for the independent-particle approximation]{}, Phys. Rev.B 76 (2007) 165106. , [Accurate quasiparticle spectra from self-consistent GW calculations with vertex corrections]{}, Phys. Rev. Lett. 99 (2007) 246403. , [Accurate band gaps of extended systems via efficient vertex corrections in GW]{}, Phys. Rev. B 92 (2015) 041115. , [Electronic structure of Pu and Am metals by self-consistent relativistic GW method]{}, Phys. Rev. B 85 (2012) 155129. , [Ground-state properties of simple elements from GW calculations]{}, Phys. Rev. B 80 (2009) 041103. , [Matsubara QSGW+DMFT: application to Mott insulator La$_{2}$CuO$_{4}$]{}, arXiv.cond.mat.:1504.07569. , Linear methods in band theory, Phys. Rev.B 12 (1975) 3060. , [Quasiparticle Self-Consistent GW Theory]{}, Phys. Rev. Lett. 96 (2006) 226402. , [Quasiparticle Band Structure of Na and Simple Metals]{}, Phys. Rev. Lett. 60 (1988) 1558. , [Self-consistent GW and higher-order calculations of electron states in metals]{}, Phys. Rev.B 54 (1996) 7758.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The Rutherford Backscattering Spectroscopy (RBS) and Particle Induced X-ray Emission (PIXE) techniques can be used to obtain complementary information about the characteristics of a sample but, traditionally, a gap has separated the available computer codes for analyzing data from each technique, being hard to simultaneously analyze data from the same sample. The recent development of a free and open source library, LibCPIXE, for PIXE simulation and analysis of arbitrary multilayered samples, has permitted to integrate this technique into the DataFurnace code which already handles many other IBA techniques such as Rutherford and non-Rutherford backscattering, elastic recoil detection, and non-resonant nuclear reaction analysis. The fitting capabilities of DataFurnace can therefore now be applied to PIXE spectra as well, including the Bayesian Inference analysis and the simultaneous and coherent fitting of multiple spectra from different techniques. Various examples are presented in which the simultaneous RBS and PIXE analysis allows us to obtain consistent results that cannot be obtained by independent analysis of the data from each technique.' address: - 'Instituto Tecnológico e Nuclear, E.N. 10, 2686-953 Sacavém Codex, Portugal' - 'CMAM, Universidad Autónoma de Madrid, E-28049 Spain' author: - 'C. Pascual-Izarra' - 'M. A. Reis' - 'N. P. Barradas' title: Simultaneous PIXE and RBS data analysis using Bayesian Inference with the DataFurnace code --- , , , Bayesian Inference, PIXE,X-rays, RBS, Rutherford Backscattering, Fit :82.80.Ej;82.80.Yc;02.70.Uu;07.05.Tp Note: This is a preprint of a work published in Nuclear Instruments and Methods in Physics Research Section B 249, 780-783 (2006). Please cite the published one. Introduction ============ Despite of the close relation of particle-based IBA techniques (RBS, ERDA,…) with photon-based ones (PIXE, PIGE,…), and despite the fact that they are often used to obtain complementary information about the same sample, the most popular IBA analysis codes are focused either on one or the other category and data cannot be analyzed in a coherent and unified way. An attempt to unify PIXE and RBS analysis had been already done [@web:DAN32] by building a common interface to the well known RUMP [@web:RUMP] and GUPIX [@web:GUPIX] programs. In this work we present a more ambitious initiative in terms of integrated analysis of IBA techniques: the implementation of the support for PIXE in the DataFurnace code (which already supported RBS, ERDA, and non-resonant NRA), achieved by using the recently developed LibCPIXE free library for PIXE yield simulation [@LibCPIXE]. This implies that all the fitting capabilities already available in DataFurnace for other techniques can now be applied to PIXE, including simulated annealing fits and the Bayesian inference error estimation. DataFurnace $+$ LibCPIXE {#s:methodology} ======================== The integration of LibCPIXE with DataFurnace, incorporates PIXE to the repertory of supported techniques in the latter code. This opens up the possibility of simultaneously analyzing data from several spectra of a given sample and thus merging the information from all of them, producing a self–consistent result. Note that the experimental spectra can be obtained with different techniques (now including PIXE) as well as with the same technique under different experimental conditions in order to obtain complementary information about the sample. In contrast with the most widely used programs for PIXE data analysis, which make use of nonlinear least squares or Marquardt fitting algorithms [@IAEA-PIXE], the DataFurnace code implements simulated annealing and Bayesian inference algorithms. The combination of these algorithms has proven to be a very powerful tool when dealing with IBA problems in which one may have multiple and/or suboptimal formal solutions [@Barradas03-BI-IBA]. In relation with PIXE analysis, this provides a very flexible fitting procedure that, e.g, does not pose any limit to the analysis of layered samples with a given element present in more than one layer (which is, currently, a limitation in the GUPIX code [@GUPIX95]). The Bayesian inference has also the advantage of providing a mathematically rigorous estimation of the errors in the fitted parameters (concentration profiles for each element in the sample). The treatment of the PIXE data in DataFurnace differs from that done for other techniques or that done by other PIXE analysis codes: instead of fitting the PIXE spectra channel-by-channel, only user-selected characteristic X-ray peak areas are fitted (avoiding the difficulties of fitting background and/or detector response functions and allowing to discard peaks which are non-apt for the fit). In order to identify the peaks and filter the background, the QXAS program [@web:QXAS] can be used (DataFurnace can directly read its output files) although a more convenient program is currently being developed and will be incorporated into the LibCPIXE library. Besides this detail, the treatment of the PIXE spectra in DataFurnace is identical to that done for the other techniques. See [@Jeynes03] for a detailed description on how the fitting process works in DataFurnace, and how the information provided by different techniques about a given sample is merged. A typical case of ambiguity found in PIXE spectra is related to the occurrence of non-resolved overlapping peaks. This can be handled either by simply ignoring them in the fit (and relaying on other peaks for those elements) or by fitting the sum instead of the separated values. Note that in many cases, the ambiguities can be also ruled out by providing additional spectra (PIXE and/or other techniques) or by setting constrains to the fit based on known information about the sample. Examples {#s:examples} ======== A first example consists in obtaining the concentration of the elements present in a bulk stoichiometric Sb$_2$S$_3$ sample. A PIXE spectrum with 1.1 MeV H$^+$ was acquired using $\phi_{inc}=15^\circ$ and $\phi_{det}=55^\circ$, where $\phi_{inc}$ and $\phi_{det}$ are, respectively, the incidence and detection angles relative to the surface normal. A Bayesian fit of the areas of the Sb $L_{\alpha_{1,2}}$ and S $K_{\alpha_{1,2}}$ peaks was performed (the only free parameters in this case being the bulk concentration of S and Sb as well as the beam fluence). The results, as shown in Table \[t:results\], reproduce the expected stoichiometric values and provide a reasonable error estimation reflecting mainly the statistical uncertainties in the experimental data. A second example consists in the analysis of a layered Mn$_4$Ir/Si sample (a Mn$_4$Ir film on top of a Si substrate). A Bayesian fit of the Mn $K_{\alpha_{1,2}}$, Ir $L_{\alpha_{1,2}}$ and Si $K_{\alpha_{1,2}}$ lines of a 1.22MeV H$^+$ PIXE spectrum (with $\phi_{inc}=7.5^\circ$ and $\phi_{det}=77.5^\circ$) was performed leaving the thickness and composition of the film as free parameters. The results are shown in Table \[t:results\] and are compatible with the nominal values (stoichiometric concentrations and a previous RBS thickness measurement) within the errors estimated by the Bayesian inference. In order to ascertain the accuracy of these results, we performed a simultaneous RBS+PIXE analysis using the same PIXE data as before and a 2 MeV He$^+$ RBS spectrum taken at normal incidence angle and 177 scattering angle. As expected, the uncertainties are greatly reduced in this case (see Table \[t:results\]) since the RBS data introduces a tight constrain in the thickness of the layer. When dealing with a more complex problem such as obtaining the concentration profiles for all elements in a GaInAsSb/GaSb sample, the information from a single PIXE spectrum (or a single RBS spectrum) is not enough to solve the ambiguities of the fit. Previous analysis of this kind of samples [@Corregidor05-GaInAsSb; @Reis05] involved the separated analysis of RBS and PIXE spectra and a manual iteration to obtain coherent results from both techniques. The DataFurnace code now allows to perform the analysis in a much more convenient way: in the third example, the simultaneous analysis of six PIXE spectra taken with different beam energies and incidence angles proves useful for characterizing the sample. The Bayesian fit (shown in Fig. \[f:GaInAsSb\]) was performed assuming a Ga$_{x}$In$_{1-x}$As$_{1-y}$Sb$_{y}$ ($x,y<1$) of unknown thickness and composition on top of a stoichiometric GaSb substrate. The results, shown in Table \[t:results\], are compatible with those of previous characterizations of the same sample (as seen in Table \[t:results\]). ![Comparison between experimental and fitted main lines of the four elements in the film. $E$ is the proton beam energy. In geometry G1, $\phi_{inc}=7.5^\circ$ and $\phi_{det}=62.5^\circ$ whereas in geometry G2, $\phi_{inc}=22.5^\circ$ and $\phi_{det}=47.5^\circ$. Spectra obtained using a proton beam of 1.05 MeV were detected without any special filter between the chamber window and the Be window of the detector. The remaining spectra were measured using a 1mm thick Mylar filter in front of the x-ray detector.[]{data-label="f:GaInAsSb"}](MS0318-Figure1.eps){width="90.00000%"} Finally, another Bayesian fit has been performed incorporating a 2 MeV H$^+$ RBS spectrum (see Fig. \[f:RBS\]) to the mentioned set of six PIXE spectra. The results in Table \[t:results\] show that less uncertainties are present in this case compared to the PIXE–only one due to the complementarity of the information provided by each technique: PIXE is good at determining the amount of each element but has poor depth sensitivity while RBS provides good depth resolution but cannot separate elements of similar mass (such as Ga–As or In–Sb). It is interesting to remark the fact that this analysis has been performed in one single step, providing all the experimental data to the DataFurnace code and obtaining a self-consistent result which is more reliable and considerably less time consuming than that obtained by using separated analysis codes. ![Experimental RBS spectrum (squares) for 2 MeV H$^+$ on the same GaInAsSb/GaSb sample as in Fig. \[f:GaInAsSb\]. The fit (solid line) was performed considering not only the RBS spectrum but also the data from six PIXE spectra (those of Fig. \[f:GaInAsSb\]). Partial spectra for each element (non-solid lines) are also shown.[]{data-label="f:RBS"}](MS0318-Figure2.eps){width="90.00000%"} [c\|cc\|c\|cc]{} Sample & & Element &\ & fit(error) & Nominal & & fit(error) & Nominal\ ------------------------------------------------------------------------ & \[(bulk)]{} & \[(bulk)]{} & Sb & 40.8(1.2) & 40\ & & & S & 59.2(1.2) & 60\ & \[551(35)]{} & \[533$^$]{} & Mn & 79.1(1.2) & 80\ & & & Ir & 20.9(1.2) & 20\ & & & Si$^\text{(S)}$& 100 & 100\ & \[534(3)]{} & \[533$^$]{} & Mn & 79.9(0.1) & 80\ & & & Ir & 20.1(0.1) & 20\ & & & Si$^\text{(S)}$& 100 & 100\ & \[7917(854)]{} & \[8442$^$]{} & Ga & 38.3(1.3) & 38.5$^$\ & & & In & 10.7(0.9) & 11.5$^$\ & & & As & 9.1(0.7) & 9.0$^$\ & & & Sb & 41.9(1.3) & 41.0$^$\ & & & Ga$^\text{(S)}$& 50 & 50\ & & & Sb$^\text{(S)}$& 50 & 50\ & \[7988(625)]{} & \[8442$^$]{} & Ga & 40.5(0.4) & 38.5$^$\ & & & In & 10.0(0.3) & 11.5$^$\ & & & As & 8.8(0.3) & 9.0$^$\ & & & Sb & 40.7(0.4) & 41.0$^$\ & & & Ga$^\text{(S)}$& 50 & 50\ & & & Sb$^\text{(S)}$& 50 & 50\ Conclusions =========== We have shown that the integration of LibCPIXE into the DataFurnace code facilitates very interesting analysis modes such as the simultaneous PIXE + RBS or the multiple (differential) PIXE analysis. These combined analysis modes are, in many cases, the only way of obtaining a consistent characterization of a specimen. The use of robust fitting —simulated annealing— and rigorous error estimation routines —Bayesian inference— makes it possible to characterize arbitrary layered samples, with no limitations on the composition of each layer. Acknowledgments {#acknowledgments .unnumbered} =============== Thanks are due to P. C. Chaves and V. Corregidor for their help with the experimental data acquisition. This work has been supported by an EU grant (HPRN-CT-2001-00199). [10]{} url \#1[`#1`]{}urlprefix , [URL]{}: [[<http://www.microbeams.co.uk/>]{}]{}. , [URL]{}: [[<http://www.genplot.com/>]{}]{}. , [URL]{}: [[<http://www.physics.uoguelph.ca/PIXE/gupix/>]{}]{}. C. Pascual-Izarra, N. P. Barradas, M. A. Reis, [LibCPIXE: a PIXE simulation open-source library for multilayered samples]{}, Nucl. Instr. and Meth. B 249 (2006) 820.822. S. Fazinić, [Intercomparison of PIXE Spectrometry Software Packages]{}, TECDOC 1342, [International Atomic Energy Agency (IAEA)]{} (2003). N. P. Barradas, [Advanced data analysis techniques for Ion Beam Analysis]{}, Surf. Interface Analysis 35 (2003) 760–769. J. A. Maxwell, W. J. Teesdale, J. L. Campbell, [The Guelph PIXE software package II]{}, Nucl. Instr. and Meth. B 95 (1995) 407–421. , [URL]{}: [[<http://www.iaea.org/OurWork/ST/NA/NAAL/pci/ins/xrf/pciXRFdown.php>]{}]{}. C. Jeynes, N. P. Barradas, P. K. Marriott, G. Boudreault, M. Jenkin, E. Wendler, R. P. Webb, Elemental thin film depth profiles by ion beam analysis using simulated annealing - a new tool, J. Phys. D: Appl. Phys. 36 (2003) R97–R126. V. Corregidor, N. P. Barradas, E. Alves, N. Franco, L. C. Alves, P. C. Chaves, M. A. Reis, [Ion beam analysis of GaInAsSb films grown by MOVPE on GaSb]{}, Nucl. Instr. and Meth. B 241 (2005) 326. M. A. Reis, P. C. Chaves, V. Corregidor, N. P. Barradas, E. Alves, F. Dimroth, A. W. Bett, [Detection angle resolved PIXE and the equivalent depth concept for thin films characterization]{}, X-ray Spectrometry 34 (2005) 372.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Extensive Monte Carlo simulations are employed in order to study the dynamic critical behavior of the one-dimensional Ising magnet, with algebraically decaying long-range interactions of the form $\frac{1}{r^{d+\sigma}}$, with $\sigma=0.75$. The critical temperature, as well as the critical exponents, is evaluated from the power-law behavior of suitable physical observables when the system is quenched from uncorrelated states, corresponding to infinite temperature, to the critical point. These results are compared with those obtained from the dynamic evolution of the system when it is suddenly annealed at the critical point from the ordered state. Also, the critical temperature in the infinite interaction limit is obtained by means of a finite-range scaling analysis of data measured with different cutoffs of the interaction range. All the estimated static critical exponents ($\gamma /\nu $, $\beta /\nu $, and $1/\nu $ ) are in good agreement with Renormalization Group (RG) predictions and previously reported numerical data obtained under equilibrium conditions. It is found that the dynamic exponent $z$ is different for quenching and annealing experiments, most likely due to the influence of the Kosterlitz-Thouless transition occurring at relatively similar algebraic decay of the interactions with $\sigma =1$. However, for annealing experiments the measured exponent $z$ is close to the RG predictions. On the other hand, the relevant exponents of the dynamic behavior ($z$ and $\theta$) are slightly different than the RG predictions, most likely due to the fact that they may depend on the especific dynamics used (Metropolis in the present paper).' address: 'Computational Physics Group, Instituto de Investigaciones Fisicoquímicas Teóricas y Aplicadas (INIFTA), Facultad de Ciencias Exactas, Universidad Nacional de La Plata, CCT-La Plata CONICET; Suc. 4, CC16 (1900) La Plata, Argentina.' author: - 'D. E. Rodríguez' - 'M. A. Bab' - 'E. V. Albano' title: 'Study of the Nonequilibrium Quenching and Annealing Dynamics for the Long-Range Ising Model ' --- Introduction ============ The study of the critical behavior of systems with long-range (LR) interactions is still a challenging topic in the field of statistical physics [@chen; @lut; @ber]. Furthermore, the understanding of the dynamic evolution of these systems, from far-from equilibrium initial states towards a final equilibrium regime, poses an additional difficulty due to the fast relaxation of relevant physical observables due to the presence of LR interactions. For these reasons, the study of relaxational processes in simple Ising and Potts models with LR interactions plays an important role for the understanding of the dynamics of second-order phase transitions. Within this context, the study of the short-time dynamics (STD) of critical systems has attracted great attention during the last two decades [@Zheng; @Oliveira; @chen; @loca]. The pioneering theoretical study of the STD, which was formulated in the context of the dynamic renormalization group [@jan], predicts the existence of a new exponent related to the initial increase of the order parameter. This prediction has subsequently been validated by a large body of numerical evidence obtained in a variety of models [@Zheng; @bab; @bab2; @santos; @silva; @grandi]. However, only few studies have been performed in order to generalize these concepts to systems with LR interactions. In fact, the field-theoretical calculations of Janssen et al. [@jan] have been extended to the case of LR interactions decaying according to a power law for the case of the continuous $n-$vector model [@chen], the random Ising model [@chen1], and the kinetic spherical model [@chen2; @bau]. On the other hand, theoretical studies of the relaxational dynamics of discrete models are still lacking, and only few preliminary numerical results on the STD of the Potts model have recently been reported [@loca]. In order to contribute to the understanding of the dynamics of phase transitions in discrete systems, the aim of this paper is to report and discuss extensive numerical simulations of the Ising model, in one dimension, with LR interactions decaying with the distance as a power law. For this purpose, we performed studies of both the STD of initially disordered states (i.e., quenching experiments) and the relaxation dynamics of initially ordered states (i.e., annealing experiments). Results obtained by applying these methods allow us to determine not only the critical temperature, but also the complete set of static and dynamic critical exponents (for the methodology used, see e.g.[@gab_bag]). In this way, we can compare our results with theoretical Renormalization Group (RG) predictions [@chen; @Binder] and with independent numerical determinations of the static exponents performed under equilibrium conditions [@ber]. The paper is organized as follows: in Section II a brief description of the model and the simulation method is presented, Section III is devoted to a brief discussion of the theoretical background subsequently applied to the analysis of the results that are discussed in Section IV. Finally, our conclusions are stated in Section V. The Ising model with LR interactions and the simulation method ============================================================== In this paper we present and discuss simulations of the Ising Model in $d = 1 $ dimensions, whose Hamiltonian $H$ is given by $$H = -J \sum_{\langle i,j \rangle} S_{i} S_{j}/r_{ij}^{d + \sigma}, \label{Hamil}$$ where $J > 0$ is the (ferromagnetic) coupling constant, $S_{i}$ is the spin variable at the site of coordinates i, which can assume two values, namely, $S_{i} = \pm 1 $, the summation is extended to all pairs of spins placed at distances $r_{i,j} = \|r_i - r_j\|$, and $\sigma$ is a parameter that controls the decay of LR interactions. Simulations are performed by using samples of length $L=10^5$ and taking periodic boundary conditions. The LR interactions described by the Hamiltonian of equation (\[Hamil\]) are evaluated up to a distance $% \|r_i-r_j\|=L/2$. Also, simulations with LR interactions truncated at the $Nth$ neighbor, i.e., $J=0$ for $r>N$, have been performed in order to apply Finite Range Scaling (FRS) analysis [@Glumac] and the results will be briefly discussed. Spin update is performed by using the standard Metropolis dynamics. Also, during a Monte Carlo time step (mcs) all the spins of the sample are updated once, on average. In order to carry out the calculations we chose $\sigma =0.75$, because for this value of the parameter the critical exponents of the Ising model are expected to be sufficiently different from mean-field values to allow a meaningful comparison with RG predictions [@nag; @mon; @uze]. Furthermore, one also likes to be as far as possible from $\sigma =1.00$, where strong Kosterlitz-Thouless behavior is known to occur [@car]. During the simulations we recorded the time dependence of the following observables: (i) The order parameter or average magnetization $M(t,\tau)$ given by $$M(t,\tau)= \frac {1}{L} \langle \sum_{i=1}^{L} S_{i}(t,\tau) \rangle, \label{magne}$$ where $\tau=\frac{T-T_c}{T_c}$ is the reduced temperature and $T_c$ is the critical temperature. \(ii) The susceptibility ($\chi(t,\tau)$) evaluated as the fluctuations of the order parameter, namely $$\chi (t,\tau )=\frac 1{LT}(M^2(t,\tau )-M(t,\tau)^2), \label{chi}$$ where $M^2(t,\tau )= \frac {1}{L^2} \langle (\sum_{i=1}^{L} S_{i}(t,\tau))^2 \rangle$. \(iii) The autocorrelation of the spin variable $$A(t,\tau )=\frac 1L\langle \sum_{i=1}^LS_i(t,\tau )S_i(0,\tau )\rangle . \label{autoc}$$ \(iv) The correlation of the order parameter at the critical point, when the initial condition corresponds to uncorrelated states, given by $$Q(t) = \frac {1}{L^2} \langle \sum_{i=1}^{L} S_{i}(t) \sum_{i=1}^{L} S_{i}(0)\rangle, \label{Qt}$$ \(v) The second-order Binder cumulant ($U(t)$), when the initial condition corresponds to the ground state, namely, $$U(t,\tau) = \frac{M^2(t,\tau)}{M(t,\tau)^2} -1 , \label{binder}$$ where in all cases the brackets indicate configurational averages performed with different samples started from equivalent (but different in the case of $T=\infty$) initial conditions. Brief theoretical background ============================ Short-time dynamics (STD): Let us now analyze the expected short-time dynamic behavior when the system starts from a disordered (uncorrelated) configuration, but with a small initial magnetization. According to the argument of Janssen et al. [@jan], the general scaling approach of the order parameter for the nonconservative dynamics of model A (according to the classification of Hohenberg and Halperin [@hoha]), is given by $$M(t,\tau,L,M_{0}) = b^{-\beta/\nu} M^{~}(t/b^{z}, b^{1/\nu}\tau, L/b, b^{x_{0}}M_{0}), \label{escamag}$$ where $b$ is a scaling parameter, and $\beta$ and $\nu$ are the order parameter and correlation length (static) critical exponents, respectively. Also, $z$ is the dynamical exponent. Furthermore, $x_{0}$ is a new exponent, introduced by Janssen et al [@jan], which accounts for the scaling dimension of the initial magnetization $M_{0}$, in the $M_{0} \rightarrow 0$ limit. For sufficiently large lattices, at the critical point ($\tau\equiv 0$), and by setting $b=t^{1/z}$, equation (\[escamag\]) becomes $$M (t, M_{o}) = t^{-\beta/ \nu z} M(t^{\frac{x_{o}}{z}}M_{o}), \label{mcero}$$ which holds for a time short enough such that the correlation length ($\xi(t) \propto t^{1/z}$) is not so large ($\xi \ll L$). Furthermore, for times even shorter than the crossover time ($t_{x} \approx M_{o}^{-z/x_{o}}$), but larger than the microscopic time that is set when the correlation length is of the order of a single lattice spacing, one has that equation (\[mcero\]) becomes $$M(t) \propto M_{0}t^{\theta}, \label{in_in}$$ which describes the (power-law) initial increase of the magnetization with exponent $\theta = x_{0}/z - \beta/ \nu z$. In the absence of an initial magnetization ($M_{0}\equiv 0$), and at criticality, the scaling behavior of the susceptibility is given by $$\chi(t) \propto t^{\gamma/ \nu z}, \label{scachi}$$ where $\gamma$ is the susceptibility exponent. Also, under these conditions ($\tau = 0$ and $M_{0} = 0)$, the time autocorrelation function is expected to follow a power law with time according to $$A(t)\propto t^{-\lambda}, \label{auto}$$ where the critical exponent is given by $\lambda = d/z-\theta$, i.e., even in the absence of an initial magnetization $\lambda$ depends on the exponent $\theta$ that describes the initial increase of the order parameter according to equation (\[in\_in\]). On the other hand, by starting with randomly generated configurations, the correlation function of the total magnetization is also expected to follow a power law with time according to $$Q(t)\propto t^{\theta}, \label{Q}$$ i.e., a relationship that allows us to obtain the initial increase exponent avoiding the numerical extrapolation $M_0\rightarrow0$ [@Oliveira]. Standard relaxation dynamics (SRD). STD measurements can be further reinforced by independent measurements of the SRD, which are started from a fully ordered or ground state configuration and are performed at criticality. In this way, one could be able not only to test the validity of some exponents evaluated by means of the STD method, as well as the critical temperature, but also obtain additional exponents and test the validity of relationships between them, e.g. the hyperscaling relationship [@Zheng]. In fact, by starting from a ground state configuration with all spins pointing in the same direction ($T=0$), upon annealing to criticality, the SRD scaling approach is given by (see also equation (\[mcero\]) $$M(t,\tau,L) = b^{-\beta/ \nu} M(t/b^z, b^{1/ \nu} \tau, L/b). \label{que}$$ For large lattices and by setting $b=t^{1/z}$ this dynamic scaling form leads to $$M(t,\tau )\propto t^{-\beta /\nu z}M(t^{1/\nu z}\tau ). \label{quen}$$ It is well known that this power-law decay of the order parameter is valid within the long-time regime, but several numerical results indicate that it also holds in the short-time regime. On the other hand, by taking the logarithmic derivative of equation (\[quen\]) with respect to the reduced temperature, evaluated at the critical point, one gets $$\frac{\partial log M(t,\tau)}{\partial \tau}\vert_{\tau=0} \propto t^{1/ \nu z}, \label{deri}$$ which allows us to evaluate the exponent $1/\nu z$, by performing measurements at and slightly away from the critical point. Furthermore, just at the critical point the second-order Binder cumulant is expected to behave according to $$U(t)\propto t^{d/z}. \label{cumu}$$ Results and discussion. ======================= Standard relaxation dynamics ---------------------------- Focusing our attention first on the relaxation dynamic behavior, at criticality, figure \[figure1\] shows the time evolution of the magnetization at different temperatures. The critical temperature $% T_c=2.630(3)$ was determined by searching the smallest standard deviation from the power law (equation (\[quen\])), and the error bar was assessed by considering closest temperatures that present noticeable but small deviations. Also, from the fit of the data the critical exponent $\beta /\nu z=0.091(2)$ was determined. Figure \[figure2\] (a) shows the time evolution of the second-order Binder cumulant that can be fitted with a power law with exponent $d/z=1.05(1)$. This value yields $z=0.95(1)$ for the dynamic exponent that is significantly larger than the RG predictions, which place $z$ between $0.7174\; ($for $\sigma =0.70)$ and $0.834\;($for $\sigma =0.80)$ [@chen; @loca]. In principle one could expect that this result may be most likely due to the fact that $z$ depends on the specific dynamics used, however, a more detailed discussion will be offered below. Finally, by using measurements of the magnetization performed at two adjacent temperature points, namely, $T=2.620$ and $T=2.640$, the logarithmic derivative of the magnetization with respect to the reduced temperature was obtained. Figure \[figure2\](b) shows that this observable also exhibits a power-law behavior with exponent $1/\nu z=0.50(2)$. Furthermore, by replacing the obtained value of $z$ in the exponent corresponding to the logarithmic derivative, one gets $1/\nu =0.47(2)$, in agreement with both the RG prediction, namely, $1/\nu =0.477$ [@Binder; @chen], and with Monte Carlo simulations performed at equilibrium, $1/\nu =0.469$ [@Binder]. ![Time evolutions obtained after annealing at $T_c=2.630$ from $T=0$ (ground state) of (a) the second-order Binder cumulant ($U(t)$), and (b) the logarithmic derivative of the magnetization with respect to the reduced temperature ($\frac{\partial Log(t)}{\partial \tau}$). The solid lines indicate the fits with equations (\[deri\]) and (\[cumu\]), respectively. []{data-label="figure2"}](Uder.eps){width="70.00000%"} Short-time Dynamics ------------------- Now we turn our attention to the short-time dynamic measurements. The autocorrelation function (figure \[figure3\](a)) exhibits a power-law behavior with time, and the exponent $\lambda =0.94(1)$ was obtained by means of a fit with equation (\[auto\]). Furthermore, the susceptibility (figure \[figure3\](b)) can also be fitted by means of a power law yielding $% \gamma /\nu z=0.869(6)$. In contrast with these measurements performed by setting $M_0\equiv 0$, the initial increase of the magnetization has to be measured for vanishingly small values of $M_0$, as shown in figure \[figure4\]. The extrapolation of the obtained results for $M_0\rightarrow 0$ yields $\theta =0.211(7)$ (see equation (\[in\_in\])), namely a figure that is close to the RG prediction given by $\theta =0.228$ [@chen]. Now, by using the relationship $\lambda =d/z-\theta $ and replacing the determined exponents one gets $z=(0.94(1)+0.211(7))^{-1}=0.868(9)$. This value of the dynamic exponent $z$ is closer to the RG prediction than the previous measurement obtained from the relaxation dynamics given by $z=0.95(1)$. Also, it interpolates between previously published STD results corresponding to a system of size $L=3000$, which are given by $z=0.81(1)$ and $0.96(4)$, for $\sigma =0.70$ and $0.80$, respectively [@loca]. So, this result strongly suggests that, for this model, both dynamic processes, the short-time and the relaxation one, are governed by different dynamic exponents. We expect that this finding may be an effect of the Kosterlitz-Thouless (KT) transition observed at $\sigma =1.00$, which even for our choice of the interaction range ($\sigma =0.75$) may influence the dynamic behavior of the system. KT effects seem to be also present in the evaluation of the exponent $1/\nu $, which for up to $\sigma \geq0.75$ exhibits a rapid decay toward zero, in agreement with the absence of a power-law temperature dependence of the correlation lengh at $\sigma =1.00$[@Binder]. On the other hand, it is well known that systems undergoing a KT transition exhibit different types of dynamic behavior upon quenching and annealing [@KT]. Additionally, one can use the values of both $\gamma /\nu z$ and $z$ in order to estimate $\gamma /\nu =0.754(11)$, in excellent agreement with the RG prediction given by $\gamma /\nu =\sigma =0.75$ [@Binder; @fish]. On the other hand, by assuming that the hyperscaling relationship ($d/z-2\,\beta /\nu z=\gamma /\nu z$) holds, on can obtain the short-time dynamic estimation of $\beta /\nu =0.122(11)$, in excellent agreement with the RG prediction given by $\beta /\nu =\frac{% d-\sigma }2=0.125$[@fish; @Binder]. ![Time evolution measured after quenching from uncorrelated (disordered) states to $T_c=2.630$ of (a) the autocorrelation $A(t)$, and (b) susceptibility $\chi(t)$. The solid lines indicate the fits with equations (\[chi\]) and (\[autoc\]), respectively.[]{data-label="figure3"}](M_A.eps){width="80.00000%"} ![Log-log plot of $M(t)$ versus time results showing the initial increase of the magnetization after quenching from uncorrelated (desordered) states, with a small magnetization $M_0$, to $T_c=2.630$. Data corresponding, from top to bottom, to $M_0=0.1, 0.05, 0.02, 0.015, 0.010$ and $0.006$, respectively. The solid lines show the best fits of the data obtained according to equation (\[in\_in\]). The inset shows the linear extrapolation of the values of the exponent $\theta$ to $M_0\rightarrow0$. []{data-label="figure4"}](Minc.eps){width="80.00000%"} Furthermore, just by starting with random configurations and measuring the correlation function of the total magnetization given by equation (\[Q\]), one can also obtain the initial increase exponent $\theta =0.207(2)$, as shown in figure \[figure5\]. This value for the exponent is in good agreement with the previous meassurement obtained by using the numerical extrapolation $M_0\rightarrow 0$, namely $\theta =0.211(7)$. Moreover, by using this independent estimation of $\theta$ and fapplying the previously described procedure, the following exponents: $z=0.872(8)$, $\gamma /\nu =0.758(10)$, and $\beta /\nu =0.121(10)$ can be obtained, which of course, are in good agreement with our previous estimations. ![Log-log plot of the time evolution of the correlation function of the total magnetization after quenching randomly generated configurations to $% T_c=2.630$. The exponent $\theta $ was determined by fitting the data with the aid of equation (\[Q\]) and the solid line shows the obtained fit. []{data-label="figure5"}](Q.eps){width="80.00000%"} Finite-Range Scaling (FRS) Analysis ----------------------------------- A FRS analysis has been developed by analogy with the finite-size scaling[@Glumac]. Here, the basic idea is to truncate the range of the interaction at a certain prefixed distance and to obtain information on the critical behavior by using scaling properties. The truncated-interaction range $N$ is defined as the number of neighbors on each side of the central spin considered in order to evaluate the Hamiltonian given by equation (\[Hamil\]). In this way, the dependence of the critical temperature on $N$ was studied. Note that due to the periodic boundary conditions used, and the largest system size studied in the present work, $L=10^4$, the maximal interaction range considered is $N=5000$. Figure \[figure6\] shows the behavior of the magnetization, when the system is annealed from the ground state up to the “critical” point ($T_c(N)$) for different values of $N$. Here one observes the power-law dependence expected from equation (\[quen\]) for slightly different ($N$-dependent) “critical” temperatures. Therefore, the inset shows the dependence of $T_c(N)$ on $N^{-1}$, which was fitted with the following scaling form $$T_c(N)=T_c(\infty )+A/N^{x_T}, \label{TcN}$$ where $T_c(\infty )$ is the critical temperature for the infinite interaction range, $x_T$ is the convergence exponent, and $A$ is a constant. The fitting was performed by imposing $x_T=1$, because this restriction does not have significant effect on the final results due to the small variation of $T_c(N)$ with $N$, and yields $T_c(\infty )=2.646(3)$. The extrapolated critical temperature $T_c(\infty )$ interpolates between the previously reported values for $\sigma =0.70$ ($T_c(\infty )=2.929$ [@Glumac] and $% T_c(\infty )=2.9269$ [@Barati]), and $\sigma =0.80$ ($T_c(\infty )=2.431$ [@Glumac] and $T_c(\infty )=2.4299$ [@Barati]), which were obtained by means of exact calculations with the transfer matrix method and FRS analysis. ![Log-log plot of the time evolution of the magnetization obtained after annealing from the ground state to the “critical” point $T_c(N)$ as measured for different values of $N$. The solid lines show the best fits of the data performed with equation (\[in\_in\]). The inset shows convergence of the “critical” temperatures $T_c(N)$ to $T_c(\infty)$, and the solid line corresponds to the best fit performed with the aid of equation (\[TcN\]).[]{data-label="figure6"}](cutoff.eps){width="80.00000%"} Conclusions =========== In this paper we present and discuss the results of extensive simulations of the dynamic behavior of the LR Ising magnet with interactions decaying as $r^{-(d+\sigma )}$, in $d=1$ dimensions and with $\sigma =0.75$. Power-law behavior of the relevant observables is found at the critical temperature $T_c=2.630(3)$ for both the relaxation and the short-time regimes. Also, the difference between the dynamic exponents $z$ determined for these regimes indicates that even for $\sigma =0.75$ and the sample size used ($L=10^4$), the effects of the Kosterlitz-Thouless transition ($\sigma =1.00$) cannot be neglected. These findings lead us to conclude that both types of dynamic measurements provide relevant information on the critical behavior. On the other hand, all the estimated static critical exponents ($\gamma /\nu $, $\beta /\nu $, and $1/\nu $ ) are in good agreement with RG predictions. However, the relevant exponents for the dynamic behavior ($z$ and $\sigma$) are slightly different from the RG predictions, most likely because they may depend on the specific Monte Carlo dynamics used (Metropolis in the present paper). This work was supported financially by CONICET, UNLP, and ANPCyT (Argentina). References ========== Y. Chen, S.H. Guo, Z. B. Li, S. Marculescu, L. Schuelke, Eur. Phys. J. B., [18]{}, 289 (2000). B. Bergersen, Z. Rácz, Phys. Rev. Lett., [67]{}, 3047 (1991). E. Luijten and H. W. J. Blote, Phys. Rev. B, [56]{}, 8945 (1997). B. Zheng, Int. J. Mod. Phys. B, [12]{}, 1419 (1998). T. Tomé and M. J. de Oliveira, Phys. Rev. E, [58]{}, 4242 (1998). K. Uzelac, Z. Glumac, and O.S. Barisíc. Eur. Phys. J. B., [63]{}, 101 (2008). H. K. Janssen, B. Schaub, and B. Schmittmann, Z. Phys. B: Cond. Matter, [73]{}, 539 (1989). M. Santos and W. Figueiredo, Phys. Rev. E, [62]{}, 1799 (2000). R. da Silva, N. A. Alves, and J. R. Drugowich de Felício, Phys. Rev. E, [66]{}, 026130 (2002). M. A. Bab, G. Fabricius, and E. V. Albano, Phys. Rev. E, [74]{}, 041123 (2006). M. A. Bab, G. Fabricius, and E. V. Albano, Europ. Lett, [81]{}, 10003 (2008). B. C. S. Grandi and W. Figueiredo, Phys. Rev. E, [70]{}, 056109 (2004). Y. Chen, Phys. Rev. E., [66]{}, 037104 (2002). Y. Chen, S. Guo, Z. Li, and A. Ye, Eur. Phys. J. B., [15]{}, 97 (2000). F. Baumann, S. B. Dutta, and M. Henkel, J. Phys. A. (Mat. & Gen.), [40]{}, 7389 (2007). G. Baglietto and E. V. Albano. Phys. Rev. E, [78]{}, 021125 (2008). K. Binder, Phys. Rep., [344]{}, 179 (2001).\] J. F. Nagle and J.C. Bonner, J. Phys. C (Cond-Mat.), [3]{}, 352 (1970). J. L. Monroe, R. Lucente, and J. P. Hourlland, J. Phys. A (Mat. & Gen.), [23]{}, 2555 (1990). K. Uzelac, and Z. Glumac, J. Phys. A (Mat. & Gen.), [21]{}, L421 (1988). J. L. Cardy, J. Phys. A (Mat. & Gen.), [17]{}, L385 (1984). P. C. Hohenberg and B.I. Halperin, Rev. Mod. Phys. [49]{}, 435 (1977). M. E. Fisher. S. K. Ma, and B. G. Nickel, Phys. Rev. Lett, [29]{}, 8945 (1972). A. Glumac and K. Uzelac, J. Phys. A (Mat. & Gen.), [22]{}, 4439 (1989). A. J. Bray, A. J. Briant, and D. K. Jervis, Phys. Rev. Lett. [84]{}, 1503 (2000). M. Barati and A. Ramazani, Phys. Rev. B, [62]{}, 12130 (2000).	{ "pile_set_name": "ArXiv" }
--- abstract: \| In this work, notion of a slant helix is extended to space E$^n$. Necessary and sufficient conditions to be a slant helix in the Euclidean $n-$space are presented. Moreover, we express some integral characterizations of such curves in terms of curvature functions.\ M.S.C. 2000: 53A04\ Keywords: Euclidean n-space; Frenet equations; Slant helices. author: - \| \ Mathematics Department\ Faculty of Science, Al-Azhar University\ Nasr City, 11448, Cairo, Egypt\ E-mail: [email protected]\ \ [^1]\ Department of Mathematics,\ Buca Educational Faculty, Dokuz Eylül University,\ 35160 Buca, Izmir, Turkey\ E-mail: [email protected], [email protected] title: 'Some Characterizations of Slant Helices in the Euclidean Space $\mathbf E^n$' --- Introduction and Statement of Results ===================================== Inclined curves or so-called general helices are well-known curves in the classical differential geometry of space curves [@mp] and we refer to the reader for recent works on this type of curves [@gl1; @sc]. Recently, Izumiya and Takeuchi have introduced the concept of slant helix in Euclidean 3-space E$^3$ saying that the normal lines makes a constant angle with a fixed direction [@izu]. They characterize a slant helix if and only if the function $$\label{eqma} \dfrac{\kappa^2}{(\kappa^2+\tau^2)^{3/2}}\Big(\dfrac{\tau}{\kappa}\Big)^{\prime}$$ is constant. In the same space, spherical images, the tangent and the binormal indicatrix and some characterizations of such curves are presented by [@ly]. With the notion of a slant helix, similar works are treated by the researchers, see [@ali; @al2; @ey; @mmhk; @ms]. In this work, we consider the generalization of the concept of a slant helix in the Euclidean n-space E$^n$.\ Let $\alpha:I\subset R \rightarrow E^n$ be an arbitrary curve in E$^n$. Recall that the curve $\alpha$ is said to be of unit speed (or parameterized by arc-length function $s$) if $\langle\alpha'(s),\alpha'(s)\rangle=1$, where $\langle,\rangle$ is the standard scalar product in Euclidean space E$^n$ given by $\langle X,Y\rangle=\sum_{i=1}^{n}\,x_i\,y_i,$ for each $X=(x_1,\ldots,x_n)$, $Y=(y_1,\ldots,y_n)\in E^n$. Let $\{V_{1}(s),\ldots,V_{n}(s)\}$ be the moving frame along $\alpha$, where the vectors $V_{i}$ are mutually orthogonal vectors satisfying $\langle V_{i}, V_{i}\rangle=1$. The Frenet equations for $\alpha$ are given by ([@gl1]) $$\left[ \begin{array}{c} V_{1}^{\prime } \\ V_{2}^{\prime } \\ V_{3}^{\prime } \\ \vdots \\ V_{n-1}^{\prime } \\ V_{n}^{\prime } \end{array} \right] =\left[ \begin{array}{ccccccc} 0 & \kappa _{1} & 0 & 0 & \cdots & 0 & 0 \\ -\kappa _{1} & 0 & \kappa _{2} & 0 & \cdots & 0 & 0 \\ 0 & -\kappa _{2} & 0 & \kappa _{3} & \cdots & 0 & 0 \\ \vdots & \vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\ 0 & 0 & 0 & 0 & \cdots & 0 & \kappa _{n-1} \\ 0 & 0 & 0 & 0 & \cdots & -\kappa _{n-1} & 0 \end{array} \right] \left[ \begin{array}{c} V_{1} \\ V_{2} \\ V_{3} \\ V_{4} \\ V_{5} \\ V_{6} \end{array} \right].$$ Recall the functions $\kappa_i(s)$ are called the i-th curvatures of $\alpha$. If $\kappa_{n-1}(s)=0$ for any $s\in I$, then $V_{n}(s)$ is a constant vector $V$ and the curve $\alpha$ lies in a $(n-1)$-dimensional affine subspace orthogonal to $V$, which is isometric to the Euclidean $(n-1)$-space E$^{n-1}$. We will assume throughout this work that all the curvatures satisfy $\kappa_i(s)\not=0$ for any $s\in I$, $1\leq i\leq n-1$. Here, recall that a regular curve with constant Frenet curvatures is called a $W-$curve [@ps]. A unit speed curve $\alpha:I\rightarrow E^n$ is called slant helix if its unit principal normal $V_2$ makes a constant angle with a fixed direction $U$. Our main result in this work is the following characterization of slant helices in Euclidean $n$-space E$^n$. \[th-main\] Let $\alpha:I\rightarrow E^n$ be a unit speed curve in E$^n$. Define the functions $$\label{u211} \begin{array}{ll} G_1=\int\kappa_{1}(s) ds,\ G_2=1,\ G_3=\dfrac{\kappa_1}{\kappa_2}G_1,\ G_{i}=\dfrac{1}{\kappa_{i-1}}\Big[\kappa_{i-2}G_{i-2}+G_{i-1}^{\prime}\Big], \end{array}$$ where $4\leq i\leq n$. Then $\alpha$ is a slant helix if and only if the function $$\label{u21} \sum_{i=1}^{n}\,G_i^2=C$$ is constant and non-zero. Moreover, the constant $C=\sec^2\theta$, being $\theta$ the angle that makes $V_2$ with the fixed direction $U$ that determines $\alpha$. This theorem generalizes in arbitrary dimensions what happens for $n=3$, namely: if $n=3$, (\[u21\]) writes $$1+\Big(1+\dfrac{\kappa_1^2}{\kappa_2^2}\Big)G_1^2=C.$$ It is easy to prove that: this equation is equivalent to $$\dfrac{\kappa_1^2}{(\kappa_1^2+\kappa_2^2)^{3/2}}\Big(\dfrac{\kappa_2}{\kappa_1}\Big)^{\prime}= \dfrac{1}{\sqrt{C-1}}$$ equation (\[eqma\]) where $C\neq 1.$ Proof of Theorem \[th-main\] ============================ Let $\alpha$ be a unit speed curve in E$^n$. Assume that $\alpha$ is a slant helix curve. Let $U$ be the direction with which $V_2$ makes a constant angle $\theta$ and, without loss of generality, we suppose that $\langle U,U\rangle=1$. Consider the differentiable functions $a_i$, $1\leq i\leq n$, $$\label{u3} U=\sum_{i=1}^{n}\,a_i(s)\,V_i(s),\ \ s\in I,$$ that is, $$a_i=\langle V_i,U\rangle,\ 1\leq i\leq n.$$ Then the function $a_2(s)=\langle V_2(s),U\rangle$ is constant, and it agrees with $\cos\theta$ as follows: $$\label{u311} a_2(s)=\langle V_2,U\rangle=\cos\theta$$ for any $s$. Because the vector field $U$ is constant, a differentiation in (\[u3\]) together (\[u211\]) gives the following system of ordinary differential equation: $$\label{u5} \left\{\begin{array}{ll} a_1^{\prime}-\kappa_1 a_2&=0\\ \kappa_1 a_1-\kappa_2 a_3&=0\\ a_{i}'+\kappa_{i-1} a_{i-1}-\kappa_{i} a_{i+1}&=0,\,\,\,\,\,3\leq i \leq n-1\\ a_n'+\kappa_{n-1} a_{n-1} &=0. \end{array}\right.$$ Let us define the functions $G_i=G_i(s)$ as follows $$\label{u51} a_i(s)=G_i(s)\,a_2, \ 1\leq i\leq n.$$ We point out that $a_2\not=0$: on the contrary, (\[u51\]) gives $a_i=0$, for $1\leq i\leq n$ and so, $U=0$, which is a contradiction. Since, the first $n$-equations in (\[u5\]) lead to $$\label{u6} \left\{ \begin{array}{ll} &G_1=\int\kappa_{1}(s) ds\\ &G_2=1\\ &G_3=\dfrac{\kappa_1}{\kappa_2}G_1\\ &G_{i}=\dfrac{1}{\kappa_{i}}\Big[\kappa_{i-2}G_{i-2}+G_{i-1}^{\prime}\Big],\,\,\,\,\,4\leq i \leq n.\\ \end{array}\right.$$ The last equation of (\[u5\]) leads to the following condition; $$\label{u62} G_n'+\kappa_{n-1} G_{n-1}=0.$$ We do the change of variables: $$t(s)=\int^s\kappa_{n-1}(u) du,\hspace{.5cm}\frac{dt}{ds}=\kappa_{n-1}(s).$$ In particular, and from the last equation of (\[u6\]), we have $$G_{n-1}'(t)=G_n(t)-\Big(\frac{\kappa_{n-2}(t)}{\kappa_{n-1}(t)}\Big)G_{n-2}(t).$$ As a consequence, if $\alpha$ is a slant helix, substituting the equation (\[u62\]) to the last equation, we express $$G_n''(t)+G_n(t)=\frac{\kappa_{n-2}(t)G_{n-2}(t)}{\kappa_{n-1}(t)}.$$ By the method of variation of parameters, the general solution of this equation is obtained $$\label{u9} G_n(t)=\Big(A-\int\frac{\kappa_{n-2}(t)G_{n-2}(t)}{\kappa_{n-1}(t)}\sin{t}\,dt\Big)\cos{t}+ \Big(B+\int\frac{\kappa_{n-2}(t)G_{n-2}(t)}{\kappa_{n-1}(t)}\cos{t}\,dt\Big)\sin{t},$$ where $A$ and $B$ are arbitrary constants. Then (\[u9\]) takes the following form $$\label{u10} \begin{array}{ll} G_n(s)=&\Big(A-\int\Big[\kappa_{n-2}(s)G_{n-2}(s)\sin{\int\kappa_{n-1}(s)ds}\Big]ds\Big)\cos{\int\kappa_{n-1}(s)ds}\\ &+ \Big(B+\int\Big[\kappa_{n-2}(s)G_{n-2}(s)\cos{\int\kappa_{n-1}(s)ds}\Big]ds\Big)\sin{\int\kappa_{n-1}(s)ds}. \end{array}$$ From (\[u62\]), the function $G_{n-1}$ is given by $$\label{u11} \begin{array}{ll} G_{n-1}(s)=&\Big(A-\int\Big[\kappa_{n-2}(s)G_{n-2}(s)\sin{\int\kappa_{n-1}(s)ds}\Big]ds\Big)\sin{\int\kappa_{n-1}(s)ds}\\ &-\Big(B+\int\Big[\kappa_{n-2}(s)G_{n-2}(s)\cos{\int\kappa_{n-1}(s)ds}\Big]ds\Big)\cos{\int\kappa_{n-1}(s)ds}. \end{array}$$ From Equation (\[u6\]), we have $$\begin{aligned} \sum_{i=1}^{n-2}G_i G_i^{\prime}&=&G_1G_1'+G_2G_2'+\sum_{i=3}^{n-2}G_i G_i^{\prime}\\ &=&\kappa_1G_1+\sum_{i=3}^{n-2}G_i\Big[\kappa_iG_{i+1}-\kappa_{i-1}G_{i-1}\Big]\\ &=&\kappa_1G_1+\sum_{i=3}^{n-2}\Big[\kappa_iG_iG_{i+1}-\kappa_{i-1}G_{i-1}G_i\Big]\\ &=&\kappa_1G_1+\kappa_{n-2}G_{n-2}G_{n-1}-\kappa_{2}G_{2}G_3\\ &=&\kappa_{n-2}G_{n-2}G_{n-1}.\end{aligned}$$ Substituting (\[u11\]) to the above equation and integrating it, we have: $$\label{u15} \begin{array}{ll} \sum_{i=1}^{n-2}G_{i}^{2}&=C-\Big(A-\int\Big[\kappa_{n-2}(s)G_{n-2}(s)\sin{\int\kappa_{n-1}ds}\Big]ds\Big)^2\\ &- \Big(B+\int\Big[\kappa_{n-2}(s)G_{n-2}(s)\cos{\int\kappa_{n-1}ds}\Big]ds\Big)^2, \end{array}$$ where $C$ is a constant of integration. Using equations (\[u10\]) and (\[u11\]), we have $$\label{u151} \begin{array}{ll} G_n^2+G_{n-1}^2&=\Big(A-\int\Big[\kappa_{n-2}(s)G_{n-2}(s)\sin{\int\kappa_{n-1}ds}\Big]ds\Big)^2\\ &+ \Big(B+\int\Big[\kappa_{n-2}(s)G_{n-2}(s)\cos{\int\kappa_{n-1}ds}\Big]ds\Big)^2, \end{array}$$ It follows from (\[u15\]) and (\[u151\]) that $$\sum_{i=3}^{n}G_{i}^{2}=C.$$ Moreover this constant $C$ can be calculated as follows. From (\[u51\]), together the $(n-2)$-equations (\[u6\]), we have $$C=\sum_{i=1}^{n}\,G_i^2=\dfrac{1}{a_2^2}\sum_{i=1}^{n}\,a_i^2=\dfrac{1}{a_2^2}=\sec^2\theta,$$ where we have used (\[u21\]) and the fact that $U$ is a unit vector field. We do the converse of Theorem. Assume that the condition (\[u6\]) is satisfied for a curve $\alpha$. Let $\theta\in R$ be so that $C=\sec^2\theta$. Define the unit vector $U$ by $$U=\cos\theta\Big[\sum_{i=1}^{n}\,G_i\,V_i\Big].$$ By taking account (\[u6\]), a differentiation of $U$ gives that $\dfrac{dU}{ds}=0$, which it means that $U$ is a constant vector field. On the other hand, the scalar product between the unit tangent vector field $V_2$ with $U$ is $$\langle V_2(s),U\rangle=\cos\theta.$$ Thus, $\alpha$ is a slant helix in the space E$^{n}$. As a direct consequence of the proof, we generalize theorem \[th-main\] in Minkowski space for timelike curves and give an another theorem which characterizes slant helices with constant curvatures. Let E$_1^n$ be the Minkowski n-dimensional space and let $\alpha:I\rightarrow E_1^n$ be a unit speed timelike curve. Then $\alpha$ is a slant helix if and only if the function $\sum_{i=i}^{n}\,G_i^2$ is constant, where the functions $G_i$ are defined as in (\[u211\]). The proof carries the same steps as above and we omit the details. We only point out that the fact that $\alpha$ is timelike means that $V_1(s)=\alpha'(s)$ is a timelike vector field. The other $V_i$ in the Frenet frame, $2\leq i\leq n$, are unit spacelike vectors and so, the second equation in Frenet equations changes to $V_2'=\kappa_1V_1+\kappa_2V_3$ (for details of Frenet equations see [@ehi]). There are no slant helices with constant and non-zero curvatures ($W-$slant helices, i.e.) in the space E$^n$. Let us suppose a slant helix with constant and non-zero curvatures. Then the equations in (\[u5\]) and (\[u6\]) hold. Since, we easily have for odd $i$, $G_{i}=\delta _{i}s$, where $\delta_{i}\in R$ and for even $i,$ $G_{i}=\delta _{i}$. Then, we form $$\sum\limits_{i=1}^{n}G_{i}^{2}=(\delta _{1}s)^{2}+\delta _{2}^{2}+(\delta _{3}s)^{2}+\delta _{4}^{2}+...$$ and it is easy to say that $\sum\limits_{i=1}^{n}G_{i}^{2}$ is nowhere constant. By the theorem \[th-main\], we arrive at that there does not exist a slant helix with constant and non-zero curvatures in the space E$^n$. Further Characterizations of Slant Helices in E$^n$ =================================================== In this section we present new characterizations of slant helix in E$^n$. The first one is a consequence of Theorem \[th-main\]. \[th-2\] Let $\alpha:I\subset R\rightarrow E^n$ be a unit speed curve in Euclidean space E$^n$. Then $\alpha$ is a slant helix if and only if there exists a $C^2$-function $G_{n}(s)$ such that $$\label{u24} G_{n}=\dfrac{1}{\kappa_{n-1}}\Big[\kappa_{n-2}G_{n-2}+G_{n-1}^{\prime}\Big],\,\,\, \dfrac{dG_{n}}{ds}=-\kappa_{n-1}(s)G_{n-1}(s),$$ where $$G_1=\int\kappa_{1}(s) ds, G_2=1, G_3=\dfrac{\kappa_1}{\kappa_2}G_1, G_{i}=\dfrac{1}{\kappa_{i-1}}\Big[\kappa_{i-2}G_{i-2}+G_{i-1}^{\prime}\Big],\ 4\leq i\leq n-1.$$ Let now assume that $\alpha$ is a slant helix. By using Theorem \[th-main\] and by differentiation the (constant) function given in (\[u21\]), we obtain $$\begin{aligned} 0&=&\sum_{i=i}^{n}G_i\,G_i^{\prime}\\ &=&G_1\kappa_1+G_3\Big(\kappa_3G_4-\kappa_2G_2\Big)+G_4\Big(\kappa_4G_5-\kappa_3G_3\Big)+...\\ & &+G_{n-1}\Big(\kappa_{n-1}G_{n}-\kappa_{n-2}G_{n-2}\Big)+G_{n}G_{n}^{\prime}\\ &=&G_{n}\Big(G_{n}'+\kappa_{n-1}G_{n-1}\Big).\end{aligned}$$ This shows (\[u24\]). Conversely, if (\[u24\]) holds, we define a vector field $U$ by $$U=\cos\theta\Big[\sum_{i=1}^{n}\,G_i\,V_i\Big].$$ By the Frenet equations, $\dfrac{dU}{ds}=0$, and so, $U$ is constant. On the other hand, $\langle V_2(s),U\rangle=\cos\theta$ is constant, and this means that $\alpha$ is a slant helix. We end giving an integral characterization of a slant helix. Let $\alpha:I\subset R\rightarrow E^n$ be a unit speed curve in Euclidean space E$^n$. Then $\alpha$ is a slant helix if and only if the following condition is satisfied $$\label{u244} \begin{array}{ll} G_{n-1}(s)=&\Big(A-\int\Big[\kappa_{n-2}G_{n-2}\sin{\int\kappa_{n-1}ds}\Big]ds\Big)\sin{\int^s\kappa_{n-1}(u)du}\\ &-\Big(B+\int\Big[\kappa_{n-2}G_{n-2}\cos{\int\kappa_{n-1}ds}\Big]ds\Big)\cos{\int^s\kappa_{n-1}(u)du}. \end{array}$$ for some constants $A$ and $B$. Suppose that $\alpha$ is a slant helix. By using Theorem \[th-2\], let define $m(s)$ and $n(s)$ by $$\phi(s)=\int^s\kappa_{n-1}(u)du,$$ $$\label{u26} \begin{array}{ll} m(s)=&G_{n}(s)\cos\phi+G_{n-1}(s)\sin\phi+\int\,\kappa_{n-2}G_{n-2}\sin\phi\,ds,\\ n(s)=&G_{n}(s)\sin\phi-G_{n-1}(s)\cos\phi-\int\,\kappa_{n-2}G_{n-2}\cos\phi\,ds. \end{array}$$ If we differentiate equations (\[u26\]) with respect to $s$ and taking into account of (\[u244\]) and (\[u24\]), we obtain $\dfrac{dm}{ds}=0$ and $\dfrac{dn}{ds}=0$. Therefore, there exist constants $A$ and $B$ such that $m(s)=A$ and $n(s)=B$. By substituting into (\[u26\]) and solving the resulting equations for $G_{n-1}(s)$, we get $$G_{n-1}(s)=\Big(A-\int\,\kappa_{n-2}G_{n-2}\sin{\phi}\,ds\Big)\sin{\phi}- \Big(B+\int\,\kappa_{n-2}G_{n-2}\cos{\phi}\,ds\Big)\cos{\phi}.$$ Conversely, suppose that (\[u244\]) holds. In order to apply Theorem \[th-2\], we define $G_n(s)$ by $$G_n(s)=\Big(A-\int\,\kappa_{n-2}G_{n-2}\sin{\phi}\,ds\Big)\cos{\phi}+ \Big(B+\int\,\kappa_{n-2}G_{n-2}\cos{\phi}\,ds\Big)\sin{\phi}.$$ with $\phi(s)=\int^s\kappa_{n-1}(u) du$. A direct differentiation of (\[u244\]) gives $$G_{n-1}^{\prime}=\kappa_{n-1}G_{n}-\kappa_{n-2}G_{n-2}.$$ This shows the left condition in (\[u24\]). Moreover, a straightforward computation leads to $G_{n}'(s)=-\kappa_{n-1}G_{n-1}$, which finishes the proof. ACKNOWLEDGEMENTS: The second author would like to thank Tübitak-Bideb for their financial supports during his Ph.D. studies. [99]{} Ali, A. Inclined curves in the Euclidean 5-space E$^5$, J. Advanced Research in Pure Math., 1* (1), 15–22, 2009. Ali, A. and López, R. Slant helices in Minkowski space E$_1^3$, preprint 2008: arXiv:0810.1464v1 \[math.DG\]. Ali, A. and López, R. Timelike B$_{2}$-slant helices in Minkowski space E$_1^4$, preprint 2008:arXiv:0810.1460v1 \[math.DG\]. Ekmekci, N., Hacisalihoglu, H.H. and Ilarslan, K. Harmonic Curvatures in Lorentzian Space, Bull. Malaysian Math. Soc. (Second Series), 23 (2), 173-179, 2000. Erdogan, M. and Yilmaz, G. Null generalized and slant helices in 4-dimensional Lorentz-Minkowski space, Int. J. Contemp. Math. Sci., 3 (23), 1113-1120, 2008. Gluck, H. Higher curvatures of curves in Euclidean space, Amer. Math. Monthly, 73, 699–704, 1966. Izumiya, S. and Takeuchi, N. New special curves and developable surfaces, Turk J. Math., 28 (2), 531–537, 2004. Kula, L. and Yayli, Y. On slant helix and its spherical indicatrix, Appl. Math. Comput. 169 (1), 600–607, 2005. Milman, R.S. and Parker, G.D. Elements of Differential Geometry, Prentice-Hall Inc., Englewood Cliffs, New Jersey, 1977. Petrovic-Torgasev, M. and Sucurovic, E. W-curves in Minkowski spacetime, Novi. Sad. J. Math. 32 (2), 55–65, 2002. Onder, M., Kazaz, M., Kocayigit, H. and Kilic, O. B$_{2}$-slant helix in Euclidean 4-space E$^4$, Int. J. Contemp. Math. Sci., 2008, 3(29): 1433-1440. Scofield, P.D. Curves of constant precession, Amer. Math. Monthly, 102, 531–537, 1995. Turgut, M. and Yilmaz, S. Some characterizations of type-3 slant helices in Minkowski space-time, Involve J. Math., 2 (1), 115-120, 2009 [^1]: Corresponding author.	{ "pile_set_name": "ArXiv" }
--- abstract: 'In a Bayesian analysis, the likelihood that specific candidate parameters govern the evolution of a quantum system are conditioned on the outcome of measurements which, in turn, cause measurement backaction on the state of the system \[M. Tsang, Phys. Rev. Lett. [108]{}, 170502 (2012)\]. Specializing to the distinction of two candidate hypotheses, we study the achievements of continuous monitoring of the radiation emitted by a quantum system followed by an optimal projective measurement on its conditioned final state. Our study of the radiative decay of a driven two-level system shows an intricate interplay between the maximum information available from photon counting and homodyne detection and the final projective measurement on the emitter. We compare the results with theory predicting a lower bound for the probability to assign a wrong hypothesis by any combined measurement on the system and its radiative environment.' author: - Alexander Holm Kiilerich - Klaus Mølmer title: Hypothesis testing with a continuously monitored quantum system --- Introduction {#sec:1} ============ Hypothesis testing is the task of assigning one of a discrete set of models to describe an observed system. Measurements on quantum systems have random outcomes and the discrimination of two hypotheses $h_0$ and $h_1$ is a statistical inference problem. Viz. there is a probability $P(m=i\|h_j)$ that measurement data processed to yield a binary outcome $m=0,1$ is (in)consistent with the true hypothesis $(i\neq j)i=j$ and a corresponding average probability that an erroneous hypothesis will be assigned, $$\begin{aligned} \label{eq:Qe}{Q_{\mathrm{e}}}= P(m=1\|h_0)P(h_0)+ P(m=0\|h_1)P(h_1),\end{aligned}$$ where $P(h_0)$ and $P(h_1)$ are the prior probabilities of each hypothesis. Distinguishing two different Hamiltonians $\hat{H}_0$ and $\hat{H}_1$, governing the evolution of a closed quantum system, is achieved by discriminating the two quantum states $\rho_0(t) = \ket{\psi_0(t)}\bra{\psi_0(t)}$ (hypothesis $h_0$) and $\rho_1(t) = \ket{\psi_1(t)}\bra{\psi_1(t)}$ (hypothesis $h_1$), resulting from time evolution under each candidate Hamiltonian from a common initial state of the system. Only orthogonal states can be discriminated unambiguously while, in general, the overlap between the candidate states defines a minimum error probability for any measurement protocol, ${Q_{\mathrm{e}}}\geq {Q_{\mathrm{e}}}^{(\text{min})}$ where [@Helstrom1969] $$\begin{aligned} \label{eq:QeMinPure}{Q_{\mathrm{e}}}^{(\text{min})} = \frac{1}{2}\left(1-\sqrt{1-4P(h_0)P(h_1)\|\bra{\psi_0}\psi_1\rangle\|^2}\right).\end{aligned}$$ As derived by Helstrom [@Helstrom1969], this bound can be saturated by performing a projective measurement of the operator $$\begin{aligned} \label{eq:PiPure}\hat{A} = P(h_0)\rho_0-P(h_1)\rho_1,\end{aligned}$$ and assigning hypothesis $h_0(h_1)$ if the outcome is one of the positive(negative) eigenvalues of $\hat{A}$. ![image](generalMeasurement.pdf){width="1.9\columnwidth"} In this article, we study the use of an open quantum system to distinguish between different Hamiltonian hypotheses. This can, in principle, be accomplished by measuring the optimal observable (\[eq:PiPure\]) where the $\ket{\psi_i(t)}$ denote the combined states of the system and its environment. We focus on the common example of a quantum system coupled to a broadband radiation reservoir. A driven system and the quantized radiation field evolves into entangled states in an infinite dimensional Hilbert space. While these states may differ significantly for different Hamiltonians the, potentially, highly non-local projective measurement (\[eq:PiPure\]) of combined system and environment observables, see [ \[fig:intro\]]{}(a), becomes difficult to achieve in practice. Instead, as illustrated in [ \[fig:intro\]]{}, one often has recourse to perform photon counting (b) or field quadrature measurements (e.g. homodyne detection (c)) on the environment. If the candidate Hamiltonians cause the system to evolve into different steady states, the mean number of emitted photons or the mean homodyne detection signal may be averaged over long enough time to suppress statistical uncertainty about their values such that the Hamiltonian can be inferred with certainty. Faster, and hence more efficient, inference can be made from observation of the correlations in the full noisy measurement record. To give an example, the steady state yields identical emission rates from atoms excited by one of two strong laser fields, while the time intervals between photon detection events follow distinct oscillatory waiting time distributions. Optimal inference from any measurement record is obtained by a Bayesian analysis which yields the probabilities $P(h_0\|{D_t})$ and $P(h_1\|{D_t})$ ascribed to each hypothesis based on their prior probabilities and on the full data record ${D_t}$ retrieved until the time $t$ [@1355-5111-8-6-002; @PhysRevA.64.042105; @PhysRevLett.108.170502; @PhysRevA.87.032115; @PhysRevA.79.022314; @PhysRevA.95.022306; @PhysRevA.94.032103; @PhysRevA.89.052110; @PhysRevA.91.012119]. In this work, we investigate to what extent supplementing continuous monitoring of the emitted radiation from the initial time $t=0$ to a final time $t=T$ by a final projective measurement on the emitter system, allows better distinction between different hypotheses governing the system dynamics. Due to the measurement backaction associated with continuous monitoring of the environment, the state of the emitter evolves in a conditional manner according to the stochastic measurement signal [@QMC]. In any particular realization of the measurement sequence, the optimal final measurement on the system (\[eq:PiPure\]) is thus conditioned on the detection record ${D_T}$ obtained up until the final time $T$, $$\begin{aligned} \label{eq:PiConditional} \hat{A}^{D_T}= P(h_0\|{D_T})\rho^{({D_T})}_0(T)-P(h_0\|{D_T})\rho^{({D_T})}_1(T).\end{aligned}$$ Here the information extracted from the environment is incorporated in the conditional candidate states $\rho_i^{(D_T)}(T)$ and their probabilities updated by Bayes rule, $P(h_i) \rightarrow P(h_i\|{D_T})$. The continuous monitoring and conditioned evolution of quantum states have for instance been realized in experiments with superconducting qubits [@murch2013observing; @PhysRevA.96.022104; @PhysRevX.6.011002] and optomechanical systems [@PhysRevLett.114.223601]. After homodyne or heterodyne detection of the radiation signal has been performed until time $T$ on for example a super conducting qubit, a final system projection can be achieved in these experiments by applying a strong, dispersively coupled probe field [@murch2013observing; @PhysRevLett.114.090403]. We compare such realistic measurement strategies with the theoretical limit for distinguishing different hypotheses. See also [@Jacobs2007FeedbackCF] for an alternative, adaptive approach to hypothesis testing and state discrimination with continuous measurements. The article is organized as follows. In Section \[sec:2\] we outline the main ideas of hypothesis testing with monitored quantum systems and we recall a lower (quantum) bound for the error probability. In Section \[sec:3\] we present numerical simulations which illustrate and exemplify different aspects of our theory. In Section \[sec:4\] we provide a conclusion and an outlook. Bayesian analysis of a measurement record {#sec:2} ========================================= We consider a system subject to a sequence of measurements or continuous monitoring from time $t=0$ to a final time $t=T$. During this phase, a signal ${dD_t}$ is recorded and by ${D_t}$ we denote the full signal obtained between time $0$ and $t$. Under hypothesis $h_i$ any given realization of ${D_t}$ has a probability $P(D_t\|h_i)$ determined from the conditional candidate quantum state ${\rho^{({D_t})}}_i(t)$. Bayes rule yields the corresponding update of the likelihood $P(h_i\|D_t)$ assigned to each hypothesis, $$\begin{aligned} \label{eq:BayesMonitor} P(h_i\|{D_t}) = \frac{P({D_t}\|h_i)P(h_i)}{\sum_j P({D_t}\|h_j)P(h_j)}.\end{aligned}$$ The backaction of the measurement associated with the outcome ${dD_t}$ applies directly on the current state of the system, ${\rho^{({D_t})}}(t) \rightarrow {\hat{M}({dD_t}){\rho^{({D_t})}}(t)\hat{M}^\dagger({dD_t})}/{P({dD_t})}$. Here the sum (integral) of the positive-operator valued measure (POVM) over all possible detection outcomes yields the identity, $\sum_{dD_t} \hat{M}^\dagger({dD_t})\hat{M}({dD_t}) = I$. The POVM formalism includes both projective measurements, in which case the $\hat{M}({dD_t})$ denote projection operators, as well as more general measurements, involving, e.g., projective measurements on ancilla systems after they have interacted with the system. Between measurements, the system evolves subject to the Hamiltonian that we want to discriminate. If the state is not renormalized after application of the POVM backaction operators, we retain the evolution of an unnormalized state ${\tilde{\rho}^{({D_t})}}(t)$, $$\begin{aligned} \label{eq:POVMunnormlized} {\tilde{\rho}^{({D_t})}}(t) \rightarrow \hat{M}({dD_t}){\tilde{\rho}^{({D_t})}}(t)\hat{M}^\dagger({dD_t}),\end{aligned}$$ whose reduction in norm is just the probability to obtain the signal ${dD_t}$. This implies that at the final time $T$, the probability $P(dD_T)\cdots P(dD_{2dt})P(dD_{dt})P(dD_0)$ for the full signal ${D_T}$ is given by the trace of ${\tilde{\rho}^{({D_t})}}(T)$. Hence, by evolving the unnormalized state under each of the two candidate hypotheses conditioned on the signal actually recorded in a given experiment, one may by [Eq. ]{}[eq:BayesMonitor]{} obtain the relative likelihoods of each hypothesis as $P(h_i\|{D_t})\propto \mathrm{Tr}({\tilde{\rho}^{({D_t})}}_i(t))$. Since any specific trajectory for ${D_t}$ is very unlikely, $\mathrm{Tr}({\tilde{\rho}^{({D_t})}}_i(T))$ becomes very small even for the true hypothesis and for numerical purposes it is favorable to propagate instead the log-likelihood $\log[P(h_i\|{D_t})]$. See [@PhysRevA.87.032115] for a detailed account of Bayesian inference with continuously monitored quantum systems. In the next subsection we specialize to cases, where the measurements are carried out continuously in time on the radiation field emitted by the quantum system of interest. The two generic setups of counting-type measurements with discrete detection events and diffusion-type measurements with continuous but infinitesimal backaction are discussed, and [Eq. ]{}[eq:POVMunnormlized]{} is replaced by stochastic master equations, suitable for numerical propagation of ${\tilde{\rho}^{({D_t})}}(t)$. For simplicity we assume that there is only a single decay channel but the expressions may readily be generalized to multi-channel cases and alternative environmental couplings. Photon counting and homodyne detection -------------------------------------- In [ \[fig:intro\]]{}(b), the florescence from the probe system is detected by a photon counter with quantum efficiency $0\leq\eta\leq1$ and the photon counting signal $N_t$ until time $t$ constitutes the detection record $D_t$. During each short time interval $dt$ there are two possible detection outcomes: no photon $dN_t = 0$ or one photon $dN_t = 1$, where $P(dN_t = 1)= \eta{\mathrm{Tr}\left({\hat{c}^\dagger}{\hat{c}}{\rho^{(N_t)}}(t)\right)}dt$ is given by the (normalized) state ${\rho^{(N_t)}}(t)$ of the system. Here ${\hat{c}}=\sqrt{\gamma}\|g\rangle \langle e\|$ denotes the quantum jump operator from an excited $\|e\rangle$ to a lower state $\|g\rangle$. The conditional evolution of the unnormalized state, in turn, obeys a linear stochastic master equation [@PhysRevLett.68.580; @QMC], $$\begin{aligned} \label{eq:meCount} \begin{split} d{\tilde{\rho}^{(N_t)}}=& \left(\mathcal{K}dt+ \mathcal{B}dN_t\right){\tilde{\rho}^{(N_t)}}, \end{split}\end{aligned}$$ where $\mathcal{K}\rho = -i[{\hat{H}},\rho] +(1-\eta){\hat{c}}\rho{\hat{c}^\dagger}-\frac{1}{2}\{{\hat{c}^\dagger}{\hat{c}},\rho\}$ and $\mathcal{B}\rho = \eta\left({\hat{c}}\rho{\hat{c}^\dagger}-\rho\right)$. As depicted in [ \[fig:intro\]]{}(c), a homodyne detector mixes the florescence with a strong local oscillator field on a beam splitter, and the signal $dY_t$ is obtained as the intensity difference between the two output ports. Homodyne detection is sensitive to the phase of the emitted radiation which may be favorable when probing certain dynamics of the system. The recorded signal $dY_t$ in each short time interval $dt$ has a mean value determined by the current state ${\rho^{(Y_t)}}(t)$ of the system, $$\begin{aligned} \label{eq:dY} dY_t = {\mathrm{Tr}\left(\mathcal{X}_\Phi{\rho^{(Y_t)}}(t)\right)}dt+dW_t.\end{aligned}$$ with $\mathcal{X}_\Phi\rho =\sqrt{\eta}\left({\hat{c}}\mathrm{e}^{-i\Phi}\rho+\rho{\hat{c}^\dagger}\mathrm{e}^{i\Phi}\right)$ where $\Phi$ is the phase of the local oscillator. Random, white-noise fluctuations around the mean are represented by infinitesimal Wiener increments which are uncorrelated, normal distributed stochastic elements with zero mean and variance $dt$. Since the signal depends only weakly on the state of the system, the backaction associated with homodyne detection is infinitesimal and [Eq. ]{}[eq:POVMunnormlized]{} is equivalent to a diffusion type linear stochastic master equation for the conditional evolution of the unnormalized state [@QMC], $$\begin{aligned} \label{eq:meHomo} d{\tilde{\rho}^{(Y_t)}}= \left(\mathcal{L} dt+\mathcal{\mathcal{X}}_\Phi dY_t \right) {\tilde{\rho}^{(Y_t)}},\end{aligned}$$ where $\mathcal{L}\rho = -i[{\hat{H}},\rho] +{\hat{c}}\rho{\hat{c}^\dagger}-\frac{1}{2}\{{\hat{c}^\dagger}{\hat{c}},\rho\}$. Upon acquiring a measurement signal, the relevant stochastic master equation, (\[eq:meCount\]) or (\[eq:meHomo\]), may be solved for each hypothesis. The corresponding candidate states are all initialized in the (known) initial state of the system, but normalized to the prior probabilities assigned the particular hypothesis, $\mathrm{Tr}(\rho_i^{D_0}(t=0)) = P(h_i)$. This way the evolving likelihood distribution over the possible hypotheses is directly given by the traces of the corresponding conditioned density matrices, $\tilde{\rho}_0^{D_t}(t),\ \tilde{\rho}_1^{D_t}(t)$. To illustrate the Bayesian inference protocol, we simulate in [ \[fig:homodyneCountingComparison\]]{} perfect monitoring of a two-level system with the purpose of discriminating two hypotheses for the resonant driving with a Rabi frequency of either $\Omega_0$ or $\Omega_1$. I.e, we test the two Hamiltonian hypotheses: $ {\hat{H}}_0 = \frac{\hbar\Omega_0}{2}{\hat{\sigma}_\mathrm{x}}$ and $ {\hat{H}}_1 = \frac{\hbar\Omega_1}{2}{\hat{\sigma}_\mathrm{x}}$. The signals, $dN_t$ from photon counting and $dY_t$ from homodyne detection, in the upper panels of (a) and (b) are sampled from the true hypothesis which we assume to be $h_0$. Conditioned on these signals, the (unnormalized) candidate states ${\tilde{\rho}^{({D_t})}}_i(t)$ with ${D_t}=N_t,Y_t$ evolve according to Eqs. (\[eq:meCount\]) and (\[eq:meHomo\]), respectively. Their traces and the condition $P(h_0\|{D_t})+P(h_1\|{D_t})=1$ yield the time evolution of the inferred probabilities for each hypothesis as shown in the lower panels of (a) and (b). We assume equal priors $P(h_0)= P(h_1) = 1/2$. ![ Simulated monitoring of a driven two-level system by (a) photon counting and (b) homodyne detection with the purpose of discriminating two hypotheses $h_0$ ($\Omega_0 = 2\gamma$) and $h_1$ ($\Omega_1 = 4\gamma$) for the Rabi frequency. The simulations are made assuming $h_0$ to be the true hypothesis and with a detector efficiency $\eta=1$ and in (b) a local oscillator phase $\Phi = -\pi/2$. The second and fourth panels show the evolution of the probabilities (\[eq:BayesMonitor\]) for the two hypotheses conditioned on (a) the photon counting signal and (b) the noisy homodyne current shown in the first and third panels. The lower panel (c), shows the $z$-component $z_{\hat{A}_t^{{D_t}}} \propto {\mathrm{Tr}({\hat{\sigma}_\mathrm{z}}\hat{A}_t^{{D_t}})}$ of the optimal Pauli measurement observable (clarified in the main text) if monitoring is stopped at any given time. We observe that this optimal system measurement differs for the three cases of counting, homodyne detection and unobserved, dissipative emitter dynamics. []{data-label="fig:homodyneCountingComparison"}](Pi_conditional.pdf){width="0.95\columnwidth"} For photon counting in (a) the probability updates are dominated by three photo detection events while periods with no detections lead to a less pronounced, continuous update. The noisy homodyne signal in (b), on the other hand, holds only very little information in each individual time-bin and here the probabilities continuously converge to reveal the true hypothesis. In both cases, at the final time $t=5\gamma^{-1}$ the accumulated signals are seen to favor the true hypothesis ($h_0$) with almost unit probability. A figure of merit for a particular measurement strategy is the speed at which we arrive at perfect distinction. Supplementing continuous monitoring by a projective measurement --------------------------------------------------------------- If the hypotheses are not sufficiently discriminated at the end of the probing at time $T$, it may be possible to extract further information by a direct measurement on the emitter system. Due to the continuous monitoring, the emitter is assigned the conditional candidate states $\rho^{({D_T})}_i(T)$, while the probabilities that we ascribe to these states, $P(h_i\|{D_T})$ are given by the traces of the unnormalized density matrices. The optimal projective measurement we can perform on the system then concerns the system observable $\hat{A}^{D_T}_T$ defined in [Eq. ]{}[eq:PiConditional]{}. For a two-level system, the projective measurement of any observable $\hat{A}$ is equivalent to the measurement of a Pauli spin component along a specific unit vector $(x_A,y_A,z_A)$ with $u_A\propto {\mathrm{Tr}(\hat{\sigma}_u A)}$. In [ \[fig:homodyneCountingComparison\]]{}(c) we visualize the optimum observable $\hat{A}^{D_T}_T$ if the continuous monitoring, yielding the signals in the upper panels of (a) and (b), is terminated at the corresponding point in time. In this example the unit vector, designating the direction of the spin measurement, is confined to the $(y,z)$-plane and we show its $z$-component During each experimental realization, $\hat{A}_t^{{D_t}}$ assumes a stochastic value, which is different from the one that optimally discriminates the states of an unobserved system governed by the corresponding Lindblad master equation $d\rho/dt = \mathcal{L}\rho$. With homodyne detection the measurement observable, represented by the blue noisy trace in [ \[fig:homodyneCountingComparison\]]{}(c), is seen to fluctuate around the full, yellow curve, pertaining to the unmonitored system, while with photon counting, large deviations arise accompany the quantum jumps of the system state. The possible eigenvalues $\lambda$ of the measurement observable $\hat{A}^{D_T}_T$ occur under hypothesis $h_i$ with probability $P(\lambda\|h_i)={\mathrm{Tr}\left(\Pi_\lambda \rho_i(t)\right)}$, where $\Pi_\lambda$ is the projector on the affiliated eigenstate of the operator $\hat{A}^{D_T}_T$. According to Bayes rule the combined information from the monitoring and from the system projection hence leads to an update of the probabilities assigned to each hypothesis $$\begin{aligned} \label{eq:assignProb} P(h_i\|{D_T},\lambda) = \frac{P(\lambda\|h_i)P(h_i\|{D_T})}{P(\lambda)}.\end{aligned}$$ The hypothesis $h_m$ with the largest likelihood $P(h_m\|{D_T},\lambda)$ is the preferred one, and averaged over many independent realizations of the final projective measurement, the fraction of erroneous assignments based on that choice will be given by the generalization of [Eq. ]{}[eq:QeMinPure]{} to mixed states, $$\begin{aligned} \label{eq:QeMin} {Q_{\mathrm{e}}}= \frac{1}{2}\left[1-\left\|P(h_0\|{D_T})\rho^{({D_T})}_0(T) -P(h_1\|{D_T}) \rho^{({D_T})}_1(T)\right\|\right],\end{aligned}$$ where $\|O\| \equiv {\mathrm{Tr}\left(\sqrt{O^\dagger O}\right)}$. To obtain the error probability of a given measurement scheme, we however still need to numerically evaluate the conditional states and probabilities and average [Eq. ]{}[eq:QeMin]{} over the random outcomes of the continuous monitoring. Note that [Eq. ]{}[eq:QeMin]{} can also be applied to the distinction of mixed states or of the (unconditioned) candidate density matrices of a system evolving under different Hamiltonian hypotheses and leaking into an un-monitored environment. A recent comparison of probing by measurements on a system alone and on both a system and its environment shows the ability of the latter to better exploit (initial) entanglement among its sub-components [@albarelli2018restoring]. The quantum bound {#sec:Qbound} ----------------- The minimum achievable error associated with any hypothetical detection of the radiation emitted by a system and a final detection on that system itself is determined by our ability to discriminate the pure states of the combined system and environment, resulting from the different Hamiltonian hypotheses. These (un-monitored) states are themselves intractable by numerical means, but if the Born-Markov approximation applies for the radiative emission process, their quantum overlap can be evaluated as the trace of an effective density matrix ${\rho_{01}}(t)$ acting only on the state space of the emitter system: $\langle \psi_0(t)\|\psi_1(t)\rangle = {\mathrm{Tr}\left({\rho_{01}}(t)\right)}$. This matrix evolves from the initial pure state of the system according to the following master equation [@PhysRevLett.114.040401; @kiilerich2018multi], $$\begin{aligned} \label{eq:2sided}\begin{split} \frac{d{\rho_{01}}}{dt} = &-i\left({\hat{H}}_0{\rho_{01}}-{\rho_{01}}{\hat{H}}_1\right) \\ &+ \sum_j\left[{\hat{c}_{0j}}{\rho_{01}}{\hat{c}_{1j}}^\dagger-\frac{1}{2}\left({\hat{c}_{0j}}^\dagger{\hat{c}_{0j}}{\rho_{01}}+{\rho_{01}}{\hat{c}_{1j}}^\dagger{\hat{c}_{1j}}\right)\right]. \end{split}\end{aligned}$$ Note that the matrix evolves under the action from the left and right with the different candidate Hamiltonians and with different relaxation operators, ${\hat{c}_{0j}}$ and ${\hat{c}_{1j}}$, representing cases where the hypotheses concern the damping of the system. Unlike the conventional Lindblad master equation, [Eq. ]{}[eq:2sided]{} does not preserve the trace, and the overlap between candidates for the full system and environment quantum states attains non-trivial values, resulting in a time dependent value of ${Q_{\mathrm{e}}}^{(\text{min})}(t)$ as given in [Eq. ]{}[eq:QeMinPure]{}. This quantity represents a lower (quantum) bound on the probability of assigning a false hypothesis based on any combined quantum measurement performed on the environment in the time interval $[0,t]$ and on the emitter system at the time $t$, corresponding to the situation depicted in [ \[fig:intro\]]{}(a). In the next section we compare the achievements of testing using continuous measurements and Bayesian discrimination with this minimum. Numerical investigations {#sec:3} ======================== Error probabilities under different detection models {#sec:EP} ---------------------------------------------------- ![Temporal evolution of the error probability in assigning one of two hypotheses $\Omega_0$ and $\Omega_1$ for the Rabi driving frequency of a two-level system. The three plots correspond to different pairs of Rabi frequency candidates as annotated in the figure windows and the system is prepared in the ground state at $t=0$. Results are shown for each of the different measurement schemes discussed in this paper. The error probabilities pertaining to monitoring protocols with perfect detection $\eta=1$ are sampled from $M=100.000$ simulations (see main text). []{data-label="fig:QeHomodyneCounting"}](homodyneCountingComparison.pdf){width="0.95\columnwidth"} To address the performance of the different monitoring schemes, we turn to the associated error probability ${Q_{\mathrm{e}}}$. We consider both the case where the probability update is based solely on the detection signal, [Eq. ]{}[eq:BayesMonitor]{}, and the case where the signal is combined with a final optimized projective measurement on the system, [Eq. ]{}[eq:assignProb]{}. The probabilities pertain to the average over many independent experimental realizations. However, they are non-linear functionals of the conditional states so there is no deterministic theory which allows their evaluation. Instead we have recourse to perform a large number $M$ of simulations of the full measurement sequence and Bayesian inference. We repeated the simulations assuming each of the two hypotheses $h_0$ and $h_1$ to be true. In testing based on the detection signal ${D_T}$ alone, hypothesis $h_i$ is assigned if $P(h_i\|{D_T})>1/2$. The probability in [Eq. ]{}[eq:Qe]{} to discard a true hypothesis $h_j$ is then estimated by $ P(m=i\|h_j) = n_{i}^{(j)}/M, $ where $n_{i}^{(j)}$ is the number of samples assigning $h_i$ when $h_j$ is true. When a final system projection with outcome $\lambda$ is included in the procedure, the assignment is dictated by $P(h_i\|{D_T},\lambda)>1/2$ and the error probability is given directly by [Eq. ]{}[eq:QeMin]{}. The resulting error probabilities for our two-level example are compared to the quantum bound and to that of a projective measurement on the open system alone in [ \[fig:QeHomodyneCounting\]]{}. Curves are shown for three pairs of Rabi frequency candidates. They are all separated by $\Omega_1-\Omega_0 = 4\gamma$, and therefore the error probabilities have the same quantum lower bound [@PhysRevLett.114.040401], but their particular offsets make either counting or homodyne detection more advantageous. All protocols yield larger error probabilities than the quantum bound. This means that none of the measurement strategies are optimal in the sense that they are able to extract all information from the full state of the system and its environment. A photon counting signal is sensitive to the intensity of the emitted radiation and hence reflects the excitation of the two-level system. As seen in (a) this makes it near ideal to distinguish $\Omega_0 = 0$, which leads to no photon emissions, from a strong drive $\Omega_1=4\gamma$. The counting signal alone generates a much smaller error probability than the homodyne signal and approaches zero on a timescale similar to that of the quantum bound. When combining the counting signal with a final system projection, the error probability follows the quantum bound closely at short times and it shows that we may at specific finite probing times distinguish the hypotheses with certainty. These are points in time where the non-zero Rabi frequency $\Omega_1$ assures an atomic or a photonic excitation. The photon count is, however, insensitive to the phase of the emitted radiation and to the coherences in the two-level system. As a consequence, the two candidates $\Omega_0=-2\gamma$ and $\Omega_1 = 2\gamma$ in (b) can not be discriminated by the photon counting signal alone; i.e. ${Q_{\mathrm{e}}}(t)=1/2$ for all times. Homodyne detection is, on the other hand, highly sensitive to the phase of the emitted radiation and when combined with a final system projection, the associated error matches the quantum bound for $\gamma t\lesssim 1.5$ after which it remains close to the bound. Since for the case studied in (b) the photon count alone holds no discriminatory power, one might expect that supplementing a counting signal with a final system projection yields an error probability identical to that pertaining to a projective measurement on the mixed state of an unmonitored system. Nevertheless, it is seen than for $\gamma t\gtrsim 1.75$, counting the photo emissions reduces the final error probability by around $10\%$. This illustrates an additional advantage of monitoring the environment. Subject to backaction, the system state remains pure and experiences a transient behavior which generally depends more strongly on the particular hypothesis than the mixed state of the unmonitored system. This allows more information to be extracted from the final system measurement. Previous works identify similar mechanisms at play in parameter estimation with monitored systems [@PhysRevA.87.032115; @PhysRevA.89.052110; @PhysRevA.91.012119; @PhysRevA.94.032103; @albarelli2018restoring]. The candidate values $\Omega_0=2\gamma,\ \Omega_1=6\gamma$ in (c) can be distinguished both by the excitation and the coherence of the system. It is evident that while homodyne detection is slightly better than counting for these particular values, they both perform well and reach within $5-10\%$ of the quantum bound. Finite detector efficiency -------------------------- While the simulations in Figures \[fig:homodyneCountingComparison\] and \[fig:QeHomodyneCounting\] assume perfect monitoring, any real experiment suffers from finite detection efficiency $\eta<1$. If the environment is monitored with perfect efficiency $\eta=1$, the system state remains pure but if, e.g., some photo emissions are missed by the detector we are unable to perfectly track the state of the system and the conditional state ${\rho^{({D_t})}}(t)$ evolves to a statistical mixture. Consequently, in addition to the direct decrease in information available from the monitoring signal, the final system measurement is performed on a mixed state with, in general, less discriminatory power. ![ Temporal evolution of the error probability in assigning one of two hypotheses $\Omega_0$ and $\Omega_1$ for the Rabi driving frequency of a two-level system. The candidate values are annotated in the figure windows and the system is prepared in the ground state at $t=0$. The full, blue curves, concerning monitoring by photo detection (a) and by homodyne detection (b) combined with a final system projection, are sampled from $M=100.000$ simulations (see main text) with different values of the detection efficiency $\eta$ as indicated on the right hand side of each plot. For comparison, we show also the quantum bound (dotted curve) and error probability associated with a projective measurement on an open system (dashed, red curve). []{data-label="fig:etaDependence"}](etaDependenceDone.pdf){width="1\columnwidth"} To probe these effects, we show in [ \[fig:etaDependence\]]{} the (sampled) error probability for different values of $\eta$. For (a) photon counting we focus on the candidates $\Omega_0=0,\, \Omega_1 = 4\gamma$ and for (b) homodyne detection $\Omega_0=-2\gamma,\, \Omega_1 = 2\gamma$ where each of the two methods work particularly well. As $\eta$ decreases, the error probability ${Q_{\mathrm{e}}}(t)$ undergoes a smooth transition from the perfect detection case studied in [ \[fig:QeHomodyneCounting\]]{} to the case of a projection measurement performed on the mixed state of the system alone in the limit $\eta\rightarrow 0$. For the parameters used in this example, the photon counting protocol in (a) is surprisingly robust to detector imperfections. This is due to the fact, that as explained in Section \[sec:EP\], even a single photo detection completely rules out the hypothesis $\Omega_0=0$. While the homodyne example in (b) shows a more linear increase in the error probability as the detector efficiency deteriorates, both plots demonstrate that even with fairly large imperfections, monitoring the environment severely improves the hypothesis testing capabilities of an open quantum system. This is due to the fact that the monitoring induces transient evolution in the system which depends more strongly on the system parameters than the steady state. Conclusion and outlook {#sec:4} ====================== We have investigated how hypothesis testing with an open quantum system may be improved by monitoring the radiative environment to which it is coupled. We propose to supplement the information retrieved directly from the monitoring signal with a final system measurement optimized according to the conditional state. For reasons of clarity, we restricted our attention to just two distinct hypotheses, but the Bayesian analysis is readily generalized to cases with multiple candidates and in Ref. [@kiilerich2018multi] we present an efficient numerical approach to evaluate the quantum bound and define the optimal system projection when multiple hypotheses are in play. It was found that, while monitoring by a photon counter or a homodyne demodulator allows the extraction of much of the information leaked from the open system into the field, the error probability in these schemes does not reach the fundamental quantum bound. As explained in the introductory section \[sec:1\], this is not surprising since generally the optimal measurement is highly non-local on the full system and environment. ![ (a) A fraction $\beta$ of the radiation emitted by a probe system is collected by a homodyne demodulator while the remaining fraction $1-\beta$ is directed to a photon counter. The system state, which defines the optimal system projection to perform at the final time $T$, is conditioned on both the photon count and the homodyne signal. (b) Temporal evolution of the error probability in assigning one of three hypotheses $\Omega=0,\pm 2 \gamma$ for the Rabi frequency of a driven two-level system based on the two monitoring signals, $N_t$ and $Y_t$ of the hybrid monitoring scheme in (a). The cases of pure counting ($\beta=0$) and pure homodyne detection ($\beta=1$) are compared to different hybrid schemes with $0 < \beta < 1$. Pure homodyne detection is only optimal for times $\gamma t\lesssim 2.3$ (shaded area). The error probabilities are sampled from $M=100.000$ simulations. []{data-label="fig:hybrid"}](hybridMeasurement.pdf){width="0.95\columnwidth"} From the results presented in [ \[fig:QeHomodyneCounting\]]{}, it is clear that homodyne detection and photon counting yield different reductions in the error probability at different stages in the evolution. I.e., at some points in time either homodyne detection or photon counting is more efficient than the other. To allow both possibilities in a single experiment, the setup illustrated in [ \[fig:hybrid\]]{} splits the radiation emitted by the system such that a fraction $1-\beta$ is monitored by a photon counter and the remaining $\beta$ fraction is subject to homodyne detection. The conditional, unnormalized state $\tilde{\rho}^{(N_t,Y_t)}(t)$ then evolves according to both monitoring signals, $$\begin{aligned} \label{eq:meCountHom} \begin{split} d\tilde{\rho}^{(N_t,Y_t)} &= \Big(\left[(1-\beta)\mathcal{K}+\beta\mathcal{L}\right]dt \\ &+(1-\beta)\mathcal{B}dN_t +\sqrt{\beta}\mathcal{X}_\Phi dY_t \Big)\tilde{\rho}^{(N_t,Y_t)}. \end{split}\end{aligned}$$ A similar scheme applies the homodyne setup, [ \[fig:intro\]]{}(c) but with a local oscillator of variable strength $\xi$ [@zhang2012mapping]. Conventional homodyne detection is realized in the limit of large $\xi$, while with a weak local oscillator the setup effectively counts photons. The significance of such hybrid schemes is more apparent in scenarios with multiple distinct hypotheses, and in [ \[fig:hybrid\]]{}(b) we illustrate this by considering the differentiation of three discrete values $\Omega=0,\pm 2\gamma$ of the Rabi frequency in our two-level model. For sake of argument, we consider only monitoring without a final system projection. As discussed in Section \[sec:EP\], pure photo detection $(\beta=0)$ is only sensitive to the absolute value of $\Omega$, and hence the error probability never reaches values lower than $Q_\mathrm{e}=1/3$, signifying perfect discrimination between $\Omega=0$ and the values $\pm 2 \gamma$ which are, on the contrary, indistinguishable. When even a small fraction $\beta>0$ of the intensity of the emission signal is monitored by a homodyne demodulator, however, the combined signal is able to perfectly distinguish the three hypotheses if sufficient time is alloted. Interestingly, while pure homodyne detection ($\beta=1$) is optimal for times $\gamma t\leq 2.3$ (shaded area), hybrid schemes with $0<\beta<0.9$ converge faster to perfect discrimination because a photon counting signal very efficiently discriminates $\Omega=0$ from any non-zero values. Notice, finally, the large reduction in the error probability from the $\beta=0$ to the $ \beta=0.01$ case. This is because just $1\%$ of the intensity amounts to $10\%$ of the amplitude, which is the relevant observable in homodyne detection, and leaves the counting signal virtually unaltered. By using a beamsplitter with a tunable transmittance $\beta(t)$ or by adjusting the local oscillator strength $\xi(t)$, the effective monitoring scheme can be updated in a time dependent manner in order to further optimize the information extracted at each point in time. Such a task may be guided by intuition or achieved by numerical optimal control based on the formalism presented in this article. Acknowledgements ================ The authors would like to thank Peng Xu for helpful discussions and acknowledge financial support from the Villum Foundation. A.H.K. further acknowledges financial support from the Danish Ministry of Higher Education and Science. [22]{}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 [**, ()](\doibase 10.1007/BF01007479) [, ()](http://stacks.iop.org/1355-5111/8/i=6/a=002) [, ()](\doibase 10.1103/PhysRevA.64.042105) [, ()](\doibase 10.1103/PhysRevLett.108.170502) [, ()](\doibase 10.1103/PhysRevA.87.032115) [, ()](\doibase 10.1103/PhysRevA.79.022314) [, ()](\doibase 10.1103/PhysRevA.95.022306) [, ()](\doibase 10.1103/PhysRevA.94.032103) [, ()](\doibase 10.1103/PhysRevA.89.052110) [, ()](\doibase 10.1103/PhysRevA.91.012119) @noop []{} (, ) @noop [**, ()]{} [, ()](\doibase 10.1103/PhysRevA.96.022104) [, ()](\doibase 10.1103/PhysRevX.6.011002) [, ()](\doibase 10.1103/PhysRevLett.114.223601) [, ()](\doibase 10.1103/PhysRevLett.114.090403) @noop [, ()]{} [, ()](\doibase 10.1103/PhysRevLett.68.580) @noop [ ()]{} [, ()](\doibase 10.1103/PhysRevLett.114.040401) @noop [ ()]{} @noop [**, ()]{}	{ "pile_set_name": "ArXiv" }
--- abstract: 'This article studies the dynamics of a finite chain with infinite components. The equation which permits us to find the probability distribution of the chain length is constructed and analysed. This research is a continuation of paper [@1].' author: - \| Elena V. Karachanskaya (Chalykh)\ [^1] title: 'Dynamics of random chains of finite size with an infinite number of elements in $ {\mathbb R}^{2}$[^2] ' --- AMS 2000 subject classification Primary: 60G60, Secondary: 60H10; 65C30 Key words: Random chain, expectation function, limit behavior, characteristic function, convergence in quadratic mean, SDE Formulation of the problem ========================== In Feller’s book [@2] the problem of the length of a random chain is considered, this chain is described in the following way: the number of the elements is equal to $n$, the length of all its elements is equal to one, the angle of one component with respect to the previous is always the same up to a sign (the probability of each angle is equal to 1/2), the distance between the end points of the chain (length of the chain) is defined by means of the average square length $${\bf M}[L_{n}^{2}]=n\,\displaystyle\frac{1+ \cos \alpha}{1-\cos \alpha}-2\cos \alpha\,\displaystyle\frac{1- \cos^{n} \alpha}{(1- \cos \alpha)^{2}}.$$ We will consider the following chain: the length of the chain is finite, the number of the components is infinite, the length of each component is a random variable, the angle of each component with respect to the previous one is also random. ![image](cep3) The physical model can be a rope in a medium of Brownian particles, as length of the chain we can understand the modulus of the vector joining the starting point and the end point of the chain. Let $l \in [0,L] $ a parameter, $L$ a constant, $ l_{1},\, l_{2},\, \ldots $ the values of the parameter, $ l_{1}<l_{2}<~\ldots~ \le~L $, $ \Delta = {L}/{N} $, $l_{j}=j\cdot \Delta$. We will consider the model of the chain described by the following system of equations: $$\label{1} \begin{array}{c}\displaystyle x_{N}(t) = \sum_{s=1} ^{N} a(l_{s})\Delta \cdot\cos{\varphi_{s} (t)} , \ \ \ \ \ \displaystyle y_{N}(t) = \sum_{s=1} ^{N} a(l_{s} ) \Delta\cdot \sin{\varphi_{s} (t)}, \end{array}$$ where $ a(l_{s}), \varphi_{s} (t) $ in general are random processes, $ a(l)>0 $, $\varphi_{s} (t)$ is the angle between the $s$-th component with respect to the previous one, $\varphi_{1} (t)$ is the angle of the first component of the chain with respect to the positive direction of the $x$-axis. If we denote by $ \|x_{n}(t) \|^{2} + \| y_{n}(t) \|^{2} $ the length of the chain consisting of the $n$ elements, then the length $\triangle (l) $ of the real component of the chain is expressed by the variable $$\triangle(l) = a(l) \Delta, \ \ a(l)> 0, \ \ \displaystyle\int_{o}^{L}a(l)dl={\mathcal L}.$$ Models of type describe the distribution of the length $ {\mathcal L}(t) $ of the chain for the case where the following inequality is satisfied: $${\mathcal L}^{2}(t) = \|x_{N}(t) \|^{2} + \| y_{N}(t) \|^{2} \le const.$$ From the point of view of the representation of the phenomenon of the turbulent diffusion, the model can be useful for some generalizations of the passive displacement under the action of vortices of different size [@3]. Let be $n<N$ (that is we will consider not the whole chain but a part of it), $N\to \infty$. Since the coordinates of the initial point and the end point of each component depend on time $t$ and from the parameter $l$, then in the model we introduce some changes. $$\label{2} \begin{array}{c} \displaystyle x_{n}(l;t) = \sum_{s=1} ^{n} a(l_{s})\Delta \cdot\cos{\varphi_{s} (t)} , \ \ \ \ \ \displaystyle y_{n}(l;t) = \sum_{s=1} ^{n} a(l_{s} ) \Delta\cdot \sin{\varphi_{s} (t)}. \end{array}$$ In this way, the random field $ \{ x_{n}(l;t); y_{n}(l;t) \} $ is a dynamical stochastic process. We will study its limit behavior for $n \to \infty$. Assumptions on the model ======================== In order to obtain the coefficient of the limit equation in analytical form we shall restrict ourself to the model satisfying the following assumptions: $$\label{3} \begin{array}{c} a(l)>0 , \ \ \ l \in [0,L],\\ \varphi_{s}(t) = \displaystyle\sum\limits_{k=1}^{s} \eta (l_{k}; t) \Delta(w(l_{k})), \ \ \ t \in [0,T], \\ \eta (l_{k}; t) = \displaystyle\int_{0}^{t} \sigma (l_{k}; \tau ) \, d w_{k} (\tau ),\\ \end{array}%$$ where $\triangle(w(l_{k})), $ $ \triangle (w_{k}(\tau )) $ are independent among themselves and for different $s$ and $\tau$ are anticipating increments of the corresponding Wiener processes defined on the product of independent probability spaces $$\{ \Omega_{1},\, \Im_{l}, \, P_{1} \} \, \times \, \{ \Omega_{2}, \, \Im_{t}(n), \, P_{2} \},$$ where $ \Im_{l} , \ $ and $ \Im_{t}(n) \ $ are the corresponding flows of sigma algebras generated by the processes $ w(l) $ and $ w(t) \in {\mathbb R}^{n} $; the functions $ a(l)\in {\mathbb C}^{1}_{[0,L]}$ and $ \sigma (l; t) \in {\mathbb C}^{2}_{[0,L]\times [0,T]} $ are deterministic functions depending on $l$ and $t$, $\eta (l_{s}; t)$ is the intensity of the angle. We have therefore, $$\label{4} \begin{array}{c} \displaystyle x_{n}(l;t) = \sum_{s=1} ^{n} a(l_{s})\Delta \cdot \cos\left[{\sum\limits_{k=1}^{s} \left(\int_{0}^{t} \sigma (l_{k}; \tau ) \, d w_{k} (\tau )\right)\triangle (w(l_{k}))}\right]\\ \displaystyle y_{n}(l;t) = \sum_{s=1} ^{n} a(l_{s})\Delta \cdot \sin\left[{\sum\limits_{k=1}^{s} \left(\int_{0}^{t} \sigma (l_{k}; \tau ) \, d w_{k} (\tau )\right)\triangle (w(l_{k}))}\right] \end{array}$$ Under the condition of bounded length of the chain for the random function $\varphi_{s}(t)$ can be defined the limit for $n \to \infty$. In this context the variable $l$ appears as a parameter. Transition to auxiliary processes ================================= Let us transform (\[2\]) by means of the Euler representation: $$\begin{array}{c} x_{n}(l;t)=\displaystyle\sum_{s=1} ^{n} a(l_{s})\Delta \cdot\cos{\varphi_{s} (t)}=\sum_{s=1} ^{n} a(l_{s})\Delta \cdot \displaystyle\frac{\exp\{i\varphi_{s} (t)\}+\exp\{-i\varphi_{s} (t)\}}{2}= \\ =\displaystyle\frac{1}{2}\sum_{s=1} ^{n} a(l_{s})\Delta \cdot \exp\{i\varphi_{s} (t)\} +\displaystyle\frac{1}{2}\sum_{s=1} ^{n} a(l_{s})\Delta \cdot \exp\{-i\varphi_{s} (t)\}, \\ y_{n}(l;t)=\displaystyle\sum_{s=1} ^{n} a(l_{s})\Delta \cdot\sin{\varphi_{s} (t)}=\sum_{s=1} ^{n} a(l_{s})\Delta \cdot \displaystyle\frac{\exp\{i\varphi_{s} (t)\}-\exp\{-i\varphi_{s} (t)\}}{2i}= \\ =\displaystyle\frac{1}{2i}\sum_{s=1} ^{n} a(l_{s})\Delta \cdot \exp\{i\varphi_{s} (t)\} -\displaystyle\frac{1}{2i}\sum_{s=1} ^{n} a(l_{s})\Delta \cdot \exp\{-i\varphi_{s} (t)\}, \end{array}$$ we now introduce the auxiliary process $$\begin{array}{c} z_{1}(s;t) = \exp \{ - i \displaystyle\sum_{j=1}^{s} \, \triangle w(l_{j}) \displaystyle\int_{0}^{t} \sigma (l_{j}; \tau ) \, d w_{j} ( \tau ) \} , \\ z_{n,1}(l,t) = \displaystyle\sum_{s=1}^{n} a(l_{s})\Delta \cdot \exp\{-i\varphi_{s} (t)\} =\sum_{s=1}^{n} a(l_{s}) \Delta \cdot z_{1}(s;t), \ \ \Delta = O(n^{-1}). \end{array}$$ By using the Euler representation we rewrite the process $ \{ x_{n}(l;t); $ $ y_{n}(l;t) \} $ in the following form $$%\label{5} x_{n}(l;t) = \frac{1}{2} (z_{n,1}(l,t) + z^{}_{n,1}(l,t)), \ \ \ y_{n}(l;t) = \frac{i}{2} (z_{n,1}(l,t) - z^{}_{n,1}(l,t)).$$ For the construction of the characteristic function of the random field $ \{ x_{n}(l;t); $ $ y_{n}(l;t) \} $ we define the form of the function $\exp\{i(\alpha x_{n}(l;t) +\beta y_{n}(l;t))\}$: $$\begin{array}{c} \displaystyle\exp\{i(\alpha x_{n}(l;t) +\beta y_{n}(l;t))\}=\exp{ \left\{ i \alpha \frac{z_{n,1}(l;t) + z_{n,1}^{}(l;t)}{2} - \beta \frac{z_{n,1}(l;t) - z_{n,1}^{}(l;t)}{2} \right\} } = \\ = \displaystyle\sum \limits _{m,r=1}^{\infty } \frac{(i\alpha - \beta)^{m} (i\alpha + \beta)^{r}}{2^{m+r} m! r!} z_{n,1}^{m}(l;t) z_{n,1}^{^{r}}(l;t). \end{array}$$ In consequence the analysis of the process $ \{ x_{n}(l; t); y_{n}(l;t) \} $ leads to the study of the process $ z_{n,1}^{m}(l;t) z_{n,1}^{^{r}}(l;t)$. Since the summation and the integration operations have the same properties we replaced (in symbolic form, when $\Delta \to 0$ and this corresponds to $n \to \infty$) the process $$z_{n,1}(l,t) = \displaystyle\sum_{s=1}^{n} a(l_{s}) \Delta \cdot \exp \left\{ - i \sum_{j=1}^{s}\left( \int\limits _{0}^{t} \sigma (l_{j}; \tau ) \, d w_{j} ( \tau )\right)\, \triangle w(l_{j}) \right\}$$ by the process $$%\label{6} \begin{array}{c} \displaystyle z_{,1} = \sum \limits_{s=1}^{n} a(l_{s}) \Delta \cdot\exp \left\{ -\sum \limits_{j=1}^{s} \eta (l_{s},t) \triangle w(l_{s})\right\}= %\\ = \displaystyle\fint\limits_{0}^{l}a(u)\exp \left\{-i \fint\limits_{0}^{u}\eta (\theta,t) d w(\theta) \right\}du, \end{array}$$ where $\eta(u,t)=\displaystyle\int_{0}^{t}\sigma (u,\tau)d\tau$. We do not loose any generality in the analysis with this assumption and in the sequel we shall use the symbol $\displaystyle\fint$ instead of $\sum$.[^3] Degree transformation ===================== By considering the continuity of the process $ z_{n,1}(l;t) $ and, consequently, of the process $ z_{\,,1}(l;t) $ with respect to both variables $ l$ and $t$, we produce the degree transformation: $$\label{7} \begin{array}{c} \displaystyle z_{,1}^{m}(l;t) = \left[ \fint_{0}^{l} a(u) \exp \left \{ -i \fint_{0}^{u} \eta (\theta;t) \, dw(\theta ) \right\} \, du \right]^{m}= \\ \displaystyle = m! \fint_{0}^{l} a(u_{1}) \, du_{1} \, \exp \left \{ -i\, m \fint_{0}^{u_{1}} \eta (\theta_{1}; t) \, dw(\theta_{1} ) \right\} \times \\ \displaystyle \times \fint_{u_{1}}^{l} a(u_{2}) \, du_{1} \,\exp \left\{ -i(m-1) \fint_{u_{1}}^{u_{2}} \eta (\theta_{2};t) \, dw(\theta_{2} ) \right\} \times \, \ldots \, \times \\ %\displaystyle \times \fint_{u_{m-2}}^{l} a(u_{m-1}) \, du_{m-1} %\,\exp \left \{ -2i \fint_{u_{m-2}}^{u_{m-1}} \eta %(\theta_{m-1};t) \, dw(\theta_{m-1} ) \right\} \times %\\ \displaystyle \times \fint_{u_{m-1}}^{l} a(u_{m}) \, du_{m} \,\exp \left \{ -i \fint_{u_{m-1}}^{u_{m}} \eta (\theta_{m};t) \, dw(\theta_{m} ) \right\} \end{array}$$ where $ 0 < u_{1} < \, \ldots \, < u_{m} <l$. In different intervals we have different $dw(\theta)$ for each time instant $t$. Determination of moments ======================== Since the process $ z^{m}_{,1}(l;t) z^{{r}}_{,1}(l;t) $ depends from two variables, for the calculation of the mean $ {\bf M} \, [ z^{m}_{,1}(l;t) z^{{r}}_{,1}(l;t) ] $ it is necessary to carry out the averaging for $t$ constant (on the space $\Omega_1$), and then $\Omega_l$. Averaging with respect to $l$ ----------------------------- On the disjoint intervals $ [u_{i}, u_{i+1} )$ for all $ i $ , for all $i$ the processes $$f(u_{i}, u_{i+1}) = \fint_{u_{i}}^{u_{i+1}} \eta (\theta_{i+1};t) \, dw(\theta_{i+1} )$$ are independent by construction (since $ dw(\theta )$ are Wiener processes, $\eta (\theta;t) $ for fixed $t$ is a non-random function depending on $\theta$). Because of this, since $a(l)$ is also a non-random function, then the mathematical mean of each factor is defined in the following way: $$\label{8} \begin{array}{c} \displaystyle{\bf M}_{t} \, [ z_{,1}^{m}(l;t) ] = m! \fint\limits_{0}^{l} a(u_{1}) \, du_{1} \,{\bf M}_{t} \left[ \exp \left \{ -i\, m \fint_{0}^{u_{1}} \eta (\theta_{1}; t) \, dw(\theta_{1} ) \right\} \right] \times \\ \displaystyle \times \fint_{u_{1}}^{l} a(u_{2}) \, du_{2} \,{\bf M}_{t} \left[ \exp \left\{ -i(m-1) \fint_{u_{1}}^{u_{2}} \eta (\theta_{2};t) \, dw(\theta_{2} ) \right\} \right] \times \, \ldots \, \times \\ %\displaystyle \times \fint_{u_{m-2}}^{l} a(u_{m-1}) \, du_{m-1} %\,{\bf M}_{t} \left[ \exp \left\{ -2i \fint_{u_{m-2}}^{u_{m-1}} %\eta (\theta_{m-1};t) \, dw(\theta_{m-1} ) \right\} \right] %\times %\\ \displaystyle \times \fint_{u_{m-1}}^{l} a(u_{m}) \, du_{m} \,{\bf M}_{t} \left[ \exp \left\{ -i \fint_{u_{m-1}}^{u_{m}} \eta (\theta_{m};t) \, dw(\theta_{m} ) \right\} \right] . \end{array}$$ In this way it is necessary to find the mean of the expression of the following type: $$\exp{ \left\{ -i (m-j) \fint_{u_{j}}^{u_{j+1}} \eta (\theta;t) \, dw(\theta) \right\} } .$$ \[lm1\] [The following equality holds $$\label{9} \displaystyle {\bf M}_{t} \, \left[ \exp{ \left\{ \alpha \fint_{a}^{b} \eta (u;t)\, dw(u) \right\} } \right] = \exp{ \left\{ \frac{1}{2} \alpha^{2} \fint_{a}^{b} \eta^{2} (u;t)\, du \right\} } .$$]{} [Proof.]{} Let us denote: $$\label{10} q(a,b;t) = \fint_{a}^{b} \eta (u;t)\, dw(u)$$ and differentiate $ q(a,b;t) $ with respect to the upper limit $b$. As a result we obtain: $$d_{b} \, q(a,b;t) = \eta (b;t) \, dw(b).$$ Therefore, by Ito formula, the stochastic differential with respect to $b$ of the expression $$\exp{ \left\{ \alpha \fint_{a}^{b} \eta (u;t)\, dw(u) \right\} }=\exp{\{ \alpha q(a,b;t) \} }$$ is equal to $$d_{b} \, \exp{\{ \alpha q(a,b;t) \} } = \exp{\{ \alpha q(b;t)\} } \alpha \eta (b;t) \, dw(b)+ \frac{1}{2} \alpha^{2} \eta^{2} (b;t) \exp{ \{ \alpha q(a,b;t)\} } db.$$ We compute the average with respect to $l$ of the obtained expression: $$\label{11} d_{b} \, {\bf M}_{t} \, \left[ \exp{\{ \alpha q(a,b;t) \} } \right]= \frac{1}{2} \alpha^{2} \eta^{2} (b;t) {\bf M}_{t} \, \left[ \exp{ \{ \alpha q(a,b;t)\} } \right]db.$$ Let us denote: $$\label{12} I_{1}(a,b;t) = {\bf M}_{t} \, \left[ \exp{\{ \alpha q(a,b;t)\} }. \right]$$ Let $ \eta (b;t) $ be independent from the stochastic process $ w(u) $. In view of we obtain the differential equation $$\displaystyle\frac{d I_{1}(a,b;t)}{db}=\frac{1}{2} \alpha^{2} \eta^{2} (b;t) I_{1}(a,b;t)$$ and its solution $$I_{1}(a,b;t) = \exp{ \left\{ \frac{1}{2} \alpha^{2} \fint_{a}^{b} \eta^{2} (u;t)\, du \right\} },$$ satisfies the initial condition $I_{1}(a,a;t)=1$. In view of , the statement of the lemma is proved. $\square$ As a consequence of Lemma \[lm1\], the mathematical mean takes the form (for $t $ constant): $$\begin{array}{c} {\bf M}_{t} \, [ z_{,1}^{m}(l;t) ] = m! \displaystyle\fint_{0}^{l} a(u_{1}) \, du_{1} \, \exp{ \left\{ - \frac{m^{2}}{2} \fint_{0}^{u_{1}} \eta^{2} (\theta; t) \, d\theta \right\} } \times \\ \times \displaystyle \fint_{u_{1}}^{l} a(u_{2}) \, du_{2} \, \exp{ \left\{ - \frac{(m-1)^{2}}{2} \fint_{u_{1}}^{u_{2}} \eta^{2} (\theta;t) \, d\theta \right\} } \times \, \ldots \, \times \\ \times \displaystyle\fint_{u_{m-1}}^{l} a(u_{m}) \, du_{m} \, \exp{ \left\{ - \frac{1}{2} \fint_{u_{m-1}}^{u_{m}} \eta^{2} (\theta;t) \, d\theta \right\} } . \end{array}$$ Averaging with respect to $t$ ----------------------------- Now we make the averaging of the process $ z_{,1}^{m}(l;t) $ on the space $\Omega_2$. Since $ w_{s}(t)$ for all $s$ are independent Wiener processes, then $ \eta^{2} (l_{s}; t) $ are independent, the average of the product is therefore equal to the product of the means. As a result we obtain: $$\begin{array}{c} \displaystyle{\bf M} \, \left[ {\bf M}_{t} \, [ z_{,1}^{m}(l;t) ] \right] = %$$ %\vskip0.5mm $$= m! \fint_{0}^{l} a(u_{1}) \, du_{1} \, {\bf M}\, \left[ \exp{ \left\{ - \frac{m^{2}}{2} \fint_{0}^{u_{1}} \eta^{2} (\theta; t) \, d\theta \right\} } \right] \times %\\ %\times \fint_{u_{1}}^{l} a(u_{2}) \, du_{2} \, {\bf M} \, \left[ %\exp{ \left\{ - \frac{(m-1)^{2}}{2} \fint_{u_{1}}^{u_{2}} \eta^{2} %(\theta;t) \, d\theta \right\} } \right] \times \, \ldots \, %\times \\ \displaystyle\times \, \ldots \,\fint_{u_{m-1}}^{l} a(u_{m}) \, du_{m} \, {\bf M} \, \left[ \exp{ \left\{ - \frac{1}{2} \fint_{u_{m-1}}^{u_{m}} \eta^{2} (\theta;t) \, d\theta \right\} } \right]. \end{array}$$ \[lm2\] [The following relationship holds $${\bf M} \, \left[ \exp{ \left\{ - \frac{\alpha^{2}}{2} \fint_{a}^{b} \eta^{2} (u;t)\, du \right\} } \right] = \exp{ \left\{ - \frac{\alpha^{2}}{4} \fint_{a}^{b} \left( \int_{0}^{t} \sigma^{2}(u; \tau ) \, d \tau \right) \, du \right\} } ,$$ where $$\eta (u;t) = \int_{0}^{t} \sigma (u; \tau ) \, d w(\tau ),$$ $ \sigma (u;t) $ is a non-random function.]{} [Proof.]{} We shall use the following representation: $$\fint_{a}^{b} \eta^{2}(u; t) \, du = \frac{b-a}{N} \sum \limits_{k=1}^{N} \eta^{2}(u_{k};t),$$ this is possible in force of the model assumptions. The processes $ \eta^{2}(u_{k};t)$ ($k=\overline{1,N} $) by definition are independent for different values of $k$. We introduce the notation: $$P_{k}(t)=\eta (u_{k};t) = \int_{0}^{t} \sigma (u_{k}; \tau ) \, d w_{k}(\tau ),$$ where $ \sigma (u_{k};t) $ is a non-random function depending on $u_k$, $t$. We consider now two cases. A. Let $ \sigma (u_{k};t) $ be constant. For the seek of simplicity in the sequel we assume that $\sigma(u_{k};t) =1$ and study the problem for the processes $$\label{13} \eta (u_{k};t) = \int_{0}^{t}dw_{k}(\tau) = {\tilde P}_{k} (t) .$$ By considering the representation of the integral in form of sums, we carry out the transformation: $$\exp{ \left\{ - \frac{\alpha^{2}(b-a)}{2N} \eta^{2}(u_{k},t) \right\}}=\exp{ \left\{ - \frac{\alpha^{2}(b-a)}{2N} \left[\int_{0}^{t} dw_{k}(\tau)\right]^{2} \right\}}=\exp{ \left\{ - \frac{\alpha^{2}(b-a)}{2N} {\tilde P}_{k}^{2} (t) \right\}}.$$ Therefore $$\exp{ \left\{ - \frac{\alpha^{2}(b-a)}{2N}\sum_{k=1}^{N} {\tilde P}_{k}^{2} (t) \right\}}=\prod\limits_{k=1}^{N}\exp{ \left\{ - \frac{\alpha^{2}(b-a)}{2N} {\tilde P}_{k}^{2} (t) \right\}}$$ We denote by: $$\exp{ \left\{ - \frac{\alpha^{2}(b-a)}{2N}\sum\limits_{k=1}^{N} {\tilde P}_{k}^{2} (t) \right\}}=I_{N}(k,\alpha^{2}).$$ Since $${\bf M} \, \left[ \exp{ \left\{\alpha q(a,b;t) \right\} } \right]={\bf M} \, \left[ \exp{ \left\{-\alpha q(a,b;t) \right\} } \right],$$ and the following relationships hold: $$\begin{array}{c} {\bf M} \, \left[\prod\limits_{k=1}^{N} \exp{ \left\{ - \displaystyle\frac{\alpha^{2}(b-a)}{2N} {\tilde P}_{k}^{2} (t) \right\}} %\right\}} \right]=\\ =\displaystyle\prod\limits_{k=1}^{N}{\bf M} \, \left[\exp{ \left\{ - \frac{\alpha^{2}(b-a)}{2N} {\tilde P}_{k}^{2} (t) %\right\}} \right\} } \right] \ \ {\mathop\rightarrow^{qm}}\ \ {\bf M} \, \left[ \exp{ \left\{\alpha q(a,b;t) \right\} } \right]. \end{array}$$ Then we carry out the Ito differentiation: $$d {\tilde P}_{k} (t) = dw_{k} (t),$$ $$\label{14} d {\tilde P}_{k}^{2} (t) = dt + 2w_{k} (t)\, dw_{k} (t)$$ and in view of we have that: $$\begin{array}{c} d_{t} \displaystyle\exp{ \left\{ - \frac{\alpha^{2}(b-a)}{2N}\, {\tilde P}_{k}^{2} (t) \right\}} = - \exp{ \left\{ - \displaystyle\frac{\alpha^{2}(b-a)}{2N}\, {\tilde P}_{k}^{2} (t) \right\}} \frac{\alpha^{2}(b-a)}{2N}\,d {\tilde P}_{k}^{2} (t) + \\ + \displaystyle\frac{\alpha^{4}(b-a)^{2}}{2N^{2}} \,{\tilde P}_{k}^{2} (t) \exp{ \left\{ -\frac{\alpha^{2}(b-a)}{2N} \, {\tilde P}_{k}^{2} (t) \right\} } \, dt. \end{array}$$ We introduce in the last differential the expression : $$\begin{array}{c} d_{t} \displaystyle\exp{ \left\{ - \displaystyle\frac{\alpha^{2}(b-a)}{2N}\, {\tilde P}_{k}^{2} (t) \right\} } = - \exp{ \left\{ - \frac{\alpha^{2}(b-a)}{2N}\, {\tilde P}_{k}^{2} (t) \right\} } \times \\ \times \displaystyle\left[ \displaystyle\frac{\alpha^{2}(b-a)}{2N} \, 2{\tilde P}_{k} (t) \, dw_{k} (t) - \frac{\alpha(b-a)}{2N} \, \left( -1 + \frac{\alpha^{2}(b-a)}{N} {\tilde P}_{k}^{2} (t) \right) \, dt \right] . \end{array}$$ We calculate the mean for the last expression by denoting $$I_{2}(t;\alpha^{2}) = {\bf M} \left[ \exp{ \left\{ - \displaystyle\frac{\alpha^{2}(b-a)}{2N}\, {\tilde P}_{k}^{2} (t) \right\} } \right] .$$ We obtain the equation: $$dI_{2}(t; \alpha^{2}) = I_{2}(t; \alpha^{2}) \, \frac{ \alpha^{2}(b-a)}{2N} \, dt + %$$ $$ + \frac{ \alpha^{4}(b-a)^{2}}{2N^{2}} \, {\bf M} \left[ {\tilde P}^{2}_{k} (t) \exp{ \left\{ - \frac{\alpha^{2}(b-a)}{2N}\, {\tilde P}_{k}^{2} (t) \right\}} \right] \, dt .$$ By considering the differentiation with respect to $ \alpha^{2} $ of the expression $$\exp{ \left\{ - \frac{\alpha^{2}(b-a)}{2N}\, {\tilde P}_{k}^{2} (t) \right\} } ,$$ the last equation can be represented as a partial differential equation with constant coefficients: $$\label{15} \frac{ dI_{2}(t; \alpha^{2})}{dt} = - \frac{ \alpha^{2}(b-a)}{2N}\, I_{2}(t; \alpha^{2}) - \frac{ \alpha^{4}(b-a)}{N} \, \frac { \partial} {\partial \alpha^{2}} \, I_{2}(t; \alpha^{2}) .$$ The solution of this equation will be obtained by exploiting the properties of the stochastic processes. With this purpose we evaluate the mean of the function $ \exp{ \left\{ - \displaystyle\frac{\alpha^{2}(b-a)}{2N}\, {\tilde P}_{k}^{2} (t) \right\} }$. By considering that the process is a Wiener process (that is a Gaussian process) we have that: $${\bf M} \left[ \exp{ \left\{ - \frac{\alpha^{2}(b-a)}{2N}\, {\tilde P}_{k}^{2} (t) \right\}} \right] = %$$ $$= \int_{-\infty}^{\infty} \exp{ \left\{ - \frac{\alpha^{2}(b-a)}{2N}\, x^{2} \right\} } \times \frac{1}{ \sqrt{2 \pi \, t}} \exp{ \left\{ - \frac{x^{2}}{2 t} \right\} } \, dx =$$ $$= \frac{1}{ \sqrt{2 \pi \, t}} \, \int_{-\infty}^{\infty} \exp{ \left\{ - \left( \frac{\alpha^{2}(b-a)}{2N}\, x^{2} + \frac{x^{2}}{2 t} \right) \right\} } \, dx .$$ Furthermore $${\bf M} \left[ \exp{ \left\{ - \frac{\alpha^{2}(b-a)}{2N}\, {\tilde P}_{k}^{2} (t) \right\}} \right] = \frac{1}{\sqrt{2 \pi \, t}} \cdot \sqrt{2 \pi} \cdot \sqrt{ t\Bigl/\left(\frac{\alpha^{2}(b-a)}{2N} t + 1\right)\Bigl.} = %$$ $$= \left( \frac{\alpha^{2}(b-a)}{2N} t + 1 \right)^{- {1/2}}.$$ In this way, the obtained expression $ \displaystyle I_{2}(t; \alpha^{2}) = \left( \frac{\alpha^{2}(b-a)}{2N} t + 1 \right)^{- {1/2}} $ is the solution of the differential equation . Besides this, in view of Lemma \[lm1\], we have that: $$\begin{array}{c} {\bf M}_{t} \, \displaystyle\left[ \exp{\{ -\alpha q(a,b;t)\}}) \right] = \exp{ \left\{ - \frac{\alpha^{2}}{2}\int_{a}^{b} \eta^{2} (u,t) \right\}} = \exp{ \left\{ - \frac{\alpha^{2}(b-a)}{2N}\sum\limits_{k=1}^{N} \eta^{2} (u_{k},t) \right\}}= \\ =\displaystyle\prod\limits_{k=1}^{N} \exp{ \left\{ - \frac{\alpha^{2}(b-a)}{2N}\eta^{2} (u_{k},t) \right\}}. \end{array}$$ Therefore $$\begin{array}{c} \displaystyle{\bf M}\left[ {\bf M}_{t} \, \left[ \exp{\{ -\alpha q(a,b;t)\}}) \right] \right]\ \ {\mathop\rightarrow^{qm}}\ \ \prod\limits_{k=1}^{N} {\bf M}\left[\exp{ \left\{ - \frac{\alpha^{2}(b-a)}{2N}{\tilde P}_{k}^{2} (t) \right\}}\right]= \\ =\displaystyle\prod\limits_{k=1}^{N}\left( \frac{\alpha^{2}(b-a)}{2N} \, t + 1 \right)^{- 1/2}=\left( \frac{\alpha^{2}(b-a)}{2N} \, t + 1 \right)^{- {N/2}} \end{array}$$ Passing to the limit we obtain the complete averaging with respect to both components: $$\displaystyle{\bf M} \, \left[ I_{1}(a,b;t) \right] = %$$ $$ = \lim \limits_{N \rightarrow\infty} \, \left( \frac{\alpha^{2}(b-a)}{2N} \, t + 1 \right)^{- {N/2}} = \exp { \left\{ - \frac{\alpha^{2}(b-a)t}{4} \right\} } .$$ B. Now we consider the case $ \sigma (u_{k};t) \not= 1$. In this case for $ P_{k} (t) $, we obtain the expression: $$d_{t} \, P_{k} (t) = \sigma (u_{k};t) \, dw_{k} (t) .$$ In this way we have that $$d_{t} P_{k}^{2} (t) = \frac {\sigma^{2} (u_{k};t)}{2} 2dt + 2 P_{k} (t) \, \sigma (u_{k};t) \, dw_{k} (t)$$ and $$\begin{array}{c} d_{t} \displaystyle \exp{ \left\{ - \frac{\alpha^{2}(b-a)}{2N}\, P_{k}^{2} (t) \right\}} = - \exp{ \left\{ - \frac{\alpha^{2}(b-a)}{2N}\, P_{k}^{2} (t) \right\} } \times \\ \times \displaystyle\frac{\alpha^{2}(b-a)}{2N} \, \left[ \sigma^{2} (u_{k};t) \, dt + 2 P_{k} (t)\, \sigma (u_{k};t) \, dw_{k} (t) \right] + \\ + \displaystyle\frac{\alpha^{4}(b-a)^{2}}{2N^{2}} \, P_{k}^{2} (t) \, \sigma^{2} (u_{k};t) \cdot \exp{ \left\{ -\frac{\alpha^{2}(b-a)}{2N} \, P_{k}^{2} (t) \right\} } \, dt . \end{array}$$ Therefore $$\begin{array}{c} \displaystyle d_{t} I_{t} (k; \alpha^{2}) = \partial_{t} \, {\bf M} \, \left[ \exp{ \left\{ -\displaystyle\frac{\alpha^{2}(b-a)}{2N} \, P_{k}^{2} (t) \right\} } \right] = \\ = - \displaystyle{\bf M} \,\left[ \exp{ \left\{ -\displaystyle\frac{\alpha^{2}(b-a)}{2N} \, P_{k}^{2} (t) \right\} } \frac{\alpha^{2}(b-a)}{2N} \, \sigma^{2} (u_{k};t) \right] \, dt . \end{array}$$ By exploiting the possibility of differentiation with respect to the parameter $ \alpha^{2} $, we arrive at the following equation: $$\label{16} \displaystyle \frac { \partial_{t} I_{t} (k; \alpha^{2}) }{ \partial t} = - \frac{\alpha^{2}(b-a)}{2N} \, \sigma^{2} (u_{k};t) \, I_{t} (k;\alpha^{2}) - %$$ \vskip0.5mm $$ - \frac{\alpha^{4}(b-a)}{N} \, \sigma^{2} (u_{k};t) \, \frac { \partial I_{t} (k; \alpha^{2}) }{ \partial \alpha^{2}} .$$ We divide both members by $ \sigma^{2} (u_{k};t)$, and denote $$\theta (t) = \int_{0}^{t} \, \sigma^{2} (u_{k};\tau) \, d \tau$$ and we pass to the auxiliary equation $$\label{17} \displaystyle \frac { \partial I_{\theta} (k; \alpha^{2}) }{\partial \theta} = - \frac{\alpha^{2}(b-a)}{2N} \, I_{\theta} (k;\alpha^{2}) - \frac{\alpha^{4}(b-a)}{N} \, \frac { \partial I_{\theta} (k; \alpha^{2}) }{\partial \alpha^{2}} .$$ Equation is a differential equation with constant coefficients stochastically equivalent to equation . Therefore its solution has the form: $$I_{\theta} (k; \alpha^{2}) = \left( \frac{\alpha^{2}(b-a)}{2N}\, \theta + 1 \right)^{- {1/2}}.$$ and we obtain, the solution of the equation : $$I_{t} (k; \alpha^{2}) = \left( \frac{\alpha^{2}(b-a)}{2N}\, \int_{0}^{t} \, \sigma^{2} (u_{k};\tau) \, d \tau + 1 \right)^{- {1/2}}.$$ As a consequence we have that $$\label{18} \displaystyle{\bf M} \, \left[ I_{N}(k;\alpha^{2}) \right] =\prod\limits_{k=1}^{N} \, I_{t}(k; \alpha^{2}) = \prod\limits_{k=1}^{N} \, \left( \frac{\alpha^{2}(b-a)}{2N} \, \int_{0}^{t} \, \sigma^{2} (u_{k};\tau) \, d \tau + 1 \right)^{- {1/2}} .$$ In order to evaluate $ I_{1}(a,b;t) $, we take the logarithm of : $$\begin{array}{c} \displaystyle\ln {\bf M} \, \left[ {I_{N}(k;\alpha^{2})}\right] = \ln \prod\limits_{k=1}^{N} \, \left( \frac{\alpha^{2}(b-a)}{2N} \, \int_{0}^{t} \, \sigma^{2} (u_{k};\tau) \, d \tau + 1 \right)^{- {1/2}} = \\ = - \displaystyle\frac {1}{2} \, \sum \limits _{k=1}^{N} \, \ln \left( \frac{\alpha^{2}(b-a)}{2N} \, \int_{0}^{t} \, \sigma^{2} (u_{k};\tau) \, d \tau + 1 \right) . \end{array}$$ By using the series expansion of $ \ln{(x+1)} $ we obtain that: $$\begin{array}{c} \displaystyle\ln{{\bf M} \, \left[ {\tilde I}_{1}(b;t) \right] } = - \frac{1}{2} \, \sum \limits _{k=1}^{N} \, \left[ \frac{\alpha^{2}(b-a)}{2N} \, \int_{0}^{t} \, \sigma^{2} (u_{k};\tau ) \, d \tau - \right. \\ \displaystyle\left.- \frac {1}{2} \frac{\alpha^{4}(b-a)^{2}}{4N^{2}} \, \left( \int_{0}^{t} \, \sigma^{2} (u_{k};\tau) \, d \tau \right)^{2} + O (N^{-3}) \right] . \end{array}$$ We calculate the limit for $N \to \infty$: $$\begin{array}{c} \displaystyle\lim_{N \to \infty} \, \ln {\bf M} \, \left[ I_{N}(k,\alpha^{2}) \right] = - \frac {1}{2} \, \lim_{N \rightarrow\infty} \, \sum \limits _{k=1}^{N} \, \left[ \frac{\alpha^{2}(b-a)}{2N} \, \int_{0}^{t} \, \sigma^{2} (u_{k};\tau) \, d \tau - \right. \\ \displaystyle\left. - \frac {1}{2} \cdot \frac{\alpha^{4}(b-a)^{2}}{4N^{2}} \, \left( \int_{0}^{t} \, \sigma^{2} (u_{k};\tau) \, d \tau \right)^{2} + O (N^{-3}) \right] = \\ = - \displaystyle\frac {1}{2} \, \lim_{N \to \infty} \, \sum \limits _{k=1}^{N} \, \frac{\alpha^{2}(b-a)}{2N} \, \int_{0}^{t} \, \sigma^{2} (u_{k};\tau) \, d \tau = - \frac{\alpha^{2}}{4} \, \int_{a}^{b} \left( \int_{0}^{t} \, \sigma^{2} (u;\tau) \, d \tau \right) \, du . \end{array}$$ By using the limit and passing to the anti-logarithm we prove that Lemma \[lm2\] holds $${\bf M} \, \left[ \exp{ \left\{ - \frac{\alpha^{2}}{2} \int_{a}^{b} \eta^{2} (u;t)\, du \right\} } \right] = \exp{ \left\{ - \frac{\alpha^{2}}{4} \int_{a}^{b} \left( \int_{0}^{t} \sigma^{2}(u; \tau ) \, d \tau \right) \, du \right\} } . \ \ \ \square$$ Passing to the limiting process ------------------------------- The field $\{ x_{n}(l;t)$; $ y_{n}(l;t) \} $ is defined by the model assumption (\[3\]) : $$\eta (l_{s};t) = \int_{0}^{t} \sigma (l_{s}; \tau ) \, d w(\tau).$$ We change the model assumption $$%\begin{equation}\label{111} \displaystyle {\tilde \eta} (l_{s};t) = \left(\frac {1}{2}\int_{0}^{t} \sigma^ {2} (l_{s}; \tau ) \, d \tau \right) ^{1/2} %\end{equation}$$ and consider the field $ \{ {\hat x}_{n}(l;t) ; $ $ {\hat y}_{n}(l;t) \} $ of the following form: $$\label{19} \begin{array}{c} \displaystyle {\hat x}_{n}(l;t) = \sum \limits_{s=1}^{n} \, a (l_{s}) \cos \left[\sum \limits_{j=1}^{s} \, \triangle (w (l_{j})) \, \left(\frac {1}{2}\int_{0}^{t} \sigma^{2} (l_{j}; \tau ) \, d \tau \right) ^{1/2} \right] \ \Delta , \\ \displaystyle {\hat y}_{n}(l;t) = \sum \limits_{s=1}^{n} \, a (l_{s}) \sin \left[\sum \limits_{j=1}^{s} \, \triangle (w (l_{j})) \, \left(\frac {1}{2}\int_{0}^{t} \sigma^{2} (l_{j}; \tau ) \, d \tau \right)^{1/2} \right] \ \Delta , \end{array}$$ where $ \triangle (w (l_{j})) $ is the increment of the Wiener process on the interval $ [l_{j};l_{j+1} ] $. This means that the variable $t$ is not a random variable and from the analysis of the process on the flow of the $\sigma$-algebras $ \displaystyle \Im_{t}(n) \, \oplus \, \Im_{l} $ it is possible to pass to the process defined on the flow of the $\sigma$-algebras $ \Im (l)$, for all $t={\it const}$ . Averaging with respect to $t$ has already been carried out. We observe that the fields $ \{ x_{n}(l;t); $ $ y_{n}(l;t) \} $ and $ \{ {\hat x}_{n}(l;t) ; $ $ {\hat y}_{n}(l;t) \} $ are defined on different spaces. We consider the processes $$z_{2} (k;t) = \exp { \left \{ - {\it i} \, \sum \limits_{j=1}^{k} \, \triangle w (l_{j}) \, \left(\frac {1}{2}\int_{0}^{t} \sigma^{2} (l_{j}; \tau ) \, d \tau \right) ^{1/2} \right \} } ,$$ $$z_{n,2} (l;t) = \sum \limits_{k=1}^{n} \, z_{2} (k;t) \, a(l_{k}) \cdot \Delta , \ \ \ \Delta = O (n^{-1}) .$$ By considering the Euler representation, the components of the fields (\[19\]) will take the form: $$\begin{array}{c} {\hat x}_{n}(l;t) = \displaystyle\frac {1}{2} \left ( z_{n,2} (l;t) + z_{n,2}^{} (l;t) \right) , \\ {\hat y}_{n}(l;t) = \displaystyle \frac {{\it i}}{2} \left ( z_{n,2} (l;t) - z_{n,2}^{} (l;t) \right) . \end{array}$$ We construct now the characteristic functions: $ g_{n} ( \alpha ; \beta ; t) $ for the field $ \left\{ x_{n}(l;t) ; y_{n}(l;t) \right\} $ and $ {\hat g}_{n} ( \alpha ; \beta ; t) $ for the field $ \left\{ {\hat x}_{n}(l;t) ; {\hat y}_{n}(l;t) \right\} $: $$g_{n} ( \alpha ; \beta; t) = {\bf M} \left[ \exp { \left \{ \frac {i}{2} ( \alpha + i \, \beta) \, z_{n,1} (l;t) + \frac {i}{2} ( \alpha - i \, \beta) \, z_{n,1}^{} (l;t) \right \} } \right] ,$$ $${\hat g}_{n} ( \alpha ; \beta; t) = {\bf M} \left[ \exp { \left \{ \frac {{\it i}}{2} ( \alpha + i \, \beta) \, z_{n,2} (l;t) + \frac {i}{2} ( \alpha - i \, \beta) \, z_{n,2}^{} (l;t) \right\} } \right] .$$ For the continuation of the research the next lemma is necessary. \[lm3\] [Under the model assumptions for the random fields $ \left\{ x_{n}(l;t) ; y_{n}(l;t) \right\} $ and $ \left\{ {\hat x}_{n}(l;t) ; {\hat y}_{n}(l;t) \right\} $ and for fixed integer $ m$ there exists a number $% n^{\prime}$ such that for all $n>n^{\prime}$ the following relationships hold: $$\label{20} {\bf M} [z_{,2}^{m}(l;t)] = {\bf M} [z_{,2}^{ \,m}(l;t)] = {\bf M} [z_{,1}^{m}(l;t)] = {\bf M} [z_{,1}^{* \, m}(l;t)] .$$* ]{} [Proof.]{} We calculate all increments in the series of the equalities by having in mind the lemmas. We introduce the following notation $\tilde{\eta}(\theta,t)= \left(\displaystyle\frac{1}{2}\int_{0}^{t}\sigma^{2}(\theta,\tau)d\tau\right)^{1/2}$. We have that $$\begin{array}{c} {\bf M} [z_{,2}^{m}(l;t)] = m ! \displaystyle\fint_{0}^{l} \, a (u_{1}) \, d u_{1} \, {\bf M} \left[ \exp { \left\{ - i\, m \, \fint_{0}^{u_{1}} \tilde{\eta}(\theta_{1},t)\, d w(\theta_{1}) \right\} } \right] \times %\\ %\times \, \displaystyle\fint_{u_{1}}^{l} \, a (u_{2}) \, du_{2} \, %{\bf M} \left[\exp { \left\{ - i\, (m-1) \, \fint_{u_{1}}^{u_{2}} %\tilde{\eta}(\theta_{2},t)\, d w(\theta_{2}) \right\} } %\right] \times \, \ldots \, \times \\ \times\, \ldots \, \times \, \displaystyle\fint_{u_{m}}^{l} \, a (u_{m}) \, du_{m}) \, {\bf M} \left[ \exp { \left \{ - i \, \fint_{u_{m-1}}^{u_{m}} \tilde{\eta}(\theta_{m},t)\, d w(\theta_{m}) \right\} } \right] = \\ =m! \displaystyle\left(\fint_{0}^{l} \, a (u_{1}) \, d u_{1} \, \exp { \left\{ - \frac{m^{2} }{2}\, \fint_{0}^{u_{1}} \tilde{\eta}^{2}(\theta_{1},t)\, d\theta_{1} \right\} } \right)\times %\\ %\times \, \displaystyle\left(\fint_{u_{1}}^{l} \, a (u_{2}) \, %du_{2} \, \exp { \left\{ - \frac{(m-1)^{2}}{2} \, %\fint_{u_{1}}^{u_{2}} \tilde{\eta}^{2}(\theta_{2},t)\, d %\theta_{2} \right\} } % \right)\times \, \ldots \, \times \\ \times \, \ldots \, \times\,\displaystyle\left(\fint_{u_{m}}^{l} \, a (u_{m}) \, du_{m}) \, \exp { \left \{ - \frac{1}{2} \, \fint_{u_{m-1}}^{u_{m}} \tilde{\eta}^{2}(\theta_{m},t)\, d \theta_{m} \right\} } \right)= \\ =m! \displaystyle\left(\fint_{0}^{l} \, a (u_{1}) \, d u_{1} \, \exp { \left\{ - \frac{m^{2} }{2}\, \fint_{0}^{u_{1}} \left(\displaystyle\frac{1}{2}\int_{0}^{t}\sigma^{2}(\theta_{1},\tau)d\tau\right)\, d\theta_{1} \right\} } \right)\times \,\ldots \, \times %\\ %\times \, \displaystyle\left(\fint_{u_{1}}^{l} \, a (u_{2}) \, %du_{2} \, \exp { \left\{ - \frac{(m-1)^{2}}{2} \, %\fint_{u_{1}}^{u_{2}} %\left(\displaystyle\frac{1}{2}\int_{0}^{t}\sigma^{2}(\theta_{2},\tau)d\tau\right)\, %d \theta_{2} \right\} } % \right)\times \, \ldots \, \times \\ \times\, \displaystyle\left(\fint_{u_{m}}^{l} \, a (u_{m}) \, du_{m}) \, \exp { \left \{ - \frac{1}{2} \, \fint_{u_{m-1}}^{u_{m}} \left(\displaystyle\frac{1}{2}\int_{0}^{t}\sigma^{2}(\theta_{m},\tau)d\tau\right)\, d \theta_{m} \right\} } \right)={\bf M} [z_{,1}^{m}(l;t)] \end{array}$$ Then, in force of and we have that: $${\bf M} [z_{,1}^{m}(l;t)] = {\bf M} [z_{,1}^{* \,m}(l;t)]. \ \$$ In this way we obtain the confirmation of Lemma. $\square$ \[lm4\] [The characteristic functions of fields $ \{ x_{n}(l;t)$; $ y_{n}(l;t) \} $ and $ \{ {\hat x}_{n}(l;t)$; $ {\hat y}_{n}(l;t) \} $ for $ n \to \infty $ coincide for all $l$ and $t$.]{} [Proof.]{} The proof is based on the coincidence of the representations for the characteristic functions $ g_{n}(\alpha , \beta , t) $ and $ {\hat g}_{n}(\alpha , \beta , t) $ by means of the Maclaurin expansion (inside the mean) with respect to $ z_{,1}(l;t) $ and $ z_{,1}^{}(l;t) $, and $ z_{,2}(l;t) $, $ z^{}_{,2}(l;t)$ respectively and also on the conclusions of Lemma \[lm3\]. $\square$ Lemma \[lm4\] permits us to pass to the study of the limit behavior of the field $ \{ {\hat x}_{n}(l;t)$; $ {\hat y}_{n}(l;t) \}$ for $n \to \infty$ exclusively. \[t1\] Let us assume that for the field $ \{ x_{n}(l;t); y_{n}(l;t) \} $ the model assumptions [(\[19\])]{} are satisfied: $$\begin{array}{c} \displaystyle {\hat x}_{n}(l;t) = \sum \limits_{s=1}^{n} \, a (l_{s}) \cos \left[\sum \limits_{j=1}^{s} \, \triangle(w (l_{j})) \, \left(\frac {1}{2}\int_{0}^{t} \sigma^{2} (l_{j}; \tau ) \, d \tau \right) ^{1/2} \right] \ \Delta , \\ \displaystyle {\hat y}_{n}(l;t) = \sum \limits_{s=1}^{n} \, a (l_{s}) \sin \left[\sum \limits_{j=1}^{s} \,\triangle (w (l_{j})) \, \left(\frac {1}{2}\int_{0}^{t} \sigma^{2} (l_{j}; \tau ) \, d \tau \right)^{1/2} \right] \ \Delta , \end{array}$$ and assume that the field $ \{ x(l;t); y(l;t) \} $ is defined in the following way $$\label{21} \begin{array}{c} x(l;t)=\displaystyle \int_{0}^{l} a(u) \cos \left[\int_{0}^{u} \frac{1}{2}\left(\int_{0}^{t}\sigma^{2}(\theta,\tau)d\tau\right)dw(\theta) \right]du,\\ y(l;t)=\displaystyle \int_{0}^{l} a(u) \sin \left[\int_{0}^{u} \frac{1}{2}\left(\int_{0}^{t}\sigma^{2}(\theta,\tau)d\tau\right)dw(\theta) \right]du. \end{array}$$ Under these conditions the characteristic functions of the processes $ \{ x(l;t); y(l;t) \} $ and $ \{ x_{n}(l;t); y_{n}(l;t) \} $ coincide. [Proof.]{} The comparison of the characteristic functions for $ \{ {\hat x}_{n}(l;t), {\hat y}_{n}(l;t) \} $ and $ \{ x(l;t), $ $ y(l;t) \} $ for all values of $t \in [0,T],$ for $n \to \infty$ leads to the proof of the theorem. $\square$ \[t2\] The stochastic process $ \{ x(l;t); y(l;t) \} $ is the solution to the Cauchy problem for the Ito stochastic differential equations: $$\label{22} \begin{array}{c} \displaystyle d_{l} p(l;t) = \left[ p(l;t) \frac{\partial }{\partial l} \ln a(l) - \frac{p(l;t)}{4} \int \limits _{0}^{t} \sigma ^{2}(l; \tau )d \tau \right] dl %- \\ - \displaystyle\left( \frac{1}{2} \int\limits _{0}^{t} \sigma ^{2}(l; \tau ) d \tau \right)^{0,5} q(l;t) dw(l), \\ \displaystyle d_{l} q(l;t) = \left[ q(l;t) \frac{\partial }{\partial l} \ln a(l) - \frac{q(l;t)}{4} \int \limits _{0}^{t} \sigma ^{2}(l; \tau ) d \tau \right] dl %+ \\ + \displaystyle\left( \frac{1}{2} \int \limits _{0}^{t} \sigma ^{2}(l; \tau ) d \tau \right)^{0,5} p(l;t) dw(l), \end{array}$$ $$d_{l}x(l;t) = q(l;t)\, dl , \ \ \ \ d_{l}y(l;t) = p(l;t)\, dl ,$$ satisfying the boundary conditions $$x(0;t)=0, \ \ y(0;t)=0, \ \ p(0;t)=a(0), \ \ q(0;t)=0.$$ [Proof.]{} We differentiate $ x(l;t) $ and $ y(l;t) $ in with respect to $l$: $$\label{23} \displaystyle \frac{\partial x(l;t)}{\partial l}=a(l) \sin \left[\int_{0}^{l} \frac{1}{2}\left(\int_{0}^{t}\sigma^{2}(\theta,\tau)d\tau\right)dw(\theta) \right]=q(l;t),$$ $$\label{24} \displaystyle \frac{\partial y(l;t)}{\partial l}=-a(l) \cos \left[\int_{0}^{l} \frac{1}{2}\left(\int_{0}^{t}\sigma^{2}(\theta,\tau)d\tau\right)dw(\theta) \right]=p(l;t).$$ The obtained expressions are now differentiated by Ito formula with respect to the variable $l$: $$\begin{array}{c} d_{l}\left(\displaystyle \frac{\partial x(l;t)}{\partial l}\right)=\displaystyle\frac{1}{a(l)}\frac{\partial a(l)}{\partial l}\,a(l)\cos \left[\int_{0}^{l} \frac{1}{2}\left(\int_{0}^{t}\sigma^{2}(\theta,\tau)d\tau\right)dw(\theta) \right]dl- \\ -\,a(l) \sin \left[\displaystyle\int_{0}^{l}\displaystyle \frac{1}{2}\left(\int_{0}^{t}\sigma^{2}(\theta,\tau)d\tau\right)dw(\theta) \right]\cdot \displaystyle\frac{1}{2} \left(\int_{0}^{t}\sigma^{2}(\theta,\tau)d\tau\right)dw(l)- \\ -\,\displaystyle\frac{1}{2}\cos \left[\int_{0}^{l} \frac{1}{2}\left(\int_{0}^{t}\sigma^{2}(\theta,\tau)d\tau\right)dw(\theta) \right]\cdot\left(\displaystyle\frac{1}{2} \int_{0}^{t}\sigma^{2}(\theta,\tau)d\tau\right)^{2}dl. \end{array}$$ Taking into account and we obtain the last equation of system . In a similar way we get the second expression of the system. The functions $x(l;t)$, $y(l;t)$, $p(l;t)$, $q(l;t)$ defined by , and satisfy the given initial conditions. $\square$ Within the framework of the given formulation ($L$ constant) we have found $ F_{t}(x;y;L) $ for different values of $t$. \[t3\] The distribution function of the process $ \{ x(l;t); y(l;t) \} $ can be obtained by integrating with respect to the variables $p$ and $q$ the Kolmogorov equation of the system [(\[22\])]{}. [Proof.]{} After the enlargement of the space obtained by introducing the new variables $p$ and $q$, the compound process $ \{ x(l;t); y(l;t); p(l;t);q(l;t) \} $ becomes a Markov process. This means that it is possible to obtain a Kolmogorov equation for the density function $\rho (x,y,p,q,l,t)$ and then by integrating with respect to $p$ and $q$ infer the density function of the distribution $\rho (x,y,l,t)$ for all $l$ and $t$. $\square$ \[t4\] The distribution function of the original process $ \{ x_{n}(l;t); y_{n}(l;t) \} $ under the model conditions [(\[4\])]{} coincide with the distribution function of the Markov process $ \{ x(l;t); y(l;t) \} $ [(\[21\])]{}. [Proof.]{} The proof is based on the conclusions of Theorem \[t1\] and Theorem \[t2\]. $\square$ The character of the analysis doesn’t substantially changes when, for example $a=a(l,t)$ (vibrating chain), $ \sigma (l;t) $ is a non anticipating measurable random function with respect to independent flows of $\sigma$-algebras governed by independent Wiener processes $ w(l) $ and $ w(t)$. In this way we arrive at a coherent representation of distribution: the parameter $t$ defines also the structure of the chain. [10]{} Doodko V.A. and Chalykh E.V. , The dynamics of finite chain which has infinite many of units in $ {\mathbb R}^{2}$. Preprint. The Inst. for Appl. Math., The FEB of Rus. Ac. Sci.– Vladivostok: Dal’nauka (1998) 18 pp. (in Russian). W. Feller, An Introduction to Probability Theory and its Applications T, Vol I, John Wiley & Sons, Inc., New-York - London -Sydney, 1968. Ventsel A.D., Course of the random processes theory, Moscow: Nauka, 1975. (in Russian). [^1]: Electronic address: [email protected] [^2]: [^3]: This symbol does not concern known designations. It is a label only.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We demonstrate trapping in a surface-electrode ion trap fabricated in a 90-nm CMOS (complementary metal-oxide-semiconductor) foundry process utilizing the top metal layer of the process for the trap electrodes. The process includes doped active regions and metal interconnect layers, allowing for co-fabrication of standard CMOS circuitry as well as devices for optical control and measurement. With one of the interconnect layers defining a ground plane between the trap electrode layer and the p-type doped silicon substrate, ion loading is robust and trapping is stable. We measure a motional heating rate comparable to those seen in surface-electrode traps of similar size. This is the first demonstration of scalable quantum computing hardware, in any modality, utilizing a commercial CMOS process, and it opens the door to integration and co-fabrication of electronics and photonics for large-scale quantum processing in trapped-ion arrays.' author: - 'K. K. Mehta' - 'A. M. Eltony' - 'C. D. Bruzewicz' - 'I. L. Chuang' - 'R. J. Ram' - 'J. M. Sage' - 'J. Chiaverini' bibliography: - 'jc\_bib1.bib' - 'jc\_bib2.bib' - 'IonLoading\_PRAfinal.bib' title: Ion Traps Fabricated in a CMOS Foundry --- [^1] [^2] [^3] Trapped atomic ions are a promising system for large-scale quantum processing[@RevModPhys.75.281; @Blatt:Wine:Nat08], as all required basic quantum operations have been demonstrated with low error[@Inns:HiFi2qubit:NatPhys:08; @Oxford:HiFiRdout:PRL:08; @NIST:HifiMicrogate:12]. However, these demonstration experiments typically consist of relatively few ions ($\lesssim10$) manipulated with optical beams and electronic signals routed from outside the ion-trap vacuum chamber. In order to scale the system to the number of quantum bits (qubits) required to provide speedups over classical computing methods, trap arrays holding orders of magnitude more ions are necessary. Additionally, each array site will require local control, readout electronics, and optics for scalability. Current microfabricated ion traps (and, in fact, realizations of any scalable quantum processing technology) depend on specialized, nonstandard processes in research clean-room facilities. The traps are often built upon non-silicon substrates[@NIST:SET:PRL:06; @Chaung:CryoHeatRates:PRL08], and where silicon is used, only a few metal layers (four maximum) have been implemented[@britton:nist:2008; @SMIT:QIC:2009; @Sandia:Bksideloading:arXiv:10; @Oxford:CaCharging:APB:11; @npl:monolithictrap:2012; @NJP:GTRI:xjunctrap:2013; @Innsbruck:CryoSiTrap:arXiv:14; @sussex:2dhexrap:2014]. None of the traps made on silicon substrates to date have had doped, active device fabrication available, and due to the idiosyncratic process steps used, the lithographic resolution is typically limited. Repeatability at different facilities is almost impossible due to local process variations and substrate processing capabilities. Here we describe the design and operation of an ion trap built into a standard high-resolution CMOS fabrication process. Based on industry-standard practices and materials, there are active and passive layers beneath the trap-electrode layer that may enable integration of electronics and photonics[@shepardAPD:APL:2010; @elecphotintegration:Orcutt:2012] for control and readout of trapped-ion quantum states. Standardization of the foundry process permits any group to produce identical devices with high yield. ![(Color online) Die and process cross-section. A micrograph of the fabricated $3\times3$ mm$^{2}$ die is shown in the upper left panel. The lower left panel shows a perspective rendering of the top aluminum trap layer and the meshed ground plane in copper below, as designed; the gaps in the trap electrodes here are 5 $\mu$m, and the ground mesh is formed of 600 nm wires with 350 nm gaps along $x$ and 10 $\mu$m gaps along $y$. A chip cross section is diagrammed at right, with approximate relevant dimensions labeled (“pSi” is polysilicon and metal interconnect layers are labelled m1 through m8). Vias shown between metal layers are only representative.\[figure:process\]](die_and_stackup7b){width="\columnwidth"} Devices were fabricated on $3\times3$ mm$^{2}$ die (Fig. \[figure:process\]) on a shared, 300-mm, multi-project wafer produced in a 90-nm CMOS process operated by IBM (9LP process designation). This process is primarily utilized for dense, high-performance digital circuits, and the trap die was one of many designs fabricated in parallel on the same wafer. The process allows for patterning of 8 copper interconnect layers, along with the top aluminum pad layer (right panel of Fig. \[figure:process\]). This $1.3$ $\mu$m thick pad layer was used for the trap electrodes, and a copper layer (m5) approximately 4 $\mu$m below the aluminum layer’s bottom surface and 2 $\mu$m above the silicon substrate was used to form a ground plane under the extent of one of the traps. Due to metal density constraints arising from chemical-mechanical polishing steps applied to these layers, this ground plane was patterned as a mesh of 600 nm strips separated by 350 nm along the $x$ direction and 10 $\mu$m along the $y$ direction (see Fig. \[figure:process\]). Metal vias connect this copper ground plane to the center electrode of the trap through the upper metal layers m6–m8. ![(Color online) Micrograph of trap chip diced from die and mounted on the sapphire interposer of a cryogenic vacuum system. Aluminum wirebonds are used to make contact from the aluminum trap electrodes to the gold interposer leads. The chip is $2.5$ mm long and $1.2$ mm wide. The inset shows two ions trapped 50 $\mu$m from the surface of the trap chip. The ions are approximately 5 $\mu$m apart.\[figure:mountedchip\]](chip_and_ions_1b){width="\columnwidth"} We designed and tested linear radio-frequency (RF) surface-electrode[@NIST:SET:QIC:05] Paul traps that confine ions 50 $\mu$m from the electrode surface. The trap has a “five-wire” trap geometry, with two RF electrodes symmetric about the trap axis. Segmented dc control electrodes are routed to the corners of the trap chip to prevent wirebonds from obstructing laser access (see Fig. \[figure:mountedchip\]). This design offers flexibility to create various trapping potentials and allows scalability to multi-zone traps with complex geometries. Other advantages include comparative ease in selecting control voltages, and relatively narrow RF electrodes, allowing for lower capacitive coupling and RF power dissipation in the trap. Some inhomogeneity in RF field along the trap axis is anticipated due to the short (2 mm) electrode length, but effects on trapping are expected to be negligible. A possible limitation to the use of high-resolution CMOS processing is breakdown at large applied potentials, especially since typical ion trap voltage amplitudes are significantly higher than those used in CMOS electronics. However, for up to 200 V static bias applied, the leakage current was below 10 pA, and no sudden increase corresponding to a dielectric breakdown was observed. We performed these tests at room temperature (in a high-dielectric-strength fluid to prevent air breakdown), applying the potential between one of the RF electrodes and either the ground plane or one of the adjacent dc electrodes in both types of trap. After commercial foundry fabrication, trap chips (diced from the full die) approximately $2.5\times1.2$ mm$^{2}$ were bonded to a larger interposer to interface with the cold stage and wiring of a cryogenic vacuum system that allows for variation of the trap-chip temperature[@PhysRevA.86.013417]. Using a quarter-wave helical resonator, a $43$ MHz RF signal of approximately 100 V amplitude was applied to the trap electrodes to produce radial trap frequencies of approximately $4$–$5$ MHz, and an axial potential with frequencies near 1 MHz was produced by application of dc potentials of up to approximately 30 V to the segmented control electrodes. We load $^{88}$Sr$^{+}$ ions by accelerating precooled Sr atoms from a magneto-optical trap toward the ion trap where they are photo-ionized and Doppler cooled[@PhysRevA.86.013417]. Although not measured precisely in this work, loading efficiency into these traps is similar to more conventionally fabricated surface-electrode traps using the same loading method. Traps without a ground plane displayed significant laser-induced photo-effects due to the excitation of carriers in the silicon by scattered light used for atom photoionization (PI, 405 nm) and ion Doppler cooling (422 nm). During trap loading, this manifested itself as variation of the RF voltage amplitude on the trap electrodes due to varying impedance of the trap when the 405-nm PI light was on. The effects on the trapping potential were visible as ion motion synchronized to the PI light switch state. We observed no photo-effects in traps with a ground plane. Traps without a ground plane also exhibit strong trap-temperature-dependent nonlinearities in the resonance response of the voltage-step-up resonator. A ground plane reduces RF leakage into the silicon substrate sufficiently to eliminate this effect, such that we observed stable trapping for chip temperatures from 300 K down to 8 K. We noticed slightly more power dissipation in the foundry traps than in traps fabricated from gold or niobium on sapphire for similar RF voltage amplitude[@PhysRevA.89.012318], most likely due to higher dissipation in the metals or dielectrics. Ion lifetimes of more than an hour were observed in the presence of Doppler cooling light, equivalent to the best lifetimes seen in other traps measured in this vacuum system. Excess micromotion (ion motion at the RF drive frequency) is caused by static electric fields that displace the ion from the RF null and can lead to ion heating. We compensate for this micromotion using the standard method of applying an additional opposing static field. Typical stray field values are on the order of 500 V/m here and appear stable over days. Although silicon oxide dielectric is exposed at the locations of gaps in the electrodes and may charge due to laser-induced photo-electron production, the stray field’s stability suggests another cause. Wirebonds, which are asymmetric with respect to the ion location and also closer to the ion here than in the case of larger trap chips, may be responsible for the steady stray electric field. The use of through-silicon-via technology can eliminate wirebonds from the chip surface, as has recently been demonstrated for surface-electrode ion traps[@gtri:invacuumelectronics:2014]. When compared to single-metal-layer traps (SMLTs) on sapphire substrates, these traps exhibit increased scatter of laser light, possibly due to higher as-deposited roughness of the aluminum layer. We examined the trap-electrode surface using atomic force microscopy and measured an RMS roughness of 35 nm, significantly larger than the 2 nm we have measured on SMLTs. Scatter from the surface can be reduced by focusing laser beams to a smaller diameter at the trap. ![(Color online) Representative measurement of heating rate in a CMOS-foundry-fabricated ion trap. Average occupation of the axial mode of vibration in the linear trap is plotted as a function of delay time after preparation in the ground state at a trap temperature of 8.4 K. A linear fit (line shown) gives a heating rate for these data of 80(5) quanta/s where the uncertainty is due to statistical errors propagated through the fit. An average of five such measurements gives a heating rate of 81(9) quanta/s where the uncertainty is due to run-to-run variability. The inset shows all five measurements with a line indicating the weighted average value.\[figure:heating\]](heatinsetv18){width="\columnwidth"} Trapped-ion multi-qubit quantum operations can be limited by electric field noise that heats the ions’ shared vibrational modes in the trap, reducing gate fidelity[@NIST:ExpIssueswithIons:JresNIST:98; @nist:muwavetechniques:2013]. Anomalously large heating rates caused by unknown noise sources have been seen in every trapped-ion experiment that has examined motional-state heating. This is particularly noticeable in small microfabricated traps as the heating rate appears to scale as $1/d^{4}$ for an ion a distance $d$ from a trap electrode surface. It is therefore important to characterize the heating rate in potentially scalable trap technologies. Using the dipole-forbidden $S_{1/2}\to D_{5/2}$ transition in $^{88}$Sr$^{+}$, we performed resolved-sideband cooling to prepare the ion in the ground state (average occupation $\bar{n}\approx0.05$) of the 1.3 MHz axial vibrational mode and then measured the heating rate using sideband amplitude spectroscopy on this transition after a varying delay[@PhysRevA.89.012318]. Results of one such measurement are presented in Fig. \[figure:heating\] for a chip temperature of 8.4 K. Five measurements were recorded over a few days for nominally the same conditions; the average heating rate is 81(9) quanta/s. When scaled by $1/d^{4}$ to compare traps of different sizes, this heating rate is lower than that reported in any other trap fabricated on a silicon substrate[@britton:nist:2008; @SMIT:QIC:2009; @Oxford:CaCharging:APB:11; @npl:monolithictrap:2012; @vittorini:043112; @Innsbruck:CryoSiTrap:arXiv:14]. Motional heating at this level would lead to an error of less than $10^{-2}$ in a 100 $\mu$s two-ion-qubit gate, below the fault-tolerance threshold for large scale quantum computing with surface-code error-correction schemes[@fowler:surfacecodereview:PRA09]. We have shown basic functionality for quantum processing using a fabrication process, without modification, that has enabled scaling to billions of transistors. This is the first demonstration, in any physical implementation, of quantum computing hardware co-fabricated with scalable classical computing hardware. The fabrication of advanced CMOS and photonic technology on the trap chip, including the extensive existing libraries of integrated circuits for digital logic and memory, offers a straightforward path to scalable, local optical and electronic control and readout of trapped-ion arrays. The demonstration of stable trapping and low electric-field noise in a foundry-process trap is therefore an initial step toward integration of the required classical computing and photonic devices for useful, large-scale quantum processing with trapped ions. We thank Peter Murphy, Chris Thoummaraj, and Karen Magoon for assistance with ion-trap-chip packaging at Lincoln Laboratory. K.K.M. and R.J.R acknowledge funding from DARPA MTO, NSF iQuISE, and a DOE Science Graduate Fellowship. I.L.C. acknowledges funding from IARPA. The work at Lincoln Laboratory is sponsored by the Assistant Secretary of Defense for Research & Engineering under Air Force contract number FA8721-05-C-0002. Opinions, interpretations, conclusions, and recommendations are those of the authors and are not necessarily endorsed by the United States Government. [^1]: K. K. Mehta, A. M. Eltony, and C. D. Bruzewicz contributed equally to this work. [^2]: K. K. Mehta, A. M. Eltony, and C. D. Bruzewicz contributed equally to this work. [^3]: K. K. Mehta, A. M. Eltony, and C. D. Bruzewicz contributed equally to this work.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Let $f=a{{\tt x}}+{{\tt x}}^{3q-2}\in\Bbb F_{q^2}[{{\tt x}}]$, where $a\in\Bbb F_{q^2}^$. We prove that $f$ is a permutation polynomial of $\Bbb F_{q^2}$ if and only if one of the following occurs: (i) $q=2^e$, $e$ odd, and $a^{\frac{q+1}3}$ is a primitive $3$rd root of unity. (ii) $(q,a)$ belongs to a finite set which is determined in the paper.' address: - 'Department of Mathematics and Statistics, University of South Florida, Tampa, FL 33620' - 'Department of Mathematics and Statistics, University of South Florida, Tampa, FL 33620' author: - 'Xiang-dong Hou\' - 'Stephen D. Lappano' title: Determination of a Type of Permutation Binomials over Finite Fields --- [^1] Introduction ============ A polynomial $f\in\Bbb F_q[{{\tt x}}]$ is called a [permutation polynomial]{} (PP) of $\Bbb F_q$ if it induces a permutation of $\Bbb F_q$. While permutation monomials of $\Bbb F_q$ are obvious ($a{{\tt x}}^n$, $a\in\Bbb F_q^$, $\text{gcd}(n,q-1)=1$), the situation for permutation binomials is much more interesting and challenging. The reason for a binomial to be a PP can be quite nontrivial despite the simple appearance of the binomial. In [@Car-Wel66], Carlitz and Wells proved that for fixed integers $e>1$ and $c> 0$, when $q$ is large enough and satisfies the conditions $e\mid q-1$ and $\text{gcd}(c,q-1)=1$, there exists $a\in\Bbb F_q^$ such that ${{\tt x}}^c({{\tt x}}^{\frac{q-1}e}+a)^k$ is a PP of $\Bbb F_q$ for all $k\ge 0$. (Note that when $k=1$, the PP is a binomial.) The special cases of this result with $c=k=1$ and $e=2,3$ appeared in [@Car62]. Carlitz and Wells’ proof of the existence result relies on a bound on the Weil sum of a multiplicative character of $\Bbb F_q$ [@Wei41], [@LN Theorem 5.39]. Using the Hasse-Weil bound on the number of degree one places of a function field over $\Bbb F_q$ [@Sti93 Theorem V.2.3], Masuda and Zieve [@Mas-Zie09] were able to make Carlitz-Wells’ existence result (with $k=1$) more precise. They proved that if $q\ge 4$ and $\frac{q-1}e>2q(\log\log q)/\log q$, then there exists $a\in\Bbb F_q^$ such that ${{\tt x}}^c({{\tt x}}^{\frac{q-1}e}+a)$ is a PP of $\Bbb F_q$. Moreover, they obtained an estimate for the number of $a$’s with this property. There are also nonexistence results on permutation binomials. Niederreiter and Robinson [@Nie-Rob82] proved that if there is a PP of $\Bbb F_q$ of the form ${{\tt x}}^m+a{{\tt x}}$, where $m>2$ and $a\in\Bbb F_q^$, then either $m$ is a power of $p$ ($p=\text{char}\,\Bbb F_q$) or $q<(m^2-4m+6)^2$. An improvement of this result was obtained by Turnwald [@Tur88]: If there is a PP of $\Bbb F_q$ of the form ${{\tt x}}^m+a{{\tt x}}^n$, where $m>n>0$ and $a\in\Bbb F_q^$, then either $\frac mn$ is a power of $p$ or $q\le(m-2)^4+4m-4$. For permutation binomials over prime fields, the nonexistence results are stronger. Wan [@Wan87] proved that if there is a PP of $\Bbb F_p$ of the form ${{\tt x}}^m+a{{\tt x}}$, where $m>1$ and $a\in\Bbb F_p^$, then $p-1\le(m-1)\text{gcd}(m-1,p-1)$. Turnwald [@Tur88] considered $f={{\tt x}}^m+a{{\tt x}}^n\in\Bbb F_p[{{\tt x}}]$, where $m>n>0$ and $a\in\Bbb F_p^$, and proved that $f$ is a PP of $\Bbb F_p$ implies $p<m\cdot\max(n,m-n)$. Masuda and Zieve [@Mas-Zie09] improved Turnwald’s bound to $p-1<(m-1)\cdot\max\{n,\text{gcd}(m-n,p-1)\}$. Let $r\ge 2$. In [@Car62], Carlitz proved that the binomial ${{\tt x}}^{1+\frac{q-1}2}+a{{\tt x}}$ ($q$ odd, $a\ne0$) cannot be a PP of $\Bbb F_{q^r}$, and he raised the same question for ${{\tt x}}^{1+\frac{q-1}3}+a{{\tt x}}$ ($q\equiv 1\pmod 3$, $a\ne 0$). Wan [@Wan87; @Wan94] answered Carlitz’s question by showing that ${{\tt x}}^{1+\frac{q-1}3}+a{{\tt x}}$ ($q\equiv 1\pmod 3$, $a\ne 0$) cannot be a PP of $\Bbb F_{p^r}$. Kim and Lee [@Kim-Lee95] proved that ${{\tt x}}^{1+\frac{q-1}5}+a{{\tt x}}$ ($q\equiv 1\pmod 5$, $a\ne 0$) cannot be a PP of $\Bbb F_{q^r}$ for $p\ne 2$. More generally, one may consider ${{\tt x}}^{1+\frac{q-1}m}+a{{\tt x}}\in\Bbb F_q[{{\tt x}}]$, where $q\equiv 1\pmod m$, $m\ge 2$, $a\ne 0$. Clearly, if $m=\frac{q-1}{p^i-1}$, where $\Bbb F_{p^i}\subset\Bbb F_q$, then ${{\tt x}}^{1+\frac{q-1}m}+a{{\tt x}}={{\tt x}}^{p^i}+a{{\tt x}}$, which is a PP of $\Bbb F_{q^r}$ if and only if $(-a)^{(q^r-1)/(p^i-1)}\ne 1$. When $1+\frac{q-1}m$ is not a power of $p$, it is not known if the binomial can be a PP of $\Bbb F_{q^r}$. Let $f={{\tt x}}^m+a{{\tt x}}^n\in\Bbb F_q[{{\tt x}}]$, where $m>n>0$ and $a\in\Bbb F_q^$. The conditions that make $f$ a PP of $\Bbb F_q$ are encoded in a simple set of parameters $m,n,q,a$ in a mysterious way that is not well understood on the whole. However, when $m$ and $n$ take certain particular forms, necessary and sufficient conditions for $f$ to be a PP of $\Bbb F_q$ have been found. Niederreiter and Robinson [@Nie-Rob82] proved that ${{\tt x}}^{\frac{q+1}2}+a{{\tt x}}\in\Bbb F_q[x]$ ($q$ odd, $a\in\Bbb F_q^$) is a PP of $\Bbb F_q$ if and only if $a^2-1$ is a square in $\Bbb F_q^$; also see [@Car62]. Akbary and Wang [@Akb-Wan06] considered binomials of the form $f={{\tt x}}^r(1+{{\tt x}}^{es})$, where $e,r,s$ are positive integers such that $s\mid q-1$, $\text{gcd}(r,s)=1$, $\text{gcd}(2e,\frac{q-1}s)=1$. They found sufficient conditions for $f$ to be a PP of $\Bbb F_q$ in terms of the period of the generalized Lucas sequence. The conditions are not entirely explicit, but their special cases do give explicit classes of permutation binomials of $\Bbb F_q$. The sufficient conditions in [@Akb-Wan06] were later extended by Wang [@Wan07] to conditions that are both necessary and sufficient. Zieve [@Zie09] considered $f={{\tt x}}^m+a{{\tt x}}^n\in\Bbb F_q[{{\tt x}}]$, where $m>n>0$ and $a\in\Bbb F_q^$, under the assumption that $\{\eta+\frac 1\eta:\eta\in\mu_{2d}\}\subset\mu_s$, where $s=\text{gcd}(m-n,q-1)$, $d=\frac{q-1}s$ and $\mu_s=\{x\in\Bbb F_q:x^s=1\}$. In this setting, it was shown that $f$ is a PP of $\Bbb F_q$ if and only $-a\notin\mu_d$, $\text{gcd}(n,s)=1$ and $\text{gcd}(2d,m+n)\le 2$. (This characterization implies the aforementioned sufficient conditions in [@Akb-Wan06]; see [@Zie09].) Recently, we determined all PPs of $\Bbb F_{q^2}$ of the form $a{{\tt x}}+{{\tt x}}^{2q-1}$ ($a\in\Bbb F_{q^2}^$) [@Hou13; @Hou-p]. The techniques introduced there are applicable to binomials of similar types. In the present paper, we determine all PPs of $\Bbb F_{q^2}$ of the form $a{{\tt x}}+{{\tt x}}^{3q-2}$. Our main result is the following. \[T1.1\] Let $f=a{{\tt x}}+{{\tt x}}^{3q-2}\in\Bbb F_{q^2}[{{\tt x}}]$, where $a\in\Bbb F_{q^2}^$. Then $f$ is a PP of $\Bbb F_{q^2}$ if and only if one of the following occurs. - $q=2^{2k+1}$ and $a^{\frac{q+1}3}$ is a primitive $3$rd root of unity. - $q=5$ and $a^2$ is a root of $({{\tt x}}+1)({{\tt x}}+2)({{\tt x}}-2)({{\tt x}}^2-{{\tt x}}+1)$. - $q=2^3$ and $a^3$ is a root of ${{\tt x}}^3+{{\tt x}}+1$. - $q=11$ and $a^4$ is a root of $({{\tt x}}-5)({{\tt x}}+2)({{\tt x}}^2-{{\tt x}}+1)$. - $q=17$ and $a^6=4,5$. - $q=23$ and $a^8=-1$. - $q=29$ and $a^{10}=-3$. [Note.]{} Shortly after the submission of the first vision of this paper, we were informed by M. Zieve of two very recent papers [@TZHL] by Tu, Zeng, Hu and Li and [@Zie13] by Zieve in the arXiv. We remark that under the assumption that $a$ is a $(q+1)$st root of unity, Theorem \[T1.1\] (i), which is the only relevant case under this assumption, follows from [@Zie13 Corollary 5.3]. If, in addition, $q$ is assumed to be even, then [@TZHL Theorem 1] also gives sufficiency part of Theorem \[T1.1\] (i). Preliminaries ============= Let $f=a{{\tt x}}+{{\tt x}}^{3q-2}\in\Bbb F_{q^2}[{{\tt x}}]$, where $a\in\Bbb F_{q^2}^$, and let $0\le\alpha,\beta\le q-1$. We have $$\label{2.1} \begin{split} \sum_{x\in\Bbb F_{q^2}}f(x)^{\alpha+\beta q} =\;&\sum_{x\in\Bbb F_{q^2}^}(ax+x^{3q-2})^\alpha(a^qx^q+x^{3-2q})^\beta\cr =\;&\sum_{x\in\Bbb F_{q^2}^}\sum_{i,j}\binom\alpha i(ax)^{\alpha-i}x^{(3q-2)i}\binom\beta j(a^qx^q)^{\beta-j}x^{(3-2q)j}\cr \end{split}$$ $$\kern 2.5cm =a^{\alpha+\beta q}\sum_{i,j}\binom\alpha i\binom\beta ja^{-i-jq}\sum_{x\in\Bbb F_{q^2}^}x^{\alpha+\beta q+3(q-1)(i-j)}.$$ The inner sum is $0$ unless $\alpha+\beta q\equiv 1\pmod{q-1}$, i.e., $\alpha+\beta=q-1$. Assume $0\le \alpha\le q-1$ and $\beta=q-1-\alpha$ in . We have $$\label{2.2} \begin{split} &\sum_{x\in\Bbb F_{q^2}}f(x)^{\alpha+(q-1-\alpha)q}\cr =\;&a^{(\alpha+1)(1-q)}\sum_{i,j}\binom\alpha i\binom{q-1-\alpha}ja^{-i-jq}\sum_{x\in\Bbb F_{q^2}^}x^{(q-1)[-\alpha-1+3(i-j)]}\cr =\;&-a^{(\alpha+1)(1-q)}\sum_{-\alpha-1+3(i-j)\equiv 0\, (\text{mod}\, q+1)}\binom\alpha i\binom{q-1-\alpha}ja^{-i-jq}. \end{split}$$ As $i$ runs over the interval $[0,\alpha]$ and $j$ over the interval $[0,q-1-\alpha]$, the range of $-\alpha-1+3(i-j)$ is $$\label{2.3} I_\alpha:=[2\alpha+2-3q,\ \alpha-1].$$ Thus we have $$\label{2.4} \sum_{x\in\Bbb F_{q^2}}f(x)^{\alpha+(q-1-\alpha)q}=-a^{(\alpha+1)(1-q)}S_q(\alpha,a),$$ where $$\label{2.4a} S_q(\alpha,a)= \sum_{-\alpha-1+3(i-j)\in I_\alpha\cap(q+1)\Bbb Z}\binom\alpha i\binom{q-1-\alpha}ja^{-i-jq}.$$ By Hermite’s criterion, $f$ is a PP of $\Bbb F_{q^2}$ if and only if $0$ is the only root of $f$ and $$\label{2.5} S_q(\alpha,a)=0\quad \text{for all}\ 0\le \alpha\le q-1.$$ \[L2.1\] If $f$ is a PP of $\Bbb F_{q^2}$, then $q+1\equiv 0\pmod 3$. Assume $f$ is a PP of $\Bbb F_{q^2}$ and assume $q\ge 3$. If $q=3$, the only multiple of $q+1$ in $I_0=[2-3q,-1]$ is $-(q+1)$. By (with $\alpha=0$), we have $$0=\sum_{-1-3j=-(q+1)}\binom{q-1}ja^{-jq}=-a^{-q},$$ which is a contradiction. Now assume $q\ge 4$. The multiples of $q+1$ in $I_0=[2-3q,-1]$ are $-2(q+1)$ and $-(q+1)$. By (with $\alpha=0$), $$\label{2.6} 0=\sum_{-1-3j=-2(q+1),-(q+1)}\binom{q-1}ja^{-jq}=\binom{q-1}{\frac{2q+1}3}^a^{-\frac{2q+1}3q}+\binom{q-1}{\frac q3}^a^{-\frac q3},$$ where $$\binom mn^= \begin{cases} \displaystyle\binom mn&\text{if}\ n\in\Bbb N,\vspace{2mm}\cr 0&\text{otherwise}. \end{cases}$$ If, to the contrary, $q+1\not\equiv 0\pmod 3$, then exactly one of $\binom{q-1}{\frac{2q+1}3}^$ and $\binom{q-1}{\frac q3}^$ is nonzero, and hence cannot hold. \[R2.2\]Assume $q+1\equiv 0\pmod 3$. - $0$ is the only root of $f$ in $\Bbb F_{q^2}$ if and only if $a^{\frac{q+1}3}\ne 1$. (Note that $(-1)^{\frac{q+1}3}=1$.) - For $b\in\Bbb F_{q^2}^$, we have $f(b{{\tt x}})=b^{3q-2}(b^{3(1-q)}a{{\tt x}}+{{\tt x}}^{3q-2})$. Thus if $f$ is a PP of $\Bbb F_{q^2}$ for $a=a_0$, the same is true for $a=\epsilon a_0$, where $\epsilon\in\Bbb F_{q^2}$, $\epsilon^{\frac{q+1}3}=1$. \[L2.3\] Assume $q+1\equiv 0\pmod 3$, $\alpha>0$, $\alpha+1\equiv 0\pmod 3$, $q\ge 2\alpha+4$. Then $$\label{2.7} S_q(\alpha,a)=(-a)^{\frac{q+1}3 q}\sum_{i=0}^\alpha(-1)^i\binom\alpha i\biggl[\binom{i+\frac{2\alpha-1}3}\alpha v^{3i}+\binom{i+\frac{2\alpha}3}\alpha v^{3i+1}+\binom{i+\frac{2\alpha+1}3}\alpha v^{3i+2}\biggr],$$ where $S_q(\alpha,a)$ is defined in and $v=a^{-\frac{q+1}3}$. Since $\alpha>0$ and $q\ge 2\alpha+4$, the multiples of $q+1$ in $I_\alpha=[2\alpha+2-3q,\,2\alpha-1]$ are $-2(q+1),-(q+1),0$. We have $$\begin{split} &\text{LHS of \eqref{2.7}}\cr =\;&\sum_{-\alpha-1+3(i-j)=-2(q+1),-(q+1),0}\binom\alpha i\binom{q-1-\alpha} ja^{-i-jq}\cr =\;&\sum_{i=0}^\alpha\binom\alpha i\sum_{l=0}^2\binom{-1-\alpha}{\frac 13(l(q+1)-\alpha-1)+i} a^{-i-[\frac13(l(q-1)-\alpha-1)+i]q}\cr =\;&\sum_{i=0}^\alpha\binom\alpha i\sum_{l=0}^2(-1)^{\frac13(l(q+1)-\alpha-1)+i}\binom{\frac13(l(q+1)-\alpha-1)+i+\alpha}\alpha a^{\frac{\alpha+1}3q-\frac{q+1}3(l+3i)}\cr &\kern5.2cm (\binom{-m}n=(-1)^n\binom{n+m-1}{m-1}\ \text{for}\ m,n\in\Bbb N)\cr =\;&(-a)^{\frac{q+1}3 q}\sum_{i=0}^\alpha(-1)^i\binom\alpha i\sum_{l=0}^2\binom{i+\frac{2\alpha-1+l}3}\alpha v^{3i+l}. \end{split}$$ Proof of Theorem \[T1.1\] ========================= \[L3.1\] Assume $q+1\equiv 0\pmod 3$ and $y:=a^{\frac{q+1}3}$ is a primitive $3$rd root of unity. Then for $1\le s\le q^2-2$, $$\label{3.1} \sum_{x\in\Bbb F_{q^2}}f(x)^s= \begin{cases} \displaystyle a^{-\frac 16(q+1)(3q-2)}(1+y)&\text{if $q$ is odd and}\ \displaystyle s=\frac{q^2-1}2,\vspace{2mm}\cr 0&\text{otherwise}. \end{cases}$$ By the observation after , we only have to consider $s=\alpha+(q-1-\alpha)q$, $0\le \alpha\le q-1$. By , we may further assume $\alpha+1\equiv 0\pmod 3$. There is nothing to prove if $q=2$. Thus we assume $q>3$. $1^\circ$ We claim that $I_\alpha=[2\alpha+2-3q,\, 2\alpha-1]$ contains exactly three consecutive multiples of $q+1$ unless $\alpha=\frac{q-1}2$ ($q$ odd); in the latter case, $I_\alpha=[-2q+1,\, q-2]$ contains exactly two multiples of $q+1$ which are $-(q+1)$ and $0$. The length of $I_\alpha$ is $2\alpha-1-(2\alpha+2-3q)=3(q-1)$. Since $2q<3(q-1)<3(q+1)$, there are at least $2$ and at most $3$ multiples of $q+1$ in $I_\alpha$. Let $k(q+1),(k+1)(q+1)\in I_\alpha$. If $I_\alpha=[2\alpha+2-3q,\,2\alpha-1]$ contains only two multiples of $q+1$, then $$\label{3.2} \begin{cases} 2\alpha+2-3q\ge k(q+1)-q,\cr 2\alpha-1\le (k+1)(q+1)+q. \end{cases}$$ (See Figure \[F1\].) $$\beginpicture \setcoordinatesystem units <3mm,3mm> point at 0 0 \arrow <4pt> [0.3, 0.67] from -1 -2 to -13 -2 \arrow <4pt> [0.3, 0.67] from 1 -2 to 13 -2 \setlinear \plot -16 0 16 0 / \plot -14 0 -14 0.3 / \plot -6 0 -6 0.3 / \plot 6 0 6 0.3 / \plot 14 0 14 0.3 / \plot -13 0 -13 -3 / \plot 13 0 13 -3 / \put {$I_\alpha$} at 0 -2 \put {$\scriptstyle k(q+1)-q$} [b] at -14 0.6 \put {$\scriptstyle k(q+1)$} [b] at -6 0.6 \put {$\scriptstyle (k+1)(q+1)$} [b] at 6 0.6 \put {$\scriptstyle (k+1)(q+1)+q$} [b] at 14 0.6 \endpicture$$ Since $2\alpha+2-3q,\; 2\alpha-1\equiv 0\pmod 3$, we have $$\label{3.3} \begin{cases} 2\alpha+2-3q\ge k(q+1)-q+2,\cr 2\alpha-1\le (k+1)(q+1)+q-2. \end{cases}$$ Taking the difference of the two inequalities in , we conclude that $$\begin{cases} 2\alpha+2-3q = k(q+1)-q+2,\cr 2\alpha-1 = (k+1)(q+1)+q-2. \end{cases}$$ Since $0\le \alpha\le q-1$, we must have $k=-1$. Thus $\alpha=\frac{q-1}2$. $2^\circ$ First assume $\alpha\ne\frac{q-1}2$. Let $I_\alpha\cap (q+1)\Bbb Z=(q+1)L$, where $L$ is a set of three consecutive integers. Then we have $$\label{3.4} \begin{split} S_q(\alpha, a)=\;&\sum_{-\alpha-1+3(i-j)\in I_\alpha\cap(q+1)\Bbb Z}\binom\alpha i\binom{q-1-\alpha}j a^{-i-jq}\cr =\;&\sum_{l\in L}\;\sum_{-\alpha-1+3(i-j)=l(q+1)}\binom\alpha i\binom{q-1-\alpha}j a^{-i-jq}\cr =\;&\sum_{l\in L}\;\sum_{i-j=\frac13[\alpha+1+l(q+1)]}\binom\alpha i\binom{q-1-\alpha}j a^{-i+j}\qquad\text{(since $a^{q+1}=1$)}\cr =\;&\sum_{l\in L}a^{-\frac 13[\alpha+1+l(q+1)]}\sum_{i-j=\frac13[\alpha+1+l(q+1)]}\binom\alpha {\alpha-i}\binom{q-1-\alpha}j\cr =\;&a^{-\frac 13(\alpha+1)}\sum_{l\in L}y^{-l}\sum_{\alpha-i+j=\frac13[2\alpha-1-l(q+1)]}\binom\alpha {\alpha-i}\binom{q-1-\alpha}j\cr =\;&a^{-\frac 13(\alpha+1)}\sum_{l\in L}y^{-l}\binom{q-1}{\frac 13[2\alpha-1-l(q+1)]}\cr =\;&-a^{-\frac 13(\alpha+1)}\sum_{l\in L}y^{-l}\cr =\;&0. \end{split}$$ In the next-to-last step, we have $\binom{q-1}{\frac 13[2\alpha-1-l(q+1)]}=(-1)^{\frac 13[2\alpha-1-l(q+1)]}=-1$ since $0\le \frac 13[2\alpha-1-l(q+1)]\le \frac 13[2\alpha-1-(2\alpha+2-3q)]=q-1$. Combining and gives $\sum_{x\in\Bbb F_{q^2}}f(x)^s=0$. $3^\circ$ Now assume $\alpha=\frac{q-1}2$. By the calculation in , we have $$\label{3.5} S_q(\alpha, a)=-a^{-\frac 13(\alpha+1)}\sum_{l=-1,0}y^{-l}=-a^{-\frac 16(q+1)}(1+y).$$ Combining and gives $$\sum_{x\in\Bbb F_{q^2}}f(x)^s=a^{\frac{q+1}2(1-q)-\frac{q+1}6}(1+y)=a^{-\frac 16(q+1)(3q-2)}(1+y).$$ ($\Leftarrow$) (ii) – (vii) are sporadic cases with small $q$. Using a computer, it is easy to verify that $f$ is a PP in each of these cases. Now assume (i). By Lemma \[L3.1\], we have $\sum_{x\in\Bbb F_{q^2}}f(x)^s=0$ for all $1\le s\le q^2-2$. Also note from Remark \[R2.2\] (i) that $0$ is the only root of $f$ in $\Bbb F_{q^2}$. Thus $f$ is a PP of $\Bbb F_{q^2}$. ($\Rightarrow$) Assume that $f$ is a PP of $\Bbb F_{q^2}$. By Lemma \[L2.1\], we have $q+1\equiv 0\pmod 3$. Let $y=a^{\frac{q+1}3}$. If $y$ is a primitive $3$rd root of unity, by Lemma \[L3.1\], $q$ must be even, i.e., $q=2^{2k+1}$, and we have case (i). Now assume that $y^2+y+1\ne 0$. We show that one of the cases (ii) – (vii) occurs. The sum in the right side of is a polynomial in $y^{-1}$ ($=v$) which can be easily computed for small $\alpha$ with computer assistance. For $\alpha=2,5,8,11,14$, we find that $$\label{3.6} S_q(\alpha,a)=(-a)^{\frac{\alpha+1}3q}y^{-3\alpha-2}(y^2+y+1) \begin{cases} 3^{-2}g_2(y)&\text{if}\ \alpha=2,\ q\ge 8,\vspace{1mm}\cr 3^{-6}g_5(y)&\text{if}\ \alpha=5,\ q\ge 14,\vspace{1mm} \cr 3^{-10}g_8(y)&\text{if}\ \alpha=8,\ q\ge 20,\vspace{1mm} \cr 3^{-15}g_{11}(y)&\text{if}\ \alpha=11,\ q\ge 26,\vspace{1mm} \cr 3^{-19}g_{14}(y)&\text{if}\ \alpha=14,\ q\ge 32, \end{cases}$$ where $$\begin{split} g_2({{\tt x}})=\,&2 {{\tt x}}^5+3 {{\tt x}}^4-23 {{\tt x}}^3-8 {{\tt x}}^2-9 {{\tt x}}+44, \cr g_5({{\tt x}})=\,&-14 {{\tt x}}^{14}-8 {{\tt x}}^{13}+22 {{\tt x}}^{12}-469 {{\tt x}}^{11}-1093 {{\tt x}}^{10}+8852 {{\tt x}}^9+6801 {{\tt x}}^8+10527 {{\tt x}}^7\cr &-61068 {{\tt x}}^6-18619 {{\tt x}}^5-25033 {{\tt x}}^4+120197 {{\tt x}}^3+13516 {{\tt x}}^2+16822 {{\tt x}}-71162, \cr g_8({{\tt x}})=\,& 130 {{\tt x}}^{23}+57 {{\tt x}}^{22}-187 {{\tt x}}^{21}+4082 {{\tt x}}^{20}+3585 {{\tt x}}^{19}-7667 {{\tt x}}^{18}+156234 {{\tt x}}^{17}\cr &+453573 {{\tt x}}^{16}-3916551 {{\tt x}}^{15}-4144622 {{\tt x}}^{14}-7594467 {{\tt x}}^{13}+48939959 {{\tt x}}^{12}\cr &+25221008 {{\tt x}}^{11}+39342423 {{\tt x}}^{10}-213366911 {{\tt x}}^9-61811112 {{\tt x}}^8-88032825 {{\tt x}}^7\cr &+422650317 {{\tt x}}^6+66303028 {{\tt x}}^5+88882095 {{\tt x}}^4-389019163 {{\tt x}}^3-25886212 {{\tt x}}^2\cr &-33211905 {{\tt x}}+135094180, \cr g_{11}({{\tt x}})=\,& -3952 {{\tt x}}^{32}-1522 {{\tt x}}^{31}+5474 {{\tt x}}^{30}-139802 {{\tt x}}^{29}-89324 {{\tt x}}^{28}+229126 {{\tt x}}^{27}\cr &-3943602 {{\tt x}}^{26}-4392909 {{\tt x}}^{25}+8336511 {{\tt x}}^{24}-180820302 {{\tt x}}^{23}-605825169 {{\tt x}}^{22}\cr &+5521784781 {{\tt x}}^{21}+7111655988 {{\tt x}}^{20}+14607372831 {{\tt x}}^{19}-101269369227 {{\tt x}}^{18}\cr &-69095625624 {{\tt x}}^{17}-119477261853 {{\tt x}}^{16}+705650100129 {{\tt x}}^{15}+303870716124 {{\tt x}}^{14}\cr &+475920749355 {{\tt x}}^{13}-2503382174319 {{\tt x}}^{12}-706243777836 {{\tt x}}^{11}-1034492806725 {{\tt x}}^{10}\cr &+4972469163636 {{\tt x}}^9+898579001889 {{\tt x}}^8+1253008322595 {{\tt x}}^7-5598768742164 {{\tt x}}^6\cr &-591556509206 {{\tt x}}^5-794043854630 {{\tt x}}^4+3339003167188 {{\tt x}}^3+157572058982 {{\tt x}}^2\cr &+205140400010 {{\tt x}}-819352075360, \cr g_{14}({{\tt x}})=\,& 41800 {{\tt x}}^{41}+14895 {{\tt x}}^{40}-56695 {{\tt x}}^{39}+1691000 {{\tt x}}^{38}+905631 {{\tt x}}^{37}-2596631 {{\tt x}}^{36}\cr &+47250150 {{\tt x}}^{35}+37894401 {{\tt x}}^{34}-85144551 {{\tt x}}^{33}+1395800990 {{\tt x}}^{32}+1826164521 {{\tt x}}^{31}\cr &-3221965511 {{\tt x}}^{30}+75566097190 {{\tt x}}^{29}+281332431561 {{\tt x}}^{28}-2683745985685 {{\tt x}}^{27}\cr &-3976231919076 {{\tt x}}^{26}-8901790877799 {{\tt x}}^{25}+65232090577890 {{\tt x}}^{24}+53701334712609 {{\tt x}}^{23}\cr &+100487514543597 {{\tt x}}^{22}-632854611825486 {{\tt x}}^{21}-347885978019711 {{\tt x}}^{20}\cr &-586551837541203 {{\tt x}}^{19}+3307822221633594 {{\tt x}}^{18}+1283881108529889 {{\tt x}}^{17}\cr &+2015859062567817 {{\tt x}}^{16}-10419893389315746 {{\tt x}}^{15}-2892546806289271 {{\tt x}}^{14}\cr &-4307726185011783 {{\tt x}}^{13}+20728564105915330 {{\tt x}}^{12}+4054215382726378 {{\tt x}}^{11}\cr &+5793391583605092 {{\tt x}}^{10}-26245535590106350 {{\tt x}}^9-3451745974770042 {{\tt x}}^8\cr &-4770402189292728 {{\tt x}}^7+20520594631893930 {{\tt x}}^6+1634454816505198 {{\tt x}}^5\cr &+2197118421394272 {{\tt x}}^4-9034762128135730 {{\tt x}}^3-330180086243950 {{\tt x}}^2\cr &-433563200685120 {{\tt x}}+1713531735146800. \end{split}$$ Combining and , we have $$\label{3.7} \begin{cases} g_2(y)=0&\text{if}\ q\ge 8,\cr g_5(y)=0&\text{if}\ q\ge 14,\cr g_8(y)=0&\text{if}\ q\ge 20,\cr g_{11}(y)=0&\text{if}\ q\ge 26,\cr g_{14}(y)=0&\text{if}\ q\ge 32. \end{cases}$$ When $q<14$, a quick computer search produces cases (ii) – (iv). So we assume $q\ge 14$. Next we compute the resultant of $g_2$ and $g_5$: $$R(g_2,g_5)=2^5\cdot 3^{35}\cdot 17^2\cdot 23\cdot 29 \cdot 103\cdot 16069.$$ By and the fact that $q+1\equiv 0\pmod 3$, we must have $p=\text{char}\,\Bbb F_q\in\{2,17,23,29\}$. When $p=2$, we have $q\ge 32$. In this case, $\text{gcd}(g_2,g_5,g_8)={{\tt x}}$, which is a contradiction to . When $p=17$, we find that $\text{gcd}(g_2,g_5,g_8)=1$. By , we must have $q=17$. A computer search results in case (v). When $p=23$, $\text{gcd}(g_2,g_5,g_8)={{\tt x}}+1$, and $g_{11}(-1)=12\ne 0$. By , we must have $q=23$. A computer search results in case (vi). When $p=29$, $\text{gcd}(g_2,g_5,g_8)={{\tt x}}+10$, and $g_{11}(-10)=0$, $g_{14}(-10)=2\ne 0$. By , we must have $q=29$. A computer search results in case (vii). [99]{} A. Akbary and Q. Wang, [A generalized Lucas sequence and permutation binomials]{}, Proc. Amer. Math. Soc. [134]{} (2006), 15 – 22. L. Carlitz, [Some theorems on permutation polynomials]{}, Bull. Amer. Math. Soc. [68]{} (1962) 120 – 122. L. Carlitz and C. Wells, [The number of solutions of a special system of equations in a finite field]{}, Acta Arith. [12]{} (1966/1967) 77 – 84. X. Hou, [A class of permutation binomials over finite fields]{}, J. Number Theory [133]{} (2013), 3549 – 3558. X. Hou, [Determination of a type of permutation trinomials over finite fields, II]{}, preprint. S. Y. Kim and J. B. Lee, [Permutation polynomials of the type $x^{1+((q-1)/m)} +ax$]{}, Commun. Korean Math. Soc. [10]{} (1995), 823 – 829. R. Lidl and H. Niederreiter, [Finite Fields]{}, 2nd ed., Cambridge Univ. Press, Cambridge, 1997. A. M. Masuda and M. E. Zieve, [Permutation binomials over finite fields]{}, Trans. Amer. Math. Soc. [361]{} (2009), 4169 – 4180. H. Niederreiter and K. H. Robinson, [Complete mappings of finite fields]{}, J. Austral. Math. Soc. Ser. A [33]{} (1982), 197 – 212. H. Stichtenoth, [Algebraic Function Fields and Codes]{}, Universitext. Springer-Verlag, Berlin, 1993. Z. Tu, X. Zeng, L. Hu, C. Li, [A class of binomial permutation polynomials]{}, [arXiv:1310.0337]{}, 2013. G. Turnwald, [Permutation polynomials of binomial type]{}, Contributions to General Algebra, 6, 281 – 286, Hölder-Pichler-Tempsky, Vienna, 1988. D. Wan, [Permutation polynomials over finite fields]{}, Acta Math. Sinica (N.S.) [3]{} (1987), 1 – 5. D. Wan, [Permutation binomials over finite fields]{}, Acta Math. Sinica (N.S.) [10]{} (1994), Special Issue, 30 – 35. Q. Wang, [Cyclotomic mapping permutation polynomials over finite fields]{}, in [Sequences, Subsequences, and Consequences]{}, S.W. Golomb, G. Gong, T. Helleseth, H.-Y. Song, (Eds.), pp. 119 – 128, Lecture Notes in Comput. Sci., vol. 4893, Springer, Berlin, 2007. A. Weil, [On the Riemann hypothesis in function fields]{}, Proc. Nat. Acad. Sci. U. S. A. [27]{} (1941), 345 – 347. M. E. Zieve, [On some permutation polynomials over $\Bbb F_q$ of the form $x^r h(x^{(q-1)/d})$]{}, Proc. Amer. Math. Soc. [137]{} (2009), 2209 – 2216. M. E. Zieve, [Permutation polynomials on $\Bbb F_q$ induced from Rédei function bijections on subgroups of $\Bbb F_q^$]{}, [arXiv:1310.0776]{}, 2013. [^1]: \ Research partially supported by NSA Grant H98230-12-1-0245.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The high-energy Universe is potentially a great laboratory for searching new light bosons such as axion-like particles (ALPs). Cosmic sources are indeed the scene of violent phenomena that involve strong magnetic field and/or very long baselines, where the effects of the mixing of photons with ALPs could lead to observable effects. Two examples are archetypal of this fact, that are the Universe opacity to gamma-rays and the imprints of astrophysical magnetic turbulence in the energy spectra of high-energy sources. In the first case, hints for the existence of ALPs can be proposed whereas the second one is used to put constraints on the ALP mass and coupling to photons.' address: 'CEA, Irfu, Centre de Saclay, F-91191 Gif-sur-Yvette \| France' author: - Pierre Brun bibliography: - 'brun.bib' title: 'Axion-like particles: possible hints and constraints from the high-energy Universe' --- Motivations for axion-like particle searches ============================================ The Standard Model of particle physics reproduces data incredibly well [@Beringer:1900zz]. Some of its foundations are however not completely understood, like for example the absence of CP violation in quantum chromodynamics (QCD). The most general QCD lagrangian includes a complex phase term which –if not exactly zero– induces CP violation. The non-observation of even a very small electric dipole moment for the neutron [@Baker:2006ts] implies that this phase is smaller than $10^{-11}$. This fact looks unnatural an calls for a explanation. A possible one is given by making this phase a dynamical field which value is driven to zero by the action of its classical potential. This is made possible by the introduction of a new U(1) global symmetry which is spontaneously broken at some scale $f$ (this is the so-called Peccei-Quinn symmetry [@Peccei:1977hh]). A new particle that is called the axion is then predicted as an associated pseudo Nambu-Goldstone boson [@Wilczek:1977pj; @Weinberg:1977ma]. In the original idea of Peccei and Quinn, $f$ was of the order of the electroweak scale (EW), inducing a mass of $\sim$100 keV for the axion, which was quickly ruled out (see [@Kim:1986ax; @Ringwald:2012hr] for details). Then it was proposed that $f$ was much greater than the EW scale, leading to a very light and weakly interacting axion (dubbed “invisible axion"). Axions are experimentally searched through their coupling to photons, from the Sun [@2009JCAP...02..008A; @2011PhRvL.107z1302A], assuming they make up the galactic dark matter [@Asztalos:2011bm] or with LASERs [@Ehret:2010mh]. Beyond the case of the strong CP problem and axions, axion-like particles (ALPs) appear in many models of physics beyond the Standard Model such as string theory [@Svrcek:2006yi; @Arvanitaki:2009fg; @Ringwald:2012cu] as pseudo Nambu-Goldstone bosons associated to the breaking of U(1) symmetries. The properties of the associated particles are similar to that of axions, but in general their mass and coupling to photons are not related, making the corresponding parameter space larger. For the general ALP case, the interaction term with photons is $$\mathcal{L}\;=\;-\frac{1}{4}g_{\gamma a} F_{\mu\nu}\tilde{F}^{\mu\nu} a \;=\; g_{\gamma a} \vec{E}\cdot\vec{B} \; a\;\;,\label{eq:1}$$ where $g_{\gamma a}$ is the dimensionful coupling between photons and ALPs, $F$ is the electromagnetic tensor and $a$ is the ALP field. On the right hand-side of Eq. \[eq:1\], the $F\tilde{F}$ term is expressed as a scalar product of the photon electric field and the magnetic field, revealing the fact that ALPs can couple to photons in the presence of an external magnetic field. In the present article, the use of natural (astrophysical) environments to search for ALPs is emphasized, like during the propagation of very high energy gamma-rays over cosmological distances, and the effect of astrophysical magnetic turbulence on high-energy photon source spectra. First the conventional view of the problem of the opacity of the Universe to gamma-rays is presented, with the discussion of a possible indication for an anomalously transparent Universe. Although the possible tensions can be solved in a conventional way, they can also be released by invoking ALPs mixing with photons. Then it is shown that this observable could be used as a signature when one tries to make a discovery, but that some uncertainties prevent from using it to derive robust constraints. It is then shown that constraints can be obtained by considering the effect of magnetic turbulence around the sources and finally some examples of constraints are given. The transparency of the Universe and ALPs ========================================= The conventional view of the Universe opacity to gamma-rays ----------------------------------------------------------- Very high energy photons (with $\sim$TeV energies) traveling through the intergalactic medium encounter different populations of background radiations. The most numerous type of background photons belong to the Cosmic Microwave Background (CMB) and a second population is the extragalactic background light (EBL). The latter has a double bump structure, that comes from direct starlight and emission re-processed by interstellar dust in the infrared band as sketched on Fig. \[fig:EBL\]. Direct measurement of the EBL is very difficult because of foregrounds and infrared radiation by the instruments. ![Left: Spectral energy density of the cosmic background photons including the CMB and the EBL (inspired from [@Dole:2006de], with permission of the authors). Right: Illustration of the pair production process.\[fig:EBL\]](EBL2.pdf "fig:"){width=".5\textwidth"} ![Left: Spectral energy density of the cosmic background photons including the CMB and the EBL (inspired from [@Dole:2006de], with permission of the authors). Right: Illustration of the pair production process.\[fig:EBL\]](pair2.pdf "fig:"){width=".4\textwidth"} TeV gamma-ray astronomy is sensitive to the EBL density and spectrum as it is responsible for the attenuation of extragalactic source fluxes at high energy. The reason for that is the pair production process $\gamma_{\rm TeV}\gamma_{\rm EBL}\rightarrow e^+e^-$, for which the threshold lies at TeV energies in the terrestrial frame. For instance considering $E_{\rm EBL}\sim 0.1\;\rm eV$, the threshold energy satisfying $E_{\rm th}E_{\rm EBL}>m_e^2$ (where $m_e$ is the mass of the electron) yields $E_{\rm th}\sim2.6\;\rm TeV$. In Fig. \[fig:EBL\] the typical range of the TeV absorption range is indicated by the horizontal arrow. Because of the pair production process, the highest energy photons have a larger optical depth. Before 2006 it was commonly admitted that is was very unlikely to detect TeV photons from sources above $z\sim 0.2$. The situation changed after HESS observations of two active galactic nuclei (AGNs) at $z=0.186$ and $z=0.165$. As reported in [@Aharonian:2005gh], when unfolded from the EBL effect, the intrinsic spectra of these sources were found too hard and in tension with the source models. This was the first indication for a Universe slightly more transparent than expected at high energy. Later, AGNs were observed at redshifts as high as 0.536 by MAGIC [@Aliu:2008ay] and possibly 0.61 by HESS [@Becherini:2012ei]. The Universe is indeed more transparent to gamma-rays that expected. That puzzle has conventional solutions, it could be for instance that spectra are actually harder, this can be realized for instance including hadronic components in the AGN jets or in relativistic shock acceleration models. The tension can be removed as well with a revision of the EBL models, mainly with a lower density as for instance in the model of [@Franceschini:2008tp] which is compatible with all observations. TeV observations are now even used to provide not only upper limits on the EBL density but actual measurements [@Abramowski:2012ry]. The current situation is illustrated in Fig. \[fig:eblconstr\], extracted from [@Wouters:2012ek; @hess]. It represents the energy above which the absorption becomes significant (defined by a optical depth $\tau(E)=1$) as a function of the redshift of the source. The lines correspond to models or lower limits for the EBL density and redshift evolution from [@Franceschini:2008tp; @Dominguez:2010bv; @Kneiske:2010pt]. Constraints from the spectral indices of different sources are shown as arrows, and the HESS measurement corresponds to the blue band. One source seems to be in tension with the measurement. However the methods that lead to the constraint and the measurements are different as the constraints rely on spectral slope measurements and the measurement comes from the observation of features in the spectra that can be related to the EBL spectral density. A more unified approach might be necessary to get a definite answer on how strong the tension is. Note also that the Fermi collaboration did the same measurement at higher redshifts and found a good compatibility with EBL models [@:2012gd]. ![Energies corresponding to $\tau=1$ for different EBL models, constraints from very high energy gamma-ray astronomy and 1-$\sigma$ measurements from HESS (figure from [@Wouters:2012ek; @hess]).\[fig:eblconstr\]](eblconstr.pdf){width="\textwidth"} Some studies however still claim for an anomaly, even with the lower EBL limits from [@Kneiske:2010pt]. It is the case in [@Horns:2012fx], where the authors claim for an anomaly, with the caveat that their claim requires to leave out some error bars. How ALPs enter the game ----------------------- The lack of opacity of the Universe to gamma-rays gave rise to the idea that ALPs could be responsible for this effect. The basic idea is that if mixing between ALPs and photons occur, the beam could travel in the form of ALPs on a significant fraction of way, not producing pairs, as sketched on Fig. \[fig:ALPs\]. If ALPs are converted back to photons before observations, this could lead to a more transparent Universe. ![Illustration of the modeling of the extragalactic magnetic field\[fig:mag\]](ALPs.pdf){width="\textwidth"} ![Illustration of the modeling of the extragalactic magnetic field\[fig:mag\]](mag.pdf){width="\textwidth"} To get an idea of the relevant masses and couplings for the ALPs that are invoked here, let us consider the most simple formalism for describing the photon/ALP mixing. The system propagation is described by a Schrödinger-like equation: $$(E-i\partial_z-\mathcal{M}) \left(\begin{array}{c} A\\a \end{array}\right ) \;=\;0 \;\; \text{with} \;\; \mathcal{M}\;=\; \left(\begin{array}{cc} -i\frac{\tau}{2z} & \Delta_B \\ \Delta_B & \Delta_a \end{array}\right ) \;\;,$$ where $\Delta_B=g_{\gamma a}B_t/2$ describes the photon/ALP coupling ($B_t$ is the transverse projection of the magnetic field), $\tau$ is the optical depth related to EBL absorption and $\Delta_a=-m_a^2/2E$ accounts for the ALP mass. The fact that the imaginary coefficient only applies to the photon part of the wavefunction leads to the change in the overall transparency. In the case of no absorption, the mixing matrix is diagonalized with a rotation angle $\theta$ such that $\tan 2\theta = -2\Delta_B/\Delta_a$. The resolution of the propagation equation in the propagation state basis leads to the probability of transition $$P_{\gamma\rightarrow a}\;=\;\frac{1}{2}\frac{1}{1+\left (E_c/E \right )^2}\sin^2\left (\frac{g_{\gamma a} B_t \;z}{2} \sqrt{1+\left (\frac{E_c}{E}\right )^2} \right ),\;\;\text{with} \;\; E_c=\frac{m_a^2}{2 g_{\gamma a} B_t}\;\;.$$ The overall 1/2 coefficient in $P_{\gamma \rightarrow a}$ accounts for the two polarizations of the photon. A critical energy $E_c$ appears, that defines the energy scale at which strong mixing occurs. From the expression of $E_c$, with cosmological magnetic fields $B\sim 1\;\rm nG$, an ALP mass $m_a\sim \rm neV$ and a coupling $g_{\gamma a}\sim 10^{-11}\;\rm GeV^{-1}$, the critical energy lies at the TeV scale. It follows that the type of ALPs that are concerned by the so-called transparency hint will fall in a region of low masses and with couplings larger than those of the corresponding axions. The full treatment of the transparency problem in the presence of ALPs requires a $3\times 3$ mixing matrix to account for the two polarization states for the photon, and a description of the magnetic field on the path from the source to the observatory. The extragalactic magnetic field is usually described as a patches of coherent domains of 1 Mpc size. The magnetic field strength is the same in all domains but from one domain to the next its orientation changes in a random way (see the sketch of Fig. \[fig:mag\]). It can be shown (see [@Grossman:2002by]) that for random orientations and a large number $N$ of domains, the transition probability is reduced to $$P_{\gamma\rightarrow a}\;=\;\frac{1}{3}\left ( 1- \exp\left( -3NP_0 \right ) \right )\;\;,$$ where $P_0$ is the transition probability in one domain. From this expression one would expect to have a $1/3$ drop in the energy spectrum above $E_c$ in the limit $NP_0\gg1$, and a flux that is boosted at high energy (typically above the pair-production related cutoff) as described in [@SanchezConde:2009wu]. At least two facts lead to revise the above statements. First, in practice the limit $NP_0\gg1$ is hardly realized. Second, due to the unknown nature of the magnetic field configuration, the prediction on the transmission has an intrinsic variance. Indeed it can happen that the ALPs do not convert back into photons before reaching the Earth, leading in that case to an even more opaque Universe. This is nicely illustrated in Fig. \[fig:stoch\] extracted from [@Mirizzi:2009aj]. Here the red dot-dashed line corresponds to the conventional opacity in the absence of ALPs and the solid black line is the average prediction with ALPs. It appears that the average transparency is indeed higher than the conventional case at high energy. However, the associated uncertainty on the prediction, in other words the variance related to the randomness of the magnetic field is such that the envelope includes the conventional case. Because of that fact, if observed without ambiguity in the future, such an effect might be seen as an indication for ALP detection but could hardly serve as a firm argument for discovery. ![Transmission of photons with and without ALPs, the ALP case is ploted with the envelope corresponding to the variance on the prediction of the transparency effect (figure from [@Mirizzi:2009aj], with permission of the authors).\[fig:stoch\]](stochastic.pdf){width=".8\textwidth"} Another limitation comes from the use of a very optimistic value for the extragalactic magnetic field. For the ALP effect to significantly affect the opacity, magnetic fields of nG strength with Mpc coherence length have to be present in the intergalactic medium. It is actually possible to generate such magnetic fields from inflation or QCD phase transition for instance, but the required strength is very close to current upper limits. This is illustrated in Fig. \[fig:magfield\] (from [@Neronov:1900zz]), where the different observational constraints on large scale magnetic field appear together with predictions from models (orange thin lines). There the red cross corresponds to the typical parameters used in the ALP analyses. It lies in a region that can be seen as fine-tuned given the size of the still open parameter space. At the moment it seems invoking such a strong magnetic field would be acceptable if the tension in the TeV observations was stronger. ![Constraints on large scale magnetic fields and prediction of the models. The red cross corresponds to the parameters used for the ALP solution to the transparency “hint" (figure adapted from [@Neronov:1900zz], with permission of the authors).\[fig:magfield\]](magfield.pdf){width=".9\textwidth"} A clever way to avoid using intergalactic magnetic fields is to remark that if the source is magnetized or embedded in a cluster, then the mixing to ALPs could occur essentially around the source. Then, the magnetic field of the Milky-Way can serve as a target magnetic field to convert back the ALPs into photons. In that case as well, a strong boost can be expected at high energy, as first proposed in [@Simet:2007sa] and then again in [@Horns:2012kw]. This effect is used in [@Meyer:2013pny] to estimate possible lower limits on the $g_{\gamma a}$ coupling, but again assuming the tension is real between observations and models for the transparency. Finally it appears that the transparency observable could be used in the future for indication or discovery, if a clear tension was observed. To do so, one might wait for the next generation of gamma-ray telescopes such as CTA to have a significantly larger sample of sources. The problem of having only a few sources can be circumvented by using an energy band for which detections are numerous. This has been proposed in [@Burrage:2009mj], where the authors remark that if the strong mixing regime is realized, the statistical properties of the observed fluxes from X-ray sources could display features distinctive of ALP effects. In that case, sources would be seen with fluxes reduced by a factor of 1/3 on average. Of course having no access to the absolute intrinsic fluxes, this overall factor is not observable. However because of the random nature of the mixing process in astrophysical magnetic fields, the first and second momentum distribution should have different shapes compared to the conventional case. In [@Burrage:2009mj] the authors claim the observation of anomalous features in the momenta distributions. That result has however shown to be questionable in [@Pettinari:2010ay] where the effect is claimed to be caused by outliers. Because in that case the detection would rely on shapes of distributions, it is difficult to infer a constraint without a deeper analysis and this effect is again used to propose a hint. Nevertheless, it illustrates one possible use of the stochastic nature of the mixing in astrophysical environments, which is no more a limitation but becomes a tool for identifying possible ALP effects. In the following, it is shown that a careful study of this randomness can lead to observable effects that are used to set constraints on the ALP parameters. Constraints on ALP parameters from observations of the high-energy sky ====================================================================== Effect of the magnetic turbulence --------------------------------- One peculiar effect of photon/ALP mixing is the fact that the magnetic field turbulence can directly imprint features in the energy spectra from high-energy sources. The exact spectral shape one gets at the end is unpredictable, but as shown in [@Wouters:2012qd] the statistical properties of the induced irregularities are a prediction of the ALP model. The authors of [@Ostman:2004eh] and [@Mirizzi:2009aj] already noticed that in principle the observed spectra should be very irregular in case of strong photon/ALP mixing, without considering the use of the irregularity as an observable. To account quantitatively for the irregularity, the 2 polarizations of the photon must be considered, so that the evolution of the system after $n$ domains is given by $$\|\psi_n\rangle\;=\;\prod_k \left ( P_k^{-1}\; \exp \left [ -i \left (E + \mathcal{M}^\star_k \right ) s_k\right ]\; P_k\right )\; \|\psi_0\rangle\;\;,\;\;\;\;\text{with}\;\;\mathcal{M}^\star_k=P_k \mathcal{M}_{k} P_k^{-1}\;\;$$ and $$\mathcal{M}_k\;\;=\;\;\left(\begin{array}{ccc} -\frac{m_\gamma^2}{2E}-i\frac{\tau}{2z} & 0 & \frac{1}{2} g_{\gamma a} B_t^{(k)}\cos\phi^{(k)} \\ \\ 0 & -\frac{m_\gamma^2}{2E}-i\frac{\tau}{2z} & \frac{1}{2} g_{\gamma a} B _t^{(k)} \sin\phi^{(k)}\\ \\ \frac{1}{2} g_{\gamma a} B_t^{(k)} \cos\phi^{(k)} & \frac{1}{2} g_{\gamma a} B_t^{(k)}\sin\phi^{(k)} & \frac{-m_{\rm a}^2}{2E} \end{array}\right)\;\;. \label{eq:33mat}$$ $k$ stands for the $k^{\rm th}$ domain, of size $s_k$, $P_k$ is the rotation matrix between the interaction eigenstates and the propagation eigenstates and the matrix $\mathcal{M}_k$ describes the mixing. The indexes $k$ are there to recall that from one magnetic domain to the next, the corresponding parameters change due to the different orientations of the magnetic field ($B_t$ is the projection of the magnetic field on the polarization plane and $\phi$ is the angle that projection makes with one of the two photon polarization). $m_\gamma = 4\pi\alpha n_e/m_e$ is the effective mass of the photon propagating in a plasma with electron density $n_e$. Examples of spectral oscillation patterns in one domain are given in Fig. \[fig:osc\] for different values of $\delta=g_{\gamma a} B_t s/2$, $s$ being the size of the coherent domain. When several domains are considered, the spectrum ends up being very irregular as shown in Fig. \[fig:irr\] in the case of an unpolarized beam. For that example, an extragalactic source is considered and the magnetic field is typical of that of a galaxy cluster. The top panel of Fig. \[fig:irr\] is the raw signal and the bottom panel is the same signal smoothed by the energy resolution of HESS ($\sim$15%). In that case the critical energy is of order 1 TeV and the effective photon mass is negligible. ![Example of ALP induced irregularity in the TeV range (top panel: Raw signal, bottom panel: Signal smeared with HESS resolution, figure from [@Wouters:2013iya; @hess2]).\[fig:irr\]](Fig1.pdf){width="\textwidth"} ![Example of ALP induced irregularity in the TeV range (top panel: Raw signal, bottom panel: Signal smeared with HESS resolution, figure from [@Wouters:2013iya; @hess2]).\[fig:irr\]](signal_irr.pdf){width="\textwidth"} Whereas in the case of the extragalactic magnetic field the naive description of the turbulence might be sufficient (essentially because its properties are very poorly known), galaxy cluster magnetic fields may deserve a better treatment. The magnetic field in that case is modeled by a gaussian field with zero mean and a distribution of modes that is described by a Kolmogorov-like spectrum as in Eq. \[eq:turb\]: $$\left (\delta B\right )^2 \propto \sigma^2\frac{k^2}{1+(kL_{\rm c})^\gamma}\label{eq:turb}\;\;.$$ The corresponding power spectrum is modeled by a function resembling that of Fig. \[fig:turb\]. In galaxy clusters, the typical coherence length of the magnetic field is 10 kpc and the strength of the field is 1 to 10 $\mu$G. ![Typical power spectrum used for modeling the magnetic turbulence in galaxy clusters.\[fig:turb\]](turb.pdf){width=".9\textwidth"} Examples of constraints ----------------------- The first example of constraints is from the HESS analysis of PKS 2155-304 [@Wouters:2013iya; @hess2], which is an AGN located at $z=0.116$. For that source, both the extragalactic magnetic field and the cluster magnetic field can be considered. In the first case, as previously, one has to assume optimistic values of the magnetic field strength for the irregularity signal to be significant. A galaxy cluster is observed around the source, but no magnetic field measurements are available. So in the case of the galaxy cluster magnetic field, conservative values for the strength and the coherence length are assumed (1 $\mu$G and 10 kpc respectively). As HESS observation ranges from hundreds of GeV to a few TeV, from the expression of the critical energy $E_c$ it is straightforward to see that the typical ALP masses that are probed are of the order of $10^{-8}$ eV. In [@Wouters:2013iya; @hess2], it is shown that the observed energy spectrum does not exhibit strong irregularities. Then an estimator of the irregularity is proposed and numerical simulations are used to exclude sets of parameters that lead to significantly too strong irregular behavior. This exclusion has to be done on a statistical basis as each realization of the magnetic field turbulence is different. The results of the analysis are presented in Fig. \[fig:excl\_hess\]. The method allows to improve the CAST limits in a limited energy range around 20 neV. Another possibility is to use a source that lies at the center of a well studied galaxy cluster. In that case, the magnetic field properties are derived observationally. This is done by studying the Faraday rotation maps of the polarized radio emission from the cluster (see [@Carilli:2001hj] for a review). These studies allow in principle a determination of the full turbulence power spectrum, yielding the intensity of the magnetic field, its coherence scale and the slope of the turbulence spectrum. A very well studied cluster is Hydra, for which a strong X-ray source is present at the center (Hydra A) [@McNamara:2000kj] . In [@Wouters:2013hua], X-ray data from the Chandra satellite are analyzed in order to derive constraints on ALP parameters. In the case of X-rays from Hydra, the diagonal terms in the matrix of Eq. \[eq:33mat\] can be simplified. Indeed the pair production related opacity is irrelevant in the case of X-rays (so $\tau=0$), and the trace of the matrix is dominated by the effective photon mass for $m_a\lesssim 10^{-11}\;\rm eV$. So the constraints are expected to extend to arbitrarily low ALP masses below that value. In [@Wouters:2013hua], the irregularity is estimated by performing $\chi^2$ tests when deriving the energy spectrum with a forward folding method. ALP parameters yielding a too high level of irregularity compared to the data are excluded. The corresponding exclusion curve is displayed in Fig. \[fig:excl\_x\]. It turns out this analysis improves the previous constraints in that mass range from the non-observation of gamma-rays associated with SN 1987 A [@Brockway:1996yr]. ![Exclusion contours from the analysis of X-ray data from Hydra A (Figure from [@Wouters:2013hua]).\[fig:excl\_x\]](constraints.pdf){width="\textwidth"} ![Exclusion contours from the analysis of X-ray data from Hydra A (Figure from [@Wouters:2013hua]).\[fig:excl\_x\]](fig3.pdf){width=".9\textwidth"} Outlook ======= The study of the high-energy Universe is potentially a nice way to search for axion-like particles. The problem of the transparency of the Universe to gamma-rays can provide a interesting observable, that requires nevertheless the observation of a large number of TeV sources to be robust. This can be achieved with the next generation of Cherenkov telescopes such as CTA. The photon/ALP mixing in astrophysical sources is intrinsically a stochastic process. That fact makes difficult the use of the transparency observations to derive constraints on the ALP parameters. It is noted however that the turbulence of the astrophysical magnetic fields has the effect of inducing irregularities in the energy spectra of sources. The statistical properties of the induced irregularity can be predicted and are used to set limits on the ALP coupling to photons. Because the method is insensitive to the polarization, these constraints go beyond classic ALPs and apply to both $F\tilde{F}$ and $F^2$ types of couplings. Two examples of limits are given in the case of TeV and X-ray observations of high-energy emitting sources inside clusters of galaxies.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The cosmic ray spectrum has been shown to extend well beyond [$10^{20}$eV]{}. With nearly 20 events observed in the last 40 years, it is now established that particles are accelerated or produced in the universe with energies near or above [$10^{21}$eV]{}. No nearby astrophysical object has been shown to correlate with the arrival directions of the highest energy events, yet the exponential cut-off in the high energy end of the spectrum one expects to see in the case of far sources is not visible. It was recently pointed out that the influence of the vacuum of quantum electrodynamics on particle propagation could explain qualitatively this mystery. This note is a critic to these ideas.' address: 'L.P.N.H.E. Paris VI-VII, 4 place Jussieu, 75252 Paris, France' author: - Olivier Deligny title: 'Comment on “The Lamb Shift and Ultra High Energy Cosmic Rays” and Comment on “Vacuum Polarization Energy Losses of High Energy Cosmic Rays”.' --- Introduction ============ The origin of Ultra High Energy Cosmic Rays observed on Earth is a long lasting mystery[@Yoshida; @Sigl; @Bertou; @Nagano]. While the cosmic ray spectrum is now shown[@HiresTaup99; @Agasa00] to extend beyond [$10^{20}$eV]{}, mechanisms producing or accelerating particles with energies near or above [$10^{21}$eV]{} are still uncertain. The detection of a large flux of events above $5\times$[$10^{19}$eV]{} is considered with great interest because of the absence of the Greisen-Zatsepin-Kuzmin (GZK) cut-off[@GZK]. This cut-off should be observed if cosmic primaries are protons originating from cosmologically distributed sources. Above $5\times$[$10^{19}$eV]{}, on their way from the sources to Earth, cosmic ray protons loose their energy photo-producing pions against the cosmic microwave background (CMB). Therefore sources further than about 50 Mpc away from Earth are not expected to contribute to the high energy end of the spectrum giving rise to the GZK cut-off. If sources are closer than 50 Mpc, protons of [$10^{20}$eV]{} should point toward them, however the present data does not indicate any correlation between arrival direction of these events and local distribution of galaxies and is mostly isotropic. Many ideas have been put forward in the literature to answer some of those problems, but the general opinion is that none of them are complety satisfactory. Recently, two papers[@Xue; @Maisheev] analysing the propagation of protons in vacuum and in the CMB claimed to qualitatively explain the absence of the GZK cut-off. After recalling a famous analogy between the dielectric properties of matter and the quantum electrodynamics vacuum, we discuss these ideas and demonstrate that one leads to (unexplicit!) Lorentz violation while the result of the semi-classical calculation of the other is not compatible with a proper QED treatment. Dielectric properties of matter =============================== When an external electric field is applied to a medium, the atomic electrons are shifted from their original equilibrium position in the field of the nuclei. Similarily if we introduce a negatively charged particle in this medium, atomic electrons will be pulled away and an excess of positive charges will surround the particle. Thus, the negative charge is screened by the positive ones and the force felt by two charges introducted in the medium becomes $$\frac {e^2}{\epsilon r^2}$$ where $r$ is the distance between the two charges and $\epsilon$ is the dielectric constant ($\epsilon > 1$). This description applies to distances greater than the interatomic distances. For smaller distances there is no screening. A unified definition was introduced by Landau[@Landau] through the integral operator $\epsilon (r)$, which approachs 1+ for $r$ small compared to the interatomic distance and the classical $\epsilon$ at larger ones. Vacuum polarization =================== The classical relativistic vacuum is Lorentz invariant. It is a reference system for the description of the motion as well as the propagation of the electromagnetic field. This vacuum is empty. This classical description was modified with the birth of statistical mechanics, and of course of quantum physics. From the statistical mechanics point of view, at non-zero temperature, the space is filled by blackbody radiation. In this approach, extracting all matter in a region of space is not sufficient to turn it into vacuum : one must also bring the temperature down to zero. The quantum theory goes further and predicts that field fluctuations subsist even at zero temperature. This is a consequence of the definition of the quantum vacuum and of the Heisenberg inequalities. Vacuum is the state for which there is no field excitations but which still contains zero-point field fluctuations. These fluctuations have observable consequences as the spontaneous emission of an isolated atom, the Lamb shift, or the Casimir force... It is instructive to consider the vacuum of quantum electrodynamics as a polarized medium under the effect of an external field, in order to understand how this vacuum feels the application of this external field. This well known analogy is quite rigourous for the electric field even from the point of view of QED. We can think of this vacuum as a stock of virtual particles which can have an existence during short periods of time according to the Heisenberg inequalities. Thus, when, for instance, a positive charge is introducted in this stock, polarization of the virtual pairs occurs : the virtual positrons are expelled from the positive charge whereas the virtual electrons are attracted by the charge. A positive vacuum surrounds the charge, and its bare value depends on the way it is measured, that is to say depends on the radius of the sphere we take around the charge to measure the surface value of the electric field. Again, screening occurs up to a distance $r$ of the order of $\hbar /mc$ where $m$ is the mass of the (screening) particle pairs. Can the vacuum accelerate a charged particle ? ============================================== In his paper relating the Lamb shift and the propagation of protons in vacuum[@Xue], the author considers that the proton field must be coupled to the vacuum fluctuations. In order to describe the propagation of the proton in the QED vacuum, fundamental fields are decomposed following their classical part and their quantum fluctuations parts; leading to the Lagrange density $L(x)$ : $$\begin{aligned} L(x)&=& -\frac{1}{4}F^2 - \frac{1}{4}F_q^2 + \overline{\Psi}(i\gamma^{\mu}\partial_{\mu}-m_p+e\gamma^{\mu}A_{\mu}^q)\Psi\\ &+&\overline{\Psi}_q(i\gamma^{\mu}\partial_{\mu}-m_e+e\gamma^{\mu}(A_{\mu}+A_{\mu}^q))\Psi_q + (c.t.)\end{aligned}$$ where the fields indexed by $q$ describe the quantum fluctuations and (c.t.) are all necessary counterterms. At this level, we wonder why the terms $-\frac{1}{2}FF_q$ and $e\overline{\Psi}\gamma^{\mu}A_{\mu}\Psi$ are absent of the decomposition : there is of course no reason to drop these terms. Next, an effective Lagrange density is computed by perturbation in term of the electromagnetic coupling: $$\begin{aligned} L_{eff}(x)&=&-\frac{1}{4}F^2 + \overline{\Psi}(i\gamma^{\mu}\partial_{\mu}-m_p)\Psi\\ &+& \emph{tr}\ln [S_F^{-1}(x)-V(x)]\end{aligned}$$ $V(x)$ is arbitrary choosen to be the same than the potential which describes the propagation of a lepton in a classical external field[@Itzykson], whereas the needed diagrams to get this potential are not permitted by the starting Lagrange density because of the absence of the term $e\overline{\Psi}\gamma^{\mu}A_{\mu}\Psi$. From there, the main argument is based on the fact that the difference of energy of the vacuum spectrum in presence of the proton can be transferred to the proton. This difference occurs because the virtual pairs are polarized along the proton path whereas these fluctuations are totally random in the absence of the electromagnetic field of the proton. According to the author, this difference accelerates the proton. In order to estimate this acceleration, he considers that virtual fermion pairs can be treated as bound states of size $\frac {1}{\alpha m}$ and density $\frac{3}{4\pi}\alpha ^3 m^3$. At this point, the explicit form of the Hamiltonian of such bound states (not well funded!) is not developped at all, and the compact pile of these bound states to estimate their density is rather surprising. Introducing a modified relation of dispersion to take into account the energy gain responsible of the acceleration $$\begin{aligned} E + \delta E = \sqrt{(\vec{p}+\delta \vec{p})^2 + m_p^2}\end{aligned}$$ the author finally obtains the numerical estimate of the energy gain (without loss of generality, we can consider a straight-line motion) $$\begin{aligned} \frac{\delta E}{\delta x} = \frac{\delta p}{\delta \tau} \sim 2.25\times 10^{-5} \frac{v}{c}\, eV cm^{-1}\end{aligned}$$ However, quantum field theory is based on the theory of the relativity. The properties of the quantum vacuum must satisfy to a precise relation with respect to the principles of Einstein’s theory. Thus, in order to satisfy the principles of relativity, it is essential that the vacuum possesses the properties required by the relativity, that is to say the vacuum must not distinguish inertial frames. If the force that the vacuum exerts on a particle moving with a uniform velocity is not zero, then the vacuum distinguishes an inertial frame with respect to a rest frame. Of course, quantum electrodynamics predict that this force vanishes. In general, this theory predicts that vacuum fluctuations seen by two inertial observers are the same, simply because the vacuum is Lorentz invariant. The acceleration by the vacuum described in[@Xue] depends on the speed of the particle and so, doesn’t preserve the properties of the vacuum for any inertial observer.The model, which makes no sense from the starting Lagrange density, is not Lorentz invariant. To go a little further about this subject, let us consider now general motion in vacuum. Fulling and Davies have shown[@Fulling] that the vacuum resists to any motion to take it back to a uniformly accelerated motion. The corresponding energy variation is radiated. Other remarkable works of Jaekel and Reynaud[@Jaekel] point out that for a uniformly accelerated motion, characterised by the acceleration $$\begin{aligned} w^{\mu} = \ddot{v}^{\mu} + v^{\mu}\dot{v}^{\mu 2}\end{aligned}$$ where $$\begin{aligned} v^{\mu}(\tau) = \frac{dx^{\mu}}{d\tau}\end{aligned}$$ and $\tau$ is the proper time satisfying $d\tau ^2 = dx ^2$, the so-called Abraham-Lorentz reaction force, which is proportional to $w^{\mu}$, vanishes. Properties of quantum vacuum are therefore invariant for a set of transformations which include uniformly accelerated frames. These are the conformal transformations.\ A forgotten damping force to explain the cosmic rays spectrum? ============================================================== In another paper describing the energy losses with a damping force by vacuum polarization[@Maisheev], a second author considers the propagation of protons in the cosmic microwave background and thus evades our last critic about the impossibility of acceleration or deceleration in vacuum because of the introduction of a preferred frame. The author calculates the energy losses by vacuum polarization in the same way used by Landau for the ionisation losses of a charged particle travelling in a continuous medium[@Landau]. Even if the damping force considered is caused by electric field of virtual pairs, his approach is classical, consisting in the introduction of a permittivity tensor of the gas of photons (the author assumes that the connection between polarization and permittivity was established by Baier and Katkov[@Baier]). This approach allows him to consider QED vacuum and matter in a unified point of view. After the resolution of the Maxwell equations mapped on what is usually done in a continuous medium, the author computes the final damping force : $$\begin{aligned} F \sim e^6 f(v^2)\end{aligned}$$ where $f$ is a function which contains the proton form factor. The numerical estimation shows that this force becomes essential in the energy range between the knee of high energy cosmic rays spectrum and the GZK cut-off \[[$10^{15}$eV]{}-[$10^{20}$eV]{}\]. Taking into account this energy loss, these new numerical results allow the author to a new interpretation of the existence of the knee, a new explanation of the absence of a clear photoabsorption threshold in the spectrum of high energy cosmic rays, as well as the observed changes in the composition of primaries in this energy range (when applied to iron nuclei). However, if this approach is made in the framework of the quantum electrodynamics as required (only the electric field of virtual pairs can cause this damping force), one observes that the computation of the damping force leads to an expression of order 3, which is compatible with the interference term between lower order of inverse Compton scattering and correction with vacuum polarization and internal photon line in the perturbation theory. In the incoherent gas of photons, this process should lead to a correction to the lower order of the inverse Compton scattering between protons (or nuclei) and photons. This Compton interaction is known to lead to only a negligibly small energy loss in this energy range[@Stecker]. Therefore, this semi-classical approximation doesn’t seem applicable in this context. Conclusions =========== Two recent ideas, proposed to explain the mysterious absence of the GZK cut-off in the UHECR spectrum, have been discussed. We have shown that the first cannot account for the observed phenomenon as the first approach does not respect fundamental principles of relativity. The second is based on a semi-classical approximation which describes coherent effects and doesn’t seem to be applicable in the context of propagation of UHECR. Consequently, the ultra high energy cosmic rays fluxes remains a mystery that current generation of detectors will hopefully solve in the near future[@Auger]. Acknowledgments {#acknowledgments .unnumbered} =============== Many thanks to P.Billoir, M.Boratav, J.Hirn, A.Letessier-Selvon and E.Parizot for helpful discussions, comments and careful read of the text. =cmti9 =cmbx9 [99]{} S. Yoshida, H. Dai, [J. Phys.]{} [G24]{} (1998) 905. P. Bhattacharjee, G. Sigl,[Phys. Rept.]{} [327]{} (2000) 109. X. Bertou, M. Boratav, A. Letessier-Selvon, [Int. J. Mod. Phys.]{} [A15]{} (2000) 2181. M. Nagano, A. A. Watson, [Rev. Mod. Phys.]{}[72]{} No. 3 (2000), 689. J.N. Matthews, C.C.H. Jui, [Nucl. Phys. Proc. Suppl.]{} [B87]{}, 2000 411. M. Takeda [et al.]{}, Astrophys. J. [552]{}, (1999) 225 K. Greisen, [Phys. Rev. Lett.]{} [16]{} (1966) 748; G.T. Zatsepin, V.A. Kuzmin, [JETP Lett.]{} [4]{} (1966) 78. C. Itzykson, J.-B. Zuber, [Quantum Field Theory]{} (1980) McGraw-Hill Inc. L.D. Landau, E.M. Lifchitz, [Courses of Theoritical Physics]{} Vol VIII (1969) Ed. Mir. S-S. Xue, [hep-ph/0207046]{} V.A. Maisheev, [hep-ph/0206185]{} V.N. Baier, V.N. Katkov, [Phys.Lett.]{} [A 216]{} (2001) 11. S.A. Fulling, P.C.W. Davies, [Proceedings Royal Society London.]{} [ A 348]{} (1976) 393. M.T. Jaekel, S. Reynaud, [Rept.Prog.Phys.]{} [60]{} (1997) 863-887, also [quant-ph/9706035]{}. J.L. Puget, F.W. Stecker, J.H. Bredekamp, [ApJ.]{} [205]{} (1976), 638-654. , Fermilab (1995), [www.auger.org.ar/admin/]{}	{ "pile_set_name": "ArXiv" }
--- abstract: 'Sound attenuation and internal friction coefficients are calculated for a realistic model of amorphous silicon. It is found that, contrary to previous views, thermal vibrations can induce sound attenuation at ultrasonic and hypersonic frequencies that is of the same order or even larger than in crystals. The reason is the internal-strain induced anomalously large Grüneisen parameters of the low-frequency resonant modes.' address: - '$^1$Department of Physics, University of Maryland at College Park, College Park, MD 20742-4111' - '$^2$Department of Physics and Astronomy, State University of New York at Stony Brook, Stony Brook, New York 11794-3800' author: - 'Jaroslav Fabian$^{1,2}$ and Philip B. Allen$^2$' title: 'Theory of sound attenuation in glasses: The role of thermal vibrations.' --- @twocolumnfalse Sound attenuation in glasses is poorly understood. This is because many competing factors lead to sound wave damping. Most important are thermally activated structural relaxation, hypothetical tunneling states, topological defects, and thermal vibrations. Sorting out different contributions for a given temperature $T$ and sound wave frequency $\nu=\Omega/2\pi$ is a difficult task. Experiments show the following features: (i) At temperatures $T\alt 10$ K and ultrasonic frequencies (10 MHz to 1 GHz) the sound attenuation coefficient $\Gamma(T)$ exhibits a small, frequency-dependent peak [@hunklinger76]. (ii) At higher temperatures, between 10 and 200 K, another peak in $\Gamma(T)$ develops whose center increases only moderately when $\nu$ increases. The peak broadens at hypersonic frequencies [@vacher81] and is not seen above 100 GHz [@morath97; @zhu91]. As a function of frequency $\Gamma(\nu)\sim\nu$ at the peak temperatures [@vacher81]. (iii) At hypersonic frequencies $\Gamma(T)$ appears to be almost independent of (or slightly increasing with) $T$ above the peak (ii) to at least 300 K [@morath97; @zhu91]. (iv) Room temperature $\Gamma(\nu)\sim\nu^2$ from at least 200 MHz [@vacher81]; this dependence continues for up to about 300 GHz [@morath97; @zhu91] and seems valid for any temperature above the peak (ii). Finally, (v) the attenuation coefficients for longitudinal ($\Gamma_L$) and transverse ($\Gamma_T$) waves are similar. [@vacher81]. While the low-temperature behavior (i) of $\Gamma$ is understood based on the interaction of sound waves with tunneling states [@hunklinger76], features (ii) through (v) lack consistent theoretical justification. The higher-temperature peak (ii) shows many attributes of a thermally activated relaxational process[@tielburger92], but a calculation shows that to fit experiment, different sets of relaxational processes are needed for different $\nu$ [@vacher81]. Also the plateau region (iii) is difficult to explain by a thermally-activated relaxation process since numerical fits require unphysically large attempt frequencies[@zhu91]. Further, thermal relaxation processes give attenuation that increases more slowly than quadratic with increasing $\nu$, contradicting (iv). Thermal vibrations have been overlooked as a sound-wave damping factor on grounds that vibrational modes would need unreasonably large Grüneisen parameters ($\gamma \approx 200$ for vitreous silica [@vacher81]) to account for the measured $\Gamma$. Until now, however, there has been no numerical study to test this argument. In this paper we examine the role thermal vibrations play in the sound attenuation in glasses. We will use the term “vibron” to refer to any quantized vibrational mode in a glass[@fabian96]. Our analysis is restricted to the region $\Omega\tau_{\rm \,in}\alt 1$ (the so called Akhiezer regime[@maris71]), where $\tau_{\rm \,in}$ is the inelastic lifetime or thermal equilibration time of a thermal vibron. We show that the unusually strong coupling (measured by Grüneisen parameters $\gamma$) between sound waves and the low-frequency resonant modes explains the features (iii) through (v). As for the interpretation of (ii), our calculation shows that confusion arises because there actually are two different peaks. One is caused by relaxational processes (not addressed here) and dominates below 1 GHz and another is due to thermal vibrations and dominates at the lowest hypersonic frequencies. A double peak structure should be expected at intermediate frequencies. There is some indication for such structure in measurements on vitreous silica[@vacher81]. Our calculation is also a prediction: the existing measurements on amorphous Si[@haumeder80] report $\Gamma$ at too low frequencies (300 MHz) to see contributions of thermal vibrations. But even at higher frequencies (say, 30 GHz) one may expect traces of thermally activated peaks due to various defects. Recently discovered amorphous Si with 1 at. % H [@liu97] in which tunneling (and perhaps also relaxational) processes are inhibited would be excellent to test our results. In the Akhiezer regime a sound wave passing through a solid can be attenuated by two processes[@gurevich86]. First, if the wave is longitudinal, periodic contractions and dilations in the solid induce a temperature wave via thermal expansion. Energy is dissipated by heat conduction between regions of different temperatures. Second, dissipation occurs as the gas of vibrons tries to reach an equilibrium characterized by a local (sound-wave induced) strain. This is the internal friction mechanism. To establish the relative importance of the two processes, consider order-of-magnitude formulas $\Gamma_h\approx (\Omega^2/\rho v^3)(\kappa T \alpha^2 \rho^2 v^2/C^2)$ and $\Gamma_i\approx(\Omega^2/\rho v^3)(CT\tau_{\rm \,in}\gamma^2)$ for the heat conductivity and internal friction processes, respectively[@gurevich86]. Here $\rho$ is density, $C$ specific heat per unit volume, $v$ sound velocity, $ \kappa$ thermal conductivity, and $\alpha$ is the coefficient of thermal expansion. The ratio $\Gamma_h/\Gamma_i\approx (\kappa\alpha^2\rho^2 v^2)/(C^3\tau_{\rm \,in}\gamma^2)$ becomes more intuitive when putting $\alpha\approx C\gamma/B$ ($B\approx \rho v^2$ is the bulk modulus) and $\kappa\approx CD$ where $D$ is diffusivity. Then $\Gamma_h/\Gamma_i\approx D/(v^2 \tau_{\rm \,in})$. The factor $v^2\tau_{\rm \,in}$ measures the ability of vibrons to absorb energy from a sound wave of velocity $v$. The difference between a glass and a crystal lies in the values of $D$ and $\tau_{\rm \,in}$. In crystals $D\approx v^2\tau_{\rm \,in}$, that is, energy is carried by phonon wave packets with group velocity $v$. The ratio $\Gamma_h/\Gamma_i$ is then of order unity. In glasses energy is transferred by diffusion (spreading rather than ballistic propagation of wave packets[@kelner98]) and $D$ is not related to $\tau_{\rm \,in}$[@allen93]. One of the reasons the contribution to $\Gamma_i$ of thermal vibrons was previously underestimated is that $\tau_{\rm \,in}$ was guessed from thermal conductivity [@vacher81]; this gave too small $\tau_{\rm \,in}$. For amorphous Si $D\approx 10^{-6}$ ${\rm m}^2/s$[@allen93], $v\approx 8\times 10^3$ m/s, and $\tau_{\rm \,in}\approx$ $10^{-12}$s[@fabian96] give $\Gamma_h/\Gamma_i\approx$ 0.02. Since these are typical values, $\Gamma_h$ can be neglected. This is consistent with experiment: compared with crystals glasses have smaller $\kappa$ and yet $\Gamma$ can be larger [@vacher81]. Internal friction leads to sound-wave energy attenuation [@gurevich86] $ \Gamma=(\Omega^2/\rho v^3 q^2)\eta_{\alpha\beta\gamma\delta} {q}_{\alpha}{e}_{\beta}{q}_{\gamma}{e}_{\delta}$, where $\eta_{\alpha\beta\gamma\delta}$ is the internal friction tensor with cartesian coordinates $\alpha...\delta$ and ${{\bf q}}$ (${{\bf e}}$) is the wave vector (polarization) of the sound wave. Summation over repeated indexes is assumed. We will evaluate $\Gamma$ for both longitudinal ($L$) and transverse ($T$) sound waves with wave vectors averaged over all directions: $$\begin{aligned} \label{eqn:1}\Gamma_L=\frac{\Omega^2}{15\rho v^3_L} (\eta_{\alpha\alpha\beta\beta}+2\eta_{\alpha\beta\alpha\beta}),\\ \label{eqn:2}\Gamma_T=\frac{\Omega^2}{30\rho v^3_T} (3\eta_{\alpha\beta\alpha\beta}-\eta_{\alpha\alpha\beta\beta}).\end{aligned}$$ The coefficients $\eta_{\alpha\beta\gamma\delta}$ are the real part of a complex tensor $\bar{\eta}_{\alpha\beta\gamma\delta}$ which can be obtained by solving a kinetic equation in relaxation time approximation [@maris71], $$\begin{aligned} \label{eqn:3} \bar{\eta}_{\alpha\beta\gamma\delta}= \sum_jTc_j\tau_j\frac{ \gamma_{\alpha\beta}^j\gamma_{\gamma\delta}^j- \left(\bar{\gamma}_{\alpha\beta} \gamma_{\gamma\delta}^j+\gamma_{\alpha\beta}^j \bar{\gamma}_{\gamma\delta}\right)/2} {1-i\Omega\tau_j}.\end{aligned}$$ The summation is over all vibrational modes $j$; $c_j$ and $\tau_j$ denote mode specific heat and relaxation time. The Grüneisen tensor $-\gamma_{\alpha\beta}^j$ is the relative shift of mode frequency $\omega_j$ per unit strain $\epsilon_{\alpha\beta}$; $\bar{\gamma}$ is the mode average of $\gamma^j$ weighted with $c_j/(1-i\Omega\tau_j)$. The applicability of kinetic theory to the problem of internal friction was justified by DeVault and coworkers [@devault65] who obtained $\eta$ from a microscopic theory as an autocorrelation function of the momentum current density operator. Remarkably, the microscopic theory shows that the momentum current in a solid is not monopolized by ballistically propagating vibrational modes as in the case of the energy current. Nonpropagating (even localized) modes can contribute as much as propagating ones to the momentum current. One consequence is that the concept of “minimum kinetic coefficient,” as introduced for electrical[@mott70] or heat[@kittel49] conductivity of disordered systems, is not realized for internal friction. We generalized [@fabian97b] DeVault’s theory to include internal strain, the atomic rearrangements in a strained solid. We found that internal strain affects internal friction only by modifying $\gamma^j$, as in the case of thermal expansion[@fabian97]: $\gamma^j$ now reflects the change between the initial mode frequency and the frequency of the mode after the rescaling (scaling parameter $1+\epsilon$) [plus]{} the rearranging of atomic positions (to achieve a new equilibrium at strain $\epsilon$). Internal strain is very important for thermal expansion of glasses[@fabian97]; we will show that it is important for $\eta$ (and $\Gamma$) as well. We calculate $\eta$ and $\Gamma$ for the model of amorphous Si based on the Wooten-Winer-Weaire atomic coordinates [@wooten85] and Stillinger-Weber interatomic forces [@stillinger85], with 1000 atoms arranged in a cube of side 27.549 Å with periodic boundary conditions. Diagonal Grüneisen parameters $\gamma_{\alpha\alpha}^j/3\equiv (\gamma_{11}^j+\gamma_{22}^j+ \gamma_{33}^j)/3 $ for this model[@feldman98] were given in Ref. [@fabian97], transverse $\gamma_{\alpha\beta}$ are calculated here. Vibrational lifetimes $\tau_j$ are extracted from their 216-atom version values[@fabian96] (see also Ref. [@bickham98]). The model has sound velocities $v_L=7640$ m/s and $v_T=3670$ m/s[@feldman91]. Figure \[fig:1\] shows the calculated $\Gamma(\nu)$ for longitudinal and transverse sound waves in amorphous Si from 10 MHz to 1 THz at 300 K. The attenuation $\Gamma\sim \nu^2$ up to about 100 GHz, where the condition for the applicability of kinetic theory, $\Omega\tau_{\rm \,in}\alt 1$ reaches its limit ($\tau_{\rm \,in}\approx$ 1 ps). Our calculation is not valid beyond this point. In comparison, the measured attenuation of longitudinal waves in vitreous silica grows quadratically with $\nu$ up to at least 400 GHz [@zhu91] suggesting that $\tau_{\rm \,in}$ in vitreous silica is several times smaller than in amorphous Si. This is not surprising since Si is remarkably harmonic: room temperature heat conductivity of crystalline Si is larger by an order of magnitude than that of quartz[@mason65], and a similar relation may hold for the corresponding $\tau_{\rm \,in}$ of the glassy phases. More surprising is the comparison with crystalline Si. Figure \[fig:1\] shows that $\Gamma_L$ is similar for the amorphous and crystalline cases (measured $\Gamma$ for vitreous silica is several times larger than for quartz[@vacher81]). One would naively expect the sound attenuation in a glass to be much smaller than in the corresponding crystal since, owing to a distribution of bond lengths and bond angles, anharmonicity of the glass is higher (and $\tau_{\rm \,in}$ smaller). The same interatomic potential, for example, yields $\tau_{\rm \,in}$ for high-frequency phonons in crystalline Si at 300 K about five times larger than in amorphous Si[@fabian96]. The reason why $\Gamma$ in glasses can be of the same order or even higher than in crystals is the internal-strain induced anomalously large Grüneisen parameters of the resonant modes[@fabian97] (see also Fig.\[fig:3\]). (Resonant modes are low-frequency extended modes whose amplitude is unusually large at a small, typically undercoordinated region [@fabian97; @biswas88].) Atomic rearrangements caused by internal strain are largest in the same regions of undercoordination where the resonant modes have largest amplitude [@fabian97]. This leads to high sensitivity (measured by $\gamma$) of the frequencies of these modes to strain. If the internal strain is neglected, the sound attenuation is an order of magnitude smaller, as seen in Fig. \[fig:1\]. (Since the resonant modes have low frequencies, their $\tau_j$ is longer than an average $\tau_{\rm \,in}$; this adds even more weight to these modes.) Fewer than one percent of the modes are capable of increasing $\Gamma$ by a decade! We believe the measured $\Gamma$ for vitreous silica is also caused by the strong coupling of sound waves and resonant modes. Vitreous silica is a much more open structure than amorphous Si so the number of resonant modes should be higher, bringing $\Gamma$ above the crystalline value. Another interesting feature in Fig.\[fig:1\] is the relative attenuation strength for longitudinal and transverse sound waves. While our model of amorphous Si gives $\Gamma_L/\Gamma_T\approx 1/3$ at 300 K, the measured ratio for crystalline Si is reversed: $\Gamma_L/\Gamma_T\approx 3$ [@mason65]. This again shows how differently is sound attenuated in glasses and in crystals. The ratio $\Gamma_L/\Gamma_T$ can be written as $(v_T/v_L)^3(\gamma_L^2/\gamma_T^2)$, where $\gamma_L$ and $\gamma_T$ are effective Grüneisen parameters. A crude way to estimate $\gamma_L^2$ and $\gamma_T^2$, suggested by Eqs. \[eqn:1\] and \[eqn:2\] is to take mode averages of $(\gamma_{\alpha\alpha}^j \gamma_{\beta\beta}^j+ 2\gamma_{\alpha\beta}^j\gamma_{\alpha\beta}^j)/15$ and $(3\gamma_{\alpha\beta}^j \gamma_{\alpha\beta}^j- \gamma_{\alpha\alpha}^j\gamma_{\beta\beta}^j)/30$. Our model gives $\gamma_L^2\approx 3$ and $\gamma_T^2\approx 1$. The ratio $\Gamma_L/\Gamma_T$ is then about 1:3, in accord with the full calculation. Assuming the same ratio $\gamma_L^2/\gamma_T^2\approx 3$ for vitreous silica ($v_L=5800$ m/s and $v_T=3800$ m/s), transverse and longitudinal waves are attenuated about equally. This is observed in experiment[@vacher81]. The explanation of the measured $\Gamma_L/\Gamma_T$ in crystalline Si can be found in Ref. [@mason65]. In Fig. \[fig:2\] we plot $\Gamma(T)$ for different $\nu$. A remarkable feature is a peak at about 20 K at 1MHz and below. As $\nu$ increases, the peak shifts towards higher $T$ and vanishes above 4-5 GHz. Two factors cause the peak. (a) The sum $\sum_j c_j(\gamma^j)^2$ saturates at much lower temperatures (about 50 K) than the model Debye temperature $T_D\approx 450$ K[@feldman91]. This is because the relevant $j$ are resonant modes with small frequencies. (b) For low-frequency modes $T\tau_j$, after increasing linearly develops a peak, before going constant \[much like $\Gamma(T)$ itself\]. As the temperature dependence of $\Gamma$ follows $\sum_jc_j(\gamma^j)^2\ T\tau_j$, the peak appears. At large $\nu$ the peak vanishes because of the factor $1/(1+\Omega^2\tau^2)$ in Eq. \[eqn:3\]. At $T$ above 100 K $\Gamma(T)$ is nearly constant, as observed in experiment as a plateau (iii). This again follows from (a) and (b). We are not aware of any experiment with which we could compare our calculations. The measurement of $\Gamma(T)$ of sputtered amorphous Si films reported in Ref. [@haumeder80], for example, was performed at 300 MHz. This is too low to see any contributions from thermal vibrations. The whole temperature spectrum is dominated by a single peak of the type (ii), except at very low temperatures. This peak is expected to increase linearly with $\nu$, until thermal vibrations become relevant (roughly at 10 GHz), causing a plateau (iv) that increases as $\nu^2$ at higher frequencies. Even at smaller frequencies one may see some vibrational contribution to $\Gamma(T)$ at large enough $T$, since the thermally activated peak decreases as $1/T$ at large $T$. Anomalous low $T$ thermal expansion already suggested[@white75] very large $\gamma$ values for low $\omega$ modes. Our large $\gamma$ values [@fabian97] agree nicely with trends in $\alpha(T)$. Like thermal expansion, $\Gamma$ should be strongly sample and model dependent. There is evidence[@feldman99] that our highly homogeneous model of amorphous Si becomes free of resonant modes when the number of atoms grows to infinity. That means an infinite model would predict $\Gamma$ about a decade smaller than calculated here. Amorphous silicon, however, can be prepared only in thin films where voids and other inhomogeneities are unavoidable. Voids loosen the strict requirements of a tetrahedral random network (for example by introducing free boundary conditions). Then, as in our finite models, regions of undercoordinated atoms will allow the formation of resonant modes. While this issue for amorphous silicon will be ultimately settled by experiment, our calculation combined with the existing data on vitreous silica strongly suggests the reality of resonant modes. Our final note concerns the mode dependence of transverse Grüneisen parameters like $\gamma_{12}$. Similarly to volumetric $\gamma_{\alpha\alpha}/3$[@fabian97], transverse $\gamma_{12}$ in Fig. \[fig:3\] ($\gamma_{13}$ and $\gamma_{23}$ look the same) is unusually large for resonant modes and have scattered values for high-frequency localized modes. (More resonant modes have $\gamma_{12}$ negative than positive which suggests that resonant modes are trapped at highly anisotropic undercoordinated regions whose sizes change under shear[@fabian97].) The $15-70$ meV vibrons (diffusons[@fabian96]) have $\gamma_{12}\approx 0$ (average magnitude 0.02), while the corresponding $\gamma_{\alpha\alpha}/3$ are of order unity[@fabian97]. Such small values (zeros in an infinite model) are characteristic for diffusons, which are extended modes whose polarization directions (atomic displacements) point, in general, at random. There remains only a short-range correlation between polarization directions which determines the diffuson’s frequency $\omega_d$. If a shear, say, $\epsilon_{12}$ is applied, $\omega_d$ changes to $\omega_d'(\epsilon_{12})$. Since long-range order in the diffuson polarization is absent, $\omega_d'(\epsilon_{12})\approx\omega_d'(-\epsilon_{12})$, and $\gamma_{12}$ which is a linear coefficient in the expansion of $\omega_d'$ in $\epsilon_{12}$ must vanish. We thank J. L. Feldman for helpful discussions. The work was supported by NSF Grant No. DMR 9725037. J. F. acknowledges also support from the U.S. ONR. S. Hunklinger and W. Arnold, in [Physical Acoustic]{}, edited by W. P. Mason and R. N. Thurston (Academic, New York, 1976), p. 155. R. Vacher, J. Pelous, F. Plicque, and A. Zarembowitch, J. Non-Cryst. Solids [45]{}, 397 (1981). C. J. Morath and H. J. Maris, Phys. Rev. B [ 54]{}, 203 (1996). T. C. Zhu, H. J. Maris, and J. Tauc, Phys. Rev. B [44]{}, 4281 (1991); H. N. Lin, R. J. Stoner, H. J. Maris, and J. Tauc, J. Appl. Phys. [69]{}, 3816 (1991). D. Tielbürger, R. Merz, R. Ehrenfels, and S. Hunklinger, Phys. Rev. B [45]{}, 2750 (1992). J. Fabian and P. B. Allen, Phys. Rev. Lett. [77]{}, 3839 (1996) H. J. Maris, in [Physical Acoustics]{}, edited by W. P. Mason and R. N. Thurston (Academic, New York, 1971), Vol. VIII. M. Von Haumeder, U. Strom, and S. Hunklinger, Phys. Rev. Lett. [44]{}, 84 (1980). X. Liu, B. E. White, R. O. Pohl, E. Iwanizcko, K. M. Jones, A. H. Mahan, B. N. Nelson, R. S. Crandall, and S. Veprek, Phys. Rev. Lett. [78]{}, 4418 (1997). V. L. Gurevich, [Transport in Phonon Systems]{} (North-Holland, Amsterdam, 1986). P. B. Allen and J. Kelner, Am. J. Phys. [66]{}, 497 (1998). P. B. Allen and J. L. Feldman, Phys. Rev. B [ 48]{}, 12 581 (1993); J. L. Feldman, M. D. Kluge, P. B. Allen, and F. Wooten, Phys. Rev. B [48]{}, 12 589 (1993). G. P. DeVault and J. A. McLennan, Phys. Rev. [138]{}, A856 (1965); G. P. DeVault, Phys. Rev. [149]{}, 624 (1966); [155]{}, 875 (1967); G. P. DeVault and R. J. Hardy, Phys. Rev. [155]{}, 869 (1967). N. F. Mott, Philos. Mag. [22]{}, 7 (1970). C. Kittel, Phys. Rev. [75]{}, 972 (1949). J. Fabian, Ph.D. Thesis, SUNY, Stony Brook, NY, 1997. J. Fabian and P. B. Allen, Phys. Rev. Lett. [79]{}, 1885 (1997). F. Wooten, K. Winer, and D. Weaire, Phys. Rev. Lett. [54]{}, 1392 (1985). F. H. Stillinger and T. A. Weber, Phys. Rev. B [31]{}, 5262 (1985). Ref. [@fabian97] used a slightly different version with a small amount of residual stress. The main results, however, remain unchanged. J. L. Feldman, private communication. S. R. Bickham and J. L. Feldman, Phys. Rev. B [57]{}, 12 234 (1998); Philos. Mag. B [77]{}, 513 (1998). J. L. Feldman, J. Q. Broughton, and F. Wooten, Phys. Rev. B [43]{}, 2152 (1991). W. P. Mason, in [Physical Acoustics]{}, edited by W. P. Mason (Academic, New York, 1965), Vol. III. R. Biswas, A. M. Bouchard, W. A. Kamikatahara, G. S. Grest, and C. M. Soukolis, Phys. Rev. Lett. [60]{}, 2280 (1988); H. R. Schober and B. Liard, Phys. Rev. B [44]{}, 6746 (1991); H. R. Schober and C. Oligschleger, Phys. Rev. B [53]{}, 11 469 (1996). G. K. White, Phys. Rev. Lett. [34]{}, 204 (1975). J. L. Feldman, P. B. Allen, S. R. Bickham (unpublished).	{ "pile_set_name": "ArXiv" }
--- abstract: 'An efficient and relatively fast algorithm for the detection of communities in complex networks is introduced. The method exploits spectral properties of the graph Laplacian matrix combined with hierarchical-clustering techniques, and includes a procedure to maximize the “modularity” of the output. Its performance is compared with that of other existing methods, as applied to different well-known instances of complex networks with a community-structure, both computer-generated and from the real-world. Our results are in all the tested cases, at least as good as the best ones obtained with any other methods, and faster in most of the cases than methods providing similar-quality results. This converts the algorithm in a valuable computational tool for detecting and analyzing communities and modular structures in complex networks.' author: - Luca Donetti - 'Miguel A. Mu[ñ]{}oz' title: 'Detecting Network Communities: a new systematic and efficient algorithm' --- Introduction {#sec:intro} ============ The outburst of activity in the field of Complex Networks in recent years has been rather spectacular and amazing. Networks of any thinkable (and sometimes “unthinkable”) type, including social, biological and technological ones have been described, and their topological as well as dynamical features studied. A whole line of research has emerged and a new perspective to tackle complex problems created. See [@Strogatz; @Laszlo; @Porto; @AleRomu; @Newman] for reviews from different perspectives and for exhaustive lists of references. One particular aspect, which has drawn much attention, is the existence of subsets of nodes highly linked among themselves but loosely connected to the rest of the network, [i.e. communities]{}. These are believed to play a central role in the functional properties of complex structures [@newman1; @marta]. Identifying communities and analyzing their nature is an important task in some fields as, for instance, computer science [@Maslov; @jenny], sociology [@newman1; @barsa], biochemistry [@bio], bibliometrics [@biblio], taxonomy, or, as a more specific instance, in the development of efficient search-engines for the WWW. According to Flake [ et al.]{} [@Flake], “as the web is self-organized into communities, search-engines implementing such a concept, would help surfers to find what they look for and avoid other contents”. The concept of “community” may be retained as rather vague and phenomenological. Indeed, depending on the network under scrutiny, it might be quite an artificial one, while, in other cases, it emerges as a very natural and useful structure-analysis tool. A way to make the concept more clear-cut and practical is through the definition of the [modularity]{}, $Q$ (see below and [@newman2; @newmanfast]), a quantity which provides a way to quantify the community-structure of a given network. Other quantities have been proposed with the same purpose [@Wagner; @marta; @roma1]. The problem of finding communities is not new and is closely related to the problem of graph-partitioning, profusely studied in the context of computer science [@Clustering; @Domany]. A review of some used techniques, including further references, can be found in [@newman1; @newmanepjb]. Related problems are image processing and pattern recognition, or more generically data-clustering: in these cases there is no underlying network, but instead some relation or similarity between existing elements can be estabilished [@image; @Pattern1; @Pattern2]. In recent years many algorithms for detecting communities have been proposed, starting with the seminal work by Girvan and Newman [@newman1; @newman2]. These authors proposed an iterative, [divisive]{} (as opposed to [ agglomerative]{}) method based on the progressive removal of links with the largest [betweenness]{}, a quantity proportional to the number of shortest paths passing through a given edge [@Freeman]. The edges (or links) with the largest betweenness have the most prominent role in connecting different parts of the graph and, therefore, by removing them recursively a good separation of the network into its components or communities can be found. This method generates very good results and has been employed by different authors in studies of various kinds [@Wilkinson]. Unluckily, as already pointed out by the authors themselves, it has a main disadvantage: its computational demand is very high. For instance, for sparse networks with $N$ nodes, the computation-time grows like $N^3$. In order to deal with large networks, for which the previous algorithm turns out to be not viable, Newman himself developed a faster method (of the order $N^2$). It is based on the iterative agglomeration of small communities, starting from isolated nodes, by locally optimizing the modularity. This method generates worse results [@worse] than the previous one. Some alternative algorithms both divisive and agglomerative (which we do not attempt to exhaustively overview here) have been proposed in the last months. Some of them are listed here in chronological order (see [@newmanepjb] for a more critical discussion of some of them): - The method by Radicchi-Castellano-Cecconi-Loreto-Parisi [@roma1] is of order $N^2$. It is a divisive algorithm that works nicely whenever triangular (or higher order) loops are present in the network. - Wu-Huberman algorithm [@huberwu]. It is a fast method (linear in $N$), based on the idea of voltage drops, which visualizes the network as an electric circuit. It can be used to locate the community to which one specific node belongs, but it requires successive iterations of the method in order to provide a global network division in communities. - Reichardt-Bornholdt method [@RB]. In this recent paper the authors introduce an algorithm inspired in the celebrated [ super-paramagnetic clustering]{} algorithm devised by Blatt, Wiseman and Domany [@BWD]. It is based on a $q$-state Potts Hamiltonian, and allows, for the first time, for the identification of fuzzy communities. - Capocci-Servedio-Colaiori-Caldarelli method [@roma2]. This algorithm combines the use of spectral properties (which are nicely reviewed and generalized to study different types of networks as, for instance, directed ones) with the use of correlation measurements to determine community closeness. - Fortunato-Latora-Marchiori method [@fortunato]. This is a variation of the method by Girvan and Newman, in which the betweenness is substituted by the alternative concept of [information centrality]{}, as a way to measure edge-centrality. The method generates good results but its performance ($N^4$ for a sparse graph) is rather poor. Apart from these techniques recently introduced in the field of complex networks, many other algorithms have been developed mainly in the context of computer science. Most of them employ spectral analysis, which provides, in a very natural way (using the first non-trivial eigenmode) a tool for bi-partitioning [@specbis] as will be illustrated along this paper. By iterative applications of bi-partitioning more elaborated divisions into communities or components can be achieved [@CS1; @CS2; @jenny]. Alternatively, some other spectral methods employ more than one eigenmode leading directly to a splitting [@Wagner; @Kleinberg; @Jordan]. Without neglecting any of these algorithms, which can be applicable depending on the situation under consideration, this paper introduces yet a new method, allowing for a systematic analysis and detection of communities. It combines the following features: [i]{}) the generation of good results in all the tested cases, [ii]{}) it is relatively fast, as compared with methods providing comparable results, [iii]{}) it includes a way to optimize the output, as will be explained in what follows. The method proposed in this paper combines spectral methods with clustering techniques, and uses the concept of modularity in order to develop a working algorithm. More precisely, the main lines of the algorithm are as follows: spectral analysis of the Laplacian matrix allows us to project the network-nodes into an [eigenvector-space]{} of variable (tunable) dimensionality. Afterwards, a [metric]{} is introduced in various possible fashions, and then, finally, by applying standard clustering techniques a [ dendrogram]{} [@newman1] is built up. The modularity of possible groupings (sections of the dendrogram) is maximized for every considered dimension of the eigenvector-space and finally, the global maximum over all possible number of eigenvectors ([i.e]{} dimensions of the space) is found. In the forthcoming sections we review some basic ideas and definitions of spectral analysis and we introduce our algorithm step by step. Then we apply it to different workbench networks, comparing its performance with that of other existing methods and, finally, the conclusions are presented. Using the Laplacian eigenvectors to detect communities {#sec:method} ====================================================== Spectral analysis: Laplacian eigenvectors ----------------------------------------- The topology of a network with $N$ vertices can be expressed through a symmetric $N \times N$ matrix $\boldsymbol{L}$, the Laplacian matrix [@Biggs]. The diagonal elements $L_{ii}$ are given by the degree $k_i$ of the corresponding vertex $i$, while off-diagonal elements $L_{ij}$ are equal to $-1$ if an edge between the corresponding vertices $i$ and $j$ exists and $0$ otherwise. The sum of elements over every fixed row or column is, trivially, equal to zero. Therefore, a “constant vector” (with all its components taking the same value) is an eigenvector with eigenvalue $0$. Furthermore, since the quadratic form $$\sum_{i,j=1}^n L_{ij} x_i x_j$$ can be written as $$\sum_{\rm{links}} (x_i - x_j)^2,$$ which is positive semidefinite, the eigenvalues of $\boldsymbol{L}$ are either zero or positive [@Mohar]. The use of other matrices, employed to study network spectral properties has been recently considered in [@roma2; @CS2; @Wagner]. If the graph under analysis is connected, there is only one zero eigenvalue corresponding to a constant eigenvector. On the contrary, for non-connected graphs (composed by $m$ connected components) the Laplacian matrix is block diagonal. Each block is the Laplacian of a subgraph and it admits a constant eigenvector with eigenvalue $0$. Therefore, the Laplacian of the whole graph has $m$ degenerate eigenvectors (corresponding to eigenvalue zero), each of them having nonzero constant components for nodes in the associated subgraph and $0$ in the rest. If the subgraphs are not fully disconnected but, instead, a few links exist between them, the degeneration disappears. This leaves only one trivial eigenvector with eigenvalue $0$ and $m-1$ approximate linear combinations of the old ones with slightly non-vanishing eigenvalues [@newmanepjb; @roma2]. As the Laplacian matrix is real-symmetric, with orthogonal eigenvectors, and since the first of them has equal components, all the other ones must have components whose total sum vanishes. In order to illustrate how these ideas can be applied to identify communities, let us take, as a particular example, the number of subgraphs to be $2$. In this case, the components of the second (first nontrivial) eigenvector are positive for one subgraph and have to be negative for the other, providing a clear-cut criterion to bisect the graph [@specbis]. If the two subgraphs are not very well separated, then this distinction between positive and negative values becomes fuzzier. In such cases, more elaborated criteria to decide how to separate into two subgroups have been profusely studied in the specialized literature. Some of them optimize purposely defined quantities as the [normalized cut]{} [@CS2] or the [conductance]{} [@CS1], which are defined as functions of the number of links that exist between the two components and their sizes [@NP]. By iterating successive bisections, techniques to obtain more elaborate splittings can be constructed [@jenny; @CS2; @CS1]. An alternative strategy is to assume that if there are more than two weakly-connected blocks it should be somehow possible to find them all by inspecting the eigenvalue spectrum more accurately, instead of considering just the first non-trivial eigenmode [@Jordan; @Kleinberg]. Let us explore this idea, which is the one we will exploit, in more detail. Figure \[fig:ev\]a shows the components of the first nontrivial eigenvector of a computer-generated graph including 4 communities, each composed by 32 nodes (see forthcoming sections for details). The group structure is clear, even if the two communities at the bottom are very near to each other and some nodes could be mis-classified. In other examples, with a number of inter-group connections larger than here, the communities become more entangled, and the prospective of extracting clear-cut subdivisions using this type of one-dimensional plot worsens. This difficulty can be circumvented by taking into account some more eigenvectors, [i.e.]{} by enlarging the projection-space. This is illustrated in figure \[fig:ev\]b, where the nodes of the same graph are plotted using the components on the first two nontrivial eigenvectors as coordinates. Simple eye-inspection shows that all communities are distinctly separated now. Actually, using three eigenvectors the nodes of the different groups fall around the vertices of a (slightly distorted) tetrahedron, with some further improvement in inter-community separation. Generalizing this idea, [each vertex in the graph is represented by a point in a $D$-dimensional space in which the coordinates are given by its projections on the first $D$ nontrivial eigenvectors]{}. Introducing a metric -------------------- Aimed at turning “eye-inspection” of communities into a more quantitative measure, the explicit introduction of a metric (or similarity measure) is required. The most straightforward choice would be the Euclidean distance. However, this is [not]{} the only possibility; another one is to consider the [angular distance]{}, defined as the angle between the vectors joining the origin of the $D$-dimensional space with the two points under consideration. This possibility is inspired by empirical observations: loosely connected nodes could be quite “Euclideanly” far from each other within a community, but still lying in the same “direction” in the eigenvector-space [@cc]. Moreover, when networks are large, nodes in the same community form a roughly one-dimensional “bundle” (see for example figure 3 in [@Maslov]). Note also, that using angular distances is tantamount to normalizing the position-vectors in the corresponding space and then measuring the Euclidean distance, similarly to what proposed in [@Jordan]. As will be shown, the angular metric generates, as a matter of fact, better results than the Euclidean one. Cluster analysis ---------------- Having introduced a way to measure distances in the eigenvector space, a method to group nodes into communities is required. Such a method is provided by standard clustering techniques [@Clustering] as, for example, [hierarchical clustering]{}. Starting from $N$ clusters, composed by individual nodes, the two closest ones are iteratively joined together. In order to define cluster-to-cluster distance or “closeness” (for a given metric) different criteria can be employed, generating among others, the following clustering algorithms [@Clustering]: - All possible pairs of nodes, taking one from each of the two clusters under examination, are considered. The minimum possible node-to-node distance is declared to be the cluster-to-cluster closeness. This leads to [single-linkage clustering]{}. - Proceeding as before, but replacing the “minimum possible node-to-node distance” between pairs by the “maximum” one, [complete-linkage clustering]{} is defined. - Another possibility consists in taking the average distance between all possible pairs. This leads to [group-average clustering]{}. - A cluster is represented by a single point located at its “center of mass”; the cluster-to-cluster distance is defined as the node-to-node distance between these two points. This leads to [centroid clustering]{}. All these criteria have been broadly studied and applied. None of them can be proved to be generically more efficient than the others. In particular, the single-linkage method, being very simple, can be useful to analyze large data sets, and possesses some further mathematical advantages [@Clustering]. On the contrary, it has a tendency to cluster together, at a relatively low level, distant nodes linked successively by a series of intermediates. This is usually called [chaining property]{}, which constitutes in some cases a serious drawback. On the other hand, a convenient advantage of both, single and complete-linkage clusterings, is that only the ordering of the similarity measure is important: every other metric which produces the same ordering of distances leads to the same results. The output of these algorithms can be represented by a hierarchical tree usually called dendrogram. The starting single-node communities are the branch-tips of such a tree, which are repeatedly joined until the whole network has been reconstructed as a single component (see, for instance, figure 2 in [@newman1]). Each level of the tree represents a possible splitting of the network into a set of communities, obtained by halting the clustering process at the corresponding level. However, the clustering algorithm gives no hint about the “goodness” of such a partition. Modularity ---------- In order to quantify the validity of possible sub-divisions (obtained as explained above) and to optimize the chosen splitting, we use, following [@newman2; @newmanfast], the concept of modularity. It is defined as follows: given a network division, let $e_{ii}$ be the fraction of edges in the network between any two vertices in the subgroup $i$, and $a_i$ the total fraction of edges with one vertex in group $i$ (where edges “internal” to each group have weight $1$ while inter-group links are weighted $1/2$). The modularity, $Q$, is then defined as $$\label{eq:modularity} Q = \sum_i (e_{ii} - a_i)^2.$$ It measures the fraction of edges that fall between communities minus the expected value of same quantity in a random graph with the same community division. The maximization of modularity has been proposed as a possible way to detect communities; since a full maximization is not possible in practice (the algorithm would take an amount of time exponential in the number of nodes to explore all possible splittings) an approximate algorithm has been suggested [@newmanfast]. In our case, modularity measurements are simply used to find the best splitting among all the possible partitions of the dendrogram obtained following the previous steps [@newman2]. Other indeces quantifying the quality of splittings have been also proposed in the literature. Some of them are the “conductance”, the “performance”, and the “coverage” to name but a few (see [@Wagner] and references therein for more details). None of these taken by itself, provides a fully useful criterion; they have to be combined somehow. It seems that the modularity is a better, more efficient, choice. Implementing a functioning algorithm ------------------------------------ Summarizing the ideas introduced in the previous sub-sections, our algorithm can be synthesized and implemented to build up a functioning algorithm as follows. First [a few]{} eigenvalues and eigenvectors of the network Laplacian matrix are computed. The question of what “a few” means will be tackled afterwards. Since the Laplacian is usually a sparse matrix and not all eigenpairs are required (that will require a time $N^3$) the relatively fast Lanczos method [@lanczos] can be employed. Nonetheless, the eigenvector computation is still the most computationally expensive step of the algorithm. For any given number $D$ of eigenvectors ([i.e.]{} for a fixed dimension of the space) a similarity measure (or metric) is chosen, providing a basis to apply one of the previously introduced clustering techniques. Typically, Euclidean or, better, angular distances are employed. Among the various hierarchical clustering methods available, we test single- and complete-linkage clustering algorithms. These two have the advantage that no new distances have to be calculated during cluster formation: when two subgroups merge to form a larger one, its distance to any other cluster is given by the shortest (single-linkage) or by the largest (complete-linkage) of the distances from the two original components. As said before, single-linkage performs poorly in many cases owing to the previously discussed “chaining” property, converting complete-linkage in the preferential choice. Other linkage methods will be explored in the future; in particular, group-average linkage could be suitable when studying tree-like graphs [@private]. An important difference between the way we apply clustering techniques and other standard applications is that we know in advance the underlying network structure. Using this knowledge we implement the constraint that [two clusters are susceptible to be merged only if there exists a link between them in the original network]{}. At every step of the clustering process the modularity is computed. Once the whole dendrogram is completed, the splitting with the maximum modularity is chosen as the output for the corresponding $D$. The optimal value of $D$ to be taken is not known a priori, but since the eigenvalue calculation is the slowest part of the algorithm, we can repeat the hierarchical clustering using all possible values of $D$, and look up for the largest value of the modularity. Typically the largest-modularity vs. $D$ curve exhibits a maximum whose corresponding splitting provides the algorithm final output. If, instead, the curve keeps on growing up to the largest $D$, the number of computed eigenpairs has to be enlarged, in order to extend the range of the curve, until a clear-cut maximum is pin-pointed. Tests of the method {#sec:test} =================== Artificial community networks {#sec:artif} ----------------------------- To prove the algorithm we first test it on computer-generated random graphs with a well-known pre-determined community structure [@newman1]. Each graph has $N=128$ nodes divided into $4$ communities of $32$ nodes each. Edges between two nodes are introduced with different probabilities depending on whether the two nodes belong to the same group or not: every node has ${k_{\rm{in}}}$ links on average to its fellows in the same community, and ${k_{\rm{out}}}$ links to the outer-world, keeping ${k_{\rm{in}}}+{k_{\rm{out}}}=16$. In figures \[fig:rc1\] and \[fig:rc2\] we plot the modularity corresponding to the best splitting identified by the algorithm normalized by the one of the known answer, and the average number of correctly classified vertices, respectively. Data for both, Euclidean and angular measures, and both, single- and complete-linkage algorithms, are shown. The number of eigenvalues leading to the largest modularity is between $3$ and $5$ for the angular distance, and between $2$ and $4$ for the Euclidean one. Let us remark that these are roughly equal to the number of communities and that the performance is much better using the angular distance. Summing up: on these computer-generated networks, our algorithm (equipped with the angular distance and complete-linkage) generates excellent results as compared with other methods (see, for instance, figure 1 in [@newmanfast] and figure 3 in [@fortunato]). Zachary karate club ------------------- Now we consider the well-known karate club friendship network studied by Zachary [@zachary], which has become a commonly used workbench for community-finding algorithms testing [@roma1; @huberwu; @newman1; @newman2; @newmanfast; @RB; @fortunato]. angular Euclidean ------------------ --------- ----------- single-linkage 0.412 0.319 complete-linkage 0.412 0.368 : Modularity of the best splitting of the Zachary club network obtained for different metrics and clustering algorithms.[]{data-label="t:karate"} Table \[t:karate\] shows the maximum modularity found by the algorithm: the best value is again obtained using angular distances combined with either single or complete-linkage clustering. The best splitting is shown in figure \[fig:zac\]; it is different from the “actual” breakdown of the club; [i.e.]{} the two groups reported by Zachary are further subdivided. Let us stress the presence of a single-node community (node $12$), and the fact that the modularity value of this splitting is larger than Zachary’s one ($0.371$), and larger that the ones found using other methods [@newmanfast; @RB; @fortunato]. In this case single and complete-linkage give the same best splitting. Nevertheless, the hierarchical structure given by the dendrogram in the two cases are quite different. Figure \[fig:den\] shows how clusters merge after the best splitting is identified, as well as the modularity value corresponding to each division. For complete-linkage the modularity value remains close to the best one until the whole network is merged in one community. On the other hand, for single-linkage it falls down rather abruptly right after the first merging, owing to the chaining problem. Moreover, in the former case, the two Zachary communities are first reconstructed and then joined together, while in the latter the merging proceeds differently. Therefore, even if the best splitting is the same one in both cases, complete-linkage produces a more reliable dendrogram, describing more accurately the hierarchical structure. Scientific collaboration networks --------------------------------- In order to test the method performance on larger networks we consider two scientific-collaboration networks first analyzed by Newman [@newman-data]. The vertices are the authors of the papers appeared in the `cond-mat` and `hep-th` archives at `ArXiv.org` between $1995$ and $1999$. Two authors are linked if they have co-authored a paper together. The `cond-mat` network contains $16726$ nodes, but we focus on its largest connected component, which contains only $13861$ authors. The computation of the first $1000$ eigenvectors takes about two hours on a personal computer. The modularity curve computation, calculated using up to $999$ eigenvectors, lasts around 15 minutes. Results for angular distance and complete-linkage clustering are plotted in figure \[fig:qmaxcm\]. The largest value of the modularity, $Q=0.736$, achieved for a splitting in $229$ communities, corresponds to a $602$-dimensional space. Obviously, we cannot compare the final splitting with a “true” one, which is not defined. As the curve in figure \[fig:qmaxcm\] is rather flat in its tail, one can legitimately wonder how does the best splitting compare to other ones obtained for similar dimensionalities. This question is difficult to answer in a rigorous way, and will deserve further analysis, which will eventually lead to a functional definition of the [community-structure robustness]{}. Analogously, the `hep-th` network has $8361$ authors with a connected component of $5835$. The largest modularity value, $Q = 0.707$, is produced by a division into $114$ communities, obtained using $416$ eigenvectors. The computation of the first $1000$ eigenvectors takes around $30$ minutes and the search of the largest modularity value about $8$ minutes. In this case, the initial number of eigenvectors could have been taken much smaller than $1000$, without affecting the final output, with the consequent time saving. As in previous cases, the number of eigenvectors used to produce the best splitting is of the order of magnitude of the number of found communities. In these cases, comparison with previous community studies is not feasible, as modularity measurements have not been (to the best of our knowledge) reported in literature. Conclusions {#sec:concl} =========== We have introduced a new algorithm aimed at detecting community structure in complex networks in an efficient and systematic way. The method combines spectral techniques, cluster analysis, and the recently introduced concept of modularity. The nodes of the network are projected into a $D$-dimensional space, where $D$ is a number of first non-trivial eigenvectors of the Laplacian matrix; their coordinates are the node-projections on each eigenvector. Then a metric (either Euclidean or angular) is introduced in such an eigenvector space. Once distances are computed, standard hierarchical clustering techniques (as, for instance, complete-linkage clustering) are employed to generate a dendrogram. The subdivision of this dendrogram giving the maximum modularity is taken as the output of the algorithm for a fixed $D$. Then, also $D$ is allowed to vary (from $1$ to some arbitrary, maximum value) providing a way to maximize the modularity and enhance the performance of the method. The best results are obtained using the angular distance and complete-linkage clustering; however, other types of distances, other clustering algorithms, or even other means to quantify the goodness of a division could be used to improve the results. In this sense our algorithm is a “block-modular” one: modifications of any of its ingredients could possibly lead to an overall improvement. Even if spectral methods have been profusely used before to analyze similar problems, we believe that our algorithm represents a step forward in studying complex-network communities, as it combines spectral techniques with (i) the novel concept of modularity, which provides a very adequate estimate of the quality of a given splitting and (ii) a way to optimize the number of eigenmodes taken into consideration. The weakest part of the method is that the maximum number of eigenvectors to be computed in order to find the one generating the maximum modularity is not known a priori. Being the calculation of eigenvectors the slowest part of the algorithm, what we do is to take a reasonable number of them and, afterwards, verify that the maximum-modularity curve as a function of $D$ decreases at its tail; [i.e.]{} we make sure that a maximum of the modularity function is located. If this is not the case, the number of eigenvectors needs to be enlarged, at the cost of higher computational effort. In the absence of a general criterion to establish the monotonicity of the modularity curve, the only possible way to decide whether the identified local maximum is the global one, would be to compute all possible eigenvalues. In practice, in all the studied cases, the best splitting is found with a relatively small number of eigenvectors, converting the algorithm in a reliable, relatively fast, and very efficient one. An open challenge would be identifying a systematic criterion to estimate, a priori, what is the order of magnitude of the number of eigenvalues to be computed to further optimize the output and efficiency. We hope that this new algorithm will be employed with success in the search and study of communities in complex networks, and will help to uncover new interesting properties. We acknowledge useful comments and discussions with F. Colaiori, A. Capocci, V. Servedio, A. Arenas, G. Caldarelli, and J. Torres. We are specially grateful to M. Newman for providing us with the data on scientific collaborations as well as for a reading of the manuscript, and to C. Castellano, for very helpful comments and suggestions. Financial support from the Spanish MCyT (FEDER) under project BFM2001-2841 and the EU COSIN project IST2001-33555 is acknowledged. [99]{} S. H. Strogatz, Nature [410]{}, 268 (2001). A. L. Barabási, Rev. Mod. Phys. [74]{}, 47 (2002). S. N. Dorogovtsev and J. F. F. Mendes, [Evolution of Networks: From Biological Nets to the Internet and WWW]{}, Oxford University Press, Oxford (2003). R. Pastor Satorras and A. Vespignani, [Evolution and Structure of the Internet: A Statistical Physics approach]{}, Cambridge University Press (2004). M. E. J. Newman, SIAM Review [45]{}, 167 (2003). M. Girvan, M. E. J. Newman, Proc. Natl. Acad. Sci. USA [ 99]{}, 7821-7826 (2002). R. Guimerá, M. Sales-Pardo, and L. A. N. Amaral, cond-mat/0403660. K. A. Eriksen, I. Simonsen, S. Maslov, and K. Sneppen, Phys. Rev. Lett. [90]{}, 148701 (2003); and cond-mat/0312476. C. Borgs, J. T. Chayes, M. Mahdian and A. Saberi, Proceedings of the 10th ACM SIGKDD International Conference on Knowledge, Discovery and Data Mining (2004). R. Guimerá, L. Danon, A. Diaz-Guilera, F. Giralt, and A. Arenas, Phys. Rev. E [68]{}, R065103 (2003); A. Arenas, L. Danon, A. Diaz-Guilera, P. M. Gleiser, and R. Guimerá, cond-mat/0312040. L. H. Hartwell, J. J. Hopfield, S. Leibler, and A. W. Murray, Nature [402]{}, C47 (1999). E. Ravasz, A. L. Somera, D. A. Mongru, Z. N. Olvai, and A. L. Barabási, Science [297]{}, 1551 (2002). L. Egghe and R. Rousseau, [Introduction to Informetrics]{}, (1990). G. W. Flake, S. Lawrence, C. L. Giles, and F. Coetzee, IEEE Computer [35]{}, 66 (2002). M. E. J. Newman, M. Girvan, Phys. Rev. E [69]{}, 026113 (2004). M. E. J. Newman, Phys. Rev. E [69]{}, 066133 (2004) U. Brandes, M. Gaertler, and D. Wagner. Proc. 11th Europ. Symp. Algorithms (ESA’03), LNCS 2832, pp. 568-579. F. Radicchi, C. Castellano, F. Cecconi, V. Loreto, and D. Parisi, Proc. Natl. Acad. Sci. USA [101]{}, 2658-2663 (2004). A. K. Jain and R. C. Dubes, [Algorithms for clustering data]{}, Prentice Hall, Englewood Cliffs, NJ (1988). B. S. Everitt, [Cluster Analysis]{}, Edward Arnold, London (1993). E. Domany, Physica A [263]{}, 158 (1999). M. E. J. Newman, Eur. Phys. J. B [38]{}, 321-330 (2004). Y. Weiss, Proceedings IEEE International Conference on Computer Vision, 975-982 (1999). R. O. Duda and P. E. Hart, [Pattern Classification and Scene Analysis]{}, Wiley. New York (1973). K. Fukunaga, [Introduction to statistical pattern recognition]{}, Academic Press, San Diego (1990). L. Freeman, Sociometry [40]{}, 35 (1977). J. R. Tyler, D. M. Wilkinson, and B. A. Huberman, in [ Proceedings of the First International Conference on Communities and Technologies]{}, Ed. M. Huysman, E. Wenger, and V. Wulf, Kluwer, Dordrecht (2003). D. Wilkinson and B. A. Huberman, Proc. Natl. Acad. Sci. USA [ 101]{}, 5241-5248 (2004). The goodness of a given division (or division method) can be decided in [absolute]{} terms (when the underlying community-structure is known, as for example, in computer-generated networks) or in [relative]{} terms (when the community structure is not known, but it maybe quantified in terms of modularity or similar measurements [@newman2; @newmanfast; @roma1]; large modularity-values corresponding to better divisions). F. Wu, B. A. Huberman, Eur. Phys. J. B [38]{}, 331-338 (2004). J. Reichardt, S. Bornholdt, cond-mat/0402349. M. Blatt, S. Wiseman, and E. Domany, Phys. Rev. Lett. [76]{}, 3251 (1996); Neural Computation [9]{}, 1805 (1997). A. Capocci, V. Servedio, F. Colaiori, and G. Caldarelli, cond-mat/0402499. S. Fortunato, V. Latora, M. Marchiori, cond-mat/0402522. M. Fiedler, Czech. Math. J. [23]{}, 298-305 (1973) A. Pothen, H. Simon, K.-P. Liou, SIAM J. Matrix Anal. Appl. [11]{}, 430-452, (1990) R. Kannan, S. Vempala, and A. Vetta, Journal of the ACM [ 51]{}, 497-515 (2004) X. He, C. H. Q. Ding, H. Zha, and H. D. Simon, [Proc. IEEE Int’l Conf. Data Mining]{}, 195 (2001). C. H. Q. Ding, X. He and H. Zha. [Proccedings 7th International Conf. on Knowledge Discovery and Data Mining (KDD 2001)]{}, 275 (2001). J. M. Kleinberg. Journal of the ACM [46]{}, 604 (1999). D. Gibson, J. M. Kleinberg, and P. Raghavan. [Proceedings of the 9th ACM Conference on Hypertext and Hypermedia]{}, 225 (1998). A. Y. Ng, M. I. Jordan, and Y. Weiss, Advances in Neural Information Processing Systems [14]{}, 849 (2002). N. L. Biggs, Algebraic Graph Theory, Cambridge University Press (1974). B. Mohar, The Laplacian spectrum of graphs, in: Y. Alavi, G. Chartrand, O. R. Ollermann, A.J. Schwenk (Eds.),[Graph Theory, Combinatorics, and Applications]{}, Wiley, New York, 1991, pp. 871-898. In principle the minimization of the conductance or the normalized-cut among all possible splits is a NP-hard problem. However, it can be shown that the cuts based on the components of the second eigenvector of the Laplacian or some related matrix give a guaranteed approximation to the optimal cut [@Jordan; @Chung]. F. Chung. [Spectral Graph Theory]{}, Number 92 in CBMS Region Conference Seriesin Mathematics. American Mathematical Society, 1997. The attentive reader could argue that figure 1b provides a counterexample to this general assertion; [i.e.]{} the two uppermost groups are nearby angularly but far apart Euclideanly. Indeed, taking the three-dimensional version of the net analyzed in such a figure, the four communities lay within the main directions on a tetrahedron, circumventing this apparent contradiction. G. H. Golub, C. F. Van Loan, [Matrix Computations]{}, Johns Hopkins University Press, Baltimore (1996). M. Newman, private communication. W. W. Zachary, J. of Anthropological Research [33]{}, 452 (1977). M. E. J. Newman, The structure of scientific collaboration networks, Proc. Natl. Acad. Sci. USA [98]{}, 404-409 (2001). See also, M. E. J. Newman, Phys. Rev. E [64]{}, 016132 (2001).	{ "pile_set_name": "ArXiv" }
--- abstract: \| A Mobile Adhoc Network (<span style="font-variant:small-caps;">manet</span>) is a cooperative engagement of a collection of mobile nodes without any centralized access point. The underlying concept of coordination among nodes in a cooperative <span style="font-variant:small-caps;">manet</span> has induced in them a vulnerability to attacks due to issues like dynamically changing network topology, cooperative algorithms and lack of centralized monitoring point. We propose a semi-distributed approach towards a reputation-based Intrusion Detection System (<span style="font-variant:small-caps;">ids</span>) that combines with the Dynamic Source Routing (<span style="font-variant:small-caps;">dsr</span>) protocol for strengthening the defense of a <span style="font-variant:small-caps;">manet</span>. Our system inherits the features of reputation from human behavior, hence making the <span style="font-variant:small-caps;">ids</span> socially inspired. It has a semi-distributed architecture as the critical observations of the system are neither spread globally nor restricted locally. The system assigns maximum priority to self observation by nodes for updating any reputation parameters, thus avoiding the need of a trust relationship between nodes. Our system is also unique in the sense that it features the concepts of Redemption and Fading with a robust Path Manager and Monitor system. Simulation studies show that <span style="font-variant:small-caps;">dsr</span> fortified with our system outperforms normal <span style="font-variant:small-caps;">dsr</span> in terms of the packet delivery ratio and routing overhead even when up to half of nodes in the network behave as malicious. Various parameters introduced such as timing window size, reputation update values, congestion parameter and other thresholds have been optimized over several simulation runs. By combining the semi-distributed architecture and other design essentials like path manager, monitor module, redemption and fading concepts, our system proves to be robust enough to counter most common attacks in <span style="font-variant:small-caps;">manet</span>s. *Keywords:* Adhoc networking, Security, promiscuous mode, Reputation based Intrusion Detection System author: - \| Animesh Kr Trivedi$^1$, Rajan Arora$^1$, Rishi Kapoor$^1$, Sudip Sanyal$^1$ and Sugata Sanyal$^2$\ $^1$Indian Institute of Information Technology, Deoghat, Jhalwa, Allahabad (U.P.), India\ {aktrivedi\_b03,rarora\_b03,rkapoor\_b03,ssanyal}@iiita.ac.in\ $^2$School of Technology and Computer Science, Tata Institute of Fundamental Research, India\ [email protected] date: 'Nov 13, 2006' title: 'A Semi-distributed Reputation-based Intrusion Detection System for Mobile Adhoc Networks' --- Introduction ============ The term adhoc networks dates back to the 1970’s where an adhoc network was first setup as a part of certain defense research projects. With advances in microelectronics technology and networking protocols, it has been possible to integrate mobile nodes and various other network devices into a single unit called an adhoc node. Further, interconnection of these nodes wirelessly is termed as an adhoc network.\ <span style="font-variant:small-caps;">manet</span>s are different from conventional networks. A <span style="font-variant:small-caps;">manet</span> is formed by an autonomous system of mobile nodes that are self-configuring and have no constraints, such as a fixed infrastructure or a central administration system. Nodes in <span style="font-variant:small-caps;">manet</span>s are both routers and terminals. They are dynamic in the sense that each node is free to join and leave the network in a nondeterministic way. In addition, they do not have a clearly defined physical boundary, and therefore, no specific entry or exit point. Such a network can thus be rapidly deployed and can provide the amount of flexibility and adaptability which is otherwise unattainable under adverse circumstances. Although <span style="font-variant:small-caps;">manet</span> is a very promising technology, challenges are slowing its development and deployment. Nodes in adhoc networks are in general limited in battery power, memory and <span style="font-variant:small-caps;">cpu</span> capacity. Hence the transmission ranges of these devices are also limited and nodes have to rely on neighbor nodes in the network to route the packet to its destination. They are sometimes referred to as multihop networks, where a hop is a direct link between two nodes. Adhoc networks have found applications in emergency rescues, battlefield operations, mobile conferencing, national crisis, home and community networking, disaster recovery etc.\ The flexible structure and volatile environment of <span style="font-variant:small-caps;">manet</span>s results in significant node misbehavior. Not only does it degrade the overall network performance, but, it also becomes difficult to detect intruders on grounds of mobility and vulnerability of the nodes. Thus, there is a serious need for a robust <span style="font-variant:small-caps;">ids</span> for <span style="font-variant:small-caps;">manet</span>s.\ Some fundamental problems of <span style="font-variant:small-caps;">manet</span>s must be kept in mind while designing any security solution. Firstly, it is often very hard to differentiate intrusions and normal operations or conditions in <span style="font-variant:small-caps;">manet</span>s because of the dynamically changing topology and volatile physical environment. Secondly, mobile nodes are autonomous units that are capable of roaming independently in an unrestricted geographical topology. This means that nodes with inadequate physical protection can be captured, compromised or hijacked. Thirdly, decision-making in adhoc networks is usually decentralized and many adhoc network algorithms rely on the cooperative participation of all nodes. Most adhoc routing protocols are also cooperative in nature and hence can be easily misguided by false routing information. Without any counter policy, the effects of misbehavior have been shown to dramatically decrease network performance. In this paper, we propose a new technique based on reputation for efficiently solving the problem of intrusion detection.\ The next section gives a brief background about routing related issues in <span style="font-variant:small-caps;">manet</span>s, section III entails a discussion of some related efforts which is followed by the system design overview in section IV. Section V describes the protocol and the following section VI talks about its implementation details. Simulation results and optimization procedures for parameters such as window size are given in the section VII. The last section presents some concluding remarks. BackGround ========== In order to understand the nature of attacks on <span style="font-variant:small-caps;">manet</span>s, we first need to look at the routing protocols for these networks. They have been classified under two main categories - Proactive and Reactive routing protocols. Proactive protocols work with tables that are used to store routing information and updates are triggered to propagate any information about changes throughout the whole network. The obvious advantage is that routes to any destination node are always available without the overhead of generating a route request whenever the need for a route arises. But, an extra overhead is always a major issue before deploying a proactive routing protocol, because it generally affects the overall throughput and power usage. Destination-Sequenced Distance Vector (<span style="font-variant:small-caps;">dsdv</span>) [@rDSDV], Wireless Routing Protocol (<span style="font-variant:small-caps;">wrp</span>) [@rWRP], Cluster Gateway Switch Routing (<span style="font-variant:small-caps;">cgsr</span>) [@rCGSR] are some common examples.\ On the other hand, Reactive routing protocols are on-demand i.e. a route discovery mechanism is initiated whenever there is a need for setting up a path for communication between a source and a destination node. The source node initiates route discovery by flooding the network successively with route queries. The destination node on receiving a route request (<span style="font-variant:small-caps;">rreq</span>) addressed to it, sends back a route reply (<span style="font-variant:small-caps;">rrep</span>) message as unicast to the source node either through the discovered route or by initiating another route request. Generally, on-demand routing requires less overhead than table-driven routing; but it incurs a path discovery delay whenever a new path is needed. Dynamic Source Routing protocol (<span style="font-variant:small-caps;">dsr</span>)[@rDSR], Adhoc On-Demand Distance Vector Routing (<span style="font-variant:small-caps;">aodv</span>) protocol [@rAODV], Temporally Ordered Routing Algorithm (<span style="font-variant:small-caps;">tora</span>) [@rTORA], Associativity-Based Routing <span style="font-variant:small-caps;">(abr)</span> [@rABR], Signal Stability Routing (<span style="font-variant:small-caps;">ssr</span>) [@rSSR], Zone-Based Hierarchical Link State Routing Protocol (<span style="font-variant:small-caps;">zrp</span>) [@rZRP] are a few examples.\ Attacks are possible on reactive protocols like <span style="font-variant:small-caps;">dsr</span> due to lack of built-in security measures and the assumption of honest coordination and cooperation among nodes and with the protocol. We will outline a few attacks by nodes below, the others are discussed in detail by Sonja Buchegger et. al. [@WDOG]: - Dropping all packets not destined to it or performing only partial dropping. Partial dropping can be restricted to specific types, such as only data packets or route control packets or packets destined to specific nodes. - Sending forged routing packets, an attacker can create a so-called black hole, a node where all packets are discarded or all packets are lost. - Modifying the nodes list in the header of a <span style="font-variant:small-caps;">rreq</span> or a <span style="font-variant:small-caps;">rrep</span> to misroute packets and adding incorrect routes in the route cache of other nodes. - Decreasing the hop count <span style="font-variant:small-caps;">dsr(ttl)</span> when receiving a packet, so that the packet will never be received by the destination. This attack could be detected by the previous node in route by enhanced passive acknowledgment. - Initiating frequent <span style="font-variant:small-caps;">rreq</span> to consume bandwidth and energy and to cause congestion. Related Work ============ Reputation-based systems are a new paradigm and are being used for enhancing security in different areas. These systems are lightweight, easy to use and are capable of facing a wide variety of attacks. Among these mechanisms, <span style="font-variant:small-caps;">core</span> [@rCORE], <span style="font-variant:small-caps;">confidant</span> [@rCONF] and <span style="font-variant:small-caps;">ocean</span> [@rOCEAN] gain a special mention.\ Reputation based systems do not rely on the conventional use of a common secret to establish confidential and secure communication between two parties. Instead, they are simply based on each other’s observations. Reputation based systems are used for enhancing security in adhoc networks as they model cooperation between the nodes which is inspired from social behavior. Such systems are used to decide whom to trust and to encourage trustworthy behavior. Resnick and Zeckhauser [@rTRUST] identify three goals for reputation systems: - To provide information to distinguish between a trustworthy principal and an untrustworthy principal. - To encourage principals to act in a trustworthy manner - To discourage untrustworthy principals from participating in the service the reputation mechanism is present to protect. Watchdog and Path-rater [@WDOG] are some essential components of any typical reputation based <span style="font-variant:small-caps;">ids</span>. Watchdog performs the activity of monitoring its neighborhood and based on these observations, pathrater ranks the available path in route cache. Misbehavior detection and reputation-based intrusion detection may be either distributed or local. Here, fully distributed means that information regarding one’s reputation change is immediately propagated in the whole network. In the latter case, called local reputation based systems, nodes are fully dependent on their personal opinion about other nodes’ reputation and behavior.\ Distributed <span style="font-variant:small-caps;">ids</span> protocols rely only on first-hand information with optional second-hand information. <span style="font-variant:small-caps;">core</span> [@rCORE] proposed by P. Michiardi and <span style="font-variant:small-caps;">confidant</span> [@rCONF] proposed by Buchegger and Le Boudec fall into this category. Some basic problems with this approach of global reputation systems are: - Every node has to maintain O(n) reputation information where n is number of nodes in network. - Extra traffic generation in reputation exchange. - Extra computation in accepting indirect reputation information (secondhand information) esp. Bayesian Estimation. - Security issues in reputation exchange such as reputation data packets can be modified. <span style="font-variant:small-caps;">confidant</span> detects misbehaving nodes by means of observation or by <span style="font-variant:small-caps;">Alarm</span> signals from neighborhood. It aggressively informs nodes in neighborhood about misbehavior of the malicious node. The weightage of <span style="font-variant:small-caps;">Alarm</span> warning signal depends upon the level of trust of receiving node about the sending node. In addition, it uses bayesian estimation for various measures and calculation of trust and reputation and thus, the <span style="font-variant:small-caps;">ids</span> becomes complex. <span style="font-variant:small-caps;">confidant</span> is vulnerable to false accusations if trusted nodes lie or if several liars collude [@FACC].\ <span style="font-variant:small-caps;">core</span> [@rCORE] proposed by P. Michiardi et. al. uses a mechanism to enforce node cooperation in <span style="font-variant:small-caps;">manet</span>s. In this mechanism, reputation is a measure of someone’s contribution to network operations. Members that have a good reputation can use available resources while members with a bad reputation cannot, because they refused to cooperate earlier and are gradually excluded from the community. <span style="font-variant:small-caps;">core</span> defines three types of reputation: - Subjective reputation is a reputation value which is locally calculated based on direct observation. - Indirect reputation is second hand reputation information which is established by other nodes. - Functional reputation is related to a certain function, where each function is given a weight as to its importance. For example, data packet forwarding may be deemed to be more important than forwarding packets with route information, so data packet forwarding will be given greater weight in the reputation calculations. <span style="font-variant:small-caps;">Core</span> reputation values range from positive (+1), through null (0), to negative (-1). <span style="font-variant:small-caps;">core</span> suffers from the problem of unwanted consequence of good reputation, where a good node may even wish to decrease its reputation by behaving badly to prevent its resources being over-used. The <span style="font-variant:small-caps;">core</span> mechanism assumes that every node will use the same reputation calculations and will also assign the same weights to the same functions. This is a potentially inappropriate assumption in heterogeneous adhoc networks, where devices with different capabilities and roles are likely to place different levels of importance on different functions depending upon <span style="font-variant:small-caps;">cpu</span> usage, battery usage etc. One can take advantage of this situation and may perform only those functions which have higher preferences in calculating reputation.\ The second type of <span style="font-variant:small-caps;">ids</span> may be categorized as local systems. They solely depend upon the first hand observation of their neighbors for reputation maintenance. <span style="font-variant:small-caps;">ocean</span> [@rOCEAN] by Bansal and Baker falls into this category. In these systems, nodes make routing decisions based only on direct observations of their neighbor nodes. This eliminates most of the trust manager complexity, but, doesn’t fit well to a highly mobile adhoc network. In such a network, it may be difficult for the reputation upgrading process to cope up with the node mobility and it might not be appropriate to depend solely upon personal observation. Also, using secondhand information can significantly accelerate the detection and subsequent isolation of malicious nodes in <span style="font-variant:small-caps;">manet</span>s [@SHAND]. System Overview =============== As stated earlier [@rRISM], the system design is based on the reputation paradigm and possesses a semi-distributed nature in terms of the reputation exchange mechanism. The term semi-distributed is used in the system observation context, which is neither restricted to the observing node nor immediately propagated to the whole network as is the case in true distributed systems. The system design has been kept simple keeping in mind the amount of traffic already in the network and constraints such as the critical amount of battery and computational power that individual nodes possess. The system runs on every node in the network and consists of the following modules: Monitor ------- In wireless networks, acknowledgements are often provided at no cost, either as an existing standard part of the <span style="font-variant:small-caps;">mac</span> protocol in use (such as the link-layer acknowledgement frame defined by <span style="font-variant:small-caps;">Ieee</span> 802.11) or by a “passive acknowledgement" (in which, a node confirms receipt at another node by overhearing the transmission from sender). The Monitor holds the responsibility of monitoring activities in the neighborhood using <span style="font-variant:small-caps;">pack</span>s (Passive <span style="font-variant:small-caps;">Ack</span>nowledgements) which have been provided as a feature in the <span style="font-variant:small-caps;">dsr</span> protocol specifications [@rDSR] as promiscuous mode. Every node registers all the data packets sent by it to its next hop neighbor and on overhearing packets in promiscuous mode, it matches those against packets registered in the queue. These packets are considered as <span style="font-variant:small-caps;">pack</span>s only if both of the following two tests succeed: - The source address, destination address, protocol, identification and fragment offset fields in the IP header of the two packets must match, and - If either packet contains a <span style="font-variant:small-caps;">dsr</span> source route header, both packets must contain one and the value in the ‘segments left’ field in the <span style="font-variant:small-caps;">dsr</span> source route header of the new packet must be less than that in the first packet. A crucial new parameter introduced in our system is the timing window that is a fixed time interval. After each timing window, nodes make a log of number of packets for which they have not received acknowledgment in the form of <span style="font-variant:small-caps;">pack</span> and communicate this information to the reputation system. In existing reputation systems, every packet is kept waiting for its <span style="font-variant:small-caps;">pack</span> for a fixed time interval. In contrast, we use the concept of timing window, which gives us the flexibility of checking timeout on fixed intervals rather than checking it on the basis of each individual packet’s timeout. Monitor maintains a log of activity of next neighbor for each window and sends it to the reputation manager. Depending upon its cooperation, performance and current environment conditions, reputation system updates the nodes’ reputation. With the help of Timing Window, the system also takes into consideration congestion state of nodes, which shall be explained in next subsection. Reputation System ----------------- Reputation system module assigns and maintains reputation of different nodes as a numeric value with a lower limit of 0 and upper limit equal to the value of $MaliciousThreshold$. Reputation of any node can change by three means, as shown in Figure \[fig:SysBehav\]: ![\[fig:SysBehav\]The System Behavior](4.pdf) - By Self observation - <span style="font-variant:small-caps;">Warning</span> Message, issued by neighboring nodes - Avoid List, appended to the <span style="font-variant:small-caps;">rreq/rrep</span> header All three means of reputation change have some associated reputation weights with them, giving maximum weightage to self observation. The reputation is updated after every timing window and the information is communicated in a sporadic way by means of avoid lists thereby avoiding much of network overhead. The concept of avoid list is inherited from <span style="font-variant:small-caps;">ocean</span> [@rOCEAN]. It facilitates easy communication among nodes by putting their malicious node list in the <span style="font-variant:small-caps;">rreq</span> header. This helps in reducing the extra network traffic which would otherwise be generated while communicating this information among the peers. A node may be tagged as normal, suspicious or malicious depending on the reputation value associated with it. Every new node starts with a normal reputation value of zero and this reputation value may be lowered by degrading its performance or it may be incremented through the positive appraisal feature on normal behavior. To add to the robustness and performance of our system, it is ensured that absolute value of the negative decrement awarded is larger than the positive appraisal. However at no point should the reputation of node go above zero to prevent the kind of attacks, where a node first gains positive reputation but later on depicts a malicious behavior, thereby bringing its reputation value back to the normal range. It also avoids the peculiar situation where a node may end up exhausting all its crucial resources in routing extra traffic faced due to the popularity gained by earning positive reputation. After each timing window, reputation system receives activity log of next hop neighbor from monitor with number of packets for which it has not received <span style="font-variant:small-caps;">pack</span>, which are classified as missing or dropped packets. The number of missing packets is then compared with the MaliciousDropThreshold and if it is comparatively lesser, then the reputation manager gives positive performance appraisal otherwise a negative one. Unlike existing systems our system does not have a rigid MaliciousDropThreshold, we introduce the concept of congestion parameter, which is given as: $$\footnotesize{Congestion\ Parameter}\ =\ \frac{Current\ queue\ length}{Total\ queue\ length}$$ With the assumption that the next node is also in same congestion state as the node in contention. Misbehavior drop threshold, that is the allowed number of packet drops in a timing window is dynamically decided as: $$\footnotesize{MaliciousDropThreshold} = \footnotesize{MaxPacketRate}\ \times \ {CongestionParameter} \times \ {WindowSize} \$$ Whenever a new node is categorized as malicious, a warning message is spread only to its immediate neighborhood, thus protecting the network flooding with reputation update messages. This can be understood from the Figure \[fig:Scenario\]: ![\[fig:Scenario\]A typical network scenario](5.pdf) If Node B categorizes A as malicious, a Warning Message is spread to all immediate neighbor nodes: C, D and E (not to F). Nodes C, D and E on receiving a warning message decrease the reputation of node A, against which the message was originally published. Lastly, another mode of reputation updating is by means of an avoid list [@SHAND]. During the route discovery phase the <span style="font-variant:small-caps;">rreq</span> sending node puts its malicious node list in <span style="font-variant:small-caps;">rreq</span> packet header and initiates discovery. When a node receives a route request packet it decreases the reputation of all those nodes quoted in the avoid list by a predefined weight. The node also appends its own malicious node list in the header and then forwards the route request packet.\ In order to deal with the attacks on a typical reputation system, like those of ‘Collusion of liars’ and ‘false warning messages’, the system has a policy that nodes can be categorized malicious only by self-observation. It helps in nullifying attacks of the aforementioned types as the false warning messages spread by nodes can only decrease reputation of the victim nodes to a certain extent, termed as suspicious threshold. Once this threshold is reached, the system solely depends upon self observation for making the final decision. Warning messages and avoid list are only effective above the suspicious threshold. Whenever a node’s reputation is in the suspicious category and a deciding node receives a new warning message or an avoid list appearance for the previous one, the system performs a knock test. Knock test is a unique test designed for checking the authenticity of a node against whom the deciding node constantly receives such information. For instance, if node A falls into the suspicious category and node B receives another warning message or an avoid list appearance corresponding to node A, the deciding node B performs a knock test over A, explained later in the protocol description section. Path Manager ------------ The path manager performs trivial path management functions in collaboration with <span style="font-variant:small-caps;">dsr</span> core. Path ranking is done according to path priority formula (3). Updating path-cache on various events such as those when new nodes are declared as malicious or a malicious node is taken back in network; taking decision on receiving route request or traffic from malicious nodes are a few responsibilities of the path manager. Concept of avoid list has been added to path manager, which is a list of malicious nodes that a certain node possesses and is appended to the <span style="font-variant:small-caps;">rreq</span> header whenever a route request is issued by some node. Nodes which find themselves in avoid list do not process the packet and may simply drop it. During <span style="font-variant:small-caps;">rrep</span>, only a path with clean nodes is preferred over those containing suspicious nodes and malicious nodes. Replies from such nodes are also dropped and nodes do not process request and/or forward data packets from such nodes. If during traffic flow, a new node is declared malicious, then, all paths containing it are deleted from route cache and a route error is generated, stating that their link to the destination node has been broken. Neighbors, after receiving a route error, clear the activity log of the node which generates a route error from the current timing window. The following function may be used to decide the path priority if need arises: $$\footnotesize{Path\ Priority}\ \propto\ \tfrac{1}{\|Min\ reputation\ of\ Node\ in\ path\| \times no\ of\ hops}$$ Redemption And Fading --------------------- Redemption and Fading are introduced in our design to allow nodes previously considered malicious to become a part of the network again. <span style="font-variant:small-caps;">manet</span>s run on cooperation and collaboration of peer nodes and no one gets benefited without cooperating with each other. Knock test is crucial for nodes in suspicious category and node may fail this test due to various reasons like transient link failures, congestion or resetting of the network interface etc. and once they fail this test, they are declared as malicious. To account for these problems, our system uses the fading mechanism. After a certain inactivity period the reputation of a node is improved by a certain predefined fading rate and finally the node is moved from the malicious list to middle of suspicious category. But, the node is not given neutral rating [@NTRAL] so that if the node again misbehaves then it is immediately put in malicious list and all transactions through that node are blocked. Here, inactivity period means no appearance in any <span style="font-variant:small-caps;">warning</span> messages or avoid list. Protocol Description ==================== This section entails a discussion of the actual working of the system and provides the flow for various activities at different types of nodes. Following algorithms give a concise idea of the route discovery phase, monitoring mode and knock test feature of our system as discussed in earlier sections. $$SENDER$$ \[1\] $\Rightarrow$ Generate <span style="font-variant:small-caps;">rreq</span> Packet\ $\Rightarrow$ Pack Malicious List in <span style="font-variant:small-caps;">rreq</span> Header as Avoid List\ $\Rightarrow$ Propagate Request $$OTHER\ NODES$$\ $\Rightarrow$ Drop Request\ $\Rightarrow$ Scan Avoid List\ $\Rightarrow$ Update Node’s Reputation\ $\Rightarrow$ Append its own malicious list to RREQ header avoiding repetition\ $\Rightarrow$ Prepare Reply\ $\Rightarrow$ Add itself in route and propagate\ The above algorithm presents a node’s behavior during route establishment phase. Sender of the <span style="font-variant:small-caps;">rreq</span> just initiates the route discovery process with avoid list of malicious nodes packed in the <span style="font-variant:small-caps;">rreq</span> packet header. The remaining nodes after receiving such requests process the avoid list attached in the received <span style="font-variant:small-caps;">rreq</span> header. If a matching entry is found for their own name in the list, the node drops the request. Otherwise, the reputation of the other nodes present in the avoid list is updated. If the receiving node is the destination for which the <span style="font-variant:small-caps;">rreq</span> has been sent, then it prepares a route reply else it appends its own malicious list in the header to the existing avoid list avoiding repetitions and propagates the route request. $$MONITOR\ MODE$$ Self Observation- \[1\] $\Rightarrow$ Negative reputation update\ $\Rightarrow$ Positive reputation update\ $\Rightarrow$ SET reputation = 0 $$WARNING\ MESSAGE\ PROPAGATION$$ $\Rightarrow$ Perform Knock Test\ $\Rightarrow$ Assign normal reputation\ $\Rightarrow$ declare as Malicious\ $\Rightarrow$ spread Warning Message\ $\Rightarrow$ decrease reputation\ The system in monitoring mode has three ways of gathering information for reputation updation: - Self Observation - Warning Message - Avoid List Some observations just monitor the neighbor with the help of <span style="font-variant:small-caps;">pack</span>. If the performance lies below the suspicious threshold, then a negative reputation update is performed over the node in consideration, otherwise a positive appraisal is given. Warning messages are only processed if they are for immediate neighbors. If the reputation of a node under consideration is below the suspicious threshold, then the knock test is performed. Otherwise, the reputation is decreased linearly. Table 2 contains actual values of these constant parameters used during system simulation. $$KNOCK\ TEST$$ \[1\] $\Rightarrow$ Identify target Node\ $\Rightarrow$ Generate fake data packet with route via target node\ $\Rightarrow$ Send packet to target node and wait for its <span style="font-variant:small-caps;">pack</span>\ $\Rightarrow$ test Passed\ $\Rightarrow$ Set reputation to default\ $\Rightarrow$ test Failed\ $\Rightarrow$ Declare node as malicious and broadcast Warning message\ Knock test is designed specifically for immediate neighbors to test whether a particular node is malicious or not and is only performed on nodes in suspicious state. In this test a dummy data packet with time to live (<span style="font-variant:small-caps;">TTL</span>) equal to 2 is sent to a node in question via last known route through that node. The sender node overhears traffic of the node in question in promiscuous mode. If the node on which knock test is being performed successfully forwards the test packet to next hop then its reputation is set to default. In case it fails, then it is immediately put into malicious category and a warning message is broadcasted in the immediate neighborhood. If in case, the dummy packet is genuinely dropped because of bad channel conditions the node may be classified as malicious. However, it still has an opportunity to become a part of network again through redemption and fading mechanism, as explained earlier. This is done because the system only trusts first hand information for putting a node into malicious category, thus, giving self observation the highest weightage. The weightage assigned to warning message and avoid list citation is comparatively less than self observation. Implementation/Simulation ========================= This section first describes the simulation environment and then we compare the throughput of our system in the presence of malicious nodes against a defenseless <span style="font-variant:small-caps;">dsr</span> protocol. The network simulator ns2 (version 2.29) [@rNS] was used to run the simulations. Mobility of nodes is characterized by a mobility model, speed and ‘pause time’. The random waypoint model is selected as a mobility model in a 1000 $\times$ 1000 m$^2$ rectangular field. Using this mobility model, each movement is a straight line between a start and an arrival point, covered at a constant speed which is a uniform distribution, between 0 and 10 m.s$^{-1}$ for each movement. The pause time is the time period between two consecutive movements. Thus, the higher the pause time, lesser is the node’s mobility. We have used 5 different pause times: 0, 100, 300, 600 and 900 seconds.\ There are two setups having a total of 10 and 20 nodes, with number of malicious nodes between 10 to 100%. We use maximum 5 and 10 <span style="font-variant:small-caps;">cbr</span> (Constant Bit Rate) connections for 10 and 20 nodes respectively, sending 64 bytes packets with a 4 pkts.s$^{-1}$ sending rate. The bandwidth is 2 Mb.s$^{-1}$. The Medium Access Control (<span style="font-variant:small-caps;">Mac</span>) protocol used is <span style="font-variant:small-caps;">Ieee</span> 802.11.\ The malicious nodes are of the following nature: dropping an average of 99% of the <span style="font-variant:small-caps;">cbr</span>-connection packets (data packets). The dropping decision is taken depending upon a number generated at random. We assume that malicious nodes do not drop the <span style="font-variant:small-caps;">dsr</span> routing packets like route request, route reply or error as they always want to be part of network. A malicious node dropping all the packets is comparatively less dangerous for the <span style="font-variant:small-caps;">manet</span> because in that case, it would drop all packets including routing packets. Following which, they would never be able to include themselves in any the communication routes. The fixed parameters for the simulation are listed in Table 1. Table 1: Fixed Parameters\ Parameter Level ------------------ ------------------------------- Area 1000 m $\times$ 1000 m Speed uniformly distributed between 0 and 10 m.s$^{-1}$ Radio Range 250 m Placement uniform Movement random waypoint model MAC 802.11 Sending capacity 2 Mbps Application CBR Packet size 64 B Simulation time 900 s Thus, if no route has been established containing these nodes, they would never be able to drop any data packet sent to them either. As a consequence, they would not affect the throughput of whole network. For evaluating the performance of our system, we account for the Packet Delivery Ratio and Routing Overhead metrics. Packet delivery ratio is calculated as the ratio of data packets received to data packets sent. For routing overhead we have taken a ratio of number of control packets generated (request, reply and error) to the number of data packets sent thus, being basically a cost v/s. gain ratio. The routing overhead ratio gives us the approximate number of control packets for each data packet sent which should not be significantly greater as compared with that of normal <span style="font-variant:small-caps;">dsr</span>. Throughput refers to the actual measured performance of the system when the delay is considered. In the simulation results, the metrics of throughput are related to the average value per node. Finally, the average delay shows the average one-way latency observed between transmitting and receiving a packet.\ Unless otherwise specified, the experiments are repeated ten times in all cases with varying random seed. The seed influences the placement and movement of the nodes. The radio range, sending capacity, and <span style="font-variant:small-caps;">mac</span> have been chosen to represent a typical adhoc mobile node; the speed is uniformly distributed between 0 and 10 m/s to represent speed of user in fixed location, walking or running. The simulation time is chosen to be long enough to potentially roam the whole area and is set to 900 seconds. The system was deployed and tested with following values of constant parameters Table 2: Values of Constant Parameters\ Constant Value --------------------------------- -------------------- Neutral Rating 0 Suspicious Threshold -35 Malicious Threshold -50 Window Size 1 second (Default) Self Observation Weightage -5 Warning Message Weightage -2 Avoid List Appearance Weightage -1 Inactivity Timeout Period 20 seconds Finally, <span style="font-variant:small-caps;">cbr</span> has been chosen for generating the traffic. The scenario and traffic connections have been randomly generated using the cbrgen and setdest utility from <span style="font-variant:small-caps;">cmu</span>’s Monarch project [@CMUS]. Results and Performance Evaluation ================================== ![\[fig:PDratio\]Packet Delivery Ratio Comparison](1.pdf) Results for Packet Delivery Ratio are shown in the Figure \[fig:PDratio\]. The system performs better than normal <span style="font-variant:small-caps;">dsr</span> comparing results taken after average over 10 iterations with different pause times. The system performance is significantly better than normal <span style="font-variant:small-caps;">dsr</span> when percentage of malicious nodes is less than 40. After which, it starts to deteriorate and significantly falls after 70%. But, in case 70% or more nodes are malicious we can simply discard the network as it is no longer of significance. There is no need to establish trust relationship and links among the nodes when 7 out of 10 are known to be malicious. Figure \[fig:ROcomp\] shows routing overhead of our system protocol compared with normal <span style="font-variant:small-caps;">dsr</span>. Number of control messages in the network are significant, as more are the number of packets, more is the time spent in establishing routes and lesser is the number of data packets sent. Our system performs better than normal <span style="font-variant:small-caps;">dsr</span> without much extra routing overhead. This extra routing overhead is generated because whenever a new node is declared as malicious, a route error is generated and the link is broadcast as broken. After which, some more time is consumed to establish a new link. This is crucial to the <span style="font-variant:small-caps;">ids</span> performance. ![\[fig:ROcomp\]Routing Overhead Comparison](2.pdf) The Figure \[fig:Wsizes\] illustrates fine tuning of the system with respect to sliding window size. Window size is a crucial parameter for the system because of its role in deciding the system performance. As depicted by the figure, in most cases, the window size 1.25 seconds scenario delivers optimal packet delivery ratio as compared to other scenarios where window size is of 0.50, 0.75, 1.00, 1.50 and 1.75 seconds. From the figure, one can infer that for a small window size, the system is too busy in various book-keeping tasks for monitoring and reputation updating. For a larger window size, the system response gets too slow. Hence, the time to identify malicious nodes increases and accordingly does the number of packet drops. Therefore, overall performance of system deteriorates in terms of packet delivery ratio. ![\[fig:Wsizes\]System performance for various window sizes](6.pdf) The Figure \[fig:Ptimes\] shows packet delivery ratio values for the system against pause times of 0, 100, 300, 600 and 900 seconds. ![\[fig:Ptimes\]System performance against Pause times](7.pdf) In a highly mobile scenario such as one with pause time of 0 seconds the system performance decreases as most of the time the system has to cope up with the mobility of nodes and tasks like updating route cache, discovering & establishing routes etc. Likewise, in a static network scenario with pause times of 600 and 900 seconds, where the system does not have many choices in terms of clean routes, once these nodes get identified, the system performance also degrades. As depicted by the Figure \[fig:ROWsizes\], the performance is optimal for a scenario with window size of 1.25 seconds in terms of routing overhead. Although, the difference between various scenarios presented is not very significant, but it is crucial for system performance to optimize this value. ![\[fig:ROWsizes\]System routing overhead for various window sizes](8.pdf) In addition, the system also performs optimally in terms of the routing overhead incurred. Thus, the proposed solution is able to strengthen the defense of <span style="font-variant:small-caps;">dsr</span> protocol without incuring much of overhead. Conclusion And Future Perspective ================================= Mobile adhoc networks have a number of significant security issues which cannot be solved alone by simple <span style="font-variant:small-caps;">ids</span>. In this paper, we have critically examined the existing systems and outlined their strength and shortcomings. We have opted a semi-distributed approach for our system in terms of mode of information propagation among nodes. The goal was to design a system incorporating the best traits of all existing systems without incurring extra routing overhead. Congestion parameter, Knock test and Timing window are some new concepts that have been introduced in this system. Detailed simulations carried out over our system using ns2 for performance evaluation have contributed significantly to some crucial design issues. Optimal values of the parameters used are obtained and critically examined for efficient performance of the system. However, some additional study is required for evaluating the adequacy and importance of congestion parameter. The system performance can also be judged by interchanging the values of weightage assigned to self observation with that of other reputation update modes such as warning message and avoid list citations. It is our belief that some interesting results are bound to come with such studies which shall justify the system design in its current stage. [0]{} C. E. Perkins, P. Bhagwat: “Highly Dynamic Destination-Sequenced Distance Vector (DSDV) for Mobile Computers" Proc. of the SIGCOMM 1994 Conference on Communications Architectures, Protocols and Applications, pp 234-244, Aug (1994). S. Murthy, J.J. Garcia-Luna-Aveces: “A Routing Protocol for Packet Radio Networks", Proc. ACM International Conference on Mobile Computing and Networking, pp. 86-95, November, (1995). Ching-Chuan Chiang, Hsiao-Kuang Wu, Winston Liy, Mario Gerla: “Routing in Clustered Multihop, Mobile Wireless Networks with Fading Channel", IEEE Singapore International Conference on Networks, (SICON’97), pp. 197-211, Singapore, 16.-17. April (1997). David B. Johnson, David A. Maltz, and Josh Broch, “DSR The Dynamic Source Routing Protocol for Multihop Wireless Ad Hoc Networks" In Ad Hoc Networking, edited by Charles E. Perkins, chapter 5, pages 139–172. Addison-Wesley, (2001). Charles E. Perkins and Elizabeth M. Royer: “Ad-hoc On-Demand Distance Vector Routing" in Proceedings of the 2nd IEEE Workshop on Mobile Computing Systems and Applications, New Orleans, LA, pp. 90-100, February (1999). V. Park, S. Corson: “Temporally-Ordered Routing Algorithm (TORA)" VERSION 1 Internet Draft, draft-ietf-manet-tora-spec- 03.txt, June (2001). Chai-Keong Toh: “A Novel Distributed Routing Protocol To Support Ad hoc Mobile Computing", Proc. IEEE 15th Annual International Phoenix Conference on Computers and Communications, IEEE IPCCC 1996, 27 March-29, Phoenix, AZ, USA, pp. 480-486 (1996). R. Dube, C. D. Rais, K. Wang and S. K. Tripathi: “Signal Stability based adaptive routing (SSR alt SSA) for ad hoc mobile networks", IEEE Personal Communication, Feb. (1997). Zygmunt J. Haas, Marc R. Pearlman, Prince Samar: “The Zone Routing Protocol (ZRP) for Ad Hoc Networks", Internet Draft, http://www.ietf.org/proceedings/02nov/I-D/draft-ietf-manet-zone-zrp-04.txt, work in progress, July (2002). Sonja Buchegger, Cedric Tissieres, Jean-Yves Le Boudec, “A Test-Bed for Misbehavior Detection in Mobile Ad-hoc Networks — How Much Can Watchdogs Really Do?," Sixth IEEE Workshop on Mobile Computing Systems and Applications (WMCSA’04), pp. 102-111, (2004). P. Michiardi, R. Molva, “Core: A COllaborative REputation mechanism to enforce node cooperation in Mobile Ad Hoc Networks", Institut EurecomResearch Report RR-02-062 - December (2001). Sonja Buchegger and Jean-Yves Le Boudec, “Performance Analysis of the <span style="font-variant:small-caps;">confidant</span> Protocol: Cooperation Of Nodes:Fairness In Dynamic Ad-hoc NeTworks" Proceedings of IEEE/ACM Symposium on Mobile Ad Hoc Networking and Computing (MobiHOC), Lausanne, June(2002). Sorav Bansal and Mary Baker, “Observation based cooperation enforcement in ad hoc networks" Technical Report, Stanford University, arXiv:cs.NI/0307012 v2 6 Jul (2003). P. Resnick and R. Zeckhauser. “Trust among strangers in internet transactions: Empirical analysis of ebay’s reputation system" In M. R. Baye, editor, The Economics of the Internet and E-Commerce, volume 11 of Advances in Applied Microeconomics. Amsterdam, Elsevier Science, (2002). Sonja Buchegger, Jean-Yves Le Boudec: “Coping with False Accusations in Misbehavior Reputation Systems for Mobile Ad-hoc Networks" EPFL Technical Report IC/2003/31 (2003). S. Buchegger and J.-Y. Le Boudec, “The effect of rumor spreading in reputation systems for mobile ad-hoc networks" Proc. WiOpt’03(Modeling and Optimization in Mobile Ad Hoc and Wireless Networks), (2003). Animesh K. Trivedi, Rishi Kapoor, Rajan Arora, Sudip Sanyal and Sugata Snayal: “RISM - Reputation Based Intrusion Detection System for Mobile Adhoc Networks", accpeted in CODEC’06, Kolkata, India (http://www.irpel.org/phpfiles/codec-06.php) to be held in Dec. (2006). P. Yau and C. J. Mitchell, “Reputation methods for routing security for mobile ad hoc networks", in Proceedings of SympoTIC ’03, Joint IST Workshop on Mobile Future and Symposium on Trends in Communications, Bratislava, Slovakia, October (2003). NS Home Page: http://www.isi.edu/nsnam/ns/ “CMU Monarch Project web site." http://www.monarch.cs.cmu.edu/ Sonja Buchegger and Jean-Yves Le Boudec, “Self-Policing Mobile Ad-Hoc Networks by Reputation Systems" IEEE Communication Magazine, vol. 43, num. 7, p. 101(2005). Sergio Marti, T.J. Giuli, Kevin Lai, and Mary Baker, “Mitigating routing misbehavior in mobile ad hoc networks" Proceedings of the 6th annual international conference on Mobile computing and networking Boston, Massachusetts. Pages: 255 - 265 (2000). Animesh Kumar Trivedi is a senior undergraduate student of Information Technology at the Indian Institute of Information Technology, Allahabad, India. His research interests include Security issues in Wireless Computer Networks, Computer Architecture and Distributed Systems. Further details about him can be had from <http://animesh.trivedi.googlepages.com/>\ Rajan Arora is a senior undergraduate student of Information Technology at the Indian Institute of Information Technology, Allahabad, India. His research interests include Wireless Computer networks, Cryptography, Network Security and Distributed systems. Further details about him can be had from <http://arorarajan.googlepages.com/>\ Rishi Kapoor is a senior undergraduate student at the Indian Institute of Information Technology, Allahabad, India pursuing his undergraduation in Information Technology. His research interests include Computer and Wireless networks, Network security and Databases. Further details about him can be had from <http://profile.iiita.ac.in/rkapoor_b03/>\ Sudip Sanyal is an Associate Professor at the Indian Institute of Information Technology, Allahabad, India. He recieved his M.S. and Ph.D. degree from the Banaras Hindu University at Varanasi, India. His current activities lie in fields of Computer networks, Software Engineering, Parallel Computing and Natural language Processing. He has published numerous papers in various national and international journals and attended many conferences.\ Sugata Sanyal is in the Faculty of the Tata Institute of Fundamental Research, India. He received his Ph.D. degree from Mumbai University, India, M.Tech from IIT, Kharagpur, India and B.E. from Jadavpur University, India. His current research interests include security in wireless and mobile ad hoc networks, distributed processing, and scheduling techniques. He has published numerous papers in national and international journals and attended many conferences. He is in the editorial board of three International Journals. He is co-recipient of ‘‘Vividhlaxi Audyogik Samsodhan Vikas Kendra Award (VASVIK)’’ for Electrical and Electronics Science and Technologies (combined) for the year 1985. He was a Visiting Professor in the Department of Electrical and Computer Engineering and Computer Science in the University of Cincinnati, Ohio, USA in 2003. He delivered a series of lectures and also interacted with the Research Scholars in the area of Network Security in USA, in University of Cincinnati, University of Iowa, Iowa State University and Oklahoma State University. He has been an Honorary Member of Technical Board in UTI (Unit Trust of India) and SIDBI (Small Industries Development Bank of India). He has also acted as a consultant to a number of leading industrial houses in India. More information about his activities is available at [http://www.tifr.res.in/sanyal](http://www.tifr.res.in/~sanyal).	{ "pile_set_name": "ArXiv" }
--- abstract: 'The methods of mode decomposition and Fourier analysis of classical and quantum fields on curved spacetimes previously available mainly for the scalar field on Friedman-Robertson-Walker (FRW) spacetimes are extended to arbitrary vector bundle fields on general spatially homogeneous spacetimes. This is done by developing a rigorous unified framework which incorporates mode decomposition, harmonic analysis and Fourier analysis. The limits of applicability and uniqueness of mode decomposition by separation of the time variable in the field equation are found. It is shown how mode decomposition can be naturally extended to weak solutions of the field equation under some analytical assumptions. It is further shown that these assumptions can always be fulfilled if the vector bundle under consideration is analytic. The propagator of the field equation is explicitly mode decomposed. A short survey on the geometry of the models considered in mathematical cosmology is given and it is concluded that practically all of them can be represented by a semidirect homogeneous vector bundle. Abstract harmonic analytical Fourier transform is introduced in semidirect homogeneous spaces and it is explained how it can be related to the spectral Fourier transform. The general form of invariant bi-distributions on semidirect homogeneous spaces is found in the Fourier space which generalizes earlier results for the homogeneous states of the scalar field on FRW spacetimes.' address: - \| Max Planck Institute for Mathematics in the Sciences, Inselstr. 22,\ 04103 Leipzig, Germany\ - \| Institut für Theoretische Physik, Universität Leipzig, Brüderstr. 16,\ 04103 Leipzig, Germany author: - 'ZHIRAYR G. AVETISYAN' bibliography: - 'lib.bib' title: A UNIFIED MODE DECOMPOSITION METHOD FOR PHYSICAL FIELDS IN HOMOGENEOUS COSMOLOGY --- Introduction ============ As long as mankind is not in possession of a successful and commonly accepted quantum theory of gravitation (and possibly even thereafter), the quantum field theory on curved spacetimes (QFT in CST) is an adequate and consistent theoretical framework for astronomy and cosmology. Vicinities of black holes and the early epoch of the universe are two prominent physical situations where gravity is sufficiently strong so that its influence on the quantum field theory cannot be neglected. At the same time, in these situations the gravity is sufficiently uniform (i.e., its local fluctuations are negligible) to be considered classical and interacting with matter quantum fields mainly macroscopically. This semiclassical picture is captured by letting quantum fields propagate on a curved spacetime. The back reaction of the matter on gravity is described by the semiclassical Einstein equation, where gravity feels only the expectation values of quantum entities. This is the domain of quantum field theory on curved spacetimes, apart from its intrinsic appeal for the beauty and variety of fundamental problems it poses in mathematical physics. QFT in CST adopts the more modern algebraic quantum field theory setup and, as appropriate to mathematical physics, attempts to be as axiomatic and deductive as possible and mathematically rigorous. The disadvantage of such an approach is the extreme difficulty of producing explicit ready-made results which can be applied on the observational level, and each such result can be considered as a remarkable success. For instance, it was not until 2010 when the authors of [@Degner_Verch_2010] obtained the first completely rigorous and at the same time explicit description of cosmological particle creation in states of low energy. This was done under several assumptions which can or cannot be considered realistic in cosmology. Namely, the Klein-Gordon field on FRW spacetimes was chosen, and homogeneity and sufficient regularity was stipulated a priori for the desired state of low energy. When one tries to step a bit beyond these restrictions one immediately faces severe mathematical difficulties along the entire way from the very setup until the final expressions. The reason is that the chain of results used in these constructions has been obtained only under the above mentioned assumptions. The aim of the current work is the extension of some of those mathematical methods to a generality where they can be applied for practically all realistic cosmological situations. To which extent this program has been successful will become clear below. A primary tool for obtaining explicit constructions are geometric symmetries. After publishing his eminent work on general relativity, Einstein declared he had no hope to see explicit solutions of his equation in the near future. It was the rich symmetry of the FRW spacetimes that allowed Friedmann to find first explicit solutions and thus to dispel the despair of Einstein shortly after his publication. This instant can be considered as the birth of modern mathematical cosmology, which until today remains one of the main appliers of explicit solutions in general relativity. One of the merits of geometric symmetries is the possibility of the separation of variables in the field equation which helps to obtain explicit solutions. The mode decomposition of the solutions of the field equation (also referred to as the Fourier method in PDE, or expansion into harmonic oscillators in physics) was probably first applied in the cosmological context by Parker [@Parker1969] who performed it on the flat FRW spacetime. The idea of the method is that one tries to separate the time variable in the field equation, and looks for solutions as linear combinations of products $X(\vec x)T(t)$ where $X$ depends only on the spatial coordinates and $T$ only on time. What Parker discovered is that this is possible on FRW spacetimes and represents a very handy tool for the analysis of the dynamics. A thorough analytical investigation of the method in the cosmological context was conducted in [@Fulling1989], where an abstract functional analytical eigenfunction expansion was introduced as a methodological background, and precise methods were suggested for the mode decomposition of regular solutions on FRW and ultrastatic spacetimes. The theory of the method does not seem to have been developed any further until nowadays. In particular, the following questions remain open. What are the precise limits of applicability of the mode decomposition by means of separation of the time variable? How many different decompositions are possible for the same geometrical setup? When and how can the decomposition be extended to weak (distributional) solutions? In the first part of this work satisfactory answers will be given to these questions and establish a unified framework for the method. Our geometrical setup will be a finite dimensional vector bundle over an arbitrary globally hyperbolic Lorentzian manifold furnished with a pseudo-Riemannian fiber metric and a fiber metric linear connection. The field equation will then be given by the field operator $D=\Box^\nabla+m^\star(x)$ where $\Box^\nabla$ is the connection d’Alambert operator acting on the smooth sections, and $m^\star(x)$ will be the variable smooth “mass term” (possibly including a coupling to scalar curvature) to which mild assumptions will be imposed. This seems to be the most general setup of a (symmetric) hyperbolic linear field on a curved spacetime, and covers most practical situations in the cosmological context. The results can be briefly described as follows. Precise geometrical necessary and sufficient conditions are obtained for the mode decomposition of smooth solutions by time separation to be realizable. This mode decomposition is given, as perhaps expected, by the time dependent Fourier transform, and is shown to be basically the only such mode decomposition possible. The decomposition is extended to all distributional solutions in a natural manner given that there exists a choice of modes fulfilling certain regularity conditions. The conditions become fully explicit once one has a Paley-Wiener theorem for the spatial Fourier transform, i.e., a precise description of the Fourier image of the test functions space. This is the situation in FRW spaces. Moreover, it is shown that if the bundle is analytic and the dynamics of the geometry sufficiently rigid (precise definitions are given) then the conditions are satisfied regardless of the harmonic analysis involved. It turns out that that the mode solutions of non scalar fields under certain circumstances experience infrared instability periods not known for scalar fields; the author is yet not sure about the physical essence of this phenomenon. Apart from this, traditionally infrared integrability issues arise when integrating modes over the spectrum ${\mbox{Spec}}(D_{\Sigma})$ of a Schrödinger operator $D_\Sigma$ if $\overline{{\mbox{Spec}}(D_{\Sigma})}$ includes the eigenvalue $0$ [@Fewster:2003ey]. Here this question of integrability of modes over the spectrum is settled by showing that the suitably chosen mode solutions remain well under control even at non positive spectral values. As an example of application and as a byproduct the explicit formula of the propagator of the field in the Fourier space is found, which generalizes one obtained in [@Lueders_Roberts_1990]. Another advantage of geometric symmetries is the possibility to apply harmonic analysis. This is particularly true for the cosmological models where a rather rich group of spatial isometries is imposed. Then the spatial sections of the spacetime can be considered as homogeneous spaces, and the spatial Fourier transform can be investigated in much more detail with many explicit consequences. These properties then can be dragged to the time dependent Fourier transform and hence to the mode decomposition. Of particular interest in the quantum field theory on cosmological spacetimes are the spatially symmetric (invariant) states, of which the 2-point functions are bi-distributional solutions of the field equation which are invariant under the full isometry group. In [@Lueders_Roberts_1990] the Fourier image of the isotropic homogeneous states of the Klein-Gordon field on FRW spacetime has been obtained, under an additional continuity requirement which has no clear physical interpretation. In contrast the most general form of invariant scalar bi-distributions on $\mathbb{R}^d$ has been obtained in [@Gelfand_Vilenkin1964] using a nice technique. What appears to be missing is a generalization of these results to sufficiently many homogeneous spaces so that practically all cosmological situations are covered. The harmonic analysis of FRW symmetry groups is well known since long, but strictly speaking isotropy is not as fundamental in cosmology as homogeneity, and one is also interested in cosmological models which are only homogeneous (Bianchi models) or in addition only partially isotropic (LRS models). The isometry groups of these spaces are described by Bianchi groups with their quotients and semidirect extensions (in case of LRS models). Some of these groups are solvable, others are semisimple, with finite or infinite center. Therefore it is not easy to establish a unified harmonic analytical approach for all cases, although one has to admit that Kirillov’s theory for the solvable groups and Helgason’s theory for semisimple groups together would cover the majority of situations. To obtain a unified theory one can adopt abstract harmonic analysis. This beautiful branch of mathematics allows to obtain many results in an admirable generality. However, apart from compact groups, it is not completely clear how to relate the abstract group Fourier transform with the eigenfunction expansion of the invariant Laplace operator. At least there seems to be no unified exposition of these techniques applied in the cosmological context. The aim of our second part will be to put together some tools from harmonic analysis which are adequate in cosmology, and to obtain useful results with their help. First a short survey on homogeneous bundles in general, and on semidirect homogeneous bundles in particular is carried out, and it is shown that they cover the vast and the most important majority of the realistic cosmological structures. Next, the abstract harmonic analytical Fourier transform is introduced with its requisites on semidirect homogeneous spaces. Although the abstract Fourier transform on groups and the representation theory in homogeneous spaces are to be considered as well studied and widely known, the abstract Fourier transform on homogeneous spaces is not that popular and deserves a better exposition (at least we were not able to find a satisfactory one in the literature). Then some properties of the Fourier transformed distributions are established. Although some of these results may be known to experts in harmonic analysis, we were not able to locate them in the for required for our purposes in the literature. And because these properties are needed to obtain our results on invariant distributions they are included here with proofs. Next an attempt is made to unify the abstract Fourier transform and the usual spatial Fourier transform given by the eigenfunction expansion of the Laplace operator. Several remarks are made on this way, which may serve as a guideline to completely build the desired correspondence once a particular structure is specified. This has been indeed performed for the purely homogeneous cosmological models, which will appear in a subsequent publication. The necessity of such an explicit correspondence consists in the ability to transfer the results obtained in the abstract setup to the situation with the usual Fourier transform, which is far more useful a tool for concrete calculations. Finally, by a generalization of the above mentioned technique in [@Gelfand_Vilenkin1964] the general form of the invariant bi-distributions in arbitrary semidirect homogeneous vector bundle is found without any additional assumptions on their regularity. It is concluded that the polynomial bound of the Fourier transformed homogeneous state as found in [@Lueders_Roberts_1990] is a consequence of the imposed regularity requirements. Mode Decomposition of Hyperbolic Fields ======================================= Linear hyperbolic fields ------------------------ It is generally believed that the forces of nature are described by tensor and spinor fields. A geometrical generalization of those are the vector bundle fields, i.e., fields as smooth sections of some vector bundles. In general relativity one works mainly on a four dimensional Lorentzian smooth manifold $(M,g)$ which is called a spacetime. We will be concerned with hyperbolic fields given by a wave equation, hence we put an additional constraint on the spacetime $(M,g)$ to be globally hyperbolic, so that the Cauchy problem of the wave equation is well-posed. For simplicity only linear fields will be discussed here. For the reduction of the Maxwell and Proca fields to linear hyperbolic fields the reader is referred to [@Fewster:2003ey],[@BarGinouxPfaffle200703]. We summarize the basic setup of the the linear hyperbolic fields in the following section. Let $V$ be an $n$-dimensional vector space. Let $\mathcal{T}\xrightarrow[]{\pi}M$ be a vector bundle with standard fiber $V$ and with a pseudo-Riemannian metric $\langle u,v\rangle_\mathfrak{g}$. Let further $\mathcal{E}(\mathcal{T})=C^\infty(\mathcal{T})$ and $\mathcal{D}(\mathcal{T})=C_0^\infty(\mathcal{T})$ be the spaces of smooth sections and of those with compact support, correspondingly. Let $\nabla$ be a metric connection on $\mathcal{T}$ and $\Box^\nabla$ the associated d’Alambert operator on $\mathcal{E}(\mathcal{T})$. Define the field operator to be the normal hyperbolic operator $D=\Box^\nabla+m^\star(x)$ acting on $\mathcal{E}(\mathcal{T})$, where $m^\star\in C^\infty(M)$ is a generalization of the usual mass term $m^2$ which now can also contain the coupling term $\xi R$. Note that because differential operators are support-decreasing, $D\mathcal{D}(\mathcal{T})\subset\mathcal{D}(\mathcal{T})$. A free linear hyperbolic field $\phi\in\mathcal{E}(\mathcal{T})$ is a solution of the field equation $D\phi=0$. Being a globally hyperbolic spacetime, $M=\mathcal{I}\times\Sigma$, where $\mathcal{I}\subseteq\mathbb{R}$ is an interval, and for each $t\in\mathcal{I}$ the hypersurface $\Sigma_t\sim\Sigma$ is a three dimensional embedded Riemannian submanifold, which is spacelike with respect to $g$ and is a Cauchy surface in the sense described below. Thanks to [@Bernal_Sanchez_2005] one can choose a smooth global time function $t$ and a coordinate atlas such that $x=(t,\vec x)=(x_0,x_1,x_2,x_3)$ where $t\in\mathcal{I}$ and $\vec x\in\Sigma$, i.e., $\Sigma_t$ are equal $t$ hypersurfaces. The restriction of the bundle $\mathcal{T}$ to the submanifold $\Sigma_t$ will be denoted by $\mathcal{T}_t=\pi^{-1}(\Sigma_t)$. The spaces of smooth sections will be $\mathcal{E}(\mathcal{T}_t)=C^\infty(\mathcal{T}_t)$ and $\mathcal{D}(\mathcal{T}_t)=C_0^\infty(\mathcal{T}_t)$. If $i_t:\mathcal{T}_t\to\mathcal{T}$ is the identical embedding, then its pullback $i_t^$ is the restriction map for objects on $\mathcal{T}$ to $\mathcal{T}_t$. In particular $i_t^:\mathcal{E}(\mathcal{T})\to\mathcal{E}(\mathcal{T}_t)$ and $i_t^:\mathcal{D}(\mathcal{T})\to\mathcal{D}(\mathcal{T}_t)$ are linear surjective maps. The embedding $\pi\circ i_t\circ\pi^{-1}:M\to\Sigma$ gives rise to a natural embedding $i_t:TM\to T\Sigma$ and of all tensor bundles (using the same symbols $i_t$, $i^_t$ for different restrictions in the spirit of polymorphism should not lead to a confusion). The Riemannian metric $h$ on $T\Sigma$ will be $h=-i^_t(g)$, with minus sign here because of the signature convention $(+,-,-,-)$. The restriction $i^_t(\nabla)=\nabla_{i^_t(.)}=\nabla^t$ is a metric connection on $\mathcal{T}_t$. The associated Laplace operator $\Delta_t=\Delta^{\nabla^t}$ is an elliptic operator on $\mathcal{E}(\mathcal{T}_t)$ (so that $-\Delta_t$ is a positive operator). The restriction of the field operator $D$ to $\mathcal{E}(\mathcal{T}_t)$ will be denoted by $D_{\Sigma_t}=-\Delta_t+m^\star(x)$. An existence and uniqueness theorem [@BarGinouxPfaffle200703],[@Gunther1988],[@Dimock1980] for wave operators tells that the Cauchy problem is well posed: there exists a bijective linear map $$\mathcal{E}(\mathcal{T}_t)\oplus \mathcal{E}(\mathcal{T}_t)\ni(f_0,f_1)\to \jmath(f_0,f_1)\in\left\{f\in\mathcal{E}(\mathcal{T})\mbox{: }Df=0\right\}$$ such that $f_0=i_t^(f)$ and $f_1=i_t^(\nabla_tf)$, where $\nabla_t=\nabla_{\frac{\partial}{\partial t}}$. Furthermore, there exist unique Green’s operators $E^\pm:\mathcal{D}(\mathcal{T})\to\mathcal{E}(\mathcal{T})$ satisfying $DE^\pm=E^\pm D=id_{\mathcal{D}(\mathcal{T})}$ and $supp\{G^\pm f\}\subset J^\pm(supp\{f\})$ for all $f\in\mathcal{D}(\mathcal{T})$. Here $J^\pm(N)$ with a subset $N\subset M$ denotes the causal future/past of $N$. Define by $E=E^+-E^-$ the [propagator]{} of $D$, which satisfies $DE=ED=0$. Now $Sol(\mathcal{T})=\jmath(\mathcal{E}(\mathcal{T}_t)\oplus \mathcal{E}(\mathcal{T})_t)$ and $Sol_0(\mathcal{T})=\jmath(\mathcal{D}(\mathcal{T}_t)\oplus \mathcal{D}(\mathcal{T}_t))$ will denote correspondingly the spaces of all smooth solutions, and of those satisfying $supp\{f\}\cap\Sigma_t$ compact for all $t\in\mathcal{I}$, respectively. Then $E\mathcal{D}(\mathcal{T})\subset Sol_0(\mathcal{T})$. There is a symplectic form on $Sol_0(\mathcal{T})$: $$\sigma(u,v)=\int_{\Sigma_t}d\mu_h\left[\langle i_t^(u),i_t^(\nabla_tv)\rangle_g-\langle i_t^(\nabla_tu),i_t^(v)\rangle_g\right]\mbox{, }\forall u,v\in Sol_0(\mathcal{T})\mbox{, }\forall t\in\mathcal{I},$$ where $h=-i_t^(g)$ is the induced Riemannian metric on $\Sigma_t$. That this is conserved (analogous to a Wronskian in ODE) can be seen by considering the Green’s identity for $u,v\in Sol_0(\mathcal{T})$ on the regular cylindric region $U=(t_1;t_2)\times\Sigma\subset M$ for any $t_1\neq t_2$, $$0=\int_Ud\mu_g\left[\langle u,Dv\rangle_g-\langle Du,v\rangle_g\right]=$$ $$=\int_{\partial U}d\mu_h\left[\langle i_t^(u),i_t^(\nabla_tv)\rangle_g-\langle i_t^(\nabla_tu),i_t^(v)\rangle_g\right]\mbox{, }\forall u,v\in Sol_0(\mathcal{T}).$$ This identity also helps us along with Green’s operators to find the explicit form of the map $\jmath$. Given any $v\in Sol_0(\mathcal{T})$, $f\in\mathcal{D}(\mathcal{T})$, we apply it two times; once for the pair $v,u=E^+(f)$ on the region $U^+=(-\inf\{\mathcal{I}\};t)$ and once for the pair $v,u=E^-(f)$ on the region $U^+=(t;\inf\{\mathcal{I}\})$. Summing up the resulting two identities and using the support properties of $E^\pm$ we arrive at $$\begin{aligned} \int_Md\mu_g\langle v,f\rangle_g=\sigma(v,E(f)).\label{CauchyExplicit}\end{aligned}$$ We see that the functional $\sigma(v,E(.)):\mathcal{D}(\mathcal{T})\to\mathbb{C}$ actually is given by a smooth integral kernel, which equals $v$. Thus we can write symbolically $$\jmath(f_0,f_1)[y]=\int_{\Sigma_t}d\mu_h\left[\langle f_0,\nabla_tE(y)\rangle_g-\langle f_1,E(y)\rangle_g\right]\mbox{, }\forall f_0,f_1\in\mathcal{D}(\mathcal{T}_t)\mbox{, }y\in M\mbox{, }t\in\mathcal{I}.$$ For full details of this last computation the reader is referred to [@DIMOCK_1992], where the argument is given for 1-forms, but is readily applicable to our more general case. \[ESurj\] The operator $E:\mathcal{D}(\mathcal{T})\to Sol_0(\mathcal{T})$ is surjective. Let $v\in Sol_0$, and let $K_v={\mbox{supp}}v\cap\left([0,1]\times\Sigma\right)$ be the compact region of its support between times 0 and 1. Let further $\chi\in\mathcal{E}(M)$ be a smooth function which equals 1 for $t<0$ and 0 for $t>1$. Denote $v^-=-v\chi$ and $v^+=v(1-\chi)$, then $v=v^+-v^-$. Let $f_v=Dv^+$, then ${\mbox{supp}}f_v\subset K_v$ is compact, hence $f_v\in\mathcal{D}(\mathcal{T})$. The equation $f_v=Dv^+$ has a unique solution with past compact support, and it is given by $v^+=E^+f_v$. Now $Dv^-=-Dv+Dv^+=f_v$, and similarly $v^-=E^-f_v$. Then $v=E^+f_v-E^-f_v=Ef_v$. The arbitrariness of $\chi$ reflects the non-injectivity of $E$. Spectral mode decomposition --------------------------- Henceforth we will use nomenclature introduced in the Appendix A without special notice. Consider the operators $D:\mathcal{D}(\mathcal{T})\to\mathcal{D}(\mathcal{T})$ and $D_\Sigma:\mathcal{D}(\mathcal{T}_t)\to\mathcal{D}(\mathcal{T}_t)$. If $m^\star(x)\in\mathbb{R}$ everywhere on $M$, then by the virtue of Green’s identity $D$ and $D_{\Sigma_t}$ are formally self-adjoint with respect to the inner products $(,)_M$ and $(,)_{\Sigma_t}$. We will not need the self-adjointness of $D$ in the current work. The constructions below will pertain mainly to $D_{\Sigma_t}$. The conditions on $m^\star(x)$ for $D_{\Sigma_t}$ to have a self-adjoint extension can be found in [@CyconFroeseKirschSimon200801]. We moreover require that the operator $D_{\Sigma_t}$ be lower semi-bounded. In practice will be mainly interested in cosmological models, where $m^\star(x)=m^\star(t)$ is a function of time only, so that no problems arise. Below we assume self-adjoint extensions for both $D$ and $D_{\Sigma_t}$, but for $D$ this is only symbolic and targets simply at coherent notations. Let $D$ and $D_{\Sigma_t}$ be extended to self-adjoint operators on $L^2(\mathcal{T})$ and $L^2(\mathcal{T}_t)$ respectively. In the rigged Hilbert spaces [@Gelfand_Vilenkin1964],[@Maurin1972],[@Maurin1968] $\mathcal{D}(\mathcal{T})\subset L^2(\mathcal{T})\subset\mathcal{D}(\mathcal{T})'$ and $\mathcal{D}(\mathcal{T}_t)\subset L^2(\mathcal{T}_t)\subset\mathcal{D}(\mathcal{T}_t)'$ operators $D$ and $D_{\Sigma_t}$ possess complete systems of eigenfunctions $\{u_\rho\}$ and $\{\zeta_\lambda\}$ satisfying $$Du_\rho=\rho u_\rho\mbox{, }u_\rho\in\mathcal{D}(\mathcal{T})'\mbox{, }\rho\in\mathbb{R},$$ $$D_{\Sigma_t}\zeta_\lambda=\lambda\zeta_\lambda\mbox{, }\zeta_\lambda\in\mathcal{D}(\mathcal{T}_t)'\mbox{, }\lambda\in\mathbb{R}.$$ Denote by $\mathcal{D}(\mathcal{T})_\rho'$ and $\mathcal{D}(\mathcal{T}_t)_\lambda'$ the linear spaces of eigenfunctions corresponding to $\rho$ and $\lambda$, respectively. Furthermore, there exists an isomorphism $$\begin{aligned} L^2(\mathcal{T}_t)=\int_\mathbb{R}^\oplus d\nu(\lambda)H(\lambda),\label{SpectralDecomp}\end{aligned}$$ where $$D_{\Sigma_t}\|_{H(\lambda)}=\lambda,$$ and $d\nu(\lambda)$ is a positive measure. Each $H(\lambda)$ is continuously embedded in $\mathcal{D}(\mathcal{T}_t)_\lambda'$. The eigenfunction expansion of $D$ will be the map $$\mathcal{D}(\mathcal{T})\ni f\to\tilde f_\rho\in(\mathcal{D}(\mathcal{T})_\rho')'^\mbox{, }\rho\in\mathbb{R},$$ ($X'^$ denotes the space of continuous antilinear functionals on the space $X$) where $\tilde f_\rho$ is defined by $$\tilde f_\rho(u_\rho)=\bar u_\rho(f)\mbox{, }\forall f\in\mathcal{D}(\mathcal{T})\mbox{, }u_\rho\in\mathcal{D}(\mathcal{T})_\rho'.$$ (Here we defer a little from Gelfand’s notations who puts $\tilde f_\rho(u_\rho)=u_\rho(f)$.)The expansion of $D_{\Sigma_t}$ on $\mathcal{D}(\mathcal{T}_t)$ is constructed similarly. Note that $D_{\Sigma_t}$ is an elliptic operator, hence $\mathcal{D}(\mathcal{T}_t)_\lambda'\subset\mathcal{E}(\mathcal{T}_t)$. If each $\mathcal{D}(\mathcal{T}_t)_\lambda'$ is finite dimensional (eigenvalue $\lambda$ has a finite multiplicity $N_\lambda$), then $$H(\lambda)=\mathcal{D}(\mathcal{T}_t)_\lambda'\mbox{, }\dim H(\lambda)=N_\lambda.$$ Choose $\{\zeta_{\lambda,i}\}_{i=1}^{N_\lambda}$ be a an orthonormal basis in $\mathcal{D}(\mathcal{T}_t)_\lambda'$ (orthonormality understood in $H(\lambda)$). Then $(\mathcal{D}(\mathcal{T}_t)_\lambda')'^\sim\mathbb{C}^{N_\lambda}$ by the bijective linear map $$\tilde f(\zeta_\lambda)=\tilde f\left(\sum_{i=1}^{N_\lambda}c_i\zeta_{\lambda,i}\right)\to\{\tilde f_i=\tilde f(\zeta_{\lambda,i})\}_{i=1}^{N_\lambda}\mbox{, }\forall \tilde f\in (\mathcal{D}(\mathcal{T}_t)_\lambda')'^,$$ where each $\tilde f_i\in\mathbb{C}$. In particular, if $\tilde f_\lambda$ is the mode expansion of $f\in\mathcal{D}(\mathcal{T}_t)$, then the map $$\begin{aligned} \mathcal{D}(\mathcal{T}_t)\ni f\to\tilde f_\lambda\to\{\tilde f_{\lambda,i}\}\in\int_\mathbb{R}^\oplus d\nu(\lambda)\mathbb{C}^{N_\lambda}\label{PreFourier}\end{aligned}$$ will serve as a Fourier transform on $\mathcal{D}(\mathcal{T}_t)$. Define $$Spec\{D_{\Sigma_t}\}={\mbox{supp}}\{d\nu\},$$ and $$\tilde\Sigma=\{(\lambda,i)\mbox{: }\lambda\in Spec\{D_{\Sigma_t}\}\mbox{, }i=1,...,N_\lambda\}.$$ Define the spectral measure on $\tilde\Sigma$ as $$d\mu(\alpha)=d\nu(\lambda)\times d\sharp(i),$$ where $d\sharp$ is the counting measure. The map (Eq.\[PreFourier\]) can be reformulated as $$\tilde f(\alpha)=\mathcal{F}[f](\alpha)\mbox{, }f\in\mathcal{D}(\mathcal{T}_t).$$ Then the formula (Eq.\[SpectralDecomp\]) arises a Plancherel formula $$(f,h)_{\Sigma_t}=\int_{\tilde\Sigma}d\mu(\alpha)\bar{\tilde{f}}(\alpha)\tilde g(\alpha),$$ and a Peter-Weyl (or Fourier inversion) formula $$\begin{aligned} f(x)=\int_{\tilde\Sigma}d\mu(\alpha)\tilde f(\alpha)\zeta_\alpha(x),\label{PeterWeyl}\end{aligned}$$ which holds in the $L^2$-sense so far. In our cases of interest this convergence will be in the compact topology. However, if $\mathcal{D}(\mathcal{T}_t)_\lambda'$ is infinite dimensional, more delicate tools are needed to obtain a Fourier transform with desired properties. Such tools naturally include an investigation of symmetries of the underlying geometrical structure, and this is the subject of the harmonic analysis. We will often use the formal structure (Eq.\[PeterWeyl\]) without mentioning a concrete realization, assuming that this is possible. For the cases of our interest we will indeed find a realization by means of adapted Fourier transform. In the theory of Fourier transform, and in particular in the Euclidean case, the Paley-Wiener theorems describe the functional analytical structure of the image $\mathcal{F}[\mathcal{D}(\mathcal{T}_t)]$ of the test function space under the action of the Fourier transform. This description is very useful when analyzing the situation in the Fourier space. Unfortunately there is no (at least known to us) general Paley-Wiener argument valid for any Fourier transform arisen in this manner, and the proofs of the existing ones are rather structure-specific. In applications we would like, however, to obtain results which are valid in a large variety of cases, and therefore we will introduce a notion of ’conventional’ Fourier transform which consists of a number of assumptions pertaining to the analytical properties of a given Fourier transform. Some of our later results will be valid under the assumption that the eigenvalue expansion of the operator $\Delta_t$ has at least some of the properties of a conventional Fourier transform. One says that a good definition is an assumption of a theorem. In this sense the following is not a good definition as we will not manage to use all properties in this work. However it seems feasible that these properties will become useful for several applications in quantum field theory. A Fourier transform $\mathcal{F}$ given by the eigenfunction expansion against a complete system $\{\zeta_\alpha\}_{\alpha\in\tilde\Sigma}$ will be called conventional if The Fourier space (or momentum space) $\tilde\Sigma$ is a manifold consisting of $n=\dim V$ components, $\tilde\Sigma=\bigcup_{i=1}^n\tilde\Sigma^i$, and each component $\tilde\Sigma^i$ is either a discreet set or an (not necessarily connected) analytical manifold The eigenvalue $\lambda(\alpha)$ is an analytic function on $\tilde\Sigma$ The range $\mathcal{F}[\mathcal{D}(\mathcal{T}_t)]$ is a subspace of the space of analytic functions $\tilde f(\alpha)$ on $\tilde\Sigma$ with rapid decay in $\lambda$ There is an involution $\alpha\to-\alpha$ on $\tilde\Sigma$ such that $\zeta_{-\alpha}=\bar\zeta_\alpha$. Note that it follows $\lambda(-\alpha)=\lambda(\alpha)$. In later sections we will give harmonic analytical justifications for such a ’conjecture’. This conjecture is anticipated, in particular, for all cosmological models. Moreover, being true for FRW spaces, it can be proven also for Bianchi I-VII spaces (to appear in a future publication). Further we will be mainly interested in the space of weak solutions of the field equation, $\mathcal{D}(\mathcal{T})_0'$, and will try to find a convenient characterization of it. In particular we will be looking for a complete system of solutions $\{u_\alpha\}$ spanning $\mathcal{D}(\mathcal{T})_0'$ and being in addition well handled (i.e., smooth, explicit etc.). One means of doing this is to look at a subspace of $\mathcal{D}(\mathcal{T})_0'$ which consist of solutions $f(x)=a(t)b(\vec x)$, $a\in C^\infty(\mathcal{I})$, $b\in\mathcal{E}(\mathcal{T}_t)$. Then under fortunate circumstances the field equation breaks apart into two lower dimensional elliptical eigenproblems, which are much easier to deal with. Which are those circumstances and whether such solutions span $\mathcal{D}(\mathcal{T})_0'$, and related questions, are the matter of the problem of variable separation. In the next sections we will find out in which cases this is possible and how to perform it. Separation of variables ----------------------- As discussed above, we would like to span the space $\mathcal{D}(\mathcal{T})_0'$ of weak solutions of the field equation by a family of easily computable smooth solutions $\{u_\alpha\}$. In this section we will see when and how one can perform this for the smooth solutions $Sol_0(\mathcal{T})$. The necessary requisites for this will be predominantly geometric requirements. In the next section we will show that under additional functional analytical assumptions the procedure can be extended to $\mathcal{D}(\mathcal{T})_0'$ in a natural way. Let $S$ be a subspace of $\mathcal{E}(\mathcal{T})$ with closure $\bar S\supseteq S$, $\mathfrak{M}$ a measure space with measure $d\mathfrak{m}$. An $\mathfrak{M}$-measurable family $\{u_\alpha\}_{\alpha\in\mathfrak{M}}$ of elements $u_\alpha\in\bar S$ will be called a [complete]{} or [spanning system]{} for $S$ if for any $v\in S$ there exists a unique (modulo null-supported functions) $\mathfrak{M}$-measurable function $a^v:\mathfrak{M}\to\mathbb{R}$ ($a^v:\mathfrak{M}\to\mathbb{C}$) such that $$v=\int_\mathfrak{M}d\mathfrak{m}(\alpha)a^v(\alpha)u_\alpha.$$ For the details on integration of nuclear Frechét space-valued functions see [@Thomas1975] and references therein. We will always take $\mathfrak{M}$ to be [minimal]{}, i.e., there exists no subset $A\subset\mathfrak{M}$ with $\mathfrak{m}(A)>0$ such that $a^v(A)=0$ for all $v\in S$. If the uniqueness requirement is relaxed, then $\{u_\alpha\}_{\alpha\in\mathfrak{M}}$ will be called a [redundant complete system]{} for $S$. Note that from the uniqueness property it follows, that for $d\mathfrak{m}$-almost all $\alpha\in\mathfrak{M}$, there exists no $\alpha\neq\beta\in\mathfrak{M}$ with $u_\alpha+p(\alpha)u_\beta=0$, $p(\alpha)\neq0$ a number. In other words, almost all $u_\alpha$ are pairwise independent. $\mathcal{E}(\mathcal{T})$ is a closed topological vector space with the topology of compact convergence, and $Sol(\mathcal{T})$ and $Sol_0(\mathcal{T})$ are linear subspaces. A spanning system $\{u_\alpha\}_{\alpha\in\mathfrak{M}}$ of $Sol_0(\mathcal{T})$ of the form $u_\alpha=T_\alpha X_\alpha$, where $T_\alpha\neq\bar T_\alpha\in C^\infty(\mathcal{I})$ ($T_\alpha$ and $\bar T_\alpha$ linearly independent) and $X_\alpha\in\mathcal{E}(\mathcal{T}_t)$, such that $DT_\alpha X_\alpha=D\bar T_\alpha X_\alpha=0$, will be called a [complete (time-)variable separated system of solutions]{} (or shorter, [separating system]{}). We will assume that a Fourier transform $\mathcal{F}$ on $\mathcal{D}(\mathcal{T}_t)$ is specified by means of the spectral decomposition of $D_{\Sigma_t}$ as described in the previous section. The system of eigenfunctions $\{\zeta^t_\alpha\}$ of $\Delta_t$, with the Fourier space $\tilde\Sigma_t$ and the spectral measure $d\mu(\alpha)$ on it, provide a spanning system for $\mathcal{D}(\mathcal{T}_t)$ by means of the Fourier inversion (or Peter-Weyl) formula. Below we will come across the question of a spectral theory of formally non-self-adjoint, i.e., asymmetric differential operators of type $a(x)D_{\Sigma_t}$. As a rule, the eigenfunction problems of asymmetric (aside from unitary) operators are ill-posed, and eigenfunctions do not comprise a complete system, but there are rare exceptions. At this point we have to admit the non-exhaustiveness of our treatment, as we do not analyze this possibility. We will loosely rule out the possibility of such operators to have a well-posed eigenfunction problem. A small remark will be useful later in the section. \[SepSysSpanRemark\] If $\{T_\alpha X_\alpha\}_{\alpha\in\mathfrak{M}}$ is a separating system for $Sol_0(\mathcal{T})$ with compact topology, then for each $t\in\mathcal{I}$, the family $\{T_\alpha(t) X_\alpha\}_{\alpha\in\mathfrak{M}}$ is a redundant complete system for $\mathcal{D}(\mathcal{T}_t)$. In particular, for each $\vec x\in\Sigma_t$, the family $\{X_\alpha(\vec x)\}_{\alpha\in\mathfrak{M}}$ contains a (possibly redundant) basis of $V$. The assertions are relatively obvious in the view of the fact, that the restriction maps $i^_t,i^_t\circ\nabla_t:Sol_0(\mathcal{T})\to\mathcal{D}(\mathcal{T}_t)$ are surjective, and hence a spanning system for $Sol_0(\mathcal{T})$ must give a redundant complete system for the Cauchy data $\mathcal{D}(\mathcal{T}_t)\oplus\mathcal{D}(\mathcal{T}_t)$ on $\Sigma_t$. \[2LinIndSolUnique\] Let two equations $\ddot T(t)+F(t)\dot T(t)+G(t)T(t)=0$ and $\ddot T(t)+H(t)\dot T(t)+J(t)T(t)=0$ have two common linearly independent solutions $T(t)$ and $S(t)$. Then by Liouville formula the Wronski determinant $\det W[T,S](t)$ evolves by $$\det W[T,S](t)=\det W[T,S](0)e^{-\int_0^td\tau F(\tau)}=\det W[T,S](0)e^{-\int_0^td\tau H(\tau)},$$ hence $F=H$ and thereby also $G=J$. \[PropVarSep\] The solution space $Sol_0(\mathcal{T})$ admits a separating system if and only if there exists a covering of $\mathcal{T}$ by local trivializations such that the following local conditions are satisfied everywhere (metric $g$ is time-separated): $g_{00}=g_{00}(t)$, the metric component $g_{00}$ depends only on time the expression $\sum_{i,j=1}^3g^{ij}(x)\frac{\partial g_{ij}}{\partial t}(x)$ is a function of time only the connection 1-form $\Gamma$ and Christoffel symbols ${\bf\Gamma}^k_{ij}$ satisfy $$\sum_{i=1}^3g^{ij}[\Gamma_0,\Gamma_i]=0\mbox{, }\forall j>0,$$ $$\sum_{i,j=1}^3g^{ij}\left[\Gamma_0,\frac{\partial\Gamma_j}{\partial x^i}+\Gamma_i\Gamma_j-\sum_{k=0}^3{\bf\Gamma}^k_{ij}\Gamma_k\right]=0,$$ $$\Gamma_0=\Gamma_0(t)\mbox{ is a function of time only}$$ the eigenfunction problem of $D_{\Sigma_t}$ on different $\Sigma_t$ can be adjusted, so that all $\tilde\Sigma_t$ are isomorphic and the eigenfunctions $\zeta^t_\alpha=\zeta_\alpha$ are time-independent. Throughout the section we will work exclusively locally, i.e., in a local trivialization $\pi^{-1}(U)\xrightarrow[]{\Psi}U\times V$, $U\subset M$. Thus we identify the sections in a bundle having a typical fiber $\mathfrak{F}$ with functions in $C^\infty(U;\mathfrak{F})$. We will not keep the flag $U$ in this section but will always understand objects as restricted to $U$. The d’Alambert operator $\Box^\nabla$ on $\mathcal{E}(\mathcal{T})$ has the following local expression in terms of the connection form coefficients $\Gamma_i$ and Christoffel symbols ${\bf\Gamma}^k_{ij}$, $$\begin{aligned} \Box^\nabla=\sum_{i,j=0}^3g^{ij}\left[\frac{\partial^2}{\partial x^i\partial x^j}+2\Gamma_i\frac{\partial}{\partial x^j}-\sum_{k=0}^3{\bf\Gamma}^k_{ij}\frac{\partial}{\partial x^k}+\frac{\partial\Gamma_i}{\partial x^j}+\Gamma_i\Gamma_j-\sum_{k=0}^3{\bf\Gamma}^k_{ij}\Gamma_k\right],\label{dAlambLocal}\end{aligned}$$ and the field operator $D$ locally looks like $$D=\sum_{i,j=0}^3g^{ij}\frac{\partial^2}{\partial x^i\partial x^j}+\sum_{i=0}^3A^i\frac{\partial}{\partial x^i}+B+m^\star,$$ where $A^i,B\in C^\infty(U,End(V))$. To achieve a time separation we need to choose a coordinate atlas such that everywhere $g^{0i}=0$ for $i>0$. Then the operator $D$ locally breaks apart into two differential operators, $D=D_t+D_{\Sigma_t}$, where $$D_t=g^{00}\frac{\partial^2}{\partial t^2}+A^0\frac{\partial}{\partial t}+B^0,$$ and $$D_{\Sigma_t}=\sum_{i,j=1}^3g^{ij}\frac{\partial}{\partial x^i}\frac{\partial}{\partial x^j}+\sum_{i=1}^3A^i\frac{\partial}{\partial x^i}+B^3+m^\star=-\Delta_t+m^\star$$ is the restricted field operator defined earlier. $B^0,B^3\in C^\infty(U,End(V))$ are to be seen explicitly from (Eq.\[dAlambLocal\]). $\Rightarrow$[Necessity:]{} Let $\{T_\alpha X_\alpha\}$ be the separating system system. Then $$\begin{aligned} DT_\alpha(t)X_\alpha(\vec x)=(D_t+D_{\Sigma_t})T_\alpha(t)X_\alpha(\vec x)=\ddot T_\alpha(t)g^{00}(x)X_\alpha(\vec x)+\nonumber\\ +\dot T_\alpha(t)A^0(x)X_\alpha(\vec x)+T_\alpha(t)\left[B^0(x)+D_{\Sigma_t}\right]X_\alpha(\vec x)=0.\label{PreModeEq}\end{aligned}$$ That the metric signature is definite it follows that $g^{00}(x)$ never vanishes. We find a family of second order linear homogeneous differential equations $$\ddot T_\alpha(t)g^{00}(x)X_\alpha^i(\vec x)+\dot T_\alpha(t)\left(A^0(x)X_\alpha(\vec x)\right)^i+T_\alpha(t)\left(\left[B^0(x)+D_{\Sigma_t}\right]X_\alpha(\vec x)\right)^i=0$$ parameterized by the spatial coordinates $\vec x\in\Sigma$ and fiber indices $i=1,...,n$. By definition we similarly have $D\bar T_\alpha(t)X_\alpha(\vec x)=0$. This means that all these equations share at least two linearly independent solutions $T_\alpha$ and $\bar T_\alpha$. If for some $\vec x$ and $i$, $X_\alpha^i(\vec x)=0$, then the existence of two linearly independent solutions for the resulting first order equation means that $$\left(A^0(x)X_\alpha(\vec x)\right)^i=\left(\left[B^0(x)+D_{\Sigma_t}\right]X_\alpha(\vec x)\right)^i=0.$$ Otherwise, by Remark \[2LinIndSolUnique\] we find that there exist functions $F_\alpha,G_\alpha\in C^\infty(\mathcal{I})$ such that $$\left(A^0(x)X_\alpha(\vec x)\right)^i=g^{00}(x)F_\alpha(t)X_\alpha^i(\vec x)\mbox{, }\left(\left[B^0(x)+D_{\Sigma_t}\right]X_\alpha(\vec x)\right)^i=g^{00}(x)G_\alpha(t)X_\alpha^i(\vec x).$$ In both cases we establish that $$\begin{aligned} g_{00}(x)A^0(x)X_\alpha(\vec x)=F_\alpha(t)X_\alpha(\vec x)\label{FDef}\end{aligned}$$ and $$\begin{aligned} g_{00}(x)\left[B^0(x)+D_{\Sigma_t}\right]X_\alpha(\vec x)=G_\alpha(t)X_\alpha(\vec x).\label{GDef}\end{aligned}$$ Thus for each $t\in\mathcal{I}$, $X_\alpha$-s must be nothing else but the joint eigenfunctions of the operators $g_{00}(x)A^0(x)$ and $g_{00}(x)\left[B^0(x)+D_{\Sigma_t}\right]$ corresponding to eigenvalues $F_\alpha(t)$ and $G_\alpha(t)$, respectively. The operator $g_{00}(x)A^0(x)$ is simply a matrix, and at each point $x\in M$ has at most $n$ independent eigenvectors. By Remark \[SepSysSpanRemark\], $X_\alpha(\vec x)$-s span $V$, and thereby $\{X_\alpha\}_{\alpha\in\mathfrak{M}}$ contains bases of all eigenspaces of $g_{00}(x)A^0(x)$. From (Eq.\[dAlambLocal\]) we find $$\begin{aligned} g_{00}A^0=2\Gamma_0-g_{00}\sum_{i,j=0}^3g^{ij}{\bf\Gamma}^0_{ij},\label{A0Simpl}\end{aligned}$$ and $$g_{00}B^0=\frac{\partial}{\partial t}\Gamma_0+\Gamma^2_0-\sum_{k=1}^3{\bf\Gamma}^k_{00}\Gamma_k-g_{00}\sum_{i,j=0}^3g^{ij}{\bf\Gamma}^0_{ij}\Gamma_0.$$ Now turn to the eigenfunction problem (Eq.\[GDef\]). As discussed above, for this problem to be well-posed it is necessary that the differential operator $g_{00}(x)\left[B^0(x)+D_{\Sigma_t}\right]$ is at least formally self-adjoint. But this is possible only if $g_{00}(x)=g_{00}(t)$, thus we have obtained the condition (i). Let us switch to an atlas, where the time function $t$ is redefined such that $g_{00}(t)=1$ (this step is not crucial, but only for convenience). It follows, that $${\bf\Gamma}^k_{00}=0\mbox{, }\forall k>0,$$ so we obtain $$\begin{aligned} A^0=2\Gamma_0-\sum_{i,j=1}^3g^{ij}{\bf\Gamma}^0_{ij},\label{A0Def}\\ B^0=\frac{\partial}{\partial t}\Gamma_0+\Gamma^2_0-\sum_{i,j=1}^3g^{ij}{\bf\Gamma}^0_{ij}\Gamma_0.\label{B0Def}\end{aligned}$$ Combining (Eq.\[FDef\]) and (Eq.\[A0Def\]) we see that $\{X_\alpha\}$-s are the eigenvectors of $\Gamma_0$, and these eigenvectors are independent of $t$. Hence they are also the eigenvectors of $\frac{\partial}{\partial t}\Gamma_0$, and thus by (Eq.\[B0Def\]) $A^0$ and $B^0$ are simultaneously triangularizable, $$B^0X_\alpha(\vec x)=H_\alpha(x)X_\alpha(\vec x),$$ for some $H_\alpha\in C^\infty(M)$. We note that $${\bf\Gamma}^0_{ij}=-\frac{1}{2}\frac{\partial g_{ij}}{\partial t},$$ and denote $$P(x)=-\sum_{i,j=1}^3g^{ij}(x){\bf\Gamma}^0_{ij}(x)=\frac{1}{2}\sum_{i,j=1}^3g^{ij}(x)\frac{\partial g_{ij}}{\partial t}(x).$$ Now (Eq.\[FDef\]) and (Eq.\[GDef\]) tell us, that for each $t\in\mathcal{I}$ the operators $A^0$ and $D_{\Sigma_t}+B^0$ have a common system of eigenfunctions spanning $\mathcal{D}(\mathcal{T}_t)$, and therefore must commute, $$\left[A^0,D_{\Sigma_t}+B^0\right]u=\left[A^0,D_{\Sigma_t}\right]u=0\mbox{, }\forall u\in\mathcal{D}(\mathcal{T}_t).$$ This requires $$A^0(x)=2\Gamma_0(x)+P(x)=A^0(t),$$ and $$\sum_{i=1}^3g^{ij}[\Gamma_0,\Gamma_i]=0\mbox{, }\forall j>0,$$ $$\sum_{i,j=1}^3g^{ij}\left[\Gamma_0,\frac{\partial\Gamma_j}{\partial x^i}+\Gamma_i\Gamma_j-\sum_{k=0}^3{\bf\Gamma}^k_{ij}\Gamma_k\right]=0,$$ exactly as the statement. Similarly, that operators $B^0$ and $B^0+D_{\Sigma_t}$ have the same eigenfunctions implies, that $[B^0,D_{\Sigma_t}]=0$, which on its turn requires $B^0(x)=B^0(t)$, and thereby $P(x)=P(t)$ and $\Gamma_0(x)=\Gamma_0(t)$. Thus we have proven parts (ii) and (iii) of the statement. It follows further, that $H_\alpha(x)=H_\alpha(t)$, and thus the eigenfunction problem (Eq.\[GDef\]) becomes $$D_{\Sigma_t} X_\alpha(\vec x)=(G_\alpha(t)-H_\alpha(t))X_\alpha(\vec x).$$ This is exactly the eigenfunction problem of $D_{\Sigma_t}$, whence we conclude, that necessarily $$\{X_\alpha\}_{\alpha\in\mathfrak{M}}\subset\{\zeta^t_\lambda\}_{\lambda\in\mathbb{R}}.$$ Therefore $$G_\alpha(t)=H_\alpha(t)+\lambda_\alpha(t),$$ where $$\lambda_\alpha(t)=\{\lambda\in\mathbb{R}\mbox{: }X_\alpha\in\mathcal{D}(\mathcal{T}_t)_\lambda'\}.$$ Now (Eq.\[PreModeEq\]) becomes $$\begin{aligned} \ddot T_\alpha(t)+F_\alpha(t)\dot T_\alpha(t)+G_\alpha(t)T_\alpha(t)=0,\label{ModEq}\end{aligned}$$ which is the mode equation for the mode $T_\alpha$. We have two spanning systems for $\mathcal{D}(\mathcal{T}_t)$: $\{X_\alpha\}_{\alpha\in\mathfrak{M}}$ and $\{\zeta^t_\alpha\}_{\alpha\in\tilde\Sigma_t}$, and hence in each eigenspace $\mathcal{D}(\mathcal{T}_t)_\lambda'$ we can choose a basis from $\{X_\alpha\}_{\alpha\in\mathfrak{M}}$. Thus a complete eigenfunction system can be chosen among $\{X_\alpha\}_{\alpha\in\mathfrak{M}}$, proving the (iv) statement of the proposition. We are complete with the necessity. $\Leftarrow$[Sufficiency: ]{} Suppose all the points of the statement are satisfied. Then, as we have seen above, by (iii) $A^0$ and $B^0$ are functions of $t$ having the same eigenvectors, and moreover, commute with $\Delta_t$. It follows that the actions of $A^0$ and $B^0$ preserve $\mathcal{D}(\mathcal{T}_t)_\lambda'$, and thus by a Gramm-Schmidt operation the representatives $\zeta_\alpha$ can be chosen such that they are eigenfunctions of $A^0$ and $B^0$. Thus each $\tilde\Sigma_\lambda$, and thereby the entire $\tilde\Sigma$, decomposes into $n$ components corresponding to the eigendirections of $A^0$, $$\tilde\Sigma=\bigcup_{i=1}^n\tilde\Sigma^i.$$ For spatially homogeneous spacetimes discussed in later sections we will give a more conceptual justification of such a subdivision in terms of the representation theory. Let for each $\alpha\in\tilde\Sigma$ choose a mode solution $T_\alpha$ of (Eq.\[ModEq\]) arbitrarily (strictly speaking, not completely arbitrarily, but such that $T_\alpha$ and $\bar T_\alpha$ are linearly independent) and consider the union of two systems $$\{u,v\}_{\alpha\in\tilde\Sigma}\doteq\{u_\alpha\}_{\alpha\in\tilde\Sigma}\cup\{v_\alpha\}_{\alpha\in\tilde\Sigma}\mbox{, }u_\alpha=T_\alpha\zeta_\alpha\mbox{, }v_\alpha=\bar T_\alpha\zeta_\alpha.$$ Choose any $\phi\in Sol_0(\mathcal{T})$. Then for each $t\in\mathcal{I}$ the restriction $i_t^(\phi)[\vec x]=\phi(t,\vec x)\in\mathcal{D}(\mathcal{T}_t)$ can be Fourier expanded as $$\begin{aligned} \phi(t,\vec x)=\int_{\tilde\Sigma}d\mu(\alpha)\hat\phi(\alpha;t)\zeta_\alpha(\vec x)\label{TimeDepFourier}\end{aligned}$$ with the integral converging in $L^2(\tilde\Sigma,\mu)$. Hence we can differentiate under the integral, $$D\phi(t,\vec x)=\int_{\tilde\Sigma}d\mu(\alpha)D\left[\hat\phi(\alpha;t)\zeta_\alpha(\vec x)\right]=\int_{\tilde\Sigma}d\mu(\alpha)\left[\ddot{\hat{\phi}}(\alpha;t)+\right.$$ $$\left.+F_\alpha(t)\dot{\hat{\phi}}(\alpha;t)+G_\alpha(t)\hat\phi(\alpha;t)\right]\zeta_\alpha(\vec x)=0,$$ where for convenience we again reparameterized $t$ to get $g_{00}=1$. Thus $\hat\phi(\alpha;t)$ is a solution of the mode equation. All solutions of the ordinary second order equation (Eq.\[ModEq\]) are smooth and comprise a two complex dimensional space, $$\hat\phi(\alpha;t)=a^\phi_\alpha T_\alpha(t)+b^\phi_\alpha\bar T_\alpha(t)\mbox{, }a^\phi_\alpha,b^\phi_\alpha\in\mathbb{C}.$$ Inserting this into (Eq.\[TimeDepFourier\]) we finally arrive at $$\phi(t,\vec x)=\int_{\tilde\Sigma}d\mu(\alpha)\left[a^\phi_\alpha T_\alpha(t)\zeta_\alpha(\vec x)+b^\phi_\alpha\bar T_\alpha(t)\zeta_\alpha(\vec x)\right],$$ which exactly means, that $\{u,v\}_{\alpha\in\tilde\Sigma}$ is a separating system for $Sol_0(\mathcal{T})$. (For compatibility with the definition one can concatenate $u_\alpha$ and $v_\alpha$ to a single function on the disjoint union $\tilde\Sigma\sqcup\tilde\Sigma$.) The assertion of this proposition can be interpreted as follows. If a mode decomposition in a reasonable fashion exists for $Sol_0(\mathcal{T})$ then it is basically the mode decomposition given by the time dependent Fourier transform which we will define a few paragraphs later. As a supplement to the proposition we make a few remarks. Let $\hat g_{ij}=g(\partial_i,\partial_j)$ and $\hat h_{ij}=h(\partial_i,\partial_j)$ be the matrices of the metrics $g$ and $h$, correspondingly, in a local chart, and $\sigma_k(\hat h)$ the eigenvalues of the symmetric matrix $\hat h$. \[RemarkCondii\] The condition (ii) of Proposition \[PropVarSep\] is equivalent to $$\det\hat g(x)=-g_{00}(t)\det\hat h(x)=-g_{00}(t)\sigma_1(\hat h)\sigma_2(\hat h)\sigma_3(\hat h)=-g_{00}(t)e^{2\int_0^tdt'P(t')}\det\hat h_0(\vec x),$$ where $\det\hat h_0(\vec x)\in C^\infty(\Sigma)$ is a positive smooth function (the notation will become clear later). The assertion follows from the combination of condition (ii) with the Laplace’s formula, $$\frac{\partial}{\partial t}\det\hat g=\det\hat g\cdot Tr[\hat g^{-1}\frac{\partial}{\partial t}\hat g].$$ Because $\nabla$ is a metric connection, the restrictions of the previous proposition imply restrictions on the fiber metric $\langle,\rangle_\mathfrak{g}$. In case of a tensor bundle of rank $(m,n)$ with Levi-Civita connection, coefficients $\Gamma_i$ are expressed in Christoffel symbols and the fiber metric is induced by the spacetime metric, thus the restrictions fall onto the spacetime $(M,g)$. \[CorMetricSep\] Let a local moving frame be chosen, such that the metric $\langle,\rangle_\mathfrak{g}$ is represented by the matrix $\hat{\mathfrak{g}}$. Conditions (iii) of Proposition \[PropVarSep\] imply the following restrictions on $\hat{\mathfrak{g}}$: $$\hat{\mathfrak{g}}(x)=\hat{\mathfrak{B}}^T(t)\hat{\mathfrak{g}}^0(\vec x)\hat{\mathfrak{B}}(t),$$ where $\hat{\mathfrak{g}}^0$ and $\hat{\mathfrak{B}}$ are matrix valued smooth functions. In particular, for the a tensor bundle of rank $(m,n)$ to allow for seprataion it is necessary that the spacetime metric be represented by a matrix $$\hat g=1\oplus\left(-\hat h_0(\vec x)\hat B(t)\right),$$ where $\hat h_0$ and $\hat B$ are matrix valued smooth functions. Locally the conservation of the metric $\nabla\langle,\rangle_\mathfrak{g}=0$ can be written as $$\frac{\partial}{\partial x^i}\hat{\mathfrak{g}}-\Gamma^T_i\hat{\mathfrak{g}}-\hat{\mathfrak{g}}\Gamma_i=0,$$ where $\Gamma_i$ are the matrices of the connection coefficients in the chosen frame. In particular, for $i=0$ we have $$\frac{\partial}{\partial t}\hat{\mathfrak{g}}(x)-\Gamma^T_0(t)\hat{\mathfrak{g}}(x)-\hat{\mathfrak{g}}(x)\Gamma_0(t)=0,$$ where $\Gamma_0=\Gamma_0(t)$ was used. The solutions of this equation are of the form $$\hat{\mathfrak{g}}(x)=\hat{\mathfrak{B}}^T(t)\hat{\mathfrak{g}}^0(\vec x)\hat{\mathfrak{B}}(t),$$ where $$\begin{aligned} \hat{\mathfrak{B}}(t)=e^{\int_0^tdt'\Gamma_0(t')},\label{BDef}\end{aligned}$$ and $\hat{\mathfrak{g}}^0(\vec x)$ is a smooth symmetric matrix field on $\Sigma$. Now if we identify the tensor space $(T_pM)^m_n$ with an $4^{n+m}$ dimensional vector space $V$ using a suitable bases, then each matrix $\Gamma_i$ will be a $4^{n+m-1}\times 4^{n+m-1}$ matrix of blocks, with blocks being the Christoffel symbols ${\bf\Gamma}_i$ for contravariant indices and $-{\bf\Gamma}^T_i$ for covariant indices. $\Gamma_0=\Gamma_0(t)$ means ${\bf\Gamma}_0={\bf\Gamma}_0(t)$. With our time-separated metric we have $$\hat g=1\oplus-\hat h.$$ One can find $${\bf\Gamma}_0=0\oplus\left(\frac{1}{2}\hat h^{-1}\frac{\partial\hat h}{\partial t}\right)={\bf\Gamma}_0(t)=0\oplus\hat A(t)$$ for some smooth $3\times3$ matrix $\hat A(t)$. The solution is $$\hat h(x)=\hat h_0(\vec x)e^{2\int_0^tdt'\hat A(t')}=\hat h_0(\vec x)\hat B(t),$$ for smooth symmetric commuting matrix fields $\hat h_0(\vec x)$ and $\hat B(t)$. Now the notation $\det\hat h_0$ of Remark \[RemarkCondii\] becomes clear, and we see that $$\det\hat B(t)=e^{2\int_0^tdt'P(t')}$$ for a tensor bundle. Note that for the scalar field conditions (iii) are trivially satisfied and do not restrict the spacetime. \[dmu\_h\] For the volume form measure $d\mu_h$ on $\Sigma_t$ we have locally $$d\mu_h(\vec x)=\sqrt{\det\hat{h}(t,\vec x)}dx^1dx^2dx^3.$$ By Remark \[RemarkCondii\] we have $$\det\hat h(t,\vec x)=e^{2\int_0^tdt'P(t')}\det\hat h_0(\vec x),$$ hence $$d\mu_h(\vec x)=e^{\int_0^tdt'P(t')}d\mu_{h_0}(\vec x),$$ where $$d\mu_{h_0}(\vec x)=\sqrt{\det\hat{h_0}(\vec x)}dx^1dx^2dx^3.$$ Henceforth by stating that a mode decomposition of $Sol_0(\mathcal{T})$ exists we will mean that the assumptions of the Proposition \[PropVarSep\] are satisfied and the corresponding covering is chosen. We are ready to formulate precisely the time dependent Fourier transform. Note that although $\zeta_\alpha$ are $t$-independent, the spatial metric $h$ and the fiber metric $\langle,\rangle_\mathfrak{g}$ depend on $t$, and $\zeta_\alpha$ are not orthonormal with respect to the measure $d\mu_h$ for all $t$ simultaneously. At this point we appoint once and forever to normalize $\zeta_\alpha$ such that they are orthonormal at $t=0$. Or equivalently, they are orthonormal with respect to the measure $d\mu_{h_0}$ of Remark \[dmu\_h\] and the fiber metric $\mathfrak{g}^0$ of Corollary \[CorMetricSep\]. For $f\in\mathcal{D}(\mathcal{T})$ we define the time dependent Fourier transform $\tilde f(t,\alpha)=\mathcal{F}[f(t,.)](\alpha)\in C_0^\infty\left(\mathcal{I},\tilde{\mathcal{D}}(\tilde\Sigma)\right)$ by $$\mathcal{F}[f(t,.)](\alpha)=\int_{\Sigma_t}d\mu_{h_0}(\vec x)\langle\bar\zeta_\alpha(\vec x),f(t,\vec x)\rangle_{\mathfrak{g}^0}.$$ Here we note another important corollary, which will be useful later. It will give the time dependent Plancherel formula. \[TDPlancherel\] Suppose the assumptions of Proposition \[PropVarSep\] are satisfied, and the corresponding covering is chosen. Then for all $f\in\mathcal{D}(\mathcal{T})$ $$(\zeta_\alpha,f(t,.))_{\Sigma_t}=I_\alpha(t)\mathcal{F}[f(t,.)](\alpha)$$ and the time dependent Plancherel formula for the time-dependent Fourier transform is given by $$(f(t,.),h(t,.))_{\Sigma_t}=\int_{\tilde\Sigma}d\mu(\alpha)I_\alpha(t)\overline{\mathcal{F}[f(t,.)]}(\alpha)\mathcal{F}[h(t,.)](\alpha),$$ where $$I_\alpha(t)=e^{\int_0^tdt'F_\alpha(t')}.$$ By Remark \[dmu\_h\] $$(f(t,.),h(t,.))_{\Sigma_t}=\int_{\Sigma_t}d\mu_h(\vec x)(f(t,.),h(t,.))_\mathfrak{g}=$$ $$=e^{\int_0^tdt'P(t')}\int_{\Sigma_t}d\mu_{h_0}(\vec x)(f(t,.),h(t,.))_\mathfrak{g}.$$ At the same time by Corollary \[CorMetricSep\] we have $$\begin{aligned} (f(t,.),h(t,.))_\mathfrak{g}=(\hat{\mathfrak{B}}(t)f(t,.),\hat{\mathfrak{B}}(t)h(t,.))_{\mathfrak{g}^0}.\label{FiberProdt}\end{aligned}$$ Because we have normalized $\zeta_\alpha$ with respect to $d\mu_{h_0}$ and $\mathfrak{g}^0$, the conventional Plancherel formula holds for them, $$\int_{\Sigma_t}d\mu_{h_0}(\vec x)(f(t,.),h(t,.))_{\mathfrak{g}^0}=\int_{\tilde\Sigma}d\mu(\alpha)\overline{\mathcal{F}[f(t,.)](\alpha)}\mathcal{F}[h(t,.)](\alpha).$$ Combining these three formulas we find $$(f(t,.),h(t,.))_{\Sigma_t}=e^{\int_0^tdt'P(t')}\int_{\tilde\Sigma}d\mu(\alpha)\overline{\mathcal{F}[\hat{\mathfrak{B}}(t)f(t,.)](\alpha)}\mathcal{F}[\hat{\mathfrak{B}}(t)f(t,.)](\alpha).$$ Meanwhile $$(\zeta_\alpha,f(t,.))_{\Sigma_t}=e^{\int_0^tdt'P(t')}\int_{\Sigma_t}d\mu_{h_0}(\vec x)(\hat{\mathfrak{B}}(t)\zeta_\alpha,\hat{\mathfrak{B}}(t)h(t,.))_{\mathfrak{g}^0}.$$ By definition $$\hat{\mathfrak{B}}(t)\zeta_\alpha=e^{\int_0^tdt'\Gamma_0(t')}\zeta_\alpha=e^{\frac{1}{2}\int_0^tdt'[A^0(t')-P(t')]}\zeta_\alpha=e^{\frac{1}{2}\int_0^tdt'[F_\alpha(t')-P(t')]}\zeta_\alpha,$$ whence $$(\zeta_\alpha,f(t,.))_{\Sigma_t}=e^{\frac{1}{2}\int_0^tdt'[F_\alpha(t')+P(t')]}\mathcal{F}[\hat{\mathfrak{B}}(t)f(t,.)](\alpha).$$ Finally $$\hat{\mathfrak{B}}(t)f(t,\vec x)=\int_{\tilde\Sigma}d\mu(\alpha)\mathcal{F}[f(t,.)](\alpha)\hat{\mathfrak{B}}(t)\zeta_\alpha(\vec x)=\int_{\tilde\Sigma}d\mu(\alpha)\mathcal{F}[f(t,.)](\alpha)e^{\frac{1}{2}\int_0^tdt'[F_\alpha(t')-P(t')]}\zeta_\alpha,$$ thus $$\mathcal{F}[\hat{\mathfrak{B}}(t)f(t,.)](\alpha)=e^{\frac{1}{2}\int_0^tdt'[F_\alpha(t')-P(t')]}\mathcal{F}[f(t,.)](\alpha).$$ The assertions now easily follow. At last we compute the spectra of operators $A^0$ and $B^0$ for the tensor bundle to find the functions $F_\alpha$ and $H_\alpha$. In view of Corollary \[CorMetricSep\] the function $P(t)$ becomes $$P(t)=\frac{1}{2}Tr\left[\hat B^{-1}(t)\frac{\partial\hat B}{\partial t}(t)\right].$$ Then $$Spec\{A^0\}=2Spec\{\Gamma_0\}+P(t),$$ $$Spec\{B^0\}=\left\{\dot \sigma(t)+\sigma^2(t)+\sigma(t)P(t)\mbox{: }\sigma\in Spec\{\Gamma_0\}\right\}.$$ As a useful example we calculate these spectra for the scalar and 1-form fields on uniformly expanding (e.g., FRW) manifolds, $$ds^2=dt^2-a^2(t)d\sigma^2(\vec x).$$ Here the matrix $\hat B(t)=a^2(t)1$, and hence $$\hat A(t)=\frac{\partial}{\partial t}\ln a(t)1=H(t)1,$$ and $$P(t)=3H(t)\mbox{, }H(t)=\frac{\dot a(t)}{a(t)}.$$ For scalar case $n=m=0$ and we have $$Spec\{\Gamma_0\}=\{0\},$$ thus $$Spec\{A^0\}=\{3H(t)\}\mbox{, }Spec\{B^0\}=\{0\},$$ as well known. For the 1-form case, $m=0$, $n=1$, we have $$Spec\{\Gamma_0\}=\{0, -H(t)\},$$ and thereby $$Spec\{A^0\}=\{3H(t),H(t)\}\mbox{, }Spec\{B^0\}=\{0,-\dot H(t)-2H^2(t)\},$$ where the first members are similar to the scalar case and represent the scalar modes, but second ones represent the transversal and longitudinal modes. As we have seen, for the separation it is necessary that the evolution of the metric be represented by linear transformations. If the connection also satisfies such a condition in a suitable sense, than the operator $D_{\Sigma_t}$ is essentially the same at every $t$ up to some scale factors. (Maybe the condition (iii) of the main proposition already implies such a restriction on the connection, but we are not sure yet.) This will be the case for all our bundles of interest, and it will provide analytical advantages. To summarize what we expect precisely we give the following definitions. We will say that the operator $D_{\Sigma_t}$ has a [strictly uniform]{} spectrum over time if there exists a lower semi-bounded function $\omega(\alpha)$ on $\tilde\Sigma$, a positive smooth function $C(t)>0$ and a smooth function $\tilde m^\star(t)$ such that $\lambda_\alpha(t)=\omega(\alpha)C(t)+\tilde m^\star(t)$, or equivalently, the expression $$\frac{d}{dt}\ln\|\lambda_\alpha(t)-\tilde m^\star(t)\|$$ does not depend on $\alpha$. This is a rather strong condition. It basically requires that the eigenspaces of $D_{\Sigma_t}$ coincide for different $t$ up to an overall shift, and that eigenvalues be linearly proportional. Such a property would be very comfortable, but it does not hold for some models of our interest. In particular, it does not hold for the Bianchi I model with distortions. Hence we will derive some of our results under a milder restriction which holds at least in all cosmological situations where the spectral theory is explicit enough so that it can be checked. (The explicit spectral theory of the scalar field on Bianchi I-VII spacetimes will appear in a subsequent publication.) We will say that the operator $D_{\Sigma_t}$ has a [loosely uniform]{} spectrum over time if $$\left\|\frac{d}{dt}\ln\|\lambda_\alpha(t)-\tilde m^\star(t)\|\right\|\le C_\mathcal{R}\mbox{, }\forall t\in\mathcal{R}, \alpha\in\tilde\Sigma,$$ for any compact interval $\mathcal{R}\subset\mathcal{I}$, and for some $0<C_\mathcal{R}\in\mathbb{R}$ and a smooth function $\tilde m^\star(t)$. If the Fourier transform is conventional, then it will be natural to require that $\omega$ be an analytic function on $\tilde\Sigma$. Some properties of the mode solutions ------------------------------------- In this section we investigate the equation (Eq.\[ModEq\]) and obtain some useful properties of the mode solutions $T_\alpha$. The mode equation is $$\ddot T_\alpha(t)+F_\alpha(t)\dot T_\alpha(t)+G_\alpha(t)T_\alpha(t)=0,$$ where $$G_\alpha(t)=H_\alpha(t)+\lambda_\alpha(t),$$ and $\lambda_\alpha(t)$ are the eigenvalues of the operator $D_{\Sigma_t}=-\Delta_t+m^\star(x)$. Note that $G_\alpha$ may become null or negative for some rates of expansion. This corresponds to the so-called positive back-reaction in a linear system and results in exponential solutions. This is an interesting phenomenon appearing in non scalar fields (for scalar fields $H_\alpha=0$), and its significance is not yet completely clear to us. To understand it one could, for instance, track its influence on the energy-momentum tensor etc. It is not obvious that this is really a physical infrared instability, because it may occur for the co-vector field but not for the vector counterpart, for instance. It is also worth mentioning, that for the co-vector (1-form) field the introduction of a conformal coupling precisely cancels this instability. It seems plausible that for each field there is a choice of the coupling constant which compensates this bad infrared behavior. We say infrared, because $\lambda_\alpha(t)$ attains arbitrarily large positive values at any $t$, thus on an unbounded subbundle of $R\times\tilde\Sigma$, $G_\alpha$ is positive. We will make this more explicit under the assumption, that $D_{\Sigma_t}$ has a strictly uniform spectrum. Then the function $C(t)$ is uniformly bounded from below, and the functions $H_\alpha$ and $\tilde m^\star$ are uniformly bounded from above on any compact interval $\mathcal{R}$. On the other hand $\omega\to+\infty$, hence is suffices to choose $\omega$ large enough to make $G_\alpha=H_\alpha+C\omega+\tilde m^\star>0$. Fix a component $\tilde\Sigma^i$ and write $H=H_\alpha$, $F=F_\alpha$ and $I=I_\alpha$ for all $\alpha\in\tilde\Sigma^i$. Define a new variable $$s(t)=\int_0^td\tau e^{-\int_0^\tau d\tau'F(\tau')}=\int_0^td\tau I^{-1}(\tau),$$ which is in a smooth monotone bijective correspondence with $t$. The inverse function will be denoted by $t(s)$. Regarding all the acting functions of $t$ as functions of $s$ we obtain $$\begin{aligned} \ddot T_\alpha(s)+\Lambda_\alpha(s)T_\alpha(s)=0,\label{ModEqs}\end{aligned}$$ where $$\Lambda_\alpha(s)=\left[G_\alpha(t)e^{2\int_0^td\tau F(\tau)}\right]_{t=t(s)}=G_\alpha(s)I^2(s).$$ This is a time dependent harmonic oscillator equation, to which the results in the appendix apply. Note that the Wronski determinant of two solutions $Q,R$ $$\det W[Q,R](s)=Q(s)\dot R(s)-\dot Q(s)R(s)=\mbox{const}$$ in variable $t$ becomes $$\det W[Q,R](t)=\frac{dt}{ds}\left(Q(t)\dot R(t)-\dot Q(t)R(t)\right)=I(t)\left(Q(t)\dot R(t)-\dot Q(t)R(t)\right)=\mbox{const}.$$ Applying Corollary \[TEstAbstract\] to (Eq.\[ModEqs\]) for different $\alpha$ we find estimates which in principle depend on $\alpha$ in a complicated way. But under the assumption of loose uniformity on $D_{\Sigma_t}$ we will be able to invoke more comfortable expressions. \[TEstLooseUni\] Suppose $D_{\Sigma_t}$ has a loosely uniform spectrum over time. Then for a family of arbitrary solutions $T_\alpha$ of (Eq.\[ModEqs\]) the following estimate holds $$\|T_\alpha(s)\|\le R_\mathcal{R}\|T_\alpha(0)\|+\frac{S_\mathcal{R}}{\max\{1,\sqrt{U_\mathcal{R}+T_\mathcal{R}\lambda_\alpha(0)}\}}\|\dot T_\alpha(0)\|\mbox{, }\forall s\in\mathcal{R},$$ with $0<R_\mathcal{R},S_\mathcal{R},T_\mathcal{R}\in\mathbb{R}$ and $U_\mathcal{R}\in\mathbb{R}$, for any compact interval $\mathcal{R}$. Fix a compact interval $\mathcal{R}$ and for each $\alpha\in\tilde\Sigma^i$ apply Corollary \[TEstAbstract\] with $\Lambda_\alpha(s)=I^2(s)H(s)+I^2(s)\lambda_\alpha(s)$. Because $\Lambda_\alpha$ is real, we get $A_\mathcal{R}(\alpha)=0$. As $\lambda_\alpha(s)$ is lower semi-bounded we have $$p_\mathcal{R}\doteq\inf_{\tilde\Sigma}\inf_\mathcal{R}\lambda_\alpha>-\infty.$$ Denote $m_\mathcal{R}=\inf_\mathcal{R}\{I^2H\}$ and $n_\mathcal{R}=\inf_\mathcal{R}\{I^2\}>0$. Then $$c_\mathcal{R}(\alpha)\ge m_\mathcal{R}+n_\mathcal{R}\inf_\mathcal{R}\lambda_\alpha\ge m_\mathcal{R}+n_\mathcal{R}p_\mathcal{R}.$$ It follows that $\kappa(\alpha)\le\sqrt{1+\|m_\mathcal{R}+n_\mathcal{R}p_\mathcal{R}\|}$ and $e_\mathcal{R}(\alpha)\ge1+\max\{0,m_\mathcal{R}+n_\mathcal{R}\inf_\mathcal{R}\lambda_\alpha\}$. Denote $M_\mathcal{R}=\sup_\mathcal{R}\{\|I^2H\|\}\ge0$ and $N_\mathcal{R}=\sup_\mathcal{R}\{I^2\}>0$. We find next $$D_\mathcal{R}(\alpha)\le1+\|m_\mathcal{R}+n_\mathcal{R}p_\mathcal{R}\|+M_\mathcal{R}+N_\mathcal{R}\left\|\sup_\mathcal{R}\lambda_\alpha\right\|.$$ Now we observe that by loose uniformity $$\left\|\ln\frac{\|\lambda_\alpha(s)-\tilde m^\star(s)\|}{\|\lambda_\alpha(s')-\tilde m^\star(s')\|}\right\|=\left\|\int_s^{s'}d\sigma\partial_s\ln\|\lambda_\alpha(\sigma)-\tilde m^\star(\sigma)\|\right\|\le\|\mathcal{R}\|\sqrt{N_\mathcal{R}}C_\mathcal{R},$$ hence $$\begin{aligned} \sup_\mathcal{R}\|\lambda_\alpha-\tilde m^\star\|\le\|\lambda_\alpha(0)-\tilde m^\star(0)\|e^{\|\mathcal{R}\|\sqrt{N_\mathcal{R}}C_\mathcal{R}}\nonumber\\ \inf_\mathcal{R}\|\lambda_\alpha-\tilde m^\star\|\ge\|\lambda_\alpha(0)-\tilde m^\star(0)\|e^{-\|\mathcal{R}\|\sqrt{N_\mathcal{R}}C_\mathcal{R}}.\label{lambda_msupinf}\end{aligned}$$ Note that whenever $\lambda_\alpha(0)-\tilde m^\star(0)>0$ then it follows by continuity that $\lambda_\alpha(s)-\tilde m^\star(s)>0$ for all $s\in\mathcal{R}$. Denote $$\lambda_{min}=\tilde m^\star(0)-\min\{0,\inf_\mathcal{R}\tilde m^\star\cdot e^{\|\mathcal{R}\|\sqrt{N_\mathcal{R}}C_\mathcal{R}}\}-\min\{0,\frac{m_\mathcal{R}}{n_\mathcal{R}}e^{\|\mathcal{R}\|\sqrt{N_\mathcal{R}}C_\mathcal{R}}\}.$$ Then from $\lambda_\alpha(0)>\lambda_{min}$ it follows $\lambda_\alpha(0)-\tilde m^\star(0)>0$, $m_\mathcal{R}+n_\mathcal{R}\inf_\mathcal{R}\lambda_\alpha>0$ and $$\inf_\mathcal{R}\lambda_\alpha\ge\inf_\mathcal{R}\tilde m^\star+(\lambda_\alpha(0)-\tilde m^\star(0))e^{-\|\mathcal{R}\|\sqrt{N_\mathcal{R}}C_\mathcal{R}}>0.$$ Now we have that $$e_\mathcal{R}(\alpha)\ge1+\chi[\lambda_\alpha(0)>\lambda_{min}]\left(m_\mathcal{R}+n_\mathcal{R}(\inf_\mathcal{R}\tilde m^\star+(\lambda_\alpha(0)-\tilde m^\star(0))e^{-\|\mathcal{R}\|\sqrt{N_\mathcal{R}}C_\mathcal{R}})\right),$$ where the characteristic function $\chi$ plays here the role of the condition checking. From (Eq.\[lambda\_msupinf\]) we find $$\sup_\mathcal{R}\|\lambda_\alpha\|\le\sup_\mathcal{R}\tilde m^\star+\|\lambda_\alpha(0)-\tilde m^\star(0)\|e^{\|\mathcal{R}\|\sqrt{N_\mathcal{R}}C_\mathcal{R}},$$ whence $$D_\mathcal{R}(\alpha)\le1+\|m_\mathcal{R}+n_\mathcal{R}p_\mathcal{R}\|+M_\mathcal{R}+N_\mathcal{R}(\sup_\mathcal{R}\tilde m^\star+\|\lambda_\alpha(0)-\tilde m^\star(0)\|e^{\|\mathcal{R}\|\sqrt{N_\mathcal{R}}C_\mathcal{R}}).$$ Thus we establish that for $\lambda_\alpha(0)\le\lambda_{min}$ $$\frac{D_\mathcal{R}(\alpha)}{e_\mathcal{R}(\alpha)}\le1+\|m_\mathcal{R}+n_\mathcal{R}p_\mathcal{R}\|+M_\mathcal{R}+N_\mathcal{R}(\sup_\mathcal{R}\tilde m^\star+(\|\lambda_{min}\|+\|\tilde m^\star(0)\|)e^{\|\mathcal{R}\|\sqrt{N_\mathcal{R}}C_\mathcal{R}}),$$ and for $\lambda_\alpha(0)>\lambda_{min}$ $$\frac{D_\mathcal{R}(\alpha)}{e_\mathcal{R}(\alpha)}\le e^{2\|\mathcal{R}\|\sqrt{N_\mathcal{R}}C_\mathcal{R}}N_\mathcal{R}\left(\frac{1}{n_\mathcal{R}}+\frac{1+\|m_\mathcal{R}+n_\mathcal{R}p_\mathcal{R}\|+M_\mathcal{R}+N_\mathcal{R}\|\sup_\mathcal{R}\tilde m^\star\|}{N_\mathcal{R}e^{\|\mathcal{R}\|\sqrt{N_\mathcal{R}}C_\mathcal{R}}}+\right.$$ $$\left.+\frac{1+\|m_\mathcal{R}\|+n_\mathcal{R}\|\inf_\mathcal{R}\tilde m^\star\|}{n_\mathcal{R}e^{-\|\mathcal{R}\|\sqrt{N_\mathcal{R}}C_\mathcal{R}}}\right).$$ Finally $$\frac{d}{ds}\ln(\kappa^2+\Lambda_\alpha)=\frac{\frac{d}{ds}(I^2(H+\tilde m^\star))+\frac{d}{ds}(I^2)(\lambda_\alpha-\tilde m^\star)+I^2(\lambda_\alpha-\tilde m^\star)\frac{d}{ds}\ln\|\lambda_\alpha-\tilde m^\star\|}{\kappa^2+I^2H+I^2\lambda_\alpha}.$$ Denote $P_\mathcal{R}=\sup_\mathcal{R}\|\partial_s(I^2(H-\tilde m^\star))\|\ge0$ and $Q_\mathcal{R}=\sup_\mathcal{R}\|\partial_s(I^2)\|\ge 0$. Again using the loose uniformity, for $\lambda_\alpha-\tilde m^\star\le1$ $$\left\|\frac{d}{ds}\ln(\kappa^2+\Lambda_\alpha)\right\|\le P_\mathcal{R}+Q_\mathcal{R}+(N_\mathcal{R})^\frac{3}{2}C_\mathcal{R},$$ and else $$\left\|\frac{d}{ds}\ln(\kappa^2+\Lambda_\alpha)\right\|\le\frac{P_\mathcal{R}+Q_\mathcal{R}+(N_\mathcal{R})^\frac{3}{2}C_\mathcal{R}}{n_\mathcal{R}}.$$ Summarizing this all we find that by Corollary \[TEstAbstract\] there exist numbers $0<R_\mathcal{R},S_\mathcal{R},T_\mathcal{R}\in\mathbb{R}$ and $U_\mathcal{R}\in\mathbb{R}$ such that for a family of arbitrary solutions $T_\alpha$ we have $$\|T_\alpha(s)\|\le R_\mathcal{R}\|T_\alpha(0)\|+\frac{S_\mathcal{R}}{\max\{1,\sqrt{U_\mathcal{R}+T_\mathcal{R}\lambda_\alpha(0)}\}}\|\dot T_\alpha(0)\|,$$ what was to be proven. The result can be strengthened under additional assumptions. These are perhaps too restrictive, but they appear to be sufficient for some important applications. Let $\mathbb{H}_a=\{z\in\mathbb{C}: \|\Im z\|<a\}$. \[TEstHol\] Suppose the bundle $\mathcal{T}$ is analytic, so that all functions figuring in (Eq.\[ModEqs\]) are real analytic functions of $s$. Suppose further that $D_{\Sigma_t}$ has a strictly uniform spectrum. Choose the initial data to be $T_\alpha(0)=p(\omega(\alpha))$ and $\dot T_\alpha(0)=q(\omega(\alpha))$, where $p(\omega)$, $q(\omega)$ are holomorphic functions on $\mathbb{H}_a$ for some $a>0$. Then for each $s$, $T_\alpha(s)=r_s(\omega(\alpha))$, where $r_s(\omega)$ is holomorphic in $\omega$ on $\mathbb{H}_a$ and real analytic in $s$, and for any compact interval $\mathcal{R}$ it holds $$\|r_s(\omega)\|\le R_\mathcal{R}\|p(\omega)\|+\frac{S_\mathcal{R}}{\max\{1,\sqrt{U_\mathcal{R}+T_\mathcal{R}\Re\omega}\}}\|q(\omega)\|\mbox{, }\forall s\in\mathcal{R},$$ with $0<R_\mathcal{R},S_\mathcal{R},T_\mathcal{R}\in\mathbb{R}$ and $U_\mathcal{R}\in\mathbb{R}$. By strict uniformity we have $\lambda_\alpha(t)=\omega(\alpha)C(t)+\tilde m^\star(t)$, and if the initial data depend only on $\omega$, then the solutions will also be such. Therefore for convenience we write $$\begin{aligned} \ddot T_\omega(s)+I^2(s)(H(s)+\omega C(s)+\tilde m^\star(s))T_\omega(s)=0\label{ModEqHol}\end{aligned}$$ with $T_\omega(0)=p(\omega)$ and $\dot T_\omega(0)=q(\omega)$. From the theory of power series it is clear that any real analytic function on $s(\mathcal{I})$ can be extended to a holomorphic function in some open neighborhood $\delta(s(\mathcal{I}))$ of $s(\mathcal{I})$. Consider (Eq.\[ModEqHol\]) as a complex differential equation, then for any $\omega\in\mathbb{H}_a$, by Satz 4.1 and Satz 4.2 of [@Herold1975] there exist neighborhoods $\delta(0)$ of 0 and $\delta(\omega)$ of $\omega$ such that $T_\omega(s)$ is holomorphic in $\delta(0)\times\delta(\omega)$. At the same time by Satz 5.3 of [@Herold1975], for any $\omega\in\mathbb{H}_a$ the solution $T_\omega$ can be analytically continued to the whole of $\delta(s(\mathcal{I}))$. Thus $T_\omega(s)$ is holomorphic in $\delta(\mathbb{R})\times\mathbb{H}_a$. Now restrict back to the real axis and fix the interval $\mathcal{R}$. The reasoning of the previous proposition can be repeated literally except that now $A_\mathcal{R}(\alpha)$ is not zero but equals $A_\mathcal{R}(\omega)=\|\Im\omega\|\sup_\mathcal{R}\{I^2C\}<a\sup_\mathcal{R}\{I^2C\}$. This results in a similar formula as in Proposition \[TEstLooseUni\] with perhaps different coefficients, and that proves our assertion. We have an immediate corollary. \[TPolBound\] Under the assumptions of Proposition \[TEstHol\], if $p,q\in\mathcal{A}(\mathbb{H}_a)$ then for each $s\in\mathcal{R}$, $r_s\in\mathcal{A}(\mathbb{H}_a)$. Mode decomposition of weak solutions ------------------------------------ The aim of this section will be to extend the mode decomposition of $Sol_0(\mathcal{T})$ obtained previously to entire $\mathcal{D}(\mathcal{T})_0'$. Here we assume all the conditions of Proposition \[PropVarSep\] are satisfied, and we have chosen the system $\{u,v\}_\alpha$ with $u_\alpha=T_\alpha\zeta_\alpha$ and $v_\alpha=\bar T_\alpha\zeta_\alpha$, which span $Sol_0(\mathcal{T})$. For convenience we will also assume at least the part (iv) of the definition of the conventional Fourier transform to hold. Unfortunately we do not have a precise analytical description of the Fourier transformed test function space $\tilde{\mathcal{D}}(\tilde\Sigma)$ even under the assumptions of the conventional Fourier transform, as it was, for instance, in the Euclidean space by Paley-Wiener theorem. In particular we need to know for which modes $T_\alpha$ it holds $$\begin{aligned} T_\alpha(t)\tilde f(\alpha)\in\tilde{\mathcal{D}}(\tilde\Sigma)\mbox{, }\forall \tilde f(\alpha)\in\tilde{\mathcal{D}}(\tilde\Sigma)\mbox{, }t\in\mathcal{I}.\label{TMultCond}\end{aligned}$$ At least we are able to find a sufficient condition under additional assumptions. \[TMultPropAnal\] Suppose the bundle $\mathcal{T}$ is analytic and $D_{\Sigma_t}$ has a strictly uniform spectrum. For each $\alpha\in\tilde\Sigma^i$ set $T_\alpha(0)=p^i(\omega(\alpha))$ and $\dot T_\alpha(0)=q^i(\omega(\alpha))$, where $p^i,q^i\in\mathcal{A}[\mathbb{H}_0]$. Then (Eq.\[TMultCond\]) holds. Choose the interval $\mathcal{R}$ such that it contains both 0 and $t$. First we note that by Corollary \[TPolBound\] for $\alpha\in\tilde\Sigma^i$ we have $T_\alpha(t)=r^i_t(\omega(\alpha))$ with $r^i_t\in\mathcal{A}[\mathbb{H}_0]$. Denote $F_t^i(\lambda)=r^i_t(\frac{\lambda}{C(t)})\in\mathcal{A}[\mathbb{H}_0]$. Obviously for any $f\in\mathcal{D}(\mathcal{T}_t)$, $$F_t^i(\lambda_\alpha(t))\tilde f(\alpha)=\widetilde{\left[F_t^i(D_{\Sigma_t})f\right]}(\alpha),$$ where $F_t^i(D_{\Sigma_t})$ is defined by functional calculus. Then by Proposition \[FuncCalcProp\] $$F_t^i(\lambda_\alpha(t))\tilde f(\alpha)\in\tilde{\mathcal{D}}(\tilde\Sigma).$$ Let $\{U_n\}$ be a covering by local trivializations of $\mathcal{T}_t$, and let $\{\imath_n\}$ be a subordinate partition of unity. The support of $f$ is covered by $N_f$ (finite) trivializing neighborhoods, and we write $f=\sum_n\imath_nf=\sum_n f_n$. It follows $\tilde f=\sum_n\tilde f_n$ and $T_\alpha(t)\tilde f(\alpha)=\sum_n T_\alpha(t)\tilde f_n(\alpha)$. Consider $f_n$ as a section in the trivial bundle $\pi^{-1}(U_n)$. As we have seen already (and as we will see even more evidently for homogeneous spacetimes in the next chapter) each component $\tilde\Sigma^i$ supports the Fourier transform of one fiber component in some local frame. Thus we can write $f_n=\sum_i f^i_n$, where $f^i_n\in\mathcal{D}(U_n)$ and $\tilde f^i_n$ is supported in $\tilde\Sigma^i$. We get $$T_\alpha(t)\tilde f(\alpha)=\sum_n\sum_i F_t^i(\lambda_\alpha(t))\tilde f^i_n(\alpha)=\sum_n\sum_i\widetilde{\left[F_t^i(D_{\Sigma_t})f^i_n\right]}(\alpha)\in\tilde{\mathcal{D}}(\tilde\Sigma),$$ which completes the proof. \[MultI\_alpha\] An argument involving local trivializations as in the proof of Proposition \[TMultPropAnal\] will show that the multiplication of $\mathcal{F}[f(t,.)]$ by $I_\alpha(t)$ amounts to multiplication of each fiber component by a number, hence $I_\alpha(t)\mathcal{F}[f(t,.)]\in\tilde{\mathcal{D}}(\tilde\Sigma)$ for all $t\in\mathcal{I}$. Two useful facts about the time dependent Fourier transform can be given by the following \[SmoothFourCompInt\] Let $\tilde f(t,\alpha)\in C_0^\infty\left(\mathcal{I},\tilde{\mathcal{D}}(\tilde\Sigma)\right)$. then $f(t,\vec x)=\mathcal{F}^{-1}[\tilde f(t,\alpha)]\in\mathcal{D}(\mathcal{T})$ $\int_\mathcal{I}dt\tilde f(t,\alpha)\in\tilde{\mathcal{D}}(\tilde\Sigma)$. Let $$f(t,\vec x)=\mathcal{F}^{-1}[\tilde f(t,\alpha)]=\int_{\tilde\Sigma}d\mu(\alpha)\tilde f(t,\alpha)\zeta_\alpha(\vec x).$$ For each $t\in\mathcal{I}$ we have $\tilde f(t,\alpha)\in\tilde{\mathcal{D}}(\tilde\Sigma)$ and therefore $f(t,\vec x)\in\mathcal{D}(\mathcal{T}_t)$. If the compact interval $A\subset\mathcal{I}$ is such that $\forall t\notin A$, $\tilde f(t,\alpha)=0$, then obviously $\forall t\notin A$, $f(t,\vec x)=0$. Because the integration converges in $L^2(\tilde\Sigma,\mu)$, differentiation can be interchanged with the integral, thus $f(t,\vec x)$ is smooth in $t$. The part (i) is proven. Now write $$\tilde f(t,\alpha)=\int_{\Sigma_t}d\mu_{h_0}(\vec x)\langle\bar\zeta_\alpha(\vec x),f(t,\vec x)\rangle_{\mathfrak{g}^0},$$ and $$\int_\mathcal{I}dt\tilde f(t,\alpha)=\int_\mathcal{I}dt\int_{\Sigma_t}d\mu_{h_0}(\vec x)\langle\bar\zeta_\alpha(\vec x),f(t,\vec x)\rangle_{\mathfrak{g}^0}=$$ $$=\int_{\Sigma_t}d\mu_{h_0}(\vec x)\langle\bar\zeta_\alpha(\vec x),\int_\mathcal{I}dtf(t,\vec x)\rangle_{\mathfrak{g}^0}=\mathcal{F}[\int_\mathcal{I}dtf(t,\vec x)],$$ where Fubini’s theorem was used with the justification that both integrals run over compact supports [@Dieudonne1976]. For the part (ii) it remains to show that $\int_\mathcal{I}dtf(t,\vec x)\in\mathcal{D}(\mathcal{T}_t)$. But this is again clear because the integral runs over a compact support. Next we want to show that the Cauchy problem can be well-posed in the distributional sense. We will do it by generalizing (Eq.\[CauchyExplicit\]) to distributional solutions. \[CauchyProbDistrib\] For any $u_0,u_1\in\mathcal{D}(\mathcal{T}_t)'$ there exists a unique $\jmath(u_0,u_1)=u\in\mathcal{D}(\mathcal{T})_0'$ such that $$u(f)=u_0(i_t^(\nabla_tE[f]))-u_1(i_t^(E[f]))\mbox{, }\forall f\in\mathcal{D}(\mathcal{T}).$$ By Proposition \[ESurj\] we know that $E$ is surjective, so we denote the bijective part of $E$ to be $E_\updownarrow:\mathcal{D}(\mathcal{T})/\ker E\to Sol_0(\mathcal{T})$. For surjectivity of $\jmath$ it suffices to set $$u_0(v_1)=u(E_\updownarrow^{-1}[\jmath(0,v_1)])\mbox{, }u_1(v_0)=-u(E_\updownarrow^{-1}[\jmath(v_0,0)])\mbox{, }\forall v_0,v_1\in\mathcal{D}(\mathcal{T}_t).$$ Indeed, $$u(f)=u(E_\updownarrow^{-1}[E_\updownarrow[f]])=u(E_\updownarrow^{-1}[\jmath(i_t^(E_\updownarrow[f]),i_t^(\nabla_tE_\updownarrow[f]))])=$$ $$=u(E_\updownarrow^{-1}[\jmath(i_t(E_\updownarrow[f]),0)])+u(E_\updownarrow^{-1}[\jmath(0,i_t^(\nabla_tE_\updownarrow[f]))])=u_0(i_t^(\nabla_tE[f]))-u_1(i_t^(E[f])).$$ For injectivity of $\jmath$ let $u_0,u_1\in\mathcal{D}(\mathcal{T}_t)'$ be given. Define $u$ as in the statement. Then obviously $u(Df)=0$ because $EDf=0$ for any $f\in\mathcal{D}(\mathcal{T})$, hence $u\in\mathcal{D}(\mathcal{T}_t)_0'$. Now suppose the same formula holds also for different $u_0',u_1'\in\mathcal{D}(\mathcal{T}_t)'$ with the same $u$. Then we have $$0=(u_0-u_0')(i_t^(\nabla_tE[f]))-(u_1-u_1')(i_t^(E[f]))\mbox{, }\forall f\in\mathcal{D}(\mathcal{T}).$$ Evaluating on $f=E_\updownarrow^{-1}(\jmath(v_0,0))$ and $g=E_\updownarrow^{-1}(\jmath(0,v_1))$ for arbitrary $v_0,v_1\in\mathcal{D}(\mathcal{T}_t)$ we find $u_0=u_0'$ and $u_1=u_1'$. Now we come to the main assertion. Let the modes $T_\alpha$ be chosen such that (Eq.\[TMultCond\]) holds. \[WeakSolDecomp\] Under the assumptions made, there exist closed topological subspaces $\tilde{\mathcal{D}}^u(\tilde\Sigma),\tilde{\mathcal{D}}^v(\tilde\Sigma)\subset\tilde{\mathcal{D}}(\tilde\Sigma)$, such that for any $\psi\in\mathcal{D}(\mathcal{T})_0'$ there are unique distributions $a^{\psi}\in\tilde{\mathcal{D}}^u(\tilde\Sigma)'$, $b^{\psi}\in\tilde{\mathcal{D}}^v(\tilde\Sigma)'$ with $$\psi(f)=a^{\psi}(u_\alpha(f))+b^{\psi}(v_\alpha(f))\mbox{, }\forall f\in\mathcal{D}(\mathcal{T}).$$ Considered as distributions, the functions $u_\alpha$ act as $$u_\alpha(f)=\langle u_\alpha,f\rangle_M=\int_\mathcal{I}dtT_\alpha(t)\langle\zeta_\alpha,f(t,.)\rangle_{\Sigma_t}=\int_\mathcal{I}dtT_\alpha(t)(\zeta_{-\alpha},\check\Gamma f(t,.))_{\Sigma_t}=$$ $$=\int_\mathcal{I}dtT_\alpha(t)I_\alpha(t)\mathcal{F}[\check\Gamma f(t,.)](-\alpha)\mbox{, }\forall f\in\mathcal{D}(\mathcal{T}).$$ (Remember that $g_{00}=1$.) The action of $v_\alpha$ is similar. By assumption (Eq.\[TMultCond\]) and Remark \[MultI\_alpha\] we find $$T_\alpha(t)I_\alpha(t)\mathcal{F}[\check\Gamma f(t,.)](-\alpha)\in\tilde{\mathcal{D}}(\tilde\Sigma).$$ Then by Proposition \[SmoothFourCompInt\] we get $u_\alpha(f)\in\tilde{\mathcal{D}}(\tilde\Sigma)$ (similarly for $v_\alpha$). In general, the maps $f\to u_\alpha(f)$ and $f\to v_\alpha(f)$ need not be surjective. Therefore we define $$\tilde{\mathcal{D}}^u(\tilde\Sigma)=u_\alpha(\mathcal{D}(\mathcal{T})).$$ By continuity of the map $f\to u_\alpha(f)$ (which is easy to establish), $\tilde{\mathcal{D}}^u(\tilde\Sigma)$ is a closed subspace of $\tilde{\mathcal{D}}(\tilde\Sigma)$. Similarly we define $\tilde{\mathcal{D}}^v(\tilde\Sigma)$. Recall the mode expansion for arbitrary $\phi\in Sol_0(\mathcal{T})$, $$\phi(t,\vec x)=\int_{\tilde\Sigma}d\mu(\alpha)\left[a^\phi_\alpha u_\alpha(x)+b^\phi_\alpha v_\alpha(x)\right].$$ Thus $Sol_0(\mathcal{T})$ can be written as a direct sum of linear subspaces, $Sol_0(\mathcal{T})=Sol_0^u(\mathcal{T})\oplus Sol_0^v(\mathcal{T})$, with $$Sol_0^u(\mathcal{T})=\{\phi\in Sol_0(\mathcal{T})\mbox{: }b^\phi=0\}\mbox{, }Sol_0^v(\mathcal{T})=\{\phi\in Sol_0(\mathcal{T})\mbox{: }a^\phi=0\},$$ and we will write $\phi=\phi^u+\phi^v$. Regarding as a distribution in $\mathcal{D}(\mathcal{T})'$, $\phi^u$ and $\phi^v$ act as $$\phi^u(f)=\int_{\tilde\Sigma}d\nu(\alpha)a^\phi_\alpha u_\alpha(f)\mbox{, }\phi^v(f)=\int_{\tilde\Sigma}d\nu(\alpha)b^\phi_\alpha v_\alpha(f).$$ The functions $a^\phi_\alpha,b^\phi_\alpha$ can be regarded as distributions $a^\phi\in\tilde{\mathcal{D}}^u(\tilde\Sigma)'$, $b^\phi\in\tilde{\mathcal{D}}^v(\tilde\Sigma)'$, and we can write $$\begin{aligned} \phi(f)=\phi^u(f)+\phi^v(f)=a^\phi(u_\alpha(f))+b^\phi(v_\alpha(f)).\label{Sol0ModDecomp}\end{aligned}$$ Now let $\varphi\in Sol(\mathcal{T})$ be a solution, which does not necessarily have $supp\{\varphi\}\cap\Sigma_t$ compact. Its Cauchy data are $$(i^_t(\varphi),i^_t(\nabla_t\varphi))=(\varphi_0,\varphi_1)\in\mathcal{E}(\mathcal{T}_t)\oplus\mathcal{E}(\mathcal{T}_t).$$ Choosing a countable (compactly finite) partition of unity on $\Sigma$ we can write $$(\varphi_0,\varphi_1)=\sum_{i=1}^\infty(\phi^i_0,\phi^i_1)\mbox{, }(\phi^i_0,\phi^i_1)\in\mathcal{D}(\mathcal{T}_t)\oplus\mathcal{D}(\mathcal{T}_t),$$ where the sum involves finite items on any compact region $U\in\Sigma$. Now for each $i$ we have $$\phi^i=i^{-1}_t(\phi^i_0,\phi^i_1)\in Sol_0(\mathcal{T}),$$ thus $$\phi^i=\phi^{i,u}+\phi^{i,v}=i^{-1}_t(\phi^{i,u}_0,\phi^{i,u}_1)+i^{-1}_t(\phi^{i,v}_0,\phi^{i,v}_1)\mbox{, }\phi^{i,u}\in Sol^u_0(\mathcal{T})\mbox{, }\phi^{i,v}\in Sol^v_0(\mathcal{T}).$$ Set $$\varphi^u=\sum_{i=1}^\infty i^{-1}_t(\phi^{i,u}_0,\phi^{i,u}_1)=\sum_{i=1}^\infty\phi^{i,u},$$ and $$\varphi^v=\sum_{i=1}^\infty i^{-1}_t(\phi^{i,v}_0,\phi^{i,v}_1)=\sum_{i=1}^\infty\phi^{i,v},$$ where the sums converge in compact topology. (This can be seen as follows. The intersection of the causal cone of any compact region with a Cauchy surface is a compact surface, and therefore only finite summands survive.) But we have $$\phi^{i,u}(f)=a^{\phi^i}(u_\alpha(f))\mbox{, }\phi^{i,v}(f)=b^{\phi^i}(v_\alpha(f))$$ for some distributions $a^{\phi^i}$ and $b^{\phi^i}$. Thus we obtain $$\varphi^u(f)=\sum_{i=1}^\infty\phi^{i,u}(f)=\sum_{i=1}^\infty a^{\phi^i}(u_\alpha(f))$$ and $$\varphi^v(f)=\sum_{i=1}^\infty\phi^{i,v}(f)=\sum_{i=1}^\infty b^{\phi^i}(v_\alpha(f)).$$ This convergence defines distributions $$a^\varphi=\sum_{i=1}^\infty a^{\phi^i}\in\tilde{\mathcal{D}}^u(\tilde\Sigma)'\mbox{, }b^\varphi=\sum_{i=1}^\infty b^{\phi^i}\in\tilde{\mathcal{D}}^v(\tilde\Sigma)',$$ such that $$\varphi^u(f)=a^\varphi(u_\alpha(f))\mbox{, }\varphi^v(f)=b^\varphi(v_\alpha(f))\mbox{, }\varphi=\varphi^u+\varphi^v\mbox{, }\forall f\in\mathcal{D}(\mathcal{T}),$$ and thus $Sol(\mathcal{T})=Sol^u(\mathcal{T})\oplus Sol^v(\mathcal{T})$, where $$Sol^u(\mathcal{T})=\{\varphi\in Sol(\mathcal{T})\mbox{: }b^\varphi=0\}\mbox{, }Sol^v(\mathcal{T})=\{\varphi\in Sol(\mathcal{T})\mbox{: }a^\varphi=0\}.$$ Now let $\psi\in\mathcal{D}(\mathcal{T})_0'$ be a weak solution, and $\{\chi_m\}$ a usual mollifier on $\Sigma_t$. Define the mollifications $\heartsuit_m\psi\in Sol(\mathcal{T})$ by $$\heartsuit_m\psi=\jmath(\chi_m\psi_0,\chi_m\psi_1),$$ where $\psi=\jmath(\psi_0,\psi_1)$ by Proposition \[CauchyProbDistrib\]. Then it is easy to see that $\heartsuit_m\psi\to\psi$ in $\mathcal{D}(\mathcal{T})'$. That $\heartsuit_m\psi\in Sol(\mathcal{T})$ it follows $$\heartsuit_m\psi=(\heartsuit_m\psi)^u+(\heartsuit_m\psi)^v\mbox{, }(\heartsuit_m\psi)^\bullet\in Sol^\bullet(\mathcal{T}).$$ The disjointness $Sol^u(\mathcal{T})\cap Sol^v(\mathcal{T})=0$ implies that $(\heartsuit_m\psi)^u\to\psi^u$ and $(\heartsuit_m\psi)^v\to\psi^v$ with some distributions $\psi^u\in\overline{Sol^u(\mathcal{T})}$, $\psi^v\in\overline{Sol^v(\mathcal{T})}$, such that $\psi=\psi^u+\psi^v$. We denote $$\psi^u(f)=\lim_{m\to\infty}(\heartsuit_m\psi)^u(f)=\lim_{m\to\infty}a^{\psi_m}(u_\alpha(f))\doteq a^{\psi}(u_\alpha(f)),$$ $$\psi^v(f)=\lim_{m\to\infty}(\heartsuit_m\psi)^v(f)=\lim_{m\to\infty}b^{\psi_m}(v_\alpha(f))\doteq b^{\psi}(v_\alpha(f)),$$ for some distributions $a^{\psi}$ and $b^{\psi}$. Finally we arrive at $$\psi(f)=a^{\psi}(u_\alpha(f))+b^{\psi}(v_\alpha(f))\mbox{, }\forall f\in\mathcal{D}(\mathcal{T}).$$ The map $$\mathcal{D}(\mathcal{T})_0'\ni\psi\to(a^{\psi},b^{\psi})\in\tilde{\mathcal{D}}^u(\tilde\Sigma)'\oplus\tilde{\mathcal{D}}^v(\tilde\Sigma)'$$ is a bijection by construction. The propagator -------------- In this section we will find the explicit form of the propagator $E$ in terms of the mode decomposition. Of course, Green’s functions can be calculated using the techniques of inverse operators. But our approach will be more concordant to the spirit of this work and will at the same time demonstrate the usefulness of the mode decomposition in general. To use the mode decomposition for weak solutions we assume that at least the condition (iv) of the conventional Fourier transform holds, and that the assumptions of Proposition \[WeakSolDecomp\] are satisfied. Choose mode solutions to be such that $T_\alpha(0)=T_{-\alpha}(0)$ and $\dot T_\alpha(0)=\dot T_{-\alpha}(0)$. Then because $\alpha\to-\alpha$ preserves both $\lambda_\alpha(t)$ and the component $\tilde\Sigma^i$, we have the same mode equations for $T_\alpha$ and $T_{-\alpha}$, hence everywhere $T_\alpha(t)=T_{-\alpha}(t)$. The function $$\det W[T_\alpha,\bar T_\alpha](t)=I_\alpha(t)\left[\dot T_\alpha(t)\bar T_\alpha(t)-T_\alpha(t)\dot{\bar{T}}_\alpha(t)\right]\in C^\infty(\mathcal{I},i\cdot \mathbb{R})$$ is the Wronskian of two independent solutions $T_\alpha$ and $\bar T_\alpha$ and is therefore an imaginary constant. For convenience we appoint once and forever to consider only the modes normalized by $$\begin{aligned} \dot T_\alpha(t)\bar T_\alpha(t)-T_\alpha(t)\dot{\bar{T}}_\alpha(t)=i\cdot I^{-1}_\alpha(t).\label{ModeTNorm}\end{aligned}$$ It can be seen that this condition is consistent with our previous assumptions for the modes $T_\alpha$. We remark that the Krein space involution $\check\Gamma$ commutes with the connection components $\Gamma_i$. Indeed, by definition $\check\Gamma=P^+-P^-$, where $P^\pm$ are the projections onto the subspaces of positive/negative definiteness of the metric $\langle,\rangle_g$. Let $\{e_i\}$ be a pseudo-orthonormal moving frame, i.e., $\langle e_i,e_i\rangle_g=\pm1$. The value of each $\langle e_i,e_i\rangle_g$ is preserved under $\nabla$, and therefore $e_i$ remains in the same eigenspace of $\check\Gamma$, although in our main frame $e_i$ experiences gradient, $$\nabla e_i=\sum_{j=1}^4\sum_{k=1}^n\Gamma^{k}_{ji}dx^j\otimes e_k.$$ Hence $\check\Gamma$ commutes with all $\Gamma_i$. We have that $$\langle u,v\rangle_{\Sigma_t}=(\check\Gamma\bar u,v)_{\Sigma_t}=\int_{\tilde\Sigma}d\mu(\alpha)s(\alpha)I_\alpha(t)\tilde u(-\alpha)\tilde v(\alpha)\mbox{, }\forall u\in\mathcal{E}(\mathcal{T}_t)\mbox{, }v\in\mathcal{D}(\mathcal{T}_t),$$ where $s(\alpha)$ is the Fourier image of the Krein involution $\check\Gamma$, which due to the remark above satisfies $s(\tilde\Sigma^i)=\{+1,-1\}$, i.e., is constant on each component $\tilde\Sigma^i$. We have used the fact that $\bar{\tilde{\bar{u}}}(\alpha)=\tilde u(-\alpha)$ which follows from the condition (iv) of the conventional Fourier transform. Now the propagator is the unique operator $E:\mathcal{D}(\mathcal{T})\to Sol_0(\mathcal{T})$ which satisfies $$\begin{aligned} v(f)=\langle v(t;.),\dot E[f](t;.)\rangle_{\Sigma_t}-\langle\dot v(t;.),E[f](t;.)\rangle_{\Sigma_t},\nonumber\\ \forall v\in Sol_0(\mathcal{T})\mbox{, }f\in\mathcal{D}(\mathcal{T})\mbox{, }t\in\mathcal{I}.\label{PropMainEq}\end{aligned}$$ As $v\in Sol_0(\mathcal{T})$ we can write $$\begin{aligned} v(x)=\int_{\tilde\Sigma}d\mu(\alpha)a^v(\alpha)u_\alpha(x)+\int_{\tilde\Sigma}d\mu(\alpha)b^v(\alpha)v_\alpha(x),\label{vModDec}\end{aligned}$$ and $$\widetilde{v(t;.)}(\alpha)=a^v(\alpha)T_\alpha(t)+b^v(\alpha)\bar T_\alpha(t),$$ $$\widetilde{\dot{v}(t;.)}(\alpha)=a^v(\alpha)\dot T_\alpha(t)+b^v(\alpha)\dot{\bar{T}}_\alpha(t).$$ Similarly for $E[f]\in Sol_0(\mathcal{T})$, $$E[f](x)=\int_{\tilde\Sigma}d\mu(\alpha)a^E[f](\alpha)u_\alpha(x)+\int_{\tilde\Sigma}d\mu(\alpha)b^E[f](\alpha)v_\alpha(x),$$ $$\widetilde{E[f](t;.)}(\alpha)=a^E[f](\alpha)T_\alpha(t)+b^E[f](\alpha)\bar T_\alpha(t),$$ and $$\widetilde{\dot{E}[f](t;.)}(\alpha)=a^E[f](\alpha)\dot T_\alpha(t)+b^E[f](\alpha)\dot{\bar{T}}_\alpha(t),$$ with some distribution fields $a^E[f](\alpha)$ and $b^E[f](\alpha)$. Using all this we compute $$\langle v(t;.),\dot E[f](t;.)\rangle_{\Sigma_t}-\langle\dot v(t;.),E[f](t;.)\rangle_{\Sigma_t}=$$ $$=-\int_{\tilde\Sigma}d\mu(\alpha)s(\alpha)I_\alpha(t)\left[\dot T_\alpha(t)\bar T_\alpha(t)-T_\alpha(t)\dot{\bar{T}}_\alpha(t)\right]\left[a^v(-\alpha)b^E[f](\alpha)-b^v(-\alpha)a^E[f](\alpha)\right]=$$ by normalization (Eq.\[ModeTNorm\]) $$\begin{aligned} =-i\int_{\tilde\Sigma}d\mu(\alpha)s(\alpha)I_\alpha(t)\left[a^v(-\alpha)b^E[f](\alpha)-b^v(-\alpha)a^E[f](\alpha)\right].\label{ECompIntRes1}\end{aligned}$$ On the other hand, we know that $ED=0$, thus $a^E[f](\alpha)$ and $b^E[f](\alpha)$ are weak solutions of the field equation and can be mode decomposed as $$\begin{aligned} a^E[f](\alpha)=a^1_\alpha(u_\beta(f))+a^2_\alpha(v_\beta(f)),\nonumber\\ b^E[f](\alpha)=b^1_\alpha(u_\beta(f))+b^2_\alpha(v_\beta(f)).\label{ECompIntRes2}\end{aligned}$$ By (Eq.\[vModDec\]) we have $$\begin{aligned} v(f)=\int_{\tilde\Sigma}d\mu(\alpha)a^v(\alpha)u_\alpha(f)+\int_{\tilde\Sigma}d\mu(\alpha)b^v(\alpha)v_\alpha(f).\label{ECompIntRes3}\end{aligned}$$ Inserting (Eq.\[ECompIntRes1\]), (Eq.\[ECompIntRes2\]) and (Eq.\[ECompIntRes3\]) into (Eq.\[PropMainEq\]) we obtain $$a^1_\alpha=b^2_\alpha=0\mbox{, }a^2_\alpha=-b^1_\alpha=i\cdot s(\alpha)\delta(\beta-\alpha).$$ And our final formula is $$E[f](x)=i\int_{\tilde\Sigma}d\mu(\alpha)s(\alpha)\left[v_{-\alpha}(f)u_\alpha(x)-u_{-\alpha}(f)v_\alpha(x)\right],$$ which is in full accord with the result obtained by [@Lueders_Roberts_1990] for scalar fields on FRW spacetimes. Aspects of Harmonic Analysis in Homogeneous Spacetimes ====================================================== Spatially homogeneous cosmological models ----------------------------------------- The main goal of the current work is to refurbish the mathematical framework of quantum field theory on classical cosmological spacetimes, in general, and to advance towards a satisfactory rigorous description of cosmological particle creation in states of low energy for hyperbolic fields, in particular. The latter would be an extension of results obtained in [@Degner_Verch_2010] for the Klein-Gordon field on specific FRW models to more general situations. Thus although some results were and will be obtained under abstract general assumptions, our attention is concentrated at the geometrical setup of most common cosmological models. Supported by observations of the universe at large scale, cosmology considers mainly spatially homogeneous, or in addition also isotropic, spacetimes. A condensed account of cosmological arguments and their geometrical implications can be found, for instance, in [@pittphilsci1507]. The essence of these geometrical restrictions is mathematically expressed by imposing the existence of a sufficiently rich system of symmetries (more precisely, a group of spatial isometries) on the spacetime. Extensive treatments of all possible isometry groups and related questions can be found in [@pittphilsci1507], [@Petrov59], [@StephaniKramerMacCallumHoenselaersHerlt200305]. An introduction to the generalities of harmonic analysis on vector bundles is given in [@Camporesi:1990wm]. In this section we will try to deductively introduce our geometrical setup with the help of the information in the above mentioned references. [Foliation by equal time Cauchy hypersurfaces.]{} Recall that we are working with a four dimensional globally hyperbolic Lorentzian manifold $(M,g)$ on which a global smooth time function and an atlas can be chosen following [@Bernal_Sanchez_2005] such that $M$ is foliated by three dimensional spacelike equal-time smooth Cauchy hypersurfaces and $$ds^2=g_{00}dt^2-d\sigma^2,$$ where $d\sigma^2$ is the line element on any of those Cauchy surfaces being Riemannian submanifolds. [The structure group.]{} Any vector bundle $\mathcal{T}$ can be considered as associated to its frame bundle $\mathcal{P}_\mathcal{T}$ with structure group $GL(n)$. If we want the fiberwise transformations to respect the fiber metric, then we have to restrict the principal bundle to the orthogonal frame bundle. All fibers $V_p$ with their respective non-degenerate pseudo-Riemannian structures $\mathfrak{g}_p$ are isomorphic, and their generalized orthogonal groups $O(\mathfrak{g}_p)$ (i.e., groups of invertible endomorphisms of $V_p$ preserving $\mathfrak{g}_p$) are isomorphic to the generalized Lorentz group $O(\pm_\mathfrak{g})$, where $\pm_\mathfrak{g}$ in this context will be understood as the signature of $\mathfrak{g}$. But the same vector bundle $\mathcal{T}$ can be associated also to another principal bundle (which we again denote by $\mathcal{P}_\mathcal{T}$) with structure group $H$ (say, for field theoretical reasons). Then we have a representation $r$ of $H$ on $V$. If $r$ also respects the metric, then $r(H)\in O(\pm_\mathfrak{g})$, so $H$ is homomorphic to $O(\pm_\mathfrak{g})$. For instance, $H=SO^+(\pm_\mathfrak{g})$ (tensor bundle) or $H=Spin^+(\pm_\mathfrak{g})$ (spinor bundle). [Isometries.]{} Let us start with reminding some definitions. An [isometry]{} of the spacetime $(M,g)$ is a diffeomorphism $\psi:M\to M$ such that $\psi^g=g$ holds on $M$, where $\psi^$ is the pullback of $\psi$. If $\psi':M\to M$ is another isometry, then obviously such is also their superposition $\psi\circ\psi'$. With the superposition as product, isometries thus constitute an abstract group, which we will denote $\mbox{\bf Iso}(M)$. If $\mathcal{T}\to M$ is a the vector bundle over $M$ as defined previously, then an [isometry]{} of the vector bundle $\mathcal{T}$ is a morphism $\Psi:\mathcal{T}\to\mathcal{T}$ covering an isometry of the base, $\pi\circ\Psi\circ\pi^{-1}\in \mbox{\bf Iso}(M)$, such that $\Psi^\mathfrak{g}=\mathfrak{g}$ and $\Psi^\nabla=\nabla$ (or more precisely $\Psi^D=D$ when a normal hyperbolic field operator $D$ is specified), where $\Psi^$ denotes pullback maps, $\mathfrak{g}$ is the pseudo-Riemannian fiber metric, and $\nabla$ is the metric connection. Again via superposition, the isometries of the bundle $\mathcal{T}$ comprise an abstract group $\mbox{\bf Iso}(\mathcal{T})$. The map $\mbox{\bf Iso}(\mathcal{T})\ni\Psi\to\pi\circ\Psi\circ\pi^{-1}\in\mbox{\bf Iso}(M)$ gives a homomorphism of $\mbox{\bf Iso}(\mathcal{T})$ into $\mbox{\bf Iso}(M)$. The image of this homomorphosm is a subgroup of $\mbox{\bf Iso}(M)$ and will be denoted by $\mbox{\bf Iso}^\mathcal{T}(M)\subset\mbox{\bf Iso}(M)$, and its kernel is a normal subgroup of $\mbox{\bf Iso}(\mathcal{T})$. This kernel $\mbox{\bf Iso}(\mathcal{T})/\mbox{\bf Iso}^\mathcal{T}(M)$ consists of isometries of the bundle $\mathcal{T}$ covering the identity map of $M$. These are precisely the smooth sections in the principle bundle $\mathcal{P}_\mathcal{T}\xrightarrow{loc}M\times H$ of $\mathcal{T}$, i.e., $\mbox{\bf Iso}(\mathcal{T})/\mbox{\bf Iso}^\mathcal{T}(M)=C^\infty(\mathcal{P}_\mathcal{T})$. The group multiplication is given by the pointwise multiplication of sections. [Homogeneous bundle structure.]{} If the sections in the bundle $\mathcal{T}$ are going to represent physical fields, than one should have a concrete picture of how they transform under the diffeomorphisms of the spacetime $M$. In case of the tensor bundle this picture is automatically encoded in the pullback map. An abstract vector bundle does not have such a structure by itself. Thus a physical field theory has to specify a homomorphism $\rho:\mbox{Diff}(M)\to C^\infty(\mathcal{P}_\mathcal{T})$. For the tangent bundle $\rho(\psi)=d\psi$, $\psi\in\mbox{Diff}(M)$. When considering arbitrary diffeomorphisms, then the structure group should be $GL(n)$ rather than a smaller $H$. But if we restrict $\rho$ to $\rho:\mbox{\bf Iso}^\mathcal{T}(M)\to C^\infty(\mathcal{P}_\mathcal{T})$, then $H$ can be chosen. For brevity denote $G=\mbox{\bf Iso}^\mathcal{T}(M)$. We have the injection $$G\ni g\to g\times\rho(g)\in\mbox{\bf Iso}(\mathcal{T}),$$ which gives sense to the left action of $G$ on $\mathcal{T}$ by isometries. The abstract group of isometries of a pseudo-Riemannian manifold of dimension $m$ is given the compact open topology, in which it becomes a Lie group of dimension at most $n(n+1)/2$ [@Helgason197901]. It can be further shown, that the compact open topology in this case is equivalent to the pointwise convergence topology of isometries. Thus we automatically obtain a Lie group structure on $\mbox{\bf Iso}(M)$. Then $G\subset\mbox{\bf Iso}(M)$ is a topological subgroup defined by $$G=\{\psi\in\mbox{\bf Iso}(M)\mbox{: }\left(\psi\times\rho(\psi)\right)^\mathfrak{g}=\mathfrak{g}\mbox{, }\left(\psi\times\rho(\psi)\right)^D=D\}.$$ If $\rho$ is a continuous homomorphism, then all the operations in the equations $$\left(\psi\times\rho(\psi)\right)^\mathfrak{g}=\mathfrak{g}\mbox{, }\left(\psi\times\rho(\psi)\right)^D=D$$ are continuous, and therefore the subspace $G$ of $\mbox{\bf Iso}(M)$ defined by this equation is a closed topological subspace. But then by Cartan’s theorem $G$ is actually a Lie subgroup, as it is a closed topological subgroup of the Lie group $\mbox{\bf Iso}(M)$. Thus we have the structure of a $G$-homogeneous vector bundle $\mathcal{T}$. [Spatially homogeneous bundle.]{} The bundle $\mathcal{T}$ will be called [spatially homogeneous]{} if the orbits of $\mbox{\bf Iso}^\mathcal{T}(M)$ are three dimensional smooth spacelike hypersurfaces which foliate $M$. (Maybe it is worth mentioning here that everywhere in this work we consider only connected spacetimes $M$.) By Theorem 8.16 in [@StephaniKramerMacCallumHoenselaersHerlt200305] there exists a parametrization of these orbits by the affine parameter of the family of normal geodesics, such that the metric takes the form $$ds^2=dt^2-d\sigma^2.$$ On the other hand, our original foliation by equal time Cauchy surfaces due to Theorem 1.1 in [@Bernal_Sanchez_2005] also yielded such a metric form. We assume that the time function can be chosen such that equal time Cauchy surfaces are the orbits of $\mbox{\bf Iso}^\mathcal{T}(M)$ (probably this can be shown to be true in general). We note that due to the transitive action of $G$ on $\Sigma_t$ for every $t$, it holds $G\subset\mbox{\bf Iso}^{\mathcal{T}_t}(\Sigma_t)$. We did not write $G=\mbox{\bf Iso}^{\mathcal{T}_t}(\Sigma_t)$ because it is possible that for some $t\neq t'\in\mathcal{I}$, $\mbox{\bf Iso}^{\mathcal{T}_t}(\Sigma_t)\neq\mbox{\bf Iso}^{\mathcal{T}_{t'}}(\Sigma_{t'})$, i.e., for some time instances the time slice may be more symmetric than usual. We will concentrate on $G$, which is the maximal guaranteed amount of symmetry which is present at any time. Thus we see that $\mathcal{T}_t$ also has the structure of a $G$-homogeneous vector bundle. Consider the principle bundle $\mathcal{P}_{\mathcal{T}_t}$ of $\mathcal{T}_t$, which is a subbundle of $\mathcal{P}_\mathcal{T}$. The smooth left action of $G$ on $\mathcal{T}_t$ gives a smooth left action of $G$ on $\mathcal{P}_{\mathcal{T}_t}$ as well. This action allows one to construct a global smooth section in $\mathcal{P}_{\mathcal{T}_t}$, whence it follows that the bundle $\mathcal{T}_t$ is trivial. Because $M\sim\Sigma_t\times\mathcal{I}$, the whole bundle $\mathcal{T}$ is also trivial. Thus spatially homogeneous vector bundles over $M$ are necessarily trivial. The requirement that the field operator $D$ is $G$-invariant implies that the function $m^\star(x)$ is in fact a function of time only. [Homogeneous space structure.]{} Now let $\mbox{\bf StabIso}^\mathcal{T}(p)\subset G$ be the stabilizer of $G$ at some fixed point $p\in M$. Then $\mbox{\bf StabIso}^\mathcal{T}(p)$ is a closed Lie subgroup by Cartan’s theorem. That for all $p\in M$, the groups $\mbox{\bf StabIso}^\mathcal{T}(p)$ are isomorphic, then we denote them all by $\mbox{\bf StabIso}^\mathcal{T}(M)$. In this case the orbits $\Sigma_t$ of $G$ are diffeomorphic to the homogeneous space $G/\mbox{\bf StabIso}^\mathcal{T}(M)\doteq\Sigma$. Denote $O=\mbox{\bf StabIso}^\mathcal{T}(M)^+$, the identity component. Then $\Gamma=\mbox{\bf StabIso}^\mathcal{T}(M)/O$ is a discrete normal subgroup of $G$. If the homogeneous space $\Sigma$ is itself a Lie subgroup of $G$, then it acts on each $\Sigma_t$ simply transitively. [The four dimensional reality.]{} As already mentioned, the isometry group $\mbox{\bf Iso}(M)$ of the $n=4$ dimensional spacetime $M$ is a Lie group of dimension at most $n(n+1)/2=10$. Thus in principle one can construct all real Lie algebras $\mathcal{G}$ of dimension up to 10, their corresponding connected simply connected Lie groups $G$, then all discrete normal subgroups $\Gamma$ of such $G$ etc., thereby exhausting all possible isometry groups of $M$. This heavy task have been done by Petrov et al [@Petrov59] and others [@StephaniKramerMacCallumHoenselaersHerlt200305], and all the possibilities are listed in tables. It turned out that only the Minkowski space has isometry group of maximal dimension 10, which is the Poincaré group. Among all possibilities we are interested in those whose orbits are $\Sigma_t$. Thus the dimension of $G$ is at least three. There are three possibilities of six dimensional such isometry groups, which correspond to FRW spacetimes. A number of possibilities are available with four dimensional groups, which correspond to the LRS spacetimes. And finally there are nine classes of three dimensional real Lie groups $Bi(N)$ (called Bianchi groups), which together with their factors $Bi(N)/\Gamma$ by discrete subgroups $\Gamma$ represent the isometry groups of the spatially homogeneous spacetimes. It turned out further, that in all these cases besides one (the so called Kantowski-Sachs model) the isometry group is the semidirect product $G=\Sigma\rtimes O$. In this case we will call $\mathcal{T}_t$ a [semidirect homogeneous vector bundle]{}. In particular, for six dimensional FRW groups, four dimensional LRS groups and three dimensional Bianchi groups $O=SO(3)$, $SO(2)$ and $\{1\}$, respectively. The normal subgroups $\Sigma$ are nothing else than $Bi(N)/\Gamma$. In next sections we will work on the semidirect homogeneous vector bundles. After establishing the necessary mathematical framework, we will obtain results concerning the structure of $G$-invariant homogeneous bi-distributions. On harmonic analysis in semidirect homogeneous vector bundles ------------------------------------------------------------- In this section we will collect information on harmonic analysis in $G$-homogeneous vector bundles $\mathcal{T}\to G/O$ where $G=\Sigma\rtimes O$ which will be useful later in the work. This does not pretend to be self-contained or systematic; quite the contrary, we will introduce mainly what we were not able to find in the literature. Otherwise references will be provided. [Semidirect homogeneous vector bundles.]{} Let $G=\Sigma\rtimes O$, where $O$ is a compact connected type I Lie subgroup, and $\Sigma$ a connected normal type I Lie subgroup. Moreover, we demand that the modular function of $\Sigma$ has a non-trivial kernel, so that the representation theories of both $\Sigma$ and $G$ are well under control by Theorem 7.50 of [@Folland199502]. We note that this is the case for all Bianchi groups which are in fact the only candidates for $\Sigma$ in our context. Let $\Sigma=G/O$ have a Riemannian structure $h$ which is invariant under the left action of $G$. Let further $\mathcal{T}\to \Sigma$ be an $n$-dimensional (real or complex) vector bundle with standard fiber $V$ and a pseudo-Riemannian fiber metric $\mathfrak{g}$. Let there be a smooth left action of $G$ on $\mathcal{T}$ covering the left multiplication of $G$ on the base, such that the fiber metric is invariant under that action. Then we will call $\mathcal{T}$ a [semidirect $G$-homogeneous vector bundle]{}. If we choose an orthonormal frame $\{X_i\}$ of $T^_1\Sigma$ (or $\{Y_i\}$ of $\mathcal{T}\|_1$), and drag it throughout $\Sigma$ using the transitive left action of $G$, we will obtain $G$-invariant global smooth frame $\{X_i\}$ in $T^\Sigma$ (similarly, $\{Y_i\}$ in $\mathcal{T}$). Thus both $T^\Sigma$ and $\mathcal{T}$ are trivial bundles. Associated to the Riemannian structure $h$ there is a Laplace operator $\Delta$ acting on sections $f\in C^\infty(\mathcal{T})$. [The regular and quasi-regular representations for the line bundle.]{} Suppose $\mathcal{T}$ from above is a line bundle, $n=1$. The left regular representation $L_g$ of $G$ on $C(G)$ acts as $$L_gf(x)=f(g^{-1}x)\mbox{, }\forall g,x\in G.$$ Because the Riemannian structure is $G$-invariant, the metric measure $dx$ is a left Haar measure on $G$, and hence $L_g$ is a unitary representation on $L^2(G)$. Now any point $x\in G$ can be uniquely written as $x=x_\Sigma x_O$, where $x_\Sigma\in \Sigma$ and $x_O\in O$. Let $dx_\Sigma$ be the metric driven left $G$-invariant measure on $\Sigma$, and $dx_O$ the Lebesgue measure on $O$ normalized to $\|O\|=1$. Then $dx=dx_\Sigma dx_O$ gives a left Haar measure on $G$. Functions $f$ on $\Sigma$ are identified with their right $O$-invariant extensions to $G$, i.e., $f(xo)=f(x)=f(xO)$, for any $x\in G$, $o\in O$. Thus $C(\Sigma)\in C(G)$ (similarly $L^2(\Sigma)\in L^2(G), etc.$) and we may consider the restriction $U_g$ of the left regular representation $L_g$ on $C(G)$ to $C(\Sigma)$. Its action will be given by $$U_gf(x_\Sigma O)=f(g^{-1}x_\Sigma O)\mbox{, }\forall x_\Sigma\in \Sigma\mbox{, }g\in G.$$ The representation $U_g$ of $G$ is the left quasi-regular representation, and it is nothing else but the induced representation $\mbox{Ind}_O^G1$. Note that for $O=\{1\}$ we simply have $G=\Sigma$ and $L_g=U_g$. Neither $L_g$ nor $U_g$ need to be irreducible. The central decomposition of $L_g$ is $$L_g=\int^\oplus_{\hat G} d\nu(\pi)L_g(\pi),$$ where $\nu(\pi)$ is the Plancherel measure and $L_g(\pi)=\pi\otimes1$ is the primary representation composed of $mult(\pi,L_g)=\dim\bar\pi\in[1,\infty]$ copies of $\pi$ [@Folland199502]. The central decomposition of $U_g$ will be $$U_g=\int^\oplus_{\hat G_\Sigma} d\mu(\pi)U_g(\pi),$$ where $\hat G_\Sigma\subset\hat G$, $d\mu$ is the spectral measure of $U_g$ and for $\mu$-almost all $\pi$, $U_g(\pi)$ is a multiple of $\pi$ (multiplicities $mult(\pi,U_g)$ and the measure $d\mu(\pi)$ need to be determined). The corresponding Hilbert space decompositions are $$L^2(G)=\int_{\hat G}^\oplus d\nu(\pi)\mathcal{H}_\pi\otimes\mathcal{H}_{\bar\pi}$$ and $$L^2(\Sigma)=\int_{\hat G_\Sigma}^\oplus d\mu(\pi)\mathcal{H}(\pi),$$ where $\mathcal{H}(\pi)=\mathcal{H}_\pi\otimes\mathbb{C}^{mult(\pi,U_g)}\subset\mathcal{H}_\pi\otimes\mathcal{H}_{\bar\pi}$. Here $\mathbb{C}^{mult(\pi,U_g)}$ symbolizes some Hilbert space of dimension $mult(\pi,U_g)$ which is finite or infinite. In the following we will deal with $U_g$ keeping in mind that in case $G=\Sigma$ everything reduces to $L_g$. [The operator $\Pi_\pi$.]{} Consider for any $\pi\in\hat G$ the bounded operator $$\Pi_\pi=\int_Odo\pi(o).$$ Then $\Pi_\pi$ is self adjoint, $$\Pi_\pi^=\int_Odo\pi(o)^=\int_Odo\pi(o^{-1})=\Pi_\pi.$$ Moreover, because $O$ is unimodular, we have $$\pi(o)\Pi_\pi=\pi(o)\int_Odo'\pi(o')=\int_Od(oo')\pi(oo')=\Pi_\pi=\Pi_\pi\pi(o)\mbox{, }\forall o\in O,$$ and hence $\Pi_\pi$ is a projection, $$\Pi_\pi^2=\int_Odo\pi(o)\Pi_\pi=\int_Odo\Pi_\pi=\Pi_\pi.$$ $\Pi_\pi$ is a projection onto an invariant subspace of $\pi\|_O$. Recall the operator $D_\pi$ of [@Folland199502] which satisfied $D_\pi\pi(x)=\Delta^{\frac{1}{2}}(x)\pi(x)D_\pi$, for all $x\in G$. In particular, we find that $D_\pi\pi(o)=\pi(o)D_\pi$ for all $o\in O$, and consequently, $D_\pi\Pi_\pi=\Pi_\pi D_\pi$. [The Fourier transform in $G/O$.]{} The Fourier transform in $\Sigma=G/O$ associated to $U_g$ is naturally the restriction of that on $G$ associated to $L_g$; for $\mu$-almost all $\pi\in\hat G_\Sigma$ $$\hat f(\pi)=\pi(f)D_\pi\in\mathcal{H}(\pi).$$ For any $f\in C_0(\Sigma)$ and $\mu$-almost all $\pi\in\hat G_\Sigma$ we have $$\begin{aligned} \pi(f)=\int_\Sigma dx_\Sigma\int_Odx_Of(x_\Sigma O)\pi(x_\Sigma)\pi(x_O)=\int_\Sigma dx_\Sigma f(x_\Sigma O)\pi(x_\Sigma)\Pi_\pi.\label{Pi_f_Def}\end{aligned}$$ As usual we have $\pi(U_g f)=\pi(L_gf)=\pi(g)\pi(f)$ for $g\in G$, $f\in C_0(\Sigma)$. The convolution $f\ast h$ has the property that if $f\in C_0(G)$ and $h\in C_0(\Sigma)$ then $f\ast h\in C_0(\Sigma)$. Moreover, it satisfies $\pi(f\ast h)=\pi(f)\pi(h)$. [The case of arbitrary $\mathcal{T}$.]{} Let now $\dim V=n\ge 1$. The left quasi-regular representation of $G$ on $C^\infty(\mathcal{T})$ acts as $$U^{\mathcal{T}}_gf(x)=g^{-1}f(g^{-1}x)\mbox{, }\forall f\in C^\infty(\mathcal{T}).$$ Recall the $G$-invariant orthonormal frame $\{Y_i\}_{i=1}^n$ in $\mathcal{T}$ and write any $f\in C^\infty(\mathcal{T})$ as $f=\sum f^iY_i$. Using that $U^{\mathcal{T}}_gY_i=Y_i$ we find $$U^{\mathcal{T}}_gf(x)=\sum_{i=1}^n U_gf^i\times Y_i,$$ where $U_g$ is the left quasi-regular representation of $G$ on $C^\infty(\Sigma)$. Thus $U^{\mathcal{T}}_g=\oplus_nU_g$, and the harmonic analysis of $U^{\mathcal{T}}_g$ is the same as that of $U_g$ except that each primary subrepresentation of $U^{\mathcal{T}}_g$ is the $n$-fold copy of the corresponding primary subrepresentation of $U_g$. Making the identification $C^\infty_0(\mathcal{T})\ni f\to \{f^i\}\in\oplus_n C^\infty_0(\Sigma)$ we find the Fourier transform of $f\in C^\infty_0(\mathcal{T})$ to be $$\hat f(\pi)=\oplus_{i=1}^n\hat f^i(\pi),$$ or to say in words, a matrix with $n$ times more columns than that of a scalar function. The inverse Fourier transform will be $$f(x)=\sum_{i=1}^n\int_{\hat G_\Sigma}d\mu(\pi)Tr\left[D_\pi\Pi_\pi\pi^(x)\hat f^i(\pi)\right]\times Y_i(x).$$ On the Fourier transform of distributions ========================================= Here we will collect miscellaneous facts about distributions and their Fourier transform, which we did not meet in the literature. We continue working with the semidirect homogeneous vector bundle $\mathcal{T}$ with notations established earlier. Let $\hat{\mathcal{D}}^{\mathcal{T}}(\hat G_\Sigma)$ be the image of $D(\mathcal{T})=C^\infty_0(\mathcal{T})$ under the harmonic analytical Fourier transform $f(x_\Sigma)\to\hat f(\pi)$. As we have already seen, $\hat f(\pi)=\oplus_n\hat f^i(\pi)$, hence $\hat{\mathcal{D}}^{\mathcal{T}}(\hat G_\Sigma)=\oplus_n\hat{\mathcal{D}}(\Sigma)$, where $\hat{\mathcal{D}}(\Sigma)$ is the image under the Fourier transform of $C^\infty_0(\Sigma)$. $\hat{\mathcal{D}}^{\mathcal{T}}(\hat G_\Sigma)$ inherits the topology of $D(\mathcal{T})$ via the Fourier transform, and one can consider the Fourier transform of distributions $D(\mathcal{T})'\ni u\to\hat u\in\hat{\mathcal{D}}^{\mathcal{T}}(\hat G_\Sigma)'$ given by $\hat u(\hat f)=u(f)$. The Fourier transform has the remarkable property that it interchanges the local and global behaviors. Namely, the local irregularities of a function $f$ are reflected in the decay properties of $\hat f(\pi)$ at large $\pi$, and conversely, the behavior at infinity of $f$ determines the local regularity of $\hat f(\pi)$. The precise description of these phenomena requires a thorough functional analytical investigation, which we, unfortunately, have no possibility to perform here. It is widely known that any distribution restricted to a compact region is of finite order. In [@Gelfand] the general structure of distributions of finite order has been found for $D(\mathbb{R}^n)$. Following a similar pattern we present here a partial generalization of that result. By Proposition \[CountNormTopEquiv\] let us choose the topology $(X_i,2,l^2)$ for convenience. \[FinOrdDistribStruct\] Let $\mathcal{T}_K$ be an $n$-dimensional (complex) pseudo-Riemannian vector bundle over a connected parallelizable (pseudo-)Riemannian manifold $K$, and let $\nabla$ be a fiber metric connection. Every $u\in D(\mathcal{T}_K)'$ of finite order has a representation $$u(f)=\sum_{q\le k}(F_{\alpha,q},P_{\alpha,q}(X_i)f)_2\mbox{, }\forall f\in D(\mathcal{T}),$$ where $F_{\alpha,q}\in L^2(\mathcal{T}_K)$ and the smallest possible such $k$ is the order of $u$. By our choice $$\\|f\\|_k=\sqrt{\sum_{q\le k}\\|P_{\alpha,q}(X_i)f\\|_2^2}.$$ Let $k$ be the order of $u$, i.e., $u$ is continuous in $\\|.\\|_k$-norm. Define the following linear injective map $$\mathcal{V}:D(\mathcal{T}_K)\to\Phi=\bigoplus_{q\le k}L^2(\mathcal{T}_K)$$ by $$\mathcal{V}(f)=\bigoplus_{q\le k}P_{\alpha,q}(X_i)f.$$ Then obviously $\\|\mathcal{V}(f)\\|_\Phi=\\|f\\|_k$. If we denote by $\Psi=\mathcal{V}\left(D(\mathcal{T}_K)\right)\subset\Phi$, then $u\circ\mathcal{V}^{-1}$ is a continuous functional on $\Psi$ with the norm $\\|.\\|_\Phi$, and thus by Hahn-Banach theorem can be extended to a continuous functional $F\in\Phi'$. But $\Phi$ is a Hilbert space, thus $\Phi'=\Phi$ and $F\in\Phi$, and for any $\phi\in\Phi$, $$F(\phi)=\sum_{q\le k}(F_{\alpha,q},\phi_{\alpha,q})_2\mbox{, }F_{\alpha,q}\in L^2(\mathcal{T}).$$ This yields our desired formula $$u(f)=\sum_{q\le k}(F_{\alpha,q},P_{\alpha,q}(X_i)f)_2.$$ If such a formula held for a smaller $k$, then obviously the order of $u$ would be smaller. Several variations of this proposition may be established by choosing different norms. Note that the order of a distribution, if finite, depends on the choice of the family of norms defining the topology. \[LocFinOrd\] As already mentioned, any distribution is [locally]{} of finite order, hence the proposition applies to the restriction $u_K\in C^\infty_0(\mathcal{T}\|_K)'$ of any $u\in D(\mathcal{T})'$ to arbitrary compact connected region $K\subset \Sigma$. We come back to our homogeneous bundle $\mathcal{T}$ and proceed to the Fourier description of distributions $u\in D(\mathcal{T})'$ of finite order, which again can be applied for restrictions to compact regions. \[FinOrdDistribFourierStruct\] Any distribution $u\in D(\mathcal{T})'$ of finite order is given by $$u(f)=\int_{\hat G_\Sigma}d\mu(\pi)Tr\left[\hat u(\pi)^\hat f(\pi)\right],$$ where $\hat u(\pi):\mathbb{C}^{mult(\pi,U_g)n}\to\mathcal{H}_\pi$ is a $\mu$-locally integrable field of Hilbert-Schmidt operators. (Note that the trace operator includes also the summation by fiber indices $i=1,..,n$, which now enumerate blocks of columns.) Let $k$ be the order of $u$. Choose $\{X_i\}$ to be the generators of left translations on $C^\infty(\mathcal{T})$ and let by Proposition \[FinOrdDistribStruct\] write $u$ as $$u(f)=\sum_{q\le k}(F_{\alpha,q},P_{\alpha,q}(X_i)f)_2.$$ Consider the Fourier transform $$\widehat{X_if}(\pi)=\int_\Sigma dx_\Sigma\left(\lim_{t\to 0}\frac{(U_{\exp(-t\xi_i)}-1)f(x_\Sigma)}{t}\right)\pi(x_\Sigma)\Pi_\pi D_\pi$$ where $\xi_i$ is the corresponding element of the Lie algebra of $\Sigma$. The integral runs over a compact region, and is therefore uniformly absolutely convergent with the Hilbert-Schmidt norm, thus we can interchange the limit with the integral, $$\widehat{X_if}(\pi)=\lim_{t\to 0}\frac{1}{t}\int_\Sigma dx_\Sigma(U_{\exp(-t\xi_i)}-1)f(x_\Sigma)\pi(x_\Sigma)\Pi_\pi D_\pi=$$ $$=\lim_{t\to 0}\frac{\pi(\exp(-t\xi_i))-1}{t}\hat f(\pi).$$ On the right hand we see nothing else but the generator of the derived representation of $\pi$, $$\lim_{t\to 0}\frac{\pi(\exp(-t\xi_i))-1}{t}=-\partial_i\pi,$$ whence we find $$\widehat{X_if}(\pi)=-\partial_i\pi\hat f(\pi).$$ As a result we have $$\widehat{P_{\alpha,q}(X_i)f}(\pi)=P_{\alpha,q}(-\partial_i\pi)\hat f(\pi),$$ and thereby $$u(f)=\sum_{q\le k}\int_{\hat G_\Sigma}d\mu(\pi)Tr\left[\hat F_{\alpha,q}(\pi)^P_{\alpha,q}(-\partial_i\pi)\hat f(\pi)\right]=\int_{\hat G_\Sigma}d\mu(\pi)Tr\left[\hat u(\pi)^\hat f(\pi)\right],$$ where $$\hat u(\pi)=\sum_{q\le k}\left[P_{\alpha,q}(-\partial_i\pi)\right]^\hat F_{\alpha,q}(\pi).$$ This completes the proof. Such a result should not be surprising. If the measurable functions $F_{\alpha,q}$ were $q$ times differentiable within the space of locally integrable functions, then we could hypothetically use integration by parts to make all the terms in the formula of Proposition \[FinOrdDistribStruct\] of order 0, which would correspond to a regular distribution. The failure of the derivatives of $F_{\alpha,q}$ to remain locally integrable is reflected in the fact, that multiplication of $\widehat{F}_{\alpha,q}(\pi)$ by $\partial_i\pi^$ makes it not square integrable any more, but possibly only locally integrable. This reflects the local-to-global interchange made by the Fourier transform: higher frequencies feel local irregularities. The image $\hat{\mathcal{D}}^{\mathcal{T}}(\hat G_\Sigma)$ of compactly supported smooth sections under the Fourier transform is of considerable interest. In harmonic analysis it is described by various Paley-Wiener type theorems. Although there are refined Paley-Wiener theorems for adapted Fourier transforms for certain classes of semisimple or solvable groups, there seems to be no such theory for the general abstract setup. Next we present a partial answer to the problem, namely, a criterion for smoothness for sufficiently decaying functions, which gives hints about how the general solution might look like. \[FourImgRapidDecPi\] For a function $f\in L^2(\mathcal{T})$ the following two conditions are equivalent: \(i) for any polynomial $P(X_i)$ of generators $\{X_i\}$ with constant coefficients, $P(X_i)f\in L^2(\mathcal{T})$ \(ii) $\hat f(\pi)$ decays at infinity of $\hat G_\Sigma$ faster than the inverse of any polynomial in the generators $\partial_i\pi^$ As we have seen in the proof of the previous proposition, $$\widehat{P(X_i)f}=P(-\partial_i\pi)^\hat f(\pi),$$ and the requirement that $\widehat{P(X_i)f}\in L^2(\hat G_\Sigma)$ for any $P(X_i)$ is equivalent to the assertion (ii) of the proposition. We can go a step further and establish a weaker necessary condition for a distribution to be given by a smooth integral kernel. For this purpose we want to remind a few definitions on a more abstract level. Let $\mathcal{D}(S)$ be a test function space. We have $\mathcal{D}(S)\subset L^\infty(S)$ and therefore $L^\infty(S)'\subset\mathcal{D}(S)'$. Let $\{\eta_i\}$ be a finite system of linear maps $\eta_i:S\to S$. A distribution $u\in\mathcal{D}(S)'$ is of [rapid decay]{} in $\{\eta_i\}$ if for any polynomial $P(\eta_i)$ of variables $\{\eta_i\}$ it holds $u(P(\eta_i).)\in L^\infty(S)'$. We will symbolically write this as $u=\mathfrak{o}(\{\eta_i\}^{-\infty})$. If $u$ is given by a locally integrable kernel, and $\{\eta_i\}$ are coordinate operators, then this definition obviously reduces to the usual criterion for functions of rapid decay. For a distribution $u\in\mathcal{D}(\mathcal{T})'$ from $\hat{\mathcal{D}}^{\mathcal{T}}(\hat G_\Sigma)'\ni\hat u=\mathfrak{o}(\{\partial_i\pi\}^{-\infty})$ it follows that $u$ has a smooth integral kernel. That $u$ is smooth means that all derivatives of all fiber components $u^j$ are continuous. In other words, for any polynomial $P$ in the generators $\{X_i\}$, the distributions $P(X_i)u^j$ can be evaluated pointwise. A precise statement can be given as follows. $u$ is smooth if and only if for any polynomial $P(X_i)$, point $m\in \Sigma$ and sequence of test functions $f_q\to\delta(x-m)$ in $C_0^\infty(\Sigma)'$, the following limit exists for all $j=1,...,n$ and is finite, $\lim_{q\to\infty}u^j(P(-X_i)f_q)$. The Fourier transform of the distribution $\delta_m=\delta(x-m)$ can be easily read from the Fourier inversion formula, $$\hat\delta_m(\hat f)=\int_{\hat G_\Sigma}d\mu(\pi)Tr\left[D_\pi\Pi_\pi\pi^(m)\hat f(\pi)\right].$$ That means $f_q\to\delta(x-m)$ is equivalent to $\hat f_q\to\pi(m)\Pi_\pi D_\pi$ in the weak sense. Hence $$\widehat{P(-X_i)f_q}\to P(\partial_i\pi)\pi(m)\Pi_\pi D_\pi$$ in the weak topology. It follows $$\begin{aligned} \label{uSmoothCondLim} \lim_{q\to\infty}u^j(P(-X_i)f_q)=\hat u^j\left(P(\partial_i\pi)\pi(m)\Pi_\pi D_\pi\right)\end{aligned}$$ whenever one of the sides converges. Now suppose $\hat u=\mathfrak{o}(\{\partial_i\pi\}^{-\infty})$. Then for any $\hat f\in L^\infty(\hat G_\Sigma)$ (i.e., $\\|\hat f(\pi)\\|\in L^\infty(\hat G_\Sigma)$ in the usual sense) we have $$\hat u^j\left(P(\partial_i\pi)\hat f^j(\pi)\right)<\infty\mbox{, }j=1,...,n.$$ In particular, $\pi(m)\Pi_\pi D_\pi\in L^\infty(\hat G_\Sigma)$, whence (Eq.\[uSmoothCondLim\]) follows. We are incline to think that this necessary condition is not far from the desirable equivalent condition. This is, however, an open problem in harmonic analysis, and we only hope to be able to give a satisfactory answer in the future at least in the context we are interested in. The adapted Fourier transform ============================= We start by noting that because the function $m^\star(t)$ is a function of time only, the eigenfunctions of $D_{\Sigma_t}$ are nothing else but the eigenfunctions of the Laplace operator $\Delta_t$. In the first chapter we introduced the eigenfunction decomposition associated to any self adjoint operator as the Laplace operator $\Delta$, $$f\to\tilde f(\alpha)=\zeta_\alpha(f),$$ where $\zeta_\alpha$-s are the generalized eigenfunctions of $\Delta$. Putting additional structure related with particular geometries one arrives at various Fourier transforms, which are very practical in many respects. On the other hand, the abstract harmonic analytical Fourier transform is a powerful tool for analyzing general problems and properties, but its machinery is functional analytically complicated for use. These two theories are, however, related, although the exact relations have not been sufficiently explored in the literature so far except for compact groups. In the compact case the eigenfunctions of $\Delta$ are the matrix elements of the irreducible representations for some choice of the basis, and the two techniques can be unified. Each choice of the basis results in a Fourier transform which is adapted to it, hence such transforms are sometimes called adapted Fourier transforms. In the non-compact case functional analytical complications arise, though intuitively the situation remains similar. In this section we will try to construct adapted Fourier transforms at least on our semidirect homogeneous bundle $\mathcal{T}$. The Laplace operator $\Delta$ is invariant under $G$ and hence commutes with $U^{\mathcal{T}}_g$. This means on each primary component it acts as a multiplication from the right by a possibly unbounded self-adjoint operator $\hat\Delta(\pi)$, $$\widehat{\Delta f}(\pi)=\hat f(\pi)\hat\Delta(\pi).$$ For any $f\in L^2(\mathcal{T})$ we have that $\Delta f$ is a distribution of order at most two. By Proposition \[FinOrdDistribFourierStruct\] it means that the multiplication of any Hilbert-Schmidt operator $\hat f(\pi)$ by $\hat\Delta(\pi)$ from the right leaves it again Hilbert-Schmidt. Let $\sigma(\pi)\subset\mathbb{R}$ be the spectrum of the self-adjoint operator $\hat\Delta(\pi)$ as acting from the right (this spectrum is non-positive, because $\Delta$ is an elliptic operator). For each $\lambda\in\sigma(\pi)$ let $\hat\xi_{\pi,\lambda,r,s}$ be the generalized eigenfunctions of $\hat\Delta(\pi)$, i.e., distributions satisfying $\hat\xi_{\pi,\lambda,r,s}\hat\Delta(\pi)=\lambda\hat\xi_{\pi,\lambda,r,s}$ which are linearly independent and complete in $\mathcal{H}(\pi)$ for $r\in R_\pi\subset\mathbb{R}$ and $s\in S^n_{\pi,\lambda}\subset\mathbb{R}$ (they can be constructed from delta functions using the spectral theorem). Now consider the following distributions in the Fourier space, $$\hat\zeta_{\pi,\lambda,r,s}(\pi')=\delta(\pi-\pi')\hat\xi_{\pi,\lambda,r,s}.$$ Their preimages are distributions $\zeta_{\pi,\lambda,r,s}\in D(\mathcal{T})'$ which are generalized eigenfunctions of $\Delta$, and by elliptic regularity theorem, are smooth sections in $\mathcal{T}$. Thus we have found, that the adapted Fourier transform $\tilde f(\pi,\lambda,r,s)$ is nothing else but the coefficients of $\hat f(\pi)$ as expended in the system $\hat\xi_{\pi,\lambda,r,s}$. It is worth noting that $r$ parameterizes $\mathcal{H}_\pi$, and $\lambda$, $s$ parameterize $\mathbb{C}^{mult(\pi,U_g)}n$. Actually, $S^n_{\pi,\lambda}$ consists of $n$ copies of some set $S_{\pi,\lambda}$. The choice of the system $\hat\xi_{\pi,\lambda,r,s}$ is rather arbitrary and leaves room for adaptations. The first adaptation we wish to make is the following. For any $\zeta_{\pi,\lambda,r,s}$ we want $\bar\zeta_{\pi,\lambda,r,s}=\zeta_{\pi',\lambda',r',s'}$ for some other parameters. Obviously $\lambda=\lambda'$, and it is easy to see from the Fourier inversion formula, that this amounts to requiring that $\bar\xi_{\pi,\lambda,r,s}$ enters the system $\xi_{\bar\pi,\lambda,r',s'}$ for the representation $\bar\pi$ with some other parameters $r'$, $s'$. The representation $\bar\pi$ may lie in the same equivalence class $[\pi]$ or not. Lie groups are analytic manifolds, and all the group and algebra structure is given by analytic functions in any analytic atlas. In particular, the eigenfunction problem $\Delta\zeta_{\pi,\lambda,r,s}=\lambda\zeta_{\pi,\lambda,r,s}$ is an analytic elliptic equation, and the solutions $\zeta_{\pi,\lambda,r,s}(x)$ are therefore analytic functions in $x$. If $\Sigma$ is compact, then $\hat G_\Sigma$ is discrete, and each $\sigma(\pi)$ is also discrete. Representations are finite dimensional, hence $r$ and $s$ run over finite sets. The set $\tilde\Sigma=\{\alpha=(\pi,\lambda,r,s)\}$ can be considered a discrete manifold symbolically divided into $n$ components as corresponding to each copy of $S_{\pi,\lambda}$. The space $\hat{\mathcal{D}}^{\mathcal{T}}(\hat G_\Sigma)$ corresponds now to the space $\tilde D(\tilde\Sigma)$ of functions on $\tilde\Sigma$, which are of rapid decay in $\lambda$, and also decay sufficiently fast in $\pi$ by Proposition \[FourImgRapidDecPi\]. If $\Sigma$ is non-compact, suppose there exists a subset $\tilde K\subset\hat G_\Sigma$ such that $\mu(\hat G_\Sigma\setminus\tilde K)=0$ and $\tilde K$ can be cast into an analytic manifold. Then we can restrict our Fourier transform from $\hat G_\Sigma$ to $\tilde K$ without violation of the Plancherel equality. Suppose further that the set $\tilde\Sigma=\{\alpha=(\pi,\lambda,r,s)\}$ can be made an analytic manifold consisting of $n$ disjoint components as in the compact case. Each component itself may have several connected components if $1<mult(\pi,U_g)<\infty$, in which case $s$ will run over a discrete set. Then we can choose $\zeta_{\pi,\lambda,r,s}$ to be analytic in all its parameters (if $s$ is discrete, analyticity in $s$ is void), so that $\hat{\mathcal{D}}^{\mathcal{T}}(\hat G_\Sigma)$ will correspond to the space $\tilde D(\tilde\Sigma)$ of some analytic functions on $\tilde\Sigma$ which have at least above mentioned decay properties in $\lambda$ and $\pi$, but also are $L^2$ in $r$, and in $s$ if the latter is continuous. Finally let us define a symbolic involution $\alpha\to-\alpha$ on $\tilde\Sigma$ satisfying $\zeta_{-\alpha}=\bar\zeta_\alpha$. Clearly this involution will preserve $\lambda$. Now if the necessary assumptions are satisfied, we arrive at a conventional Fourier transform. In the next publication we will see that in the majority of situations in cosmology these assumptions are valid, and that will enable us to exploit the machinery of mode decomposition to our cosmological models. Invariant bi-distributions -------------------------- In this section we will try to analyze the structure of bi-distributions $w\in\left(\mathcal{D}(\mathcal{T})\otimes\mathcal{D}(\mathcal{T})\right)'$ which are invariant under the left quasi-regular action $U^{\mathcal{T}}_g$ of $G$ on $\mathcal{D}(\mathcal{T})$, $$w(U^{\mathcal{T}}_gf,U^{\mathcal{T}}_gh)=w(f,h)\mbox{, }\forall f,h\in\mathcal{D}(\mathcal{T}),$$ and compare with results obtained earlier in the literature. Decomposing each $f=\sum f^iY_i$, $f^i\in C^\infty_0(\Sigma)$, we find for $u\in\mathcal{D}(\mathcal{T})'$ and $w\in\left(\mathcal{D}(\mathcal{T})\otimes\mathcal{D}(\mathcal{T})\right)'$ $$u(f)=\sum_{i=1}^nu^i(f^i)\mbox{, }w(f,h)=\sum_{i,j=1}^nw^{ij}(f^i,h^j),$$ $$u^i\in C^\infty_0(\Sigma)'\mbox{, }w^{ij}\in\left(C^\infty_0(\Sigma)\otimes C^\infty_0(\Sigma)\right)',$$ so that the problem reduces to that for scalar distributions. The following proposition establishes the general form of the $G$-invariant (or homogeneous) bi-distributions. Our approach is greatly inspired by [@Gelfand_Vilenkin1964] where this analysis is performed for $\mathbb{R}^n$. \[InvBiDistrib\] Every $w\in \left(C^\infty_0(\Sigma)\otimes C^\infty_0(\Sigma)\right)'$ satisfying $w(U_gf,U_gh)=w(f,h)$, $\forall f,h\in C^\infty_0(\Sigma)$, $g\in G$, has the form $$w(f,h)=u_w\left(\bar f^\ast h\right)$$ for some $u_w\in C^\infty_0(\Sigma)'$. And conversely, any $u_w\in C^\infty_0(\Sigma)'$ gives rise to such an invariant bi-distribution $w$. Recall that for scalar functions $U_gf(x_\Sigma O)=f(g^{-1}x_\Sigma O)$. By the nuclear theorem $w$ can be uniquely extended to $\tilde w\in C^\infty_0(\Sigma\times \Sigma)'$ via embedding $$C^\infty_0(\Sigma)\otimes C^\infty_0(\Sigma)\ni f(x_\Sigma)\otimes h(y_\Sigma)\to f(x_\Sigma)h(y_\Sigma)\in C^\infty_0(G\times G).$$ That $$w(f(g^{-1}x_\Sigma O),h(g^{-1}y_\Sigma O))=w(f,h)$$ by continuity implies that $$\tilde w(\phi(g^{-1}x_\Sigma O,g^{-1}y_\Sigma O))=\tilde w(\phi(x_\Sigma,y_\Sigma))\mbox{, } \forall\phi\in C^\infty_0(\Sigma\times \Sigma).$$ Define the linear automorphism $$C^\infty_0(\Sigma\times \Sigma)\ni\phi(x_\Sigma,y_\Sigma)\to\psi_\phi(x_\Sigma,y_\Sigma)\in C^\infty_0(\Sigma\times \Sigma)$$ by $$\psi_\phi(x_\Sigma,y_\Sigma)=\int_Odx_O\phi(x_\Sigma,x_\Sigma x_Oy_\Sigma O)=\int_Odx_O\phi(x_\Sigma,x_\Sigma Ox_Oy_\Sigma O),$$ and the pullback of $\tilde w$ under this automorphism by $\tilde v$, $\tilde v(\psi_\phi)=\tilde w(\phi)$. If $\phi_g(x_\Sigma,y_\Sigma)=\phi(g^{-1}x_\Sigma O,g^{-1}y_\Sigma O)$ then $$\psi_{\phi_g}(x_\Sigma,y_\Sigma)=\int_Odx_O\phi(g^{-1}x_\Sigma O,g^{-1}x_\Sigma Ox_Oy_\Sigma O)=\psi_\phi(g^{-1}x_\Sigma O,y_\Sigma).$$ Now $$\tilde w(\phi_g)=\tilde v(\psi_{\phi_g})=\tilde v(\psi_\phi)=\tilde w(\phi),$$ thus $$\tilde v(\psi_\phi(x_\Sigma,y_\Sigma))=\tilde v(\psi_\phi(g^{-1}x_\Sigma O,y_\Sigma))\mbox{, }\forall g\in G.$$ Consider the restriction $v$ of $\tilde v$ to $C^\infty_0(\Sigma)\otimes C^\infty_0(\Sigma)$. The last equation implies $v(f(g^{-1}x_\Sigma O),h(y_\Sigma))=v(f,h)$, $\forall f,h\in C^\infty_0(\Sigma)$. If we fix $h$, then $v(.,h)\in C^\infty_0(\Sigma)'$ is a distribution which is invariant under all translations, and is thus given by a constant kernel, $v(f,h)=u_w(h)\int_\Sigma dx_\Sigma f(x_\Sigma)$, for some $u_w:C^\infty_0(\Sigma)\to\mathbb{C}$. On the other hand, if we fix $f$, then continuity in $h$ implies $u_w\in C^\infty_0(\Sigma)'$. Because the integral $\int_\Sigma dx_\Sigma f(x_\Sigma)$ runs over a compact region, it can be transferred into $u_w$, i.e., $v(f,h)=u_w\left(\int_\Sigma dx_\Sigma f(x_\Sigma)h(y_\Sigma)\right)$. This in turn implies by continuity, that $\tilde v(\psi(x_\Sigma,y_\Sigma))=u_w\left(\int_\Sigma dx_\Sigma\psi(x_\Sigma,y_\Sigma)\right)$. Finally we arrive at $$w(f,h)=\tilde w(f(x_\Sigma)h(y_\Sigma))=\tilde v(f(x_\Sigma)\int_Odx_Oh(x_\Sigma x_oy_\Sigma O))=$$ $$=u_w\left(\int_\Sigma dx_\Sigma f(x_\Sigma)\int_Odx_Oh(x_\Sigma x_Oy_\Sigma O)\right)=u_w(\bar f^\ast h).$$ The converse statement is obvious. For a distribution $w\in\left(\mathcal{D}(\mathcal{T})\otimes\mathcal{D}(\mathcal{T})\right)'$ this will mean $$w(f,h)=\sum_{i,j=1}^nu_w^{ij}\left((\bar f^i)^\ast h^j\right).$$ Note that every $G$-invariant bi-distribution $w\in\left(C^\infty_0(\Sigma)\otimes C^\infty_0(\Sigma)\right)'$ is in particular $\Sigma$-invariant. Let $f\ast h$ ($f\star h$) and $f^$ ($f^\star$) denote the convolution and the involution with respect to $G$ ($\Sigma$), respectively. Then $$w(f,h)=u_w\left(\int_\Sigma dx_\Sigma f(x_\Sigma)\int_Odx_Oh(x_\Sigma x_Oy_\Sigma O)\right)=$$ $$=u_w\left(\int_Odx_OL_{x_O^{-1}}\bar f^\star\star h(y_\Sigma O)\right)=u'_w(\bar f^\star\star h)$$ for some other $u'_w\in\left(C^\infty_0(\Sigma)\otimes C^\infty_0(\Sigma)\right)'$ as expected. Let $\hat{\mathcal{D}}(\hat G_\Sigma)$ be the image of $C^\infty_0(\Sigma)$ under the harmonic analytical Fourier transform $f(x_\Sigma)\to\hat f(\pi)$. As an obvious corollary we arrive at the form of an invariant bi-distribution in the Fourier space. \[InvBiDistribFourier\] A $G$-invariant bi-distribution $w\in\left(C^\infty_0(\Sigma)\otimes C^\infty_0(\Sigma)\right)'$ in the Fourier space is given by $$w(f,h)=\hat u_w(\pi(\bar f^)\hat h(\pi))=\hat u_w(\pi(\bar f)^\hat h(\pi))$$ for some $\hat u_w\in\hat{\mathcal{D}}(\hat G_\Sigma)'$. An immediate consequence of Proposition \[FinOrdDistribFourierStruct\] is the following \[InvFinOrdBiDistribFourier\] Under the assumptions of Proposition \[FinOrdDistribFourierStruct\], a $G$-invariant bi-distribution $w_K\in\left(C^\infty_0(K)\otimes C^\infty_0(K)\right)'$ with $K\subset \Sigma$ compact is given by $$w_K(f,h)=\int_{\hat G_\Sigma}d\mu(\pi)Tr\left[\hat u_K(\pi)^\pi(\bar f)^\hat h(\pi)\right].$$ It suffices to note that $$\mbox{supp}\{f\ast h\}\subset O(\mbox{supp}\{f\})^{-1}\mbox{supp}\{h\}O,$$ and to apply Proposition \[FinOrdDistribFourierStruct\]. Finally we establish a generalization of the results by [@Lueders_Roberts_1990] for FRW spacetimes. \[FinMultUnimodScalar\] Suppose that the group $G$ is such that all multiplicities $mult(\pi,U_g)$ are finite. Then any $G$-invariant bi-distribution $w\in\left(D(\mathcal{T})\otimes D(\mathcal{T})\right)'$ has the form $$w(f,h)=\int_{\hat G_\Sigma}d\mu(\pi)Tr\left[(\hat{\bar{f}}(\pi)\hat u(\pi))^\hat h(\pi)\right],$$ where $\hat u(\pi)$ is a $\mu$-locally measurable field of $\left[mult(\pi,U_g)\cdot n\right]\times\left[mult(\pi,U_g)\cdot n\right]$ complex matrices. Let start with the case $w\in\left(C^\infty_0(\Sigma)\otimes C^\infty_0(\Sigma)\right)'$. The condition that the modular function of $\Sigma$ has a nontrivial kernel ensures that the formula (7.49) of (Folland) is valid, so that for $\mu$-almost all $\pi$ the operator $D_\pi$ is invertible (injective). Therefore we can write $\pi(f)=\hat f(\pi)D_\pi^{-1}$, so that $\pi(\bar f)^\hat h(\pi)=D_\pi^{-1}\hat{\bar{f}}(\pi)^\hat h(\pi)$ where $\hat{\bar{f}}(\pi)^\hat h(\pi)$ is a $mult(\pi,U_g)\times mult(\pi,U_g)$ complex matrix. Now for any compact $K\subset \Sigma$ by Corollary \[InvFinOrdBiDistribFourier\] we find that the restriction $w_K$ of $w$ to $C^\infty_0(K)\otimes C^\infty_0(K)$ is given by $$w_K(f,h)=\int_{\hat G_\Sigma}d\mu(\pi)Tr\left[\hat u_K'(\pi)^D_\pi^{-1}\hat{\bar{f}}(\pi)^\hat h(\pi)\right]=\int_{\hat G_\Sigma}d\mu(\pi)Tr\left[\hat u_K(\pi)^\hat{\bar{f}}(\pi)^\hat h(\pi)\right],$$ where $\hat u_K(\pi)$ is a $mult(\pi,U_g)\times mult(\pi,U_g)$ complex matrix. Choosing a larger compact $K\subset K'\subset \Sigma$ we will arrive at another matrix $\hat u_{K'}(\pi)$. But when restricted to $K$, $w_{K'}$ must coincide with $w_K$, hence $\hat u_{K'}(\pi)=\hat u_K(\pi)$. Thus the matrix $\hat u_K(\pi)$ is the same for any $K$, and the formula holds for the entire $w$. Now for $w\in\left(D(\mathcal{T})\otimes D(\mathcal{T})\right)'$ we have $$w(f,h)=\int_{\hat G_\Sigma}d\mu(\pi)\sum_{i,j=1}^nTr\left[(\hat{\bar{f}}^i(\pi)\hat u^{ij}(\pi))^\hat h^j(\pi)\right]=\int_{\hat G_\Sigma}d\mu(\pi)Tr\left[(\hat{\bar{f}}(\pi)\hat u(\pi))^\hat h(\pi)\right],$$ which completes the proof. In the case of FRW spacetimes all the assumptions of the last proposition are satisfied. In particular all $mult(\pi,U_g)=1$ and for the scalar case we find that any $G$-invariant bi-distribution is given by a locally measurable scalar field $\hat u(\pi)$. Acknowledgements ================ The author expresses his thank to the Institute for Mathematics in the Sciences and in particular to the International Max Planck Research School for the organizational and material support of the PhD project which this work is a part of. Invaluable is the scientific patronage of the supervisor of this PhD project Prof. Dr. Rainer Verch, whose remarks and suggestions used in this work are too many to be listed here. Special gratitude is addressed to Prof. Dr. Gerald Folland for extremely helpful e-mail correspondence. Space structures. Distributions =============================== Let us start with introducing symmetric metric products $$\langle f,h\rangle_M=\int_Md\mu_g(x)\langle f(x),h(x)\rangle_g\mbox{, }f\in\mathcal{E}(\mathcal{T})\mbox{, }h\in\mathcal{D}(\mathcal{T}),$$ $$\langle f,h\rangle_{\Sigma_t}=\int_\Sigma d\mu_h(\vec x)\langle f(\vec x),h(\vec x)\rangle_g\mbox{, }f\in\mathcal{E}(\mathcal{T}_t)\mbox{, }h\in\mathcal{D}(\mathcal{T}_t).$$ The pseudo-Riemannian metric $\langle,\rangle_\mathfrak{g}$ induces a Krein space structure on $V$, the typical fiber of $\mathcal{T}$. Whence there is a Krein involution $\check\Gamma$, such that $(u,v)_\mathfrak{g}=\langle\bar u,\check\Gamma v\rangle_\mathfrak{g}$, $u,v\in V$, is a positive definite hermitian inner product. This gives rise to positive definite hermitian inner products $$(f,h)_M=\int_Md\mu_g(x)(f(x),h(x))_\mathfrak{g}\mbox{, }f\in\mathcal{E}(\mathcal{T})\mbox{, }h\in\mathcal{D}(\mathcal{T}),$$ $$(f,h)_{\Sigma_t}=\int_\Sigma d\mu_h(\vec x)(f(\vec x),h(\vec x))_\mathfrak{g}\mbox{, }f\in\mathcal{E}(\mathcal{T}_t)\mbox{, }h\in\mathcal{D}(\mathcal{T}_t).$$ The completion of spaces $\mathcal{D}(\mathcal{T})$ and $\mathcal{D}(\mathcal{T}_t)$ with respect to these products becomes the Hilbert spaces $L^2(\mathcal{T})$ and $L^2(\mathcal{T}_t)$, respectively. The tangent space $T_pM$ at a point $p\in M$ with the Lorentzian metric $g$ is another example of a Krein space. In the same spirit one defines the positive definite inner product $(,)_g$ on $T_pM$. The metric $h$ on $T_p\Sigma$ is Riemannian, so the construction of $(,)_h$ is straightforward. Note that $\langle,\rangle_\mathfrak{g}$ and $g$ together give pseudo-Riemannian metrics on all product bundles $T^M\otimes...\otimes T^M\otimes\mathcal{T}$ (respectively, $\langle,\rangle_\mathfrak{g}$ and $h$ on $T^\Sigma\otimes...\otimes T^\Sigma\otimes\mathcal{T}_t$). All the resulting standard fibers are again Krein spaces, and can be given inner products $(,)_g$ in the same fashion. These in their turn produce products $(,)_M$ and $(,)_{\Sigma_t}$ on the respective sections. The perfect countably Banach topology of the test function spaces $\mathcal{D}(\mathcal{T})$ and $\mathcal{D}(\mathcal{T}_t)$ can be given as usual (e.g.,[@BarGinouxPfaffle200703]). However, as we are going to perform a spectral analysis, we will need nuclear countably Hilbert space structure, to which we proceed [@Maurin1972]. Let $\mathcal{O}\subset M$ be a compact region. Let $$\mathcal{D}_\mathcal{O}(\mathcal{T})=\left\{f\in\mathcal{D}(\mathcal{T})\mbox{: }supp\{f\}\subset\mathcal{O}\right\}$$ and define the family of positive definite inner products $(,)_{\mathcal{O},p}$ on $\mathcal{D}_\mathcal{O}(\mathcal{T})$ by $$(f,h)_{\mathcal{O},p}=\sum_{q\le p}((\nabla)^qf,(\nabla)^qh)_M\mbox{, }\forall f,h\in\mathcal{D}_\mathcal{O}(\mathcal{T})\mbox{, }p,q\in\mathbb{N},$$ which induces a family of norms $\\|.\\|_{\mathcal{O},p}$. One can show that this family of norms is growing and consistent, and gives the same topology as the usual one. Let us give $\mathcal{D}_\mathcal{O}(\mathcal{T})$ a countably Hilbert space structure in the following sense, $$\mathcal{D}_\mathcal{O}(\mathcal{T})=\bigcap_{\mathbb{N}}\overline{\mathcal{D}_\mathcal{O}(\mathcal{T})}^{(,)_{\mathcal{O},p}}.$$ It can be shown, that thus constructed countably Hilbert space $\mathcal{D}_\mathcal{O}(\mathcal{T})$ is nuclear. Let now $$\mathcal{O}_1\subset...\subset\mathcal{O}_n\subset...\subset M$$ be an infinite family of growing compact regions. Then give $\mathcal{D}(\mathcal{T})$ the inductive limit topology $$\mathcal{D}(\mathcal{T})=\lim_{n\to\infty}\mathcal{D}_{\mathcal{O}_n}(\mathcal{T}).$$ Here we are done. Distributions $\mathcal{D}(\mathcal{T})'$ and operations on them can be defined as usual. The same construction can be done for $\mathcal{D}(\mathcal{T}_t)$ with minor modifications. At the end let us consider the choice of the topology in detail. In the literature one usually chooses the family of norms $\\|.\\|_p$ (or sometimes a family of seminorms $\|\!(.)\!\|_p$; from these seminorms one can make norms by $\\|.\\|_p=\sum_{q<p}\|\!(.)\!\|_q$ or $\\|.\\|_p=\sup_{q<p}\|\!(.)\!\|_q$ etc.) rather arbitrarily in accordance with the setup of the problem, and it is tacitly assumed but not everywhere proven, that all such choices give equivalent topologies. Let us for consistency present here a proof of this fact. The zest of the proof (the usage of the Sobolev embedding theorem) was suggested by G. Folland. \[CountNormTopEquiv\] Let $\mathcal{T}\xrightarrow{\pi}M$ be an $n$ dimensional pseudo-Riemannian vector bundle over the $d$-dimensional parallelizable pseudo-Riemannian manifold $M$ with positive metric product $(,)_\mathfrak{g}$ constructed as above, so that we have well defined $L^m$ norms $\|\\|.\\|\|_m$ for $1\le m\le\infty$ on $\mathcal{D}(\mathcal{T})$. Let $\nabla$ be a connection on $\mathcal{T}$. Let $X_1...X_d$ be a system of first order smooth differential operators on $C^\infty(\mathcal{T})$ which span the tangent space $T^M$ everywhere the seminorms be given by $\|\!( f)\!\|_{\alpha,q}=\|\\|P_{\alpha,q}(X_i)f\\|\|_m$, where $P_{\alpha,q}(X_i)$ are various monomials of order $q$ in $\{X_i\}$, $f\in\mathcal{D}(\mathcal{T})$ the family of norms be given as $\\|f\\|_p=\|\\|\{\|\!(f)\!\|_{\alpha,q}\}_{q\le p}\\|\|_{l^k}$, or by a superposition of different $\|\\|.\\|\|_{l^k}$, $1\le k\le\infty$. Then the topology of $\mathcal{D}(\mathcal{T})$ defined by this family of norms is independent of the decisions (i) to (iii). For convenience denote by $(X_i,m,\ast)$ the triple of choices at points (i),(ii) and (iii). Then $(X_i,m,\ast)\sim(X_i',m',\ast')$ will mean that this two topologies are equivalent. As the topology of $\mathcal{D}(\mathcal{T})$ is the inductive limit of various $\mathcal{D}(\mathcal{T}_K)$ with $\mathcal{T}_K=\pi^{-1}(K)$, $K\subset M$ compact, it suffices to prove the assertion for an arbitrary $\mathcal{D}(\mathcal{T}_K)$. The topologies given by two families of norms $\{\\|.\\|_p\}$ and $\{\\|.\\|_p'\}$ are equivalent if and only if these two systems of norms are themselves equivalent, i.e., $\forall p$, $\exists q(p),r(p)>0$, $0<C_p,C_p'\in\mathbb{R}$ such that $\\|.\\|_p\le C_p\\|.\\|_{q(p)}$ and $\\|.\\|_p'\le C_p'\\|.\\|_{r(p)}'$. Let us start with the point (iii). Suppose the choices (i) and (ii) are fixed, i.e., consider $(X_i,m,\ast)$ and $(X_i,m,\ast')$. Then all possible choices in (iii) give equivalent systems of norms because of the elementary inequalities $$\|\\|\{\|\!(f)\!\|_{\alpha,q}\}_I\\|\|_{l^\infty}\le...\le\|\\|\{\|\!(f)\!\|_{\alpha,q}\}_I\\|\|_{l^k}\le...\le\|\\|\{\|\!(f)\!\|_{\alpha,q}\}_I\\|\|_{l^1}\le N_I\|\\|\{\|\!(f)\!\|_{\alpha,q}\}_I\\|\|_{l^\infty},$$ where $N_I$ is the number of terms in the index set $I$. These inequalities can be applied consecutively to estimate any composite norm by, say, $\|\\|.\\|\|_{l^\infty}$. An example of a composite norm is $\\|f\\|_p=\sup_{q\le p}\|\\|\nabla^q f\\|\|_\infty$. We found that $(X_i,m,\ast)\sim(X_i,m,\ast')$. Now let $1\le m\le\infty$ at (ii) and $k=\infty$ at (iii) be chosen, and choose two systems of operators $\{X_i\}$ and $\{Y_i\}$ at point (i) to construct the families of norms $\{\\|.\\|_p\}$ and $\{\\|.\\|_p'\}$, respectively. This corresponds to $(X_i,m,l^\infty)$ and $(Y_i,m,l^\infty)$. Because $\{X_i\}$ spans $T^M$, there are functions $c_{ij}(x)\in C^\infty(M)$ and smooth fields of homomorphisms $\tilde\Gamma_i\in C^\infty(Hom(\mathcal{T},\mathcal{T}))$ with $Y_i(x)=\sum_jc_{ij}(x)X_j(x)+\tilde\Gamma_i$. Using this for any monomial $P_{\alpha,q}(Y_i)$ we get $$P_{\alpha,q}(Y_i)f=\sum_\beta c_{\alpha,q}^\beta(x)Q_{\alpha,q}^\beta(X_i)f,$$ where $c_{\alpha,q}^\beta(x)\in C^\infty(M)$ and $Q_{\alpha,q}^\beta(X_i)$ are monomials of order less or equal $q$. The number of summands is less than, say, $(4d)^q$. It follows by Minkowsky inequality $$\|\!( f)\!\|_{\alpha,q}'=\|\\|P_{\alpha,q}(Y_i)f\\|\|_m\le\sum_\beta\|\\|c_{\alpha,q}^\beta(x)Q_{\alpha,q}^\beta(X_i)f\\|\|_m,$$ and then by Hölder inequality $$\sum_\beta\|\\|c_{\alpha,q}^\beta(x)Q_{\alpha,q}^\beta(X_i)f\\|\|_m\le C_{\alpha,q}\sum_\beta\|\\|Q_{\alpha,q}^\beta(X_i)f\\|\|_m=C_{\alpha,q}\sum_\beta\|\!( f)\!\|_{\alpha(\alpha,q,\beta),q(\alpha,q,\beta)},$$ where $0<C_{\alpha,q}=\sup_\beta\|\\|c_{\alpha,q}^\beta\\|\|_\infty$. In other words, the seminorms of order $q$ of the second system can be estimated by linear combinations of seminorms of the first system of the same or lower order. Then $$\\|f\\|_p'=\sup_{q\le p}\|\!(f)\!\|_{\alpha,q}'\le C_p\sup_{q\le p}\sum_\beta\|\!( f)\!\|_{\alpha(\alpha,q,\beta),q(\alpha,q,\beta)}\le$$ $$\le C_p(4d)^p\sup_{q\le p}\|\!( f)\!\|_{\alpha(\alpha,q,\beta),q(\alpha,q,\beta)}\le C_p(4d)^p\sup_{q\le p}\|\!(f)\!\|_{\alpha,q}=C_p(4d)^p\\|f\\|_p,$$ where $0<C_p=\sup_{q\le p}C_{\alpha,q}$. For the other direction of the estimate we simply need to switch $\{X_i\}$ and $\{Y_i\}$. Thus these two topologies are equivalent, $(X_i,m,l^\infty)\sim(Y_i,m,l^\infty)$. Finally let $X_i=\nabla_i$ (components with respect to a global orthonromal frame in $T^M$) be chosen at (i), and $\\|.\\|=\|\\|\{\|\!(.)\!\|_{\alpha,q}\}_{q\le p}\\|\|_{l^2}$ at (iii). We construct two families of norms by choosing $1\le m<\infty$ and $m'=\infty$ at (ii) for $\\|.\\|_p$ and $\\|.\\|_p'$, respectively. This can be symbolized as $(\nabla_i,m,l^2)$ and $(\nabla_i,\infty,l^2)$. Because $K$ is compact, by an application of Hölder inequality we obtain $$\|\\|.\\|\|_m\le C_m\|\\|.\\|\|_\infty$$ for some $0<C_m\in\mathbb{R}$, and hence obviously $$\\|.\\|_p\le C_m\\|.\\|_p'\mbox{, }p\in\mathbb{N}_0.$$ The opposite inequality requires an application of Sobolev embedding theorem for compact manifolds [@Hebey1999],[@Taylor1991]. Denote the Sobolev norms (which are equivalent to those in [@Hebey1999]) $$\|\\|f\\|\|_{W^{p,m}}=\sqrt{\sum_{q\le p}\|\\|\nabla^q f\\|\|_m^2}.$$ Then an application of Sobolev embedding theorem gives $$\|\\|.\\|\|_{W^{0,\infty}}=\|\\|.\\|\|_\infty\le D\|\\|.\\|\|_{W^{d,1}}$$ for some $0<D\in\mathbb{R}$. By another application of Hölder inequality we find $$\|\\|.\\|\|_{W^{d,1}}\le\|\\|.\\|\|_{W^{d,2}},$$ and therefore $$\|\\|.\\|\|_\infty\le D\sqrt{\sum_{q\le d}\|\\|\nabla^q f\\|\|_2^2}.$$ Next $$\|\\|\nabla^q f\\|\|_2^2=\sum_\alpha\|\\|P_{\alpha,q}(X_i)f\\|\|_2^2,$$ and finally $$\\|f\\|_p'=\sqrt{\sum_{q\le p}\|\\|P_{\alpha,q}(X_i)f\\|\|_\infty^2}\le D\sqrt{\sum_{q\le p}\sum_{j\le d}\|\\|\nabla^j P_{\alpha,q}(X_i)f\\|\|_2^2}=$$ $$=D\sqrt{\sum_{q\le p}\sum_{j\le d}\|\\|P_{\beta,j}(X_i)P_{\alpha,q}(X_i)f\\|\|_2^2}\le D\sqrt{\sum_{q\le p+d}\|\\|P_{\alpha,q}(X_i)f\\|\|_2^2}\le$$ $$\le DE\sqrt{\sum_{q\le p+d}\|\\|P_{\alpha,q}(X_i)f\\|\|_m^2}=DE\\|f\\|_{p+d}',$$ where in the last inequality again Hölders inequality was used with some $0<E\in\mathbb{R}$. Thus we have shown that choosing any $1\le m<\infty$ is equivalent to choosing $m=\infty$ at point (ii), i.e., $(\nabla_i,m,l^2)\sim(\nabla_i,\infty,l^2)$. Write $$(X_i,m,\ast)\sim(X_i,m,l^\infty)\sim(\nabla_i,m,l^\infty)\sim(\nabla_i,m,l^2)\sim(\nabla_i,\infty,l^2)\sim(\nabla_i,m',l^2)\sim$$ $$\sim(\nabla_i,m',l^\infty)\sim(X_i',m',l^\infty)\sim(X_i',m',\ast').$$ The proof is complete. On the time dependent harmonic oscillator ========================================= Here we will concentrate on some properties of the solutions of the smooth complex time dependent harmonic oscillator equation $$\begin{aligned} \ddot T(s)+\Lambda(s)T(s)=0\label{THDO}\end{aligned}$$ where $\Lambda(s)$ is a smooth complex function on the real line. This equation is under attention since a long time, but some results are not that easily available today (at least for us). We start with an easy remark. Denote by $$W[Q,R](s)=\begin{pmatrix} Q(s) & \dot Q(s)\\ R(s) & \dot R(s) \end{pmatrix}$$ the Wronski matrix of two solutions $Q$ and $R$. \[TbyW\_QR\] Let $Q,R$ be two linearly independent solutions of (Eq.\[THDO\]), and $T$ an arbitrary solution. Then from the conservation of $\det W[Q,T]$ and $\det W[R,T]$ it is easy to find $$\begin{pmatrix} \dot T(s)\\ -T(s) \end{pmatrix} =W[Q,R]^{-1}(s)\times W[Q,R](0)\times \begin{pmatrix} \dot T(0)\\ -T(0) \end{pmatrix}=$$ $$=\det W[Q,R]^{-1}(0) \begin{pmatrix} \dot R(s) & -\dot Q(s)\\ - R(s) & Q(s) \end{pmatrix}\times \begin{pmatrix} Q(0) & \dot Q(0)\\ R(0) & \dot R(0) \end{pmatrix}\times \begin{pmatrix} \dot T(0)\\ -T(0) \end{pmatrix}.$$ Thus having at hand two such particular solutions $Q,R$, we have a control over arbitrary solutions $T$ in terms of their initial data. Our first task is to obtain a control over the magnitude of the solution $T$ on a given compact interval $\mathcal{R}$ in terms of its initial data $T(0)$ and $\dot T(0)$. This is done by the so called energy estimate. Define the energy of a solution $T$ by $$\mathcal{W}[\breve T](s)=\frac{1}{2}\|\dot T\|^2(s)+\frac{1}{2}\Re\Lambda(s)\|T(s)\|^2.$$ If $\Re\Lambda>0$ on $\mathcal{R}$ then $2\mathcal{W}[\breve T]$ dominates $\Re\Lambda\|T\|^2$ and $\|\dot T\|^2$, and obtaining bounds on $\mathcal{W}[\breve T]$ we automatically get bounds on $\|T\|$ and $\|\dot T\|$. \[EnergyEstCompl\] For arbitrary solution $T$ of $$\ddot T(s)+\Lambda(s)T(s)=0,$$ with smooth complex valued $\Lambda(s)$ having a positive real part (i.e., $\Re\Lambda(s)>0$) on a compact interval $\mathcal{R}$, the energy function $\mathcal{W}[\breve T](s)$ satisfies the estimate $$\mathcal{W}[\breve T](0)e^{-\int_0^sd\sigma(\frac{2\|\Im\Lambda(\sigma)\|}{\sqrt{\Re\Lambda(\sigma)}}+\|\partial_s\ln\Re\Lambda(\sigma)\|)}\le\mathcal{W}[\breve T](s)\le\mathcal{W}[\breve T](0)e^{\int_0^sd\sigma(\frac{2\|\Im\Lambda(\sigma)\|}{\sqrt{\Re\Lambda(\sigma)}}+\|\partial_s\ln\Re\Lambda(\sigma)\|)}$$ for all $s\in\mathcal{R}$. Write $T(s)=R(s)+iS(s)$, $\Lambda(s)=\Theta(s)+i\Xi(s)$, and insert into the equation. We will get the following system of real equations, $$\begin{cases} \ddot R(s)+\Theta(s)R(s)-\Xi(s)S(s)=0,\\ \ddot S(s)+\Theta(s)S(s)+\Xi(s)R(s)=0. \end{cases}$$ We can cast this into a real vector equation $$\ddot{\breve{T}}(s)+\hat\Lambda(s)\breve T(s)=0$$ by denoting $$\breve T(s)=(R(s),S(s))^\top,$$ and $$\hat\Lambda(s)= \begin{pmatrix} \Theta(s) & -\Xi(s)\\ \Xi(s) & \Theta(s) \end{pmatrix}=\hat\Lambda^+(s)+\hat\Lambda^-(s)= \Theta(s){{\bf 1}}+\Xi(s)\begin{pmatrix} 0 & -1\\ 1 & 0 \end{pmatrix},$$ where $\hat\Lambda^\pm$ denote the symmetric and antisymmetric parts. The energy function equals $$\mathcal{W}[\breve T](s)=\frac{1}{2}\dot{\breve{T}}^2(s)+\frac{1}{2}\breve T^\top(s)\hat\Lambda(s)\breve T(s)=\frac{1}{2}\dot{\breve{T}}^2(s)+\frac{1}{2}\breve T^\top(s)\hat\Lambda^+(s)\breve T(s).$$ On the interval $\mathcal{R}$ we have $\mathcal{W}[\breve T](s)>0$ as by the assumption $\Theta(s)>0$. One can easily find that $$\dot{\mathcal{W}}[\breve T](s)=\breve T^\top(s)\hat\Lambda^-(s)\dot{\breve{T}}(s)+\frac{1}{2}\breve T^\top(s)\dot{\hat{\Lambda}}^+(s)\breve T(s),$$ whence it follows $$\left\|\dot{\mathcal{W}}[\breve T](s)\right\|\le\|\Xi(s)\|\|\breve T(s)\|\|\dot{\breve{T}}(s)\|+\|\partial_s\ln\Theta(s)\|\mathcal{W}[\breve T](s).$$ By definition of $\mathcal{W}[\breve T]$ and positivity of $\Theta$ we have $\|\dot{\breve{T}}(s)\|\le\sqrt{2\mathcal{W}[\breve T](s)}$ and $\|\breve T(s)\|\le\sqrt{2\mathcal{W}[\breve T](s)/\Theta(s)}$ on $\mathcal{R}$. It follows then $$\left\|\partial_s\ln\mathcal{W}[\breve T](s)\right\|\le\frac{2\|\Xi(s)\|}{\sqrt{\Theta(s)}}+\|\partial_s\ln\Theta(s)\|,$$ and integrating this we finally arrive at $$\mathcal{W}[\breve T](0)e^{-\left\|\int_0^sd\sigma(\frac{2\|\Xi(\sigma)\|}{\sqrt{\Theta(\sigma)}}+\|\partial_s\ln\Theta(\sigma)\|)\right\|}\le\mathcal{W}[\breve T](s)\le\mathcal{W}[\breve T](0)e^{\left\|\int_0^sd\sigma(\frac{2\|\Xi(\sigma)\|}{\sqrt{\Theta(\sigma)}}+\|\partial_s\ln\Theta(\sigma)\|)\right\|},$$ precisely as in the statement. If however $\Lambda$ is not guaranteed to be positive, then on those regions where it is negative the magnitude of the solutions is expected to behave exponentially. We are able to capture that exponential factor by the following beautiful trick. \[KappaTrick\] For any $0<\kappa\in\mathbb{R}$, any solution of the equation $$\ddot T(s)+\Lambda(s)T(s)=0$$ can be represented as $T(s)=\tau(\frac{1}{\kappa}{\operatorname{th}}(\kappa s)){\operatorname{ch}}(\kappa s)$, where $\tau(z)$ is a solution of the equation $$\ddot\tau(z)+\Omega(z)\tau(z)=0$$ with $$\Omega(z)=\frac{\kappa^2+\Lambda(\frac{1}{\kappa}{\operatorname{ath}}(\kappa z))}{(1-\kappa^2z^2)^2}\mbox{, }z\in(-\frac{1}{\kappa},\frac{1}{\kappa}).$$ The proof is elementary once we already know the clue: the substitution of variables $\kappa z={\operatorname{th}}(\kappa s)$. The substitution $T(s)=\tau(s){\operatorname{ch}}(\kappa s)$ into the original equation gives $$\ddot\tau(s)+2\kappa{\operatorname{th}}(\kappa s)\dot\tau(s)+(\kappa^2+\Lambda(s))\tau(s)=0,$$ then the substitution $s\to z$ yields the final formulas. Let us say a couple of words about this. If $\Re\Lambda$ has a minimal negative value $-c$ on some domain, then it suffices to set $\kappa=\sqrt{c}$ to reduce the problem to an oscillatory equation for $\rho$. The upper bound of the rate of exponential expansion is precisely given by the square root of the minimal negative value of $\Re\Lambda$. Finally we combine these two statements to find an explicit uniform bound on an arbitrary solution $T$. Let the compact interval $\mathcal{R}$ containing 0 be fixed, and set $$A_\mathcal{R}=\sup_\mathcal{R}\|\Im\Lambda\|\mbox{, }c_\mathcal{R}=\inf_\mathcal{R}\Re\Lambda\mbox{, }\kappa=\sqrt{1+\|\min\{0,c_\mathcal{R}\}\|}\mbox{, }B_\mathcal{R}=\sup_\mathcal{R}\left\|\partial_s\ln\left(\kappa^2+\Re\Lambda\right)\right\|,$$ $$D_\mathcal{R}=\sup_\mathcal{R}(\kappa^2+\Re\Lambda)\mbox{, }e_\mathcal{R}=\inf_\mathcal{R}(\kappa^2+\Re\Lambda)=1+\max\{0,c_\mathcal{R}\},$$ $$L_\mathcal{R}=\left(2A_\mathcal{R}+{\operatorname{ch}}^2(\kappa\|\mathcal{R}\|)B_\mathcal{R}+2\kappa{\operatorname{sh}}(2\kappa\|\mathcal{R}\|)\right)$$ (we suppressed the index $\mathcal{R}$ of $\kappa$ for convenience). \[TEstAbstract\] For an arbitrary solution $T$ it holds $$\|T(s)\|\le\|T(0)\|\sqrt{\frac{D_\mathcal{R}}{e_\mathcal{R}}}e^{L_\mathcal{R}}{\operatorname{ch}}(\kappa\|\mathcal{R}\|)+\|\dot T(0)\|\frac{1}{\sqrt{e_\mathcal{R}}}e^{L_\mathcal{R}}{\operatorname{ch}}(\kappa\|\mathcal{R}\|)$$ for all $s\in\mathcal{R}$. Consider the linearly independent solutions $Q$ and $R$ given by initial data $$Q(0)=1\mbox{, }\dot Q(0)=0\mbox{, }R(0)=0\mbox{, }\dot R(0)=1.$$ Using Proposition \[KappaTrick\] represent them as $Q(s)=\xi(z(s)){\operatorname{ch}}(\kappa s)$ and $R(s)=\rho(z(s)){\operatorname{ch}}(\kappa s)$, where $\xi(z)$ and $\rho(z)$ are solutions of the equation $$\ddot\tau(z)+\Omega(z)\tau(z)=0$$ with $$\Omega(z)=\frac{\kappa^2+\Lambda(\frac{1}{\kappa}{\operatorname{ath}}(\kappa z))}{(1-\kappa^2z^2)^2}.$$ Using $z(0)=0$ and $$\begin{aligned} \frac{d}{ds}\left[\tau(z(s)){\operatorname{ch}}(\kappa s)\right]=\frac{\dot\tau(z(s))}{{\operatorname{ch}}(\kappa s)}+\tau(z(s)){\operatorname{sh}}(\kappa s)\kappa,\label{dTbydtau}\end{aligned}$$ we find $$\xi(0)=1\mbox{, }\dot \xi(0)=0\mbox{, }\rho(0)=0\mbox{, }\dot \rho(0)=1.$$ Note that $\Re\Omega(s)=\kappa^2+\Re\Lambda\ge1$, thus Proposition \[EnergyEstCompl\] is applicable for $\xi$ and $\rho$. We have $\mathcal{W}[\xi](0)=\frac{1}{2}\Re\Omega(0)$ and $\mathcal{W}[\rho](0)=\frac{1}{2}$. Now $$\frac{d}{dz}\ln\Re\Omega(z)=\frac{ds}{dz}(s)\frac{d}{ds}\ln\Re\Omega(z(s))={\operatorname{ch}}^2(\kappa s)\frac{d}{ds}\ln\left((\kappa^2+\Re\Lambda(s)){\operatorname{ch}}^4(\kappa s)\right)=$$ $$={\operatorname{ch}}^2(\kappa s)\frac{d}{ds}\ln\left(\kappa^2+\Re\Lambda(s)\right)+2\kappa{\operatorname{sh}}(2\kappa s).$$ Then it follows $$\left\|\int_0^zd\sigma(\frac{2\|\Im\Omega(\sigma)\|}{\sqrt{\Re\Omega(\sigma)}}+\|\partial_z\ln\Re\Omega(\sigma)\|)\right\|\le\frac{2}{\kappa}{\operatorname{th}}(\kappa\|\mathcal{R}\|)\left(2A_\mathcal{R}+{\operatorname{ch}}^2(\kappa\|\mathcal{R}\|)B_\mathcal{R}+2\kappa{\operatorname{sh}}(2\kappa\|\mathcal{R}\|)\right)\le$$ $$\le2\left(2A_\mathcal{R}+{\operatorname{ch}}^2(\kappa\|\mathcal{R}\|)B_\mathcal{R}+2\kappa{\operatorname{sh}}(2\kappa\|\mathcal{R}\|)\right)=2L_\mathcal{R}.$$ By Proposition \[EnergyEstCompl\] we have $$\mathcal{W}[\xi](z)\le\mathcal{W}[\xi](0)e^{2L_\mathcal{R}}\mbox{, }\mathcal{W}[\rho](z)\le\mathcal{W}[\rho](0)e^{2L_\mathcal{R}},$$ which entails $$\|\xi(z)\|\le\sqrt{\frac{\kappa^2+\Re\Lambda(0)}{\kappa^2+\Re\Lambda(s(z))}}e^{L_\mathcal{R}}\mbox{, }\|\dot\xi(z)\|\le\sqrt{\kappa^2+\Re\Lambda(0)}e^{L_\mathcal{R}},$$ $$\|\rho(z)\|\le\frac{1}{\sqrt{\kappa^2+\Re\Lambda(s(z))}}e^{L_\mathcal{R}}\mbox{, }\|\dot\rho(z)\|\le e^{L_\mathcal{R}}.$$ For $Q$ and $R$ we get $$\|Q(s)\|\le\sqrt{\frac{D_\mathcal{R}}{e_\mathcal{R}}}e^{L_\mathcal{R}}{\operatorname{ch}}(\kappa\|\mathcal{R}\|)\mbox{, }\|R(s)\|\le\frac{1}{\sqrt{e_\mathcal{R}}}e^{L_\mathcal{R}}{\operatorname{ch}}(\kappa\|\mathcal{R}\|),$$ and using (Eq.\[dTbydtau\]) $$\|\dot Q(s)\|\le\sqrt{D_\mathcal{R}}e^{L_\mathcal{R}}\left(1+\frac{\kappa{\operatorname{sh}}(\kappa\|\mathcal{R}\|)}{\sqrt{e_\mathcal{R}}}\right)\mbox{, }\|\dot R(s)\|\le e^{L_\mathcal{R}}\left(1+\frac{\kappa{\operatorname{sh}}(\kappa\|\mathcal{R}\|)}{\sqrt{e_\mathcal{R}}}\right).$$ Finally let $T$ be an arbitrary solution of the original equation. Applying Remark \[TbyW\_QR\] for $Q$,$R$ and $T$ we find $$T(s)=T(0)Q(s)+\dot T(0)R(s),$$ and hence $$\|T(s)\|\le\|T(0)\|\sqrt{\frac{D_\mathcal{R}}{e_\mathcal{R}}}e^{L_\mathcal{R}}{\operatorname{ch}}(\kappa\|\mathcal{R}\|)+\|\dot T(0)\|\frac{1}{\sqrt{e_\mathcal{R}}}e^{L_\mathcal{R}}{\operatorname{ch}}(\kappa\|\mathcal{R}\|),$$ as asserted. A result from functional calculus ================================= In this section we will obtain a result using the theory of holomorphic functional calculus of strip type operators. We are grateful to M. Haase for very useful comments on this theory, and refer to his book [@Haase2006] for all the information necessary in this section. Let $\mathbb{H}_a=\{z\in\mathbb{C}: \|\Im z\|<a\}$ denote the symmetric strip of height $a>0$. If for an (unbounded) operator $A$ on the Banach space $\mathcal{X}$ we have $A\in{\mbox{Strip}}(a)$, then we can apply the holomorphic functional calculus of $A$ given by $$F(A)=\frac{1}{2\pi i}\int_{\gamma_a}dzf(z)\mathfrak{R}(z,A)\mbox{, }\forall F\in\mathcal{M}[\mathbb{H}_a],$$ where $\gamma_a=\partial\mathbb{H}_a$ oriented positively (counterclockwise), and $\mathfrak{R}(z,A)$ is the resolvent of $A$ for $z\in\mathbb{C}$. Define $$\mathcal{A}(\mathbb{H}_a)=\{F\in{\mbox{Hol}}(\mathbb{H}_a): \exists N\in\mathbb{N}\mbox{ s.t. }F=O(\|\Re z\|^N)\},$$ and $$\mathcal{A}[\mathbb{H}_a]=\bigcup_{b>a}\mathcal{A}(\mathbb{H}_b).$$ Now let $D_{\Sigma_t}=-\Delta+m^\star(t,\vec x)$ be the known real lower semi-bounded operator acting on the vector bundle $\mathcal{T}_t$ over a Riemannian manifold $\Sigma_t$, and let $K\subset\Sigma_t$ be a compact region. Denote $$\mathcal{D}(K)=\{f\in\mathcal{D}(\mathcal{T}_t): {\mbox{supp}}f\subset K\}.$$ Then we have the following result. \[FuncCalcProp\] For any $F\in\mathcal{A}[\mathbb{H}_0]$ and $f\in\mathcal{D}(K)$ it follows $$F(D_{\Sigma_t})f\in\mathcal{D}(K).$$ Let the nuclear topology be given by $(X_i,2,l^2)$, i.e., for any $p\in\mathbb{N}_0$ we set $$(f,h)_p=\sum_{q\le p}(Q_{\alpha,q}(X_i)f,Q_{\alpha,q}(X_i)h)_{L^2}$$ and consider the induced norms $\\|.\\|_p$. Define the Hilbert spaces $$\mathcal{H}_p=\overline{\mathcal{D}(K)}^{(,)_p},$$ then by the property of the countably normed spaces we have $$\mathcal{H}_p\subset\mathcal{H}_q\mbox{, }q<p,$$ $$\mathcal{D}(K)=\bigcap_{p=0}^\infty\mathcal{H}_p.$$ Fix $p$, and define the operator $D_p$ on $\mathcal{H}_p$ by setting $D_p f=D_{\Sigma_t} f$ for all $f\in{\mbox{Dom}}(D_{\Sigma_t})\cap\mathcal{H}_p$, then ${\mbox{Dom}}(D_p)\supset\mathcal{H}_{p+2}$ is a dense subspace of $\mathcal{H}_p$. Then $D_p$ is a real symmetric operator, and hence by von Neumann’s theorem possesses a self-adjoint extension $A_p$ which needs not be lower semi-bounded. The self-adjoint operator $A_p$ has a purely real spectrum, thus $A_p\in{\mbox{Strip}}(0)$. Let $\mathcal{A}(\mathbb{H}_a)\ni F(z)=O(\|\Re z\|^N)$, then for a sufficiently large $a<\lambda\in\mathbb{R}$, the function $e(z)=(z-i\lambda)^{-(N+2)}$ will regularize $F$ on $\mathbb{H}_a$. In particular, we will have $[eF](A_p)\in\mathcal{B}(\mathcal{H}_p)$. Then $F(A_p)=(A_p-i\lambda)^{N+2}[eF](A_p)=[eF](A_p)(A_p-i\lambda)^{N+2}$, from where it follows that ${\mbox{Dom}}(A_p^{N+2})\subset{\mbox{Dom}}(F(A_p))$. From the definition of $A_p$ it is clear that $\mathcal{H}_{p+2(N+2)}\subset{\mbox{Dom}}(A_p^{N+2})$, whence $\mathcal{H}_{p+2(N+2)}\subset{\mbox{Dom}}(F(A_p))$. Thus we have established, that whenever $f\in\mathcal{H}_{p+2(N+2)}$, then necessarily $F(A_p)f\in\mathcal{H}_p$. Now if $f\in\mathcal{D}(K)$, then for any $p\ge0$ we have $f\in\mathcal{H}_{p+2(N+2)}$, and hence $F(A_p)f\in\mathcal{H}_p$. Meanwhile for any $p\ge0$, the self-adjoint operator $D_{\Sigma_t}$ agrees with $A_p$ on $\mathcal{D}(K)$. Therefore also their functional calculi agree, $F(D_{\Sigma_t})f=F(A_p)f\in\mathcal{H}_p$. Thus $$F(D_{\Sigma_t})f\in\bigcap_{p=0}^\infty\mathcal{H}_p=\mathcal{D}(K),$$ which completes the proof.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We give characterizations of affine transformations and affine vector fields in terms of the spray. By utilizing the Jacobi type equation that characterizes affine vector fields, we prove some rigidity theorems of affine vector fields on compact or forward complete non-compact Finsler manifolds with non-positive total Ricci curvature.' address: - 'School of Mathematical Sciences and LPMC, Nankai University, Tianjin 300071, P. R. China' - 'School of Science, Wuhan University of Technology, Wuhan 430070, P. R. China' author: - Libing Huang - Qiong Xue title: Affine Vector Fields on Finsler Manifolds --- Introduction ============ It is well kown that every affine vector field on a compact orientable Riemannian manifold is a Killing vector field [@Kobayashi]. In the noncompact case, Junichi Hano proved that if the length of an affine vector field in a complete Riemannian manifold is bounded, then its affine vector field is a Killing vector field [@Hano]. This result was generalized later by Shinuke Yorozu, who proved that every affine vector field with finite norm on a complete noncompact Riemannian manifold is a Killing vector field. Moreover, if the Riemannian manifold has non-positive Ricci curvature, then every affine vector field with finite global norm on it is a parallel vector field [@Yorozu]. We generalize this result to Finsler manifolds. We investigate affine vector fields on Finsler manifolds and prove some rigidity theorems of affine vector fields on compact and forward complete non-compact Finsler manifolds with non-positive total Ricci curvature. \[thm:main1\] Let $(M,F)$ be an $n$-dimensional compact Finsler manifold with non-positive total Ricci curvature. Then every affine vector field $V$ on $M$ is a linearly parallel vector field. \[thm:main2\] Let $(M,F)$ be an $n$-dimensional forward complete non-compact Finsler manifold with non-positive total Ricci curvature and bounded reversibility. Then every affine vector field $V$ on $M$ with finite global norm is parallel. The proofs of the above theorems follow a typical Finslerian style. We extensively use knowledge of the tangent bundle and sphere bundle, which are by no means neccessary in Riemannian geometry. Moreover, since the sphere bundle is always orientable, we can drop the orientable condition of the manifold that is used in Riemannian case. In Riemannian geometry, affine vector field $V$ is characterized by the fact that the Lie derivative of the Riemannian metric $g$ is paralle, namely, $$\nabla(\mathcal{L}_Vg) = 0,$$ where $\nabla$ is the Levi-Civita connection [@Poor]. In Finsler geometry, there are many choices of connections but the characterization of an affine vector field is much more concise and without any connection. We show that a vector field $V$ is affine if and only if its flow commutes with the geodesic flow. Equivalently, $\hat{V}$, the complete lift of $V$, commutes with the Finsler spray $G$. i.e., $$\mathcal{L}_{\hat{V}}G = 0.$$ By expanding the above Lie bracket, we obtain a Jacobi type equation, thus Bochner type techniques could be applied. The crucial difference between the Riamannian case and Finsler case is that, in Riemannian geometry one needs to compute the Laplacian of the energy of $V$, but in Finsler geometry we don’t need further computation. The organization of the paper is as follows. Firstly, we review some basic facts of Finsler geometry and give a definition of the total Ricci curvature in Section \[sec:pre\]. Then we define affine transformations and affine vector fields on Finsler manifolds respectively in Section \[sec:aff\]. Several characterizations of affine vector fields are provided, which are useful to the theorems. Finally, the proofs of the above theorems are presented in Section \[sec:prf\]. Preliminaries {#sec:pre} ============= In this section, we give a brief description of some basic materials that are needed to prove Theorem \[thm:main1\]. Spray and Riemann curvature tensor ---------------------------------- Let $M$ be an $n$-dimensional smooth manifold. A smooth function $F$ on the punctured tangent bundle $TM_0:= TM\backslash\{0\}$ is called a Finsler metric, if the restriction $F\|_{T_xM\backslash\{0\}}$ is a Minkowski norm for every $x$ in $M$. The natural projection $\pi:TM_0\to M$ induces a pull back bundle $\pi^TM$ over $TM_0$, whose fiber at each point $y\in TM_0$ is just a copy of $T_xM$, where $x=\pi(y)$. For each fixed $y\in TM_0$, one can define an inner product $g_y$ on the fibre $T_xM$ as follows $$g_y = g_{ij} \operatorname{d}x^i\otimes\operatorname{d}x^j,\quad g_{ij} = \frac{1}{2}[F^2]_{y^iy^j},$$ where we have used the natural local coordinate system $(x^i, y^i)$ on $TM_0$. When $y$ varies, the inner products $g_y$ become a globally defined tensor field on $\pi^TM$, called the fundamental tensor [@ZShen]. The following set of functions are defined locally and are called the spray coefficients $$G^i = \frac{1}{4}g^{il}\big([g_{jl}]_{x^k} + [g_{lk}]_{x^j} - [g_{jk}]_{x^l}\big)y^j y^k.$$ It can be verified that the vector field $G = y^i{\frac{\partial{}}{\partial{x^i}}} - 2G^i{\frac{\partial{}}{\partial{y^i}}}$ is globally defined and it is called the spray induced by $F$. A curve $\gamma:(a,b)\to M$ is called a geodesic, if the curve $\dot\gamma:(a,b)\to TM$ is an integral curve of $G$. Locally, the coordinates $(\gamma^i(t))$ of a geodesic $\gamma(t)$ satisfy $$\ddot{\gamma}^i(t) + 2G^i(\gamma(t),\dot{\gamma}(t)) = 0.$$ Hence, for any $y\in TM_0$, there is a unique geodesic $\gamma(t)$ defined on a maximal open interval containing $0$, and it satisfies $\dot\gamma(0)=y$. The local functions $N^i_j = [G^i]_{y^j}$ are called the (nonlinear) connection coefficients. Let $$\frac{\delta}{\delta x^i} = {\frac{\partial{}}{\partial{x^i}}} - N_i^j{\frac{\partial{}}{\partial{y^j}}},$$ then $HTM = \mathrm{span}\big\{\frac{\delta}{\delta x^i}\big\}$ and $VTM = \mathrm{span}\big\{{\frac{\partial{}}{\partial{y^i}}}\big\}$ are well-defined subbundles of $TTM_0$ and $TTM_0 = HTM\oplus VTM$. Direct computation yields $$\label{eq:Gv} \Big[G, {\frac{\partial{}}{\partial{y^i}}}\Big] = -\frac{\delta}{\delta x^i} + N_i^j {\frac{\partial{}}{\partial{y^j}}},$$ and $$\label{eq:Gh} \Big[G, \frac{\delta}{\delta x^i}\Big] = R^j{}_i {\frac{\partial{}}{\partial{y^j}}} + N^j_i \frac{\delta}{\delta x^j},$$ where the functions $R^j{}_i$ are given by $$R^j{}_i = 2[G^j]_{x^i} - G(N^j_i) - N_k^j N^k_i.$$ For each fixed $y\in TM_0$, the $(1,1)$ tensor $R_y = R^j{}_i {\frac{\partial{}}{\partial{x^j}}}\otimes \operatorname{d}x^i$ is called the Riemann curvature tensor. Sphere bundle and volume form ----------------------------- The set $SM = \{y\in TM_0\,\|\, F(y)=1\}$ is called the unit sphere bundle or indicatrix bundle. Let $\omega = F_{y^i}\operatorname{d}x^i = g_{ij} y^j/F \operatorname{d}x^i$, then $\omega$ is a globally defined contact form on $SM$ and it is called the Hilbert form. By straightforward computation, one can show that the vector field $\xi = G/F$ satisfies $$\label{eq:Reeb} \omega(\xi) = 1,\quad \operatorname{d}\omega(\xi,\;\cdot\;) = 0.$$ Henceforth, $\xi$ is called the Reeb field according to the contact terminology. Although the manifold $M$ could be non-oriented, the sphere bundle $SM$ is always oriented because it carries the following volume form $$\operatorname{d}\nu = c_n\;\omega\wedge(\operatorname{d}\omega)^{n-1},$$ where the constant $c_n = (-1)^{-1+n(n+1)/2}/(n-1)!$. The mean Ricci curvature $\widetilde{Ricci}$ is defined in [@BShen] with the help of an auxiliary Riemannian metric. For our purposes, we define the total Ricci curvature $\mathcal{T}(V)$ as follows $$\mathcal{T}(V) = \int_{SM} \frac{1}{F^2}g_y(R_y(V), V)\;\operatorname{d}\nu.$$ where $V$ is any vector field on $M$. When $M$ is oriented, $\mathcal{T}(V)$ is the integration of $\widetilde{Ricci}(V)$ over $M$, where $M$ has been assigned the volume form of the auxiliary Riemannian metric. Dynamical derivative -------------------- Berwald connection is a linear connection on the vector bundle $\pi^TM$ over $TM_0$, whose connection coefficients are given by $\Gamma^i_{jk} = [G^i]_{y^jy^k}$. Using Berwald connection, one can take covariant derivatives of any tensor fields on $TM_0$. For example, if $T=T^i_j {\frac{\partial{}}{\partial{x^i}}}\otimes\operatorname{d}x^j$, then the horizontal covariant derivative is given by $$T^i_{j\|k} = \frac{\delta T^i_j}{\delta x^k} + T^l_j\Gamma^i_{lk} - T^i_l\Gamma^l_{jk}.$$ The dynamical derivative* is just the horizontal covariant derivative along the direction $G=y^k\frac{\delta}{\delta x^k}$. For example, the dynamical derivative of the above tensor field $T$ is given by $T_{\|0} = T^i_{j\|0}{\frac{\partial{}}{\partial{x^i}}}\otimes\operatorname{d}x^j$, where $$T^i_{j\|0} = G(T^i_j) + T^l_j N^i_l - T^i_l N^l_j.$$ It should be remarked that, although we introduce the dynamical derivative using Berwald connection, this special derivative is actually independent of any named connection. One may consult [@Foulon] for another treatment of this concept. For every smooth function $f$ on $TM_0$, the dynamical derivative of $f$ is given by $f_{\|0} = G(f)$. In other words, $f_{\|0}$ is just the derivative of $f$ along geodesics, i.e., $$f_{\|0}(x,y) = \left.\frac{\operatorname{d}}{\operatorname{d}t} f(\gamma(t),\dot\gamma(t))\right\|_{t=0}, \quad\forall (x,y)\in TM_0,$$ where $\gamma$ is the unique geodesic with $\gamma(0)=x$, $\dot\gamma(0)=y$. As a concrete example, the dynamical derivative of the Finsler function $F$ is zero. Another example is the fundamental tensor; its dynamical derivative is also zero, because $$g_{ij\|0} = G(g_{ij}) - g_{kj}N^k_i - g_{ik}N^k_j = 0.$$ Sometimes, the dynamical derivative is also defined to be horizontal covariant derivative along the direction $\xi = G/F$. When using this convention, we shall denote the dynamical derivative of a tensor field $T$ by $\dot{T}$. For example, if $V=V^i{\frac{\partial{}}{\partial{x^i}}}$, then $$\dot{V} = \frac{V^i_{\|0}}{F}{\frac{\partial{}}{\partial{x^i}}}.$$ Since every vector field $V$ on $M$ can be thought of as a smooth section of $\pi^TM$, the symbol $V_{\|0}$ or $\dot{V}$ makes sense. In this case $\dot{V}=0$ if and only if $V_{\|0}=0$, if and only if $V_{\|i}= 0$, i.e., $V$ is a linearly parallel vector field [@ChernShen]. Affine transformation and affine vector field {#sec:aff} ============================================= Affine transformation --------------------- In affine differential geometry, affine transformation is a special kind of projective transformation, which preserves the (parametrized) geodesics on a manifold with an affine connection. In Euclidean space, affine transformation consists of translation, scaling, homothety, similarity transformation, reflection, rotation, shear mapping, and compositions of the above in any combination or sequence. This concept can also be studied in the realm of spray geometry. Let $M$ be an $n$ dimensional manifold with a spray $G$. A diffeomorphism $\phi:M\to M$ is called affine transformation* if for any geodesic $\gamma:(a,b)\to M$, the curve $\phi\circ\gamma$ is also a geodesic. Since the geodesics are defined by the spray, we can prove affine transformation preserves the spray. \[lem:affinetrans\] A diffeomorphism $\phi$ is an affine transformation iff $$\hat{\phi}_G = G,$$ where $\hat{\phi}: TM\to TM$ is the lift of $\phi$ defined by $$\hat{\phi}(x,y) = (\phi(x),\phi_(y)),\quad \forall x\in M, \quad y\in T_xM.$$ For any $(x,y)\in TM_0$, let $\gamma$ be the unique geodesic with $\gamma(0)=x$, $\dot\gamma(0) = y$. Let $\sigma = \dot\gamma$, then $\sigma$ is an integral curve of $G$, namely, $G_{\sigma} = \dot\sigma$. If $\phi$ is an affine transformation, then $\phi\circ\gamma$ is also a geodesic. In this case, $(\phi\circ\gamma)' = \hat\phi\circ\sigma$ is also an integral curve of $G$, i.e., $$G_{\hat{\phi}\circ\sigma} = (\hat{\phi}\circ\sigma)' = \hat{\phi}_\dot\sigma = \hat{\phi}_ G_{\sigma}.$$ Taking $t=0$ in the above identity, then we have $\hat{\phi}_G_{(x,y)} = G_{\hat{\phi}(x,y)}$. Conversely, if $\hat{\phi}_G = G$, then $\hat{\phi}$ maps every integral curve of $G$ to an integral curve. Consequently it will map every geodesic to a geodesic. As a result, $\phi$ is affine. Affine vector field ------------------- Let $V$ be a vector field on $M$ which generates a local one-parameter group $\phi_t$, $t\in(-\varepsilon,\varepsilon)$. Let $\hat{\phi}_t$ be the lift of $\phi_t$, then $\hat{\phi}_t$ is also a local one-parameter group on $TM$. So there is a vector field $\hat{V}$ on $TM$ induced by $\hat{\phi}_t$; it is called the complete lift of $V$. Locally, if $V=V^i{\frac{\partial{}}{\partial{x^i}}}$, then we have $\hat{V} = V^i{\frac{\partial{}}{\partial{x^i}}} + y^j{\frac{\partial{V^i}}{\partial{x^j}}}{\frac{\partial{}}{\partial{y^i}}}$. On a spray manifold $(M,G)$, a vector field $V$ is called affine vector field, if the local one-parameter group $\phi_t$ generated by $V$ consists of affine transformations. Based on the property of affine transformation, we can deduce the following characterizations of affine vector fields. \[prp:charaff\] Let $V$ be a vector field on a spray manifold $(M,G)$. Then the following assertions are mutually equivalent. - $V$ is an affine vector field; - $\mathcal{L}_{\hat{V}}G = 0$; - $V^i_{\|0\|0} + V^k R^i{}_k = 0$. By definition, $V$ is an affine vector field, if and only if the corresponding one-parameter group $\phi_t$ consists of affine transformations. This is also equivalent to $\hat{\phi}_{t}G = G$ by Lemma \[lem:affinetrans\]. Taking derivative with respect to $t$ at $t=0$, we have $\mathcal{L}_{\hat{V}}G = 0$. So the implication (1)$\Longrightarrow$(2) is proved. The converse is clear by the definition of Lie derivative (or one may consult [@KN Prop. 1.7]). Now we prove the equivalence of (2) and (3) by some local computation. First, note that $$\begin{aligned} \hat{V} &= V^i{\frac{\partial{}}{\partial{x^i}}} + y^j{\frac{\partial{V^i}}{\partial{x^j}}}{\frac{\partial{}}{\partial{y^i}}} = V^i\big(\frac{\delta}{\delta x^i}+N_i^j{\frac{\partial{}}{\partial{y^j}}}\big) + y^j{\frac{\partial{V^i}}{\partial{x^j}}}{\frac{\partial{}}{\partial{y^i}}}\\ &=V^i\frac{\delta}{\delta x^i} + \big(y^j\frac{\delta V^i}{\delta x^j} + V^k N_k^i\big){\frac{\partial{}}{\partial{y^i}}} = V^i\frac{\delta}{\delta x^i} + V^i_{\|0}{\frac{\partial{}}{\partial{y^i}}}.\end{aligned}$$ Using the above equation, we deduce $$\begin{aligned} [G, \hat{V}] &= \big[G, V^i\frac{\delta}{\delta x^i} + V^i_{\|0} {\frac{\partial{}}{\partial{y^i}}} \big] \\ &= G(V^i)\frac{\delta}{\delta x^i} + G(V^i_{\|0}){\frac{\partial{}}{\partial{y^i}}} + V^i\cdot \big[G, \frac{\delta}{\delta x^i}\big] + V^i_{\|0}\cdot\big[G, {\frac{\partial{}}{\partial{y^i}}}\big]\\ &= G(V^i)\frac{\delta}{\delta x^i} + G(V^i_{\|0}){\frac{\partial{}}{\partial{y^i}}} + V^i\big(R^j{}_i{\frac{\partial{}}{\partial{y^j}}} + N^j_i\frac{\delta}{\delta x^j}\big)\\ &\qquad +V^i_{\|0}\big(-\frac{\delta}{\delta x^i} + N^j_i {\frac{\partial{}}{\partial{y^j}}}\big)\\ &= \big(G(V^i_{\|0}) + V^k_{\|0}N^i_k + V^kR^i{}_k\big){\frac{\partial{}}{\partial{y^i}}} \\ &= (V^i_{\|0\|0} + V^kR^i{}_k){\frac{\partial{}}{\partial{y^i}}},\end{aligned}$$ where we have used (\[eq:Gh\]) and (\[eq:Gv\]). It is clear that $\mathcal{L}_{\hat{V}}G = 0$ if and only if $V^i_{\|0\|0}+V^k R^i{}_k = 0$. Thus the proposition is proved. The third assertion can be written as $$V_{\|0\|0} + R_y(V) = 0, \quad\text{or}\quad \ddot{V} + R_y(V)/F^2 = 0.$$ This equation is the same as Jacobi equation when restricted to a geodesic. So one can conclude that $V$ is an affine vector field, if and only if the restriction of $V$ to any geodesic is a Jacobi field. Affine vector fields on Finsler manifolds {#sec:prf} ========================================= The main purpose of this section is to prove Theorems \[thm:main1\] and \[thm:main2\]. To that end, we need a technical lemma. \[lem:key\] Let $(M,F)$ be a Finsler manifold without boundary. Let $\operatorname{d}\nu=c_n\,\omega\wedge(\operatorname{d}\omega)^{n-1}$ be the volume form of $SM$. Then for any compactly supported smooth function $f$ on $SM$, we have $$\int_{SM} \dot{f}\; \operatorname{d}\nu = 0.$$ It is well-known [@BaoChernShen] that we can choose local coframe field $$\omega^1,\omega^2,\cdots,\omega^{n-1},\omega^n=\omega, \omega^{n+1},\cdots,\omega^{2n-1},$$ on $SM$, such that $\operatorname{d}\omega = \sum_{\alpha=1}^{n-1}\omega^{\alpha}\wedge\omega^{n+\alpha}$. Thus we have $$(\operatorname{d}\omega)^{n-1} = (n-1)!\;\omega^1\wedge\omega^{n+2}\wedge\cdots\wedge\omega^{n-1}\wedge\omega^{2n-1}.$$ Let $e_1$, $\cdots$, $e_n$, $e_{n+1}$, $\cdots$, $e_{2n-1}$ be the dual frame field, then $e_n = \xi$ (compare (\[eq:Reeb\])). If we write $$\operatorname{d}f = f_1\omega^1 + \cdots + f_n \omega^n + \cdots + f_{2n-1}\omega^{2n-1},$$ then it is clear that $f_n = \xi(f) = \dot{f}$. So we have $$\dot{f}\;\omega\wedge(\operatorname{d}\omega)^{n-1} = \operatorname{d}f\wedge(\operatorname{d}\omega)^{n-1} = \operatorname{d}(f\wedge (\operatorname{d}\omega)^{n-1}).$$ In other words, the form to be integrated is exact. By Stokes theorem, the integration is zero since $M$ is boundariless. Compact case ------------ We restate Theorem \[thm:main1\] as follows. Let $(M,F)$ be an $n$-dimensional compact Finsler manifold and let $V$ be an affine vector field. If $\mathcal{T}(V)\leq 0$, then $V$ is a linearly parallel field. We first do some computation on $TM_0$. Let $f = \frac{1}{F} g_{ij}V^iV^j_{\|0}$, then we have $$f_{\|0} =\frac{1}{F}( g_{ij}V^i_{\|0}V^j_{\|0} + g_{ij} V^i V^j_{\|0\|0}) = \frac{1}{F}(g_{ij}V^i_{\|0}V^j_{\|0} - g_{ij}V^iV^kR^j{}_k),$$ where we have used the characterization (3) of affine vector field in Prop. \[prp:charaff\]. Since $f$ is $y$-homogeneous of order $0$, it can be thought of as a function on $SM$. The above equation can be interpreted as $$\dot f = g_y(\dot{V},\dot{V}) - \frac{1}{F^2} g_y(R_y(V), V).$$ Taking integral of both sides on $SM$ and using Lemma \[lem:key\] yield $$0 = \int_{SM} g_y(\dot{V},\dot{V}) \operatorname{d}\nu - \int_{SM} \frac{1}{F^2}g_y(R_y(V),V)\operatorname{d}\nu \geq 0.$$ Therefore, $\dot{V} = 0$, which means $V$ is a linearly parallel field. Forward complete non-compact case --------------------------------- In this section, we consider affine vector fields on forward complete non-compact Finsler manifolds. Again, we state a precise version of Theorem \[thm:main2\] as follows. Let $(M,F)$ be an $n$-dimensional forward complete non-compact Finsler manifold. Assume that - $V$ is an affine vector field with finite global norm, i.e., $$\int_{SM} g_y(V,V)\;\operatorname{d}\nu < +\infty;$$ - The total Ricci curvature is non-positive, i.e., $\mathcal{T}(V)\leq 0;$ - The reversibility $\lambda(F):=\sup_{y\in SM}\frac{F(x,-y)}{F(x,y)} < +\infty$. Then $V$ is a linearly parallel vector field. Let $p$ be a fixed point of $M$. For each point $x\in M$, we denote by $d(p,x)$ the forward geodesic distance from $p$ to $x$. Let $\sigma : [0, +\infty) \to [0,1]$ be a smooth cut-off function* such that $$\sigma(t) = \begin{cases} 1, & t \in [0,1],\\ 0, & t \in [2,+\infty). \end{cases}$$ Fix a positive real number $\alpha$ and let $\mu(x) = \sigma\big(\frac{d(p,x)}{\alpha}\big)$. Then we have the following Claim. There is a positive constant $A$ such that $\dot{\mu}^2 \leq \frac{A\cdot\lambda(F)^2}{\alpha^2}$. Proof of the claim. Set $\rho(x) = d(p,x)$. We first show that $\dot\rho$ is bounded. Recall that the value of $\dot{\rho}$ at $(x,y)\in SM$ is given by $$\begin{aligned} \dot{\rho}(x,y) &= \left.\frac{\operatorname{d}}{\operatorname{d}t}\rho(\gamma(t))\right\|_{t=0} = \left.\frac{\operatorname{d}}{\operatorname{d}t}d(p, \gamma(t))\right\|_{t=0}\\ &= \lim_{\epsilon\to 0}\frac{1}{\epsilon}(d(p,\gamma(\epsilon))-d(p,\gamma(0))) \leq\lim_{\epsilon\to 0}\frac{1}{\epsilon}d(\gamma(0),\gamma(\epsilon))=F(x,y)=1.\end{aligned}$$ In a similar manner, we have $$\begin{aligned} \dot{\rho}(x,y) &= \lim_{\epsilon\to 0}\frac{1}{\epsilon}(d(p,\gamma(\epsilon))-d(p,\gamma(0)))\\ &\geq -\lim_{\epsilon\to 0}\frac{1}{\epsilon}d(\gamma(\epsilon),\gamma(0)) = -F(x,-y)\geq -\lambda(F).\end{aligned}$$ So $\dot\mu = \sigma'(\rho/\alpha)\cdot \dot\rho/\alpha$ satisfies $\dot\mu^2 \leq A\cdot\lambda(F)^2/\alpha^2$, where $A$ is the upper bound of $\|\sigma'(t)\|^2$. Thus the claim is proved. Now, consider the function $f = \mu^2 g_y(V, \dot{V})$ on $SM$. We have $$\label{eq:dotf} \dot f = 2\mu\dot{\mu} g_y(V, \dot{V}) + \mu^2 g_y(\dot{V},\dot{V}) + \mu^2 g_y(V,\ddot{V}).$$ Expanding the inequality $\frac{1}{2}g_y(\mu\dot{V} + 2\dot{\mu} V, \mu\dot{V} + 2\dot{\mu} V) \geq 0$ yields $$2\mu\dot{\mu}g_y(V, \dot{V}) \geq - 2\dot{\mu}^2 g_y(V,V) - \frac{1}{2}\mu^2 g_y(\dot{V},\dot{V}).$$ Substituting this inequality into (\[eq:dotf\]), we get $$\begin{aligned} \dot f &\geq -2\dot{\mu}^2 g_y(V, V) + \frac{1}{2}\mu^2 g_y(\dot{V},\dot{V}) + \mu^2 g_y(V,\ddot{V})\\ &\geq -\frac{2A\cdot\lambda(F)^2}{\alpha^2} g_y(V,V) + \frac{1}{2}\mu^2 g_y(\dot{V},\dot{V}) - \frac{\mu^2}{F^2} g_y(V,R_y(V)).\end{aligned}$$ Taking integral of both sides on $SM$ and using Lemma \[lem:key\] yield $$\begin{aligned} 0\geq -\frac{2A\cdot\lambda(F)^2}{\alpha^2}\int_{SM}g_y(V,V)\operatorname{d}\nu &+ \frac{1}{2}\int_{SM}\mu^2 g_y(\dot{V},\dot{V})\operatorname{d}\nu \\ &- \int_{SM}\frac{\mu^2}{F^2}g_y(V,R_y(V))\operatorname{d}\nu.\end{aligned}$$ It follows that $$\begin{aligned} &\frac{2A\cdot\lambda(F)^2}{\alpha^2}\int_{SM}g_y(V,V)\operatorname{d}\nu \\ \geq& \frac{1}{2}\int_{SM}\mu^2 g_y(\dot{V},\dot{V})\operatorname{d}\nu -\int_{SM}\frac{\mu^2}{F^2}g_y(V,R_y(V))\operatorname{d}\nu.\end{aligned}$$ Letting $\alpha\to\infty$, then the left hand side approaches $0$ by the conditions (1) and (3), while the right hand side approaches $\frac{1}{2}\int_{SM}g_y(\dot{V},\dot{V})\operatorname{d}\nu -\mathcal{T}(V)\geq 0$. Consequently, we must have $\dot{V}=0$. So $V$ is linearly parallel. Acknowledgements {#acknowledgements .unnumbered} ================ The second author is supported by NFSC (no. 61573021) and the Fundamental Research Funds for the Central Universities (no. 2017IA006). The authors are happy to acknowledge the hospitality and great working conditions provided by the Department of Mathematics, IUPUI. They also want to thank Zhongmin Shen and Wei Zhao for useful conversations. [99]{} D. Bao, S.-S. Chern, Z. Shen, An introduction to Riemann-Finsler geometry, Graduate Texts in Mathematics, vol. 200, Springer-Verlag, New York, 2000. S.-S. Chern, Z. Shen, Riemann-Finsler geometry, World Scientific, Singapore, 2005. P. Foulon, Curvature and global rigidity in Finsler manifolds, Houston J. Math. 28(2002), 263–292. J. Hano, On affine transformations of a Riemannian manifold, Nagoya Math. Journal 9(1955), 99–109. S. Kobayashi, Transformation groups in differential geometry, Classics in Mathematics, Springer-Verlag, Berlin Heidelberg, 1995. B. Shen, Vanishing of Killing vector fields on Finsler manifolds, Kodai Math. J. 41(2018), 1–15. Z. Shen, Lectures on Finsler Geometry, World Scientific, Singapore, 2001. S. Kobayashi, K. Nomizu, Foundations of Differential Geometry, Interscience Publishers, New York, 1963. S. Yorozu, Affine and projective vector fields on complete noncompact Riemannian manifolds, Yokohama Math. J. 31(1983)1–2, 41–46. W. A. Poor, Differential geometric structures, McGraw-Hill Book Co., New York, 1981.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Electron-spin nitrogen-vacancy color centers in diamond are a natural candidate to act as a quantum memory for superconducting qubits because of their large collective coupling and long coherence times. We report here the first demonstration of strong coupling and coherent exchange of a single quantum of energy between a flux-qubit and an ensemble of nitrogen-vacancy color centers.' author: - Xiaobo Zhu - Shiro Saito - Alexander Kemp - Kosuke Kakuyanagi - 'Shin-ichi Karimoto' - Hayato Nakano - 'William J. Munro' - Yasuhiro Tokura - 'Mark S. Everitt' - Kae Nemoto - Makoto Kasu - Norikazu Mizuochi - Kouichi Semba title: 'Coherent coupling of a superconducting flux-qubit to an electron spin ensemble in diamond' --- During the last decade, research into superconducting quantum bits (qubits) based on Josephson junctions has made rapid progress [@CW08]. Many foundational experiments have been performed [@Nakamura99; @Vion02; @Irinel03; @NIST07; @Yale07; @DiCarlo09; @Ansmann09; @Neeley10; @DiCarlo10] and superconducting qubits are now considered one of the most promising systems for quantum information processing. However, the experimentally reported coherence times are likely to be insufficient for future large scale quantum computation. A natural solution is a dedicated [engineered]{} quantum memory based on atomic and molecular systems. Since macroscopic quantum coherence was first demonstrated in Josephson junction circuits [@Nakamura99], the question of whether or not coherent quantum coupling between a single macroscopic artificial atom and an ensemble of natural atoms or molecules is possible has attracted significant attention [@Sorensen04; @Tian04; @Rabl06; @Marcos10]. In this letter we present for the first time evidence of coherent strong coupling between a single macroscopic superconducting artificial atom (a flux qubit) and an ensemble of electron-spin nitrogen-vacancy color centers (NV$^{-}$ centers) in diamond. Furthermore, we have observed coherent exchange of a single quantum of energy between a flux qubit and a macroscopic ensemble consisting of $\sim 3\times10^{7}$ of NV$^{-}$ centers. This provides a foundation for future quantum memories and hybrid devices coupling microwave and optical systems. With the early successes of single atom quantum state manipulation [@Brune96], research in quantum information processing with atomic and solid-state systems has progressed largely in a separate fashion. In recent years, significant effort has been devoted to coupling atomic and molecular system to solid-state qubits to form hybrid quantum devices [@Sorensen04; @Tian04; @Rabl06]. Hybrid devices involving the integration of an atomic system with a superconducting transmission line resonator have been realized [@Schuster10; @Kubo10; @TU-Wien11]. Such schemes have the potential to couple superconducting solid-state qubits to optical fields via atomic systems, thus enabling quantum media conversion. The coupling strength $g$ of an individual atomic system to one electromagnetic mode in a resonator circuit is usually too small for the coherent exchange of quantum information. However, the coupling strength of an ensemble of $N$ such atomic systems will be enhanced by a factor of $\sqrt{N}$ [@Kimble89], allowing one to reach the strong coupling regime ($g\sqrt{N}\gg \kappa, \gamma$, where $\kappa$ and $\gamma$ are the damping rates of resonator circuit and atomic system). Of the many possible hybrid systems, coupling a flux-qubit to an (NV$^{-}$) center in diamond is particularly appealing. Firstly, the magnetic coupling between a flux-qubit and a single NV$^{-}$ center can be three orders of magnitude larger than that associated with a superconducting transmission line resonator [@Marcos10]. Second, the ground state of an NV$^{-}$ center is a spin 1 triplet due to its $C_{3v}$ symmetry (Figure 1b). The $S=1$ spin triplet $\|m_{s}=0\rangle$ state is separated by 2.88 GHz from the near degenerate excited states $\|m_{s}=\pm 1\rangle$ under zero magnetic field (Figure 1c). This energy separation is ideal for a gap tunable flux-qubit to be brought on and off resonance with it. ![ [Experimental set-up of an NV-diamond sample attached to a flux-qubit system. ]{} (a) A diamond crystal is glued on top of a flux-qubit and its superconducting circuits (under the red box) with the diamonds $^{12}$C implanted (001) surface facing the chip. The distance between the flux-qubit and surface of the diamond crystal is carefully adjusted to be less than a micrometer using 100 nm height mesa structure on the diamond surface, a circle adjacent to the red square, and the optical interference pattern (Newton’s ring). (b) A sketch of a NV color center of diamond with its vacancy (V) and nitrogen (N) atom, as well neighboring carbon atoms. Four equivalent NV-axis exist depicted in purple color dangling bonds, all making the same angle with \[001\] direction to which the magnetic field generated by the flux-qubit points. (c) Energy diagram of the NV center, with the spin triplet $\|m_{s}=0\rangle$ ground state separated by 2.88 GHz from the degenerated $\|m_{s}=\pm 1\rangle$ excited states under zero magnetic field[@Jelezko04]. (d) Optical micrograph and the circuit scheme of the aluminum made flux-qubit, the magnified view of the chip under the red box region shown in Fig1a. The central M shaped circuit contains a flux-qubit and a SQUID detector. Two high-bandwidth (20 GHz) MW-control lines located both sides of the qubit circuit. (e) The H-shaped gap tunable flux-qubit and the edge-shared SQUID used as a switching qubit state detector. The flux-qubit contains two loops, the main loop (blue) and the $\alpha$-control loop (magenta) which controls tunneling energy gap of the flux-qubit. Magnified view of Josephson junctions are also shown. The mutual inductance of control line-1 to the $\alpha$-loop and main loop are 90 fH and 256 fH, and those of control line-2 are 0.5 fH and 549 fH. The magnetic flux penetrating these two loops can be controlled in ns time scale by applying synchronized current pulses to these control lines in situ.[]{data-label="FIG_1"}](Fig1_v1_3_st){width="1.0\columnwidth"} In this Letter, we report on the first observation of vacuum Rabi oscillations between a flux-qubit and an ensemble of approximately three million NV$^{-}$ centers in diamond. This demonstrates strong coherent coupling between two dissimilar quantum systems with an effective collective coupling constant of $g_{\rm ens}\sim70$ MHz. We begin by describing our experimental setup as depicted in (Figure 1). An NV$^{-}$ diamond sample was prepared by ion implantation of $^{12}C^{2+}$ at 700 keV under high vacuum into high-pressure, high-temperature (HPHT)-synthesized type Ib (001) surface orientation single crystal diamond. The $^{12}C^{2+}$ ions, with a dose condition of $3\times10^{13}$ cm$^{-2}$, were stopped at a depth of $ 600^{+50}_{-100}$ nm. This generated on the order of $5\times10^{18}$ cm$^{-3}$ vacancies over a depth of $\sim0.7 \mu$m. After implantation, the crystals were annealed at $900 ^{\circ}$C under vacuum for 3 hours. This high dose carbon implantation method enhances the yield of generated NV$^{-}$ centers [@Naydenov10]. Photoluminescence (PL) optical spectroscopy (shown in Figure 2) established that NV$^{-}$ centers were generated with a density of $\sim 1.1\times10^{18}$ cm$^{-3}$ over a $1 \mu$m depth. We can describe the ground state of a single NV$^{-}$ center by the Hamiltonian [@Neumann09], $$\begin{aligned} H_{ \rm NV}=hDS_{z}^2+hE(S_{x}^2-S_{y}^2)+ h g_{\rm NV}\mu_{\rm B}{\bf B}\cdot{\bf S},\end{aligned}$$ where $S_x, S_y, S_z$ are the usual Pauli spin 1 operators, $D$ the zero-field splitting (2.878 GHz), $E$ the strain-induced splitting ($<$1 MHz), the N-V Landé factor $g_{NV}=2$, and $\mu_{B}=14$ MHz$/$mT. The last term represents the Zeemen splitting, which is negligible in our case as the magnetic field applied perpendicular to the surface of the chip to prepare the flux qubit is less than $0.1$ mT. ![ [Photoluminescence spectra ]{} of (a) ensembles of color centers in the highly carbon implanted sample and (b) a single NV- center in pure diamond at room temperature. Their signal intensities are normalized for the comparison of the spectra. In the spectrum of the single NV center, the contributions of phonon Raman scattering from bulk diamond at 573 and between 600 and 620 nm are subtracted[@Gru]. The zero phonon line of NV- at 637 nm[@Gru; @Kur] is clearly observed in both spectra. Broad spectrum of phonon replicas is also very similar to each other and to the reported ones[@Gru; @Kur].These indicate that the NV- center is produced as a major color center in the highly carbon implanted sample. The signal intensity of the ensemble is about $6.5\times10^{4}$ times stronger than that of the single one. From this result, the concentration of the NV- center in the highly carbon implanted sample was estimated to be $1.1\times10^{18}$ cm$^{-3}$. []{data-label="FIG_2"}](Fig2n2){width="0.87\columnwidth"} A diamond crystal was glued on top of the superconducting circuit with the $^{12}$C implanted surface facing the flux-qubit (Figure 1a). We used a gap tunable flux-qubit [@Zhu10; @Fedorov10] (Figure 1d, 1e) where the smallest junction of the three Josephson junction qubit is replaced by a low inductance superconducting quantum interference device (dc-SQUID) loop (the magenta loop in Figure 1e). The flux qubit - NV ensemble coupled system is measured by the qubit state using a built-in dc-SQUID (the biggest loop in Fig.1e sharing edges with the flux-qubit) which is inductively coupled to the qubit. When biasing the main loop close to half a flux quantum, the device is an effective two-level system [@Mooij99] described by the Hamiltonian $$\begin{aligned} H_{\rm qb}=\frac{h}{2}(\Delta\sigma_{x}+\epsilon\sigma_{z}),\end{aligned}$$ which is given in the basis of clockwise and counter-clockwise currents. Here, $\sigma_{x,z}$ are the Pauli spin $\frac{1}{2}$ matrices, $h\epsilon=2I_{\rm P}(\Phi_{\rm ex}-3\Phi_{0}/2)$ is the energy bias ($I_{\rm P}\approx300$ nA is the persistent current in the qubit, $\Phi_{\rm ex}$ is the external flux threading the qubit loop, and $\Phi_{0}=h/2e$ is the flux quantum), and $\Delta$ is the tunnel splitting. The energy splitting of the gap tunable flux-qubit is $hF=h\sqrt{\epsilon^{2}+\Delta^{2}}$ where $\epsilon$ and $\Delta$ can be controlled independently by the external magnetic flux threading the two loops. This type of flux-qubit can be tuned into resonance with an NV$^{-}$ ensemble in-situ at a base temperature of $\sim$12 mK while keeping the qubit at its optimum flux bias (degeneracy point). The total Hamiltonian of the coupled system is $$\begin{aligned} H&=& \frac{h}{2}(\Delta\sigma_{x}+\epsilon\sigma_{z}) \nonumber \\ &+& h \sum_{i} \left[ D S_{z,i}^2 + E (S_{x,i}^2-S_{y,i}^2)\right]\nonumber \\ &+& \frac{h}{2} \sum_i g_i \sigma_z S_{x,i}, \end{aligned}$$ where $i$ runs over the NV$^{-}$ centers which couple to the flux qubit. The corresponding coupling constant can be estimated using the Biot-Savart law at $g_i \sim 8.8$ kHz. In our situation here, the $\|\pm 1\rangle_i$ states of the NV$^{-}$ electronic spin are near degenerate and so our flux-qubit couples to both the $\left\|0\right\rangle_i \rightleftharpoons\left\|1\right\rangle_i $ and $\left\|0\right\rangle_i \rightleftharpoons\left\|-1\right\rangle_i$ transitions. This results in an effective coupling constant $\sqrt{2} g_i$ larger that generally anticipated. ![ [Energy spectrum of the flux-qubit coupled to ensemble of NV-centers.]{} a) Resonant frequencies indicated by peaks in the SQUID detector switching probability when a 500 ns long microwave pulse excites the system before the readout pulse. Data are represented as a function of the external magnetic flux through the effective qubit area ($\Phi_{ex}=\Phi_{m}+\Phi_{\alpha}/2$, where $\Phi_{m}$ and $\Phi_{\alpha}$ are the flux through the qubit main-loop and $\alpha$-loop, respectively). Inset: resonant frequency spectrum of the same flux-qubit over the same region without mounting the NV-diamond crystal. (b) Magnified view of the white dotted box region in (a). A vacuum Rabi splitting as large as $70$ MHz is clearly observed. Since the qubit phase relaxation time measured by spin-echo $T_{2}^{\rm echo}\approx0.25$ ${\mu}s$, the strong coupling condition is satisfied by the $\sqrt{N}$ enhancement in $g$. $N\approx 3\times10^{7}$ is the estimated number of electron spins in the ensemble strongly coupled to the flux-qubit.[]{data-label="FIG_3"}](Fig3){width="0.95\columnwidth"} ![[Vacuum Rabi oscillation of the flux-qubit NV-ensemble coupled system.]{} (a) Damped oscillation (blue curve) of an initially excited flux coupled to an NV-ensemble with the results from a phenomenological model shown as a thin red curve. Inset, schematic measurement sequence for the vacuum Rabi measurement. An Initial ground state $\|0\rangle_{\rm qubit}\|0\rangle_{\rm ens}$ is prepared by letting the system to relax longer than 800 $\mu$s under the base temperature (T=12 mK) of the dilution refrigerator at the optimal readout flux bias. Immediately following this a $\pi$-pulse is applied to the flux qubit resulting in the state $\|1\rangle_{\rm qubit}\|0\rangle_{\rm ens}$. The system is then brought into resonance non-adiabatically by a flux-bias shift current pulse through the control line 2. On resonance vacuum Rabi oscillation takes place for a given time; $\|1\rangle_{\rm qubit}\|0\rangle_{\rm ens} \rightleftharpoons \|0\rangle_{\rm qubit}\|1\rangle_{\rm ens}$. Finally, the SQUID detector reads the qubit state as a function of the time keeping the system on resonance. We repeat these measurement typically $2\times10^4$ times under the same condition to obtain good statistics. (b) 2D plot of the SQUID detector switching probability as a function of both the detuning by flux bias shift and the time keeping the flux-qubit NV-ensemble system at a given detuning. The white broken line corresponds to the switching probability shown in a). []{data-label="FIG_4"}](Fig4){width="0.95\columnwidth"} From the spectroscopic measurements, a clear anti-crossing was observed (Figure 3a) near the degeneracy point of the flux-qubit, while no gap was observed in the same flux-qubit prior to the mounting of the ensemble (inset in Figure 3a). We also note a narrow resonance at 2.878GHz of less than 1MHz width near these anti-crossings. This can be ascribed to the near degenerate excited states of the NV$^{-}$ ensemble and so indicates a strain-induced zero-field splitting coefficient $E$ of less than 1MHz. From the fine scan spectroscopy shown in (Figure 3b), a vacuum Rabi splitting near $g_{\rm ens} \sim 70$MHz was clearly observed confirming strong coupling between the flux-qubit and the NV$^{-}$ ensemble. Next from the measured vacuum Rabi splitting and our calculated value of $g_i$ we can estimate the number of NV$^{-}$ centers in the ensemble at $N=g^{2}_{\rm ens}/2g^{2}\approx 3.2\times10^{7}$, where the factor of $2$ in the denominator is due to the two-fold degeneracy of the excited $\|\pm 1\rangle_i$ states of an NV$^{-}$ center. This estimate is consistent with the density of NV$^{-}$ centers measured by PL spectroscopy in the whole sample ($1.1\times10^{18}$cm$^{-3}$) multiplied by the volume of centers coupling to the flux qubit (area 40$\mu$m$^{2}\times$ effective thickness 0.7$\mu m$). The PL spectroscopy approach gives the number of coupled centers as $\approx3.1\times10^{7}$. Next, we investigated the dynamics of our system in the time domain using a similar measurement cycle to that performed in qubit-LC resonator coupled systems [@Jan06]. We first excited the flux-qubit and then brought it into resonance with the NV$^{-}$ ensemble. Single energy quantum exchange between the flux-qubit and NV$^{-}$ ensemble at resonance manifests itself as the vacuum Rabi oscillations $$\|1\rangle_{\rm qb}\|0\rangle_{\rm ens} \rightleftharpoons \|0\rangle_{\rm qb}\|1\rangle_{\rm ens}$$ where $\|1\rangle_{\rm ens}=\frac{1}{\sqrt{N}}\sum_{i} S_{+,i}\|00\cdots0\rangle$, with $S_{+,i}=\|1\rangle_i \langle 0\|_i+\|-1\rangle_i \langle 0 \|_i$ being the raising operator of the $i$-th NV$^{-}$ spin to both the $\|\pm 1\rangle$ states. (Figure 4a) clearly shows vacuum Rabi oscillations between the flux-qubit and ensemble of electronic spins at the 2.878 GHz resonance. The decay time of the oscillations however is approximately $20$ ns. This is much shorter than the relaxation time of both the flux-qubit ($T_{1,\rm qb}\sim150$ ns) and the NV$^{-}$ ensemble ($T_{1,\rm NV} \gg 10 \mu$s). As we tune the flux-qubit away from the 2.878 GHz resonance, the decay time associated with vacuum Rabi measurement becomes significantly longer (Figure 4b). From these results, one must conclude that a source of strong dephasing of unknown origin exists in the system near resonance. There are several likely sources. The most probable is a large electron spin $\frac{1}{2}$ bath from the P1 (nitrogen atom substituting a carbon atom) centers present in our HPHT Ib-type diamond crystal used to prepare the NV$^{-}$ ensemble. In our situation, where there is no external magnetic field, the NV$^{-}$ centers and P1 centers naturally couple[@Marcos10]. Hanson et. al [@Hanson08] have shown an enhanced decay may result. The P1 center issue can be eliminated to a large extent by applying an external magnetic field to split the $\left\|\pm 1 \right \rangle$ NV$^{-}$ states. A 1 mT field could split these by approximately 15 MHz, detuning the P1 centers and thus significantly improving the dephasing time of the coupled system. We can also decrease the number of P1 centers in the sample (from 100 ppm to 1 ppm) by using different synthesized diamond crystals. In addition, by using non HPHT Ib-type crystals we can remove the effect of other natural defects that may be present. Finally there is also a strong hyperfine interaction ($\sim 100$ MHz) between the NV$^{-}$ electron spin and $^{13}$C nuclear spins. Without the nuclear spins being initially polarized, unwanted dephasing will result. By polarizing the nuclear spins this source of dephasing can be removed. This should allow us to observe vacuum Rabi oscillations where we are limited by $T_2$ of the flux-qubit. In conclusion, we have experimentally demonstrated strong coherent coupling between a flux-qubit and an ensemble of nitrogen-vacancy color centers in single crystal diamond. Furthermore, we have observed, via vacuum Rabi oscillations, the coherent exchange (transfer) of a single quantum of energy. This is the first step towards the realization of a long lived quantum memory for condensed matter systems with an additional potential future application as an interface between the microwave and optical domains. [Acknowledgments]{}: We thank T. Tawara, H. Gotoh and T. Sogawa for optical measurement in the early stage of this work. This work was supported in part by the Funding Program for World-Leading Innovative R[&]{}D on Science and Technology (FIRST), Scientific Research of Specially Promoted Research, Grant No.18001002 by MEXT, Grant-in-Aid for Scientific Research on Innovative Areas Grant No.22102502, and Scientific Research(A) Grant No.22241025 by Japanese Society for the Promotion of Science (JSPS). M.S.E. was supported with a JSPS fellowship. Clarke, J. & Wilhelm, F. K. [Nature]{} [453]{}, 1031 (2008). Nakamura,Y., Pashkin,Yu.A. & Tsai,J. S. [Nature]{} (London) [398]{}, 786 (1999). Vion, D., Aassime, A., Cottet, A., Joyez, P., Pothier, H., Urbina, C., Esteve, D. & Devoret, M. H. [Science]{} [296]{}, 886 (2002). Chiorescu, I., Nakamura, Y., Harmans, C. J. P. M. & Mooij, J. E. [Science]{} [299]{}, 1869 (2003). Sillanpää, M. A., Park, J. I. & Simmonds, R. W. [Nature]{} [449]{}, 438 (2007). Majer, J., Chow, J. M., Gambetta, J. M., Koch, J., Johnson, B. R., Schreier, J. A., Frunzio, L., Schuster, D. I., Houck, A. A., Wallraff, A., Blais, A., Devoret, M. H., Girvin, S. M. & Schoelkopf, R. J. [Nature]{} [449]{}, 443 (2007). DiCarlo, L., Chow, J. M., Gambetta, J. M., Bishop, L. S., Johnson, Schuster, D. I., Majer, J., Blais, A., Frunzio, L., Girvin, S. M. & Schoelkopf, R. J. [Nature]{} [460]{}, 240 (2009). Ansmann, M., Wang, H., Bialczak, R. C., Hofheinz, M., Lucero, E., Neeley, M., O’Connell, A. D., Sank, D., Weides, M., Wenner, J., Cleland, A. N. & Martinis, J. M. [Nature]{} [461]{}, 504 (2009). Neeley, M., Bialczak, R. C., Lenander, M., Lucero, E., Mariantoni, M., O’Connell, A. D., Sank, D., Wang, H., Weides, M., Wenner, J., Yin, Y., Yamamoto, T., Cleland, A. N. & Martinis, J. M. [Nature]{} [467]{}, 570 (2010). DiCarlo, L., Reed, M. D., Sun, L., Johnson, B. R., Chow, J.M., Gambetta, J. M., Frunzio, L., Girvin, S. M., Devoret, M. H. & Schoelkopf, R. J. [Nature]{} [467]{}, 574 (2010). S[ø]{}rensen, Anders S., van der Wal, Caspar H., Childress, Lilian I. & Lukin, Mikhail D. [Phys. Rev. Lett.]{} [92]{}, 063601 (2004). Tian, L., Rabl, P., Blatt, R. & Zoller, P. [Phys. Rev. Lett.]{} [92]{}, 247902 (2004). Rabl, P. , DeMille, D., Doyle, J. M., Lukin, M. D., Schoelkopf, R. J. & Zoller, P. [Phys. Rev. Lett.]{} [97]{}, 033003 (2006). Marcos, D., Wubs, M. , Taylor, J. M., Aguado, R., Lukin, M. D. & Sørensen, A. S. [Phys. Rev. Lett.]{} [105]{}, 210501 (2010). Brune, M., Schmidt-Kaler, F., Maali, A., Dreyer, J., Hagley, E., Raimond, J. M. & Haroche, S. [Phys. Rev. Lett.]{} [76]{}, 1800 (1996). Schuster, D. I., Sears, A. P., Ginossar, E., DiCarlo, L., Frunzio, L., Morton, J. J. L., Wu, H. , Briggs, G. A. D., Buckley, B. B., Awschalom, D. D. & Schoelkopf, R. J. [Phys. Rev. Lett.]{} [105]{}, 140501 (2010). Kubo, Y., Ong, F. R., Bertet, P., Vion, D., Jacques, V., Zheng, D., Dréau, A., Roch, J.-F., Auffeves, A., Jelezko, F. ,Wrachtrup, J., Barthe, M. F., Bergonzo, P. & Esteve, D. [Phys. Rev. Lett.]{} [105]{}, 140502 (2010). Amsüss, R., Koller, C., Nöbauer, T., Putz, S., Rotter, S., Sandner, K., Schneider,S., Schramböck, M., Steinhauser, G., Ritsch, H., Schmiedmayer, J. & Majer, J. [arXiv:1103.1045]{} Raizen, M. G., Thompson, R. J., Brecha, R. J., Kimble, H. J. & Carmichael, H. J. [Phys. Rev. Lett.]{} [63]{}, 240 (1989). Naydenov, B., Richter, V., Beck, J., Steiner, M., Neumann, P., Balasubramanian, G., Achard, J., Jelezko, F., Wrachtrup, J. & Kalish, R. [Appl. Phys. Lett.]{} [96]{}, 163108 (2010). Neumann, P., Kolesov, R., Jacques, V., Beck, J., Tisler, J., Batalov, A., Rogers, L., Manson, N. B., Balasubramanian, G., Jelezko, F. & Wrachtrup, J. [New J. Phys.]{} [11]{}, 013017 (2009). Zhu, X., Kemp, A., Saito, S., & Semba, K. [Appl. Phys. Lett.]{} [97]{}, 102503 (2010). Fedorov, A., Feofanov, A. K., Macha, P., Forn-Dı´az, P., Harmans,C. J. P. M. & Mooij, J. E. [Phys. Rev. Lett.]{} [105]{}, 060503 (2010). Mooij, J. E., Orlando, T. P., Levitov, L., Tian, L., van der Wal, C. H. & Lloyd S. [Science]{} [285]{}, 1036, (1999). Johansson, J., Saito, S., Meno, T., Nakano, H., Ueda, M., Semba, K. & Takayanagi, H. [Phys. Rev. Lett.]{} [96]{}, 127006 (2006). Hanson, R., Dobrovitski, V. V., Feiguin, A. E., Gywat, O. & Awschalom, D. D. [Science]{} [320]{}, 352, (2008). Mizuochi, N., Neumann, P., Rempp, F., Beck, J., Jacques, V., Siyushev, P., Nakamura, K., Twitchen, D., Watanabe, H., Yamasaki, S., Jelezko & F., Wrachtrup, J. [Phys. Rev.]{} [B80]{}, 041201(R) (2009). Jelezko, F., Gaebel, T., Popa, I., Gruber, A. & Wrachtrup, J. [Phys. Rev. Lett.]{} [92]{}, 076401 (2004). Gruber, A., Dräbenstedt, A., Tietz, C., Fleury, L., Wrachtrup, J. & von Borczyskowski, C. [Science]{} [276]{}, 2012, (1997). Kurtsiefer,Ch., Zarda, P., Mayer, S. & Weinfurter, H. [Phys. Rev. Lett.]{} [85]{}, 290 (2000).	{ "pile_set_name": "ArXiv" }
--- abstract: 'In our previous work we have introduced the logic STL\, an extension of Signal Temporal Logic (STL) that allows value freezing. In this paper, we define robustness measures for STL\ by adapting the robustness measures previously introduced for Metric Temporal Logic (MTL). Furthermore, we present an algorithm for STL\* robustness computation, which is implemented in the tool Parasim. Application of STL\* robustness analysis is demonstrated on case studies.' author: - 'L. Brim, T. Vejpustek, D. Šafránek, and J. Fabriková [^1]' bibliography: - 'hsb2013.bib' title: 'Robustness Analysis for Value-Freezing Signal Temporal Logic' --- Introduction ============ A particular place among formalisms adopted by systems biology is occupied by temporal logics, which serve as a language for description of biological systems behaviour. Resulting temporal formulae can be used during computer-aided system analysis, such as model checking [@modelchecking], which automatically verifies whether a model satisfies given temporal formula. Methods based on temporal logics have been successfully employed to study biological phenomena [@STL-use; @biocham-practice; @eziocav11] (see [@SFM] for review). Since most of current models developed in computational systems biology have the form of ordinary differential equations, model checking cannot be directly employed and is typically replaced with a non-exhaustive procedure of monitoring [@Maler_STL]. In this setting, a (finite) set of signals representing individual time-courses of the model is monitored wrt a given temporal specification. In particular, the respective temporal logics are interpreted over individual signals that are most typically simplified to discrete timed state sequences (time series) approximating the continuous trajectories by means of numerical simulation. Temporal logics fitting this interpretation are Metric Temporal Logic (MTL) [@MTL] and Signal Temporal Logic (STL) [@Maler_STL], which allow quantifying modalities with the time frame represented by a closed time interval. MTL possesses both discrete and continuous semantics, as it can be interpreted over both infinite timed state sequences and continuous signals. STL is practically focused and is defined for piece-wise linear approximations of continuous signals. Temporal logics are satisfactorily used in systems biology to express statements about a single instance of system behaviour such as in five minutes, concentration of glucose will be greater than 0.8. However, many biological hypotheses contain relative temporal references, e.g., after protein $P$ reaches the maximum concentration, a steady concentration of $P$ is reached which is less than half of the maximum. Such a scenario can be found, e.g., in feed-forward genetic regulatory circuits generating pulses in expression signals [@kaplanthe2008]. In common temporal logics, such a general query cannot be expressed. This is because the values in different time points cannot be compared, i.e., the property in five minutes, concentration of glucose will rise by 0.2, which relates glucose concentration at current time and in the future, cannot be specified. Of specific interest is oscillatory behaviour, e.g., a sequence of gradually increasing peaks followed by a limit cycle with a stable amplitude [@limitcycles]. In order to express the increasing amplitude, it is necessary to detect local extremes in signals and compare respective signal values. This cannot be achieved using common temporal logics. Signals with a series of increasing local maxima have been observed, e.g., in response of FGF signalling pathways transferring stimuli from mutated FGFR3 receptors to target effectors affecting bone cells growth [@Krejci2004152]. Since the mentioned behaviour correlates with the phenotype of dysplasia, it is necessary to develop models that mechanistically capture the respective signalling pathways and to analyse circumstances under which the undesired behaviour occurs. This makes a necessary step before designing a targeted medical treatment. To this end, temporal logics and verification procedures which allow to capture and analyse such complex phenotypes have to be developed. In [@Dluhos-stlstar], we have introduced a new temporal logic [STL\]{} which alleviates limitations mentioned above. Expressiveness of [STL\]{} is enhanced by signal-value freeze operator which stores values at certain time, which may be referred to in the future. This allows [STL\]{} to specify and distinguish various dynamic aspects which occur in biological systems, in addition to the phenomena mentioned above, these can be, e.g., damped oscillations [@sybi-periodic] or local extremes in species concentration. It is worth noting that some more complex queries can be expressed in traditional temporal logic by including signal derivatives into atomic propositions. However, this does not directly apply to queries mentioned above. One can express the presence and shape of a local extreme by using the first and second derivative, but still the values in particular time points have to be compared in order to express the complex queries. An important concept associated with biological systems and temporal logics is robustness, the ability of a system to maintain its function against perturbations [@Kitano-robustness]. Since system function can be expressed in the terms of temporal logic, we speak of robustness with respect to a temporal logic formula, which can be quantified and computed [@FP-robustness; @robustness-property]. Robustness significantly enhances model analysis and gives an optimization goal for model parameter estimation/synthesis [@STL-robustness; @STL-parameters; @FagesRobustness]. This paper introduces the notion of robustness in the value-freezing logic [STL\]{} setting. In particular, we extend the continuous and discrete measure defined for MTL by [Fainekos et al.]{} [@FP-robustness] to the semantic domain of [STL\]{}. Robustness of the input signal with respect to [STL\]{} formula delineates the robust neighbourhood of the signal (the maximal “tube” around the signal where the formula is satisfied). The robustness measure we propose (Section \[sec:robustness\]) is defined inductively wrt the formula structure and is based on a distance metrics employed on the signal domain extended with (multiple) dimensions representing the frozen time points. The theoretical framework is computationally supported with an algorithm based on solving the optimization problem (Section \[sec:computation\]) provided that the logic is restricted to linear predicates. Special consideration is given to optimization of the formula to overcome unnecessary computational overhead. Implementation of our algorithm is included as a part of Parasim [@parasim], a tool aimed as a modular environment for monitoring and robustness analysis of kinetic models. To demonstrate the usage and evaluate the performance, we present case studies of two simple kinetic models (Section \[sec:casestudy\]). Related Work ------------ Robustness measures have been defined for three temporal logics targeting deterministic continuous systems: STL [@STL-robustness], MTL [@FP-robustness] and QFLTL [@robustness-property]. We adopt the concept of behaviour-based robustness introduced on a fragment of MTL by Fainekos et al. [@FP-robustness], who define robustness measure for MTL formulae with discrete [@FP-robustness0] and continuous [@FP-robustness] semantics. In [@FP-robustness], Fainekos et. al prove a theorem connecting discrete and continuous robustness, which is valuable for robustness computation. A recent tool [@FP-taliro] implements the method. [Donzé et al.]{} [@Maler_STL] use STL to define a distinct robustness measure, albeit constructed from [@FP-robustness], and propose its application for space exploration [@STL-robustness; @STL-parameters], which was implemented in the Breach Toolbox [@STL-breach]. The work is further improved from the computational point of view in [@DFM13]. Our implementation (Parasim) is based on a simplified version of the robustness analysis algorithm for STL where the sensitivity-based computation of local robustness is replaced with direct computation of trajectories distance. The extension for [STL\]{} as presented in Section \[sec:computation\] is implemented in this setting. Fages et al. [@robustness-property] introduced property-based approach to robustness that fixes input behaviour and examines the formula. Basically, it measures the extent to which the formula can be modified while preserving its satisfaction. The tool BioCham implements this idea [@biocham]. Extended LTL logic with constraints over real numbers (quantifier-free LTL) is employed being defined for finite discrete time-series. It is worth noting that the problem of formula satisfiability is undecidable for MTL [@MTL]. To achieve decidability, Alur and Henzinger specified further conditions on intervals associated with temporal operators [@MITL]. The result, metric interval temporal logic, requires all intervals to be non-singular and is interpreted over timed state-sequences where time points are replaced with consecutive time intervals. STL was introduced by Maler and Nickovic in [@Maler_STL] as a basis for their monitoring procedure. Technically, it comprises a variant of MITL interpreted over real signals. Because of its practical purpose, in [@Dluhos-stlstar] we selected STL as a good candidate for extension with value-freezing. Background ========== [STL\]{} is evaluated over finite time continuous signals (finite signals for short). \[def:signal\] Let $n\in{\mathbb{N}}$ and $T=[0,r]$ where $r\in{\mathbb{R}}^+$. Then $s:T\to{\mathbb{R}}^n$ is a bounded continuous-time signal and $T$ its time domain. We denote $l(s)=r$ the length of signal $s$. Signal value freezing is facilitated by the following structure which is used to store time values at various time points which then can be referred to in predicates. \[def:frozen\_time\] Let ${\mathcal{I}}$ be a finite index set. Frozen time vector is a function: $$t:{\mathcal{I}}\to{\mathbb{R}}^+_0$$ The symbol $t_i=t(i)$ is referred to as $i$-th frozen time. For convenience reasons and without loss of generality, we will henceforth assume that an index set ${\mathcal{I}}=\{1,\ldots,m\}$ is given, where $m\in{\mathbb{N}}$. Predicates comprise Boolean expressions over values of a signal $s$ at time $t$ and each frozen time $t_i$, where $x_j$ denotes the $j$-the component of the signal at time $t$, i.e. $s(t)=(x_1,\ldots,x_j,\ldots,x_n),$ and $x_j{\ensuremath{\sp{\ast_{i}{}}}}$ the $j$-th component at time $t_i$. When $\|{\mathcal{I}}\|=1$, we usually omit the index of asterisk, e.g. $x_i=x_i{\ensuremath{\sp{\ast_{1}{}}}}$. We consider only predicates given by linear inequalities, so that analytic expressions of predicate robustness is possible. \[def:stl\predicate\] Let $n\in{\mathbb{N}}$, $b\in{\mathbb{R}}$ and $a_{ij}\in{\mathbb{R}}$ where $i\in \{0\}\cup{\mathcal{I}}$, $j\in\{1,\ldots,n\}$ and not all $a_{ij}$ are zero. A predicate* is defined as a subset of ${\mathbb{R}}^n\times\left({\mathbb{R}}^n\right)^{\mathcal{I}}$ such that: $$\sum_{j=1}^na_{0j}x_j+\sum_{i=1}^{\|{\mathcal{I}}\|}\sum_{j=1}^na_{ij}x_j{\ensuremath{\sp{\ast_{i}{}}}}+b\ge 0$$ Predicates are specified by the set of associated coefficients $a_{ij},b$ (where coefficients $a_{0j}$ are connected with the current time $t$). Therefore, for convenience reasons, we will use these coefficients to represent predicates. Predicates with all coefficients $a_{ij}$ zero were omitted since they are of the form $b\ge 0$ and, therefore, trivially true or false. Predicates with equality (i.e. having $=$ in place of $\ge$), although theoretically possible, lack practical value, as they are not robust (small perturbation may invalidate the property). This has been already argued in [@Dluhos-stlstar], albeit without defining the concept of robustness. Since robustness of predicates with strict and non-strict inequalities does not differ, we consider only non-strict inequalities. Freeze operator is used to store the time point into frozen time vector, thus facilitating signal value freezing. The following definition introduces an auxiliary concept of storing the current time $t$ as the $i$th component of the frozen time vector. \[def:stl\store\] Let $t$ be frozen time vector, $i,j\in{\mathcal{I}}$ and $t\in{\mathbb{R}}^+_0$. Freezing $i$th component of $t$ in $t$* is denoted as ${\ensuremath{t[i\leftarrowt]}}$ and defined: $${\ensuremath{t[i\leftarrowt]}}(j)=\begin{cases} t &i=j\\ t_j&i\neq j\\ \end{cases}$$ \[def:stl\syntax\] Syntax of [STL\]{} is defined by the following grammar: $${\varphi}::= \mu \mid \top \mid \neg{\varphi}\mid {\varphi}_1\lor{\varphi}_2 \mid {\varphi}_1{\@ifstar{{\ensuremath{\operatorname{\textbf{U}}_{}}}}{{{\ensuremath{\operatorname{\textbf{U}}_{[}}}]}}}I{\varphi}_2 \mid {\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}i{\varphi}$$ where $i\in{\mathcal{I}}$, $\top$ denotes the true constant, $\mu$ is a predicate as of Definition \[def:stl\predicate\] and $I\subseteq{\mathbb{R}}^+_0$ a closed non-singular interval. Note that all Boolean connectives and temporal operators ${\@ifstar{{\ensuremath{\operatorname{\textbf{F}}_{}}}}{{{\ensuremath{\operatorname{\textbf{F}}_{[}}}]}}}{}$ and ${\@ifstar{{\ensuremath{\operatorname{\textbf{G}}_{}}}}{{{\ensuremath{\operatorname{\textbf{G}}_{[}}}]}}}{}$ can be defined using the basic operators defined above. Similarly to predicates, when $\|{\mathcal{I}}\|=1$, we usually omit the index of freeze operator, as in ${\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}{\@ifstar{{\ensuremath{\operatorname{\textbf{G}}_{}}}}{{{\ensuremath{\operatorname{\textbf{G}}_{[}}}]}}}I(x>x)={\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}1{\@ifstar{{\ensuremath{\operatorname{\textbf{G}}_{}}}}{{{\ensuremath{\operatorname{\textbf{G}}_{[}}}]}}}I(x>x{\ensuremath{\sp{\ast_{1}{}}}})$. Henceforth, let $i,\mu,{\varphi},{\varphi}_1,{\varphi}_2$ be the same as in Definition \[def:stl\syntax\]. \[def:stl\semantics\] Let $s\in\left({\mathbb{R}}^n\right)^T$ be a signal, $t\in T$ a time point and $t\in T^{\mathcal{I}}$ a frozen time vector. Formula satisfaction is defined inductively: $$\begin{array}{lcl} (s,t,t)\models\top && \\ (s,t,t)\models\mu &\iff& (s(t),s \circ t)\in\mu\\ (s,t,t)\models\neg{\varphi}&\iff& (s,t,t)\not\models{\varphi}\\ (s,t,t)\models{\varphi}_1\lor{\varphi}_2 &\iff& (s,t,t)\models{\varphi}_1\lor(s,t,t)\models{\varphi}_2\\ (s,t,t)\models{\varphi}_1{\@ifstar{{\ensuremath{\operatorname{\textbf{U}}_{}}}}{{{\ensuremath{\operatorname{\textbf{U}}_{[}}}]}}}I{\varphi}_2 &\iff& \exists\ t'\in t\oplus I:(s,t',t)\models{\varphi}_2\,\land\\ && \forall\ t''\in[t,t']:(s,t'',t)\models{\varphi}_1\\ (s,t,t)\models{\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}i{\varphi}&\iff&(s,t,{\ensuremath{t[i\leftarrowt]}})\models{\varphi}\end{array}$$ Operator $\circ$ is used to denote function composition, i.e. $(s\circ t)\in\left({\mathbb{R}}^n\right)^{\mathcal{I}}$ and $(s\circ t)(i)=s(t_i)$ and $t\oplus I$ stands for $\{t+u\mid u \in I\}$. Let $s\in\left({\mathbb{R}}^n\right)^T$ be signal and ${\varphi}$ formula. Formula satisfaction by signal is given: $$s\models{\varphi}\iff (s,0,{\mathbf{0}})\models{\varphi}$$ where ${\mathbf{0}}$ denotes the zero frozen time vector, i.e. $\{(i,0)\|i\in{\mathcal{I}}\}$. Intuitively, interpretation of ${\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}i{\varphi}$ is the following: freeze operator stores signal values at the time of ${\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}i{\varphi}$ evaluation, which can then be referred to using index $i$ in predicates of ${\varphi}$. An example property, “in the next five time units, $x$ increases by $8$”* can be specified as: $${\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}{\@ifstar{{\ensuremath{\operatorname{\textbf{F}}_{}}}}{{{\ensuremath{\operatorname{\textbf{F}}_{[}}}]}}}{0,5}(x\ge x+8)$$ where $x$ refers to value of $x$ at time $0$. When intervals associated with until operators are bounded, satisfaction of a given formula can be decided on any finite signal of sufficient length. This length can be determined from the formula structure in a way similar to [@Maler_STL] and corresponds to the furthest time point (among all possible signals) which has to be examined in order to determine formula satisfaction. This clearly also holds for frozen time values. \[def:necessary\_length\] Let ${\varphi}$ be a formula. The necessary input length* for ${\varphi}$, $l({\varphi})$ is defined inductively: $$\begin{aligned} l(\top)=l(\mu) &= 0\\ l(\neg{\varphi})=l({\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}i{\varphi}) &= l({\varphi})\\ l({\varphi}_1\lor{\varphi}_2) &= \max(l({\varphi}_1),l({\varphi}_2))\\ l({\varphi}_1{\@ifstar{{\ensuremath{\operatorname{\textbf{U}}_{}}}}{{{\ensuremath{\operatorname{\textbf{U}}_{[}}}]}}}I{\varphi}_2) &= \max(l({\varphi}_1),l({\varphi}_2))+\sup I \end{aligned}$$ When $l(s)<l({\varphi})$ we state that $s\not\models{\varphi}$. Frozen time indices and freeze operators share some similarities with variables and quantifiers of predicate logic. We may distinguish free and bound indices, where index $i$ is free if it is used in a predicate (i.e. coefficient $a_{ij}$ is not zero for some $j$) and is not in the scope of operator ${\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}i$. Naturally, whenever $i$ is free in ${\varphi}$, then $s\models{\varphi}$ iff $s\models{\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}i{\varphi},$ since $t_i$ is zero in both cases. Additionally, we may substitute for free indices of a formula in a manner similar to variable substitution. However, it only makes sense to substitute one index for another, which we will denote index renaming and express as ${\varphi}[\pi]$ where $\pi$ is a total function on ${\mathcal{I}}$ (but not necessarily a permutation – two indices can be renamed to one) or ${\varphi}[k/l]$, where $k$ is renamed to $l$. To preserve formula semantics, renaming is only safe when no free index becomes bound after renaming in any subformula. Robustness Measures for STL\* {#sec:robustness} ============================= Following from [STL\]{} semantics, robustness of signal $s$ with respect to formula ${\varphi}$ is given for each time point $t$ and frozen time vector $t$ and denoted by $\rho({\varphi},s,t,t)$. We also define $\rho({\varphi},s)=\rho({\varphi},s,0,{\mathbf{0}})$. Robustness of signal $s$ with respect to formula ${\varphi}$ is a value, which under-approximates the distance of $s$ from the set of signals where ${\varphi}$ has different truth value [@FP-robustness]. To express this formally, we first need to define certain basic concepts (where $S$ is a set of signals): - Distance of signals is given by their maximum pointwise distance: $d(s,s')=\max_{t\in{\mathbb{R}}^+_0}d(s(t),s'(t))$ - Set distance* is given by minimum distance to the set: ${{\ensuremath{\operatorname{\mathbf{dist}}}}}(s,S)=\min\{d(s,s')\mid s'\in S\}$ - Set depth is given by set distance to the complement: ${{\ensuremath{\operatorname{\mathbf{depth}}}}}(s,S)={{\ensuremath{\operatorname{\mathbf{dist}}}}}\left(s,\overline{S}\right)$ - Signed distance is given: ${{\ensuremath{\operatorname{\mathbf{Dist}}}}}(s,S)=\begin{cases}-{{\ensuremath{\operatorname{\mathbf{dist}}}}}(s,S)&s\notin S\\{{\ensuremath{\operatorname{\mathbf{depth}}}}}(s,S)&s\in S\end{cases}$ The value $\rho({\varphi},s)$ underapproximates the signed distance of $s$ from the set of all signals satisfying ${\varphi}$, ${\ensuremath{\mathcal{L}}}({\varphi})$, i.e. $\|\rho({\varphi},s)\|\le\|{{\ensuremath{\operatorname{\mathbf{Dist}}}}}(s,{\ensuremath{\mathcal{L}}}({\varphi}))\|$ holds while their signs are identical. The absolute value of $\rho({\varphi},s)$ thus delineates an equidistant tube where all signals satisfy ${\varphi}$ if and only if $s$ does – the robust neighbourhood of $s$ (see Figure \[fig:tube\]). ![Signal $s$ (blue, thick) and borders of its robust neighbourhood (blue, dashed) with an example of a signal (red) contained in the robust neighbourhood (adapted from [@FP-robustness]).[]{data-label="fig:tube"}](tube.pdf){width="40.00000%"} It would be desirable to define the robustness equal to the signed distance; however, by [@FP-robustness], the robustness computation would not be feasible then. In order to be sound, the robustness definition has to satisfy the following property (for any ${\varphi}$, $s$, $t$ and $t$): $$\label{eq:robust_sound} -{{\ensuremath{\operatorname{\mathbf{dist}}}}}\left(s,{\ensuremath{{\ensuremath{\mathcal{L}}}_{t,t}}}({\varphi})\right)\le\rho({\varphi},s,t,t)\le{{\ensuremath{\operatorname{\mathbf{depth}}}}}\left(s,{\ensuremath{{\ensuremath{\mathcal{L}}}_{t,t}}}({\varphi})\right),$$ where ${\ensuremath{{\ensuremath{\mathcal{L}}}_{t,t}}}({\varphi})=\{s\mid (s,t,t)\models{\varphi}\}$. Since ${{\ensuremath{\operatorname{\mathbf{depth}}}}}(s,{\ensuremath{{\ensuremath{\mathcal{L}}}_{t,t}}}({\varphi}))=0$ when $(s,t,t)\not\models{\varphi}$ (and analogously for ${{\ensuremath{\operatorname{\mathbf{dist}}}}}$), this actually requires that: 1. $s\models{\varphi}\implies 0\le\rho({\varphi},s,t,t)\le{{\ensuremath{\operatorname{\mathbf{depth}}}}}\left(s,{\ensuremath{{\ensuremath{\mathcal{L}}}_{t,t}}}({\varphi})\right)$, 2. $s\not\models{\varphi}\implies -{{\ensuremath{\operatorname{\mathbf{dist}}}}}\left(s,{\ensuremath{{\ensuremath{\mathcal{L}}}_{t,t}}}({\varphi})\right)\le\rho({\varphi},s,t,t)\le 0$. Robustness is defined inductively for each logical connective from its semantics in such manner that Boolean functions $\land$ and $\lor$ are replaced by real functions $\min$ and $\max$ (respectively). Quantifiers in the semantics of operator ${\@ifstar{{\ensuremath{\operatorname{\textbf{U}}_{}}}}{{{\ensuremath{\operatorname{\textbf{U}}_{[}}}]}}}{}$ can then be expressed by infinite disjunction or conjunction. Robustness wrt predicate $\mu$ is defined as ${{\ensuremath{\operatorname{\mathbf{Dist}}}}}(s,{\ensuremath{{\ensuremath{\mathcal{L}}}_{t,t}}}(\mu))$, i.e. the ideal value without underapproximation. If $\rho(\mu,s)$ was lower, it would diminish resulting robustness value, for robustness wrt formula cannot be greater than robustness wrt any of its predicates. Soundness of this definition (property ) is, naturally, proved inductively wrt formula structure. This has already been established by [Fainekos et al.]{} in [@FP-robustness], albeit for MTL which does not allow signal value freezing. Nevertheless, their definition can be directly extended for [STL\]{}. Intuitively, this is due to frozen time values being only stored by freeze operators and retrieved in predicates, which does not affect other logical connectives. The full proof can be found in [@src] (page 83). Consequently, we have to define robustness for the freeze operator. It follows from its semantics: $$\label{eq:op:frz1} {\ensuremath{{\ensuremath{\mathcal{L}}}_{t,t}}}({\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}i{\varphi})=\{s \mid (s,t,t)\models{\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}i{\varphi}\}=\{s \mid (s,t,{\ensuremath{t[i\leftarrowt]}})\models{\varphi}\}={\ensuremath{{\ensuremath{\mathcal{L}}}_{t,{\ensuremath{t[i\leftarrowt]}}}}}({\varphi})$$ Thus, robustness of freeze operator can be defined in the following manner: $$\label{eq:op:frz2} \rho({\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}i{\varphi},s,t,t)=\rho({\varphi},s,t,{\ensuremath{t[i\leftarrowt]}})$$ Assume $-{{\ensuremath{\operatorname{\mathbf{dist}}}}}(s,{\ensuremath{{\ensuremath{\mathcal{L}}}_{t,t}}}({\varphi}))\le\rho({\varphi},s,t,t)\le{{\ensuremath{\operatorname{\mathbf{depth}}}}}(s,{\ensuremath{{\ensuremath{\mathcal{L}}}_{t,t}}}({\varphi}))$ for any $t,t$. Therefore, it also holds for $t$ and ${\ensuremath{t[i\leftarrowt]}}$ and thus: $$-{{\ensuremath{\operatorname{\mathbf{dist}}}}}(s,{\ensuremath{{\ensuremath{\mathcal{L}}}_{t,{\ensuremath{t[i\leftarrowt]}}}}}({\varphi}))\le\rho({\varphi},s,t,{\ensuremath{t[i\leftarrowt]}})\le{{\ensuremath{\operatorname{\mathbf{depth}}}}}(s,{\ensuremath{{\ensuremath{\mathcal{L}}}_{t,{\ensuremath{t[i\leftarrowt]}}}}}({\varphi}))$$ From which follows the validity of for $\rho({\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}i{\varphi},s,t,t)$. [STL\]{} robustness for logical connectives is presented in Figure \[fig:robustness\]. $$\begin{array}{l@{\ =\ }l} \rho(\top,s,t,t) & +\infty\\ \rho(\neg{\varphi},s,t,t) & -\rho({\varphi},s,t,t)\\ \rho({\varphi}_1\lor{\varphi}_2,s,t,t) & \max\left(\rho({\varphi}_1,s,t,t),\rho({\varphi}_2,s,t,t)\right)\\ \rho({\varphi}_1{\@ifstar{{\ensuremath{\operatorname{\textbf{U}}_{}}}}{{{\ensuremath{\operatorname{\textbf{U}}_{[}}}]}}}I{\varphi}_2,s,t,t) & \max\limits_{t'\in t\oplus I}\min\left(\rho({\varphi}_2,s,t',t),\min\limits_{t''\in[t,t']}\rho({\varphi}_1,s,t'',t)\right)\\ \rho({\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}i{\varphi},s,t,t) & \rho({\varphi},s,t,{\ensuremath{t[i\leftarrowt]}}) \end{array}$$ Robustness of Predicates ------------------------ Finding ${{\ensuremath{\operatorname{\mathbf{Dist}}}}}(s,{\ensuremath{{\ensuremath{\mathcal{L}}}_{t,t}}}(\mu))$ generally constitutes a convex analysis problem [@FP-robustness]. Thus, it could be solved using convex programming for each $t$ and $t$, which would, however, greatly increase computation time, and therefore, analytic solution is preferable. To this end, we have restricted [STL\]{} predicates to be linear. For predicate $\mu$ with coefficients $a_{ij},b$, the problem of finding ${{\ensuremath{\operatorname{\mathbf{Dist}}}}}(s,{\ensuremath{{\ensuremath{\mathcal{L}}}_{t,t}}}(\mu))$ can be reduced to optimization of $f({\mathbf{d}})=\max_i\sum_jd_{ij}^2$ (where $i\in{\mathcal{I}}$ and $j\in\{1,\ldots,n\}$) under the constraint $\sum_i\sum_ja_{ij}d_{ij}+{\varepsilon}=0$ for some positive ${\varepsilon}$. This is a non-trivial problem, since $f$ is not differentiable at point ${\mathbf{d}}$ where $f({\mathbf{d}})=\sum_jd_{kj}^2=\sum_jd_{lj}^2$ for some $k\neq l$. To solve it, generalized method of Lagrange multipliers from [@nonsmooth_analysis] was used, resulting in the following definition of the robustness $\rho$ (detailed derivation can be found in [@src] (page 47)). \[def:predicate\_robustness\] Let $\mu$ be a predicate with coefficients $a_{ij},b$. Then $$\rho(\mu,s,t,t)=\frac{\sum_ja_{0j}s_j(t)+\sum_i\sum_ja_{ij}s_j\left(t_i\right)+b}{\sum_i\sqrt{\sum_ja_{ij}^2}}$$ for arbitrary $s$, $t$, $t$, $i$ ranging over ${\mathcal{I}}$, $j$ ranging over $\{1,\ldots,n\}$. The numerator corresponds to the left-hand side value of the predicate. It holds that $\rho(\mu,s,t,t)={{\ensuremath{\operatorname{\mathbf{Dist}}}}}(s,{\ensuremath{{\ensuremath{\mathcal{L}}}_{t,t}}}(\mu))$, unless some time points given by $t$ and $t$ are equal. This originates from the optimization problem, where $t_k=t_l$ (or $t=t_k$) would constitute another constraint, which might change the solution. Suppose that $t_k=t_l$ (reasoning for $t=t_k$ is similar). We can merge (sum) coefficients $a_{kj}$ and $a_{lj}$ for any given $j$, which effectively reduces the number of considered frozen times. Robustness of predicates with merged coefficients is greater, since the denominator of definition \[def:predicate\_robustness\] becomes smaller as $\sqrt{\sum_j\left(a_{kj}+a_{lj}\right)^2}\le\sqrt{\sum_ja_{kj}^2}+\sqrt{\sum_ja_{lj}^2}$ due to triangle inequality. Therefore, even if we disregard possible time point equality, property still holds. However, the greater the value of $\rho(\mu,s,t,t)$ is, the better approximation of ${{\ensuremath{\operatorname{\mathbf{Dist}}}}}(s,{\ensuremath{{\ensuremath{\mathcal{L}}}_{t,t}}}({\varphi}))$ is obtained. Therefore, we will investigate two distinct cases when time points can be equal: 1. It happens consistently for given formula ${\varphi}$ and predicate $\mu$, i.e. ${\varphi}$ is built in such way that the same time value is stored by freeze operator associated with both indices, such as: $$\psi={\@ifstar{{\ensuremath{\operatorname{\textbf{G}}_{}}}}{{{\ensuremath{\operatorname{\textbf{G}}_{[}}}]}}}{I_1}({\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}i\neg {\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}j{\@ifstar{{\ensuremath{\operatorname{\textbf{F}}_{}}}}{{{\ensuremath{\operatorname{\textbf{F}}_{[}}}]}}}{I_2}(x{\ensuremath{\sp{\ast_{i}{}}}}+x{\ensuremath{\sp{\ast_{j}{}}}}\ge x))$$ 2. It is a result of ${\varphi}\equiv{\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}i\left({\varphi}_1{\@ifstar{{\ensuremath{\operatorname{\textbf{U}}_{}}}}{{{\ensuremath{\operatorname{\textbf{U}}_{[}}}]}}}I{\varphi}_2\right)$ (or similar formula) evaluation: $$\begin{split} (s,t,t)\models{\varphi}&\iff(s,t,{\ensuremath{t[i\leftarrowt]}})\models{\varphi}_1{\@ifstar{{\ensuremath{\operatorname{\textbf{U}}_{}}}}{{{\ensuremath{\operatorname{\textbf{U}}_{[}}}]}}}{a,b}{\varphi}_2\iff\\ &\exists t'\in[a+t,b+t]:(s,t',{\ensuremath{t[i\leftarrowt]}})\models{\varphi}_2\land\forall t''\in[t,t']:(s,t'',{\ensuremath{t[i\leftarrowt]}})\models{\varphi}_1 \end{split}$$ When $a=0$, it may occur that $t'=t$. Additionally, $t''\in[t,t']$, therefore, satisfaction of ${\varphi}_1$ by $(s,t,{\ensuremath{t[i\leftarrowt]}})$ has to be evaluated. The equality of $t$ and $i$-th frozen time may be propagated to predicates. We have decided to omit this case in order to simplify robustness computation. Improving Approximation {#sec:merge_indices} ----------------------- The formula $\psi$ (see above) is obviously badly written, since it can be reformulated with only one frozen time index: ${\@ifstar{{\ensuremath{\operatorname{\textbf{G}}_{}}}}{{{\ensuremath{\operatorname{\textbf{G}}_{[}}}]}}}{I_1}(\neg{\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}{\@ifstar{{\ensuremath{\operatorname{\textbf{F}}_{}}}}{{{\ensuremath{\operatorname{\textbf{F}}_{[}}}]}}}{I_2}(2x\ge x))$. This eliminates time point equality and thus improves robustness approximation. We have formulated three rules which can be used to automatically rewrite formula so that it does not induce consistent time point equality (while preserving its meaning): 1. Freeze operator is distributive over Boolean connectives. Consequently, freeze operators can be moved down along the formula syntax tree until they reach a temporal operator, predicate or another freeze operator. 2. Freeze operator preceding predicate can be merged with the predicate (associated coefficients being merged with coefficients for unfrozen time). 3. \[rule:freeze-merge\] Two consecutive freeze operators and their associated indices can be merged. However, in order to preserve the formula meaning, a completely new index has to be chosen as the result of merging. Subsequently, all [STL\]{} formulae can be written in such manner that each freeze operator is followed by until operator, which also ensures that all frozen time indices generally refer to distinct time points. Indeed, all meaningful formulae (i.e. not serving to illustrate semantic peculiarities) in [@Dluhos-stlstar] are specified in this manner. This reinforces the connection between temporal operators and freeze operators expressiveness. Subsequently, it may be practical to define an alternate [STL\]{} syntax, where signal value freezing is directly tied to the until operator, such as ${\varphi}_1{\@ifstar{{\ensuremath{\operatorname{\textbf{U}}_{}}}}{{{\ensuremath{\operatorname{\textbf{U}}_{[}}}]}}}I{\ensuremath{\sp{\ast_{i}{}}}}{\varphi}_2\equiv{\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}i\left({\varphi}_1{\@ifstar{{\ensuremath{\operatorname{\textbf{U}}_{}}}}{{{\ensuremath{\operatorname{\textbf{U}}_{[}}}]}}}I{\varphi}_2\right)$. However, we do not deem it necessary, seeing that it entails no expressiveness gain. Moreover, the current syntax of [STL\]{} may permit shorter and more transparent formulae. It should be noted that although application of previous rules may increase number of indices used in a formula (due to the rule which introduces one new index), it does not increase the number of free indices in each subformula. On the contrary, the number of free indices may decrease. Computation {#sec:computation} =========== To compute (or monitor) robustness of continuous signal, we use the approach of [Fainekos et al.]{} [@FP-robustness], which is based on discrete robustness semantics. The following procedure is used: 1. Sample input signal $s:T\to{\mathbb{R}}^m$ into a timed state sequence $(\tau,\sigma):{\mathbb{N}}\to T\times{\mathbb{R}}^m$. 2. Compute robustness over points of the resulting timed state sequence (i.e. the discrete robustness). This only approximates continuous robustness of $s$. When MTL robustness is concerned, [Fainekos et al.]{} give bound for error introduced by this approximation under certain conditions, which can be summarized as signal sampling being sufficiently dense with respect to given formula. We assume this strong theorem translates to [STL\]{} (as [STL\]{} robustness extends MTL robustness) and deem the previous procedure good approximation for an input signal with large enough sampling rate. Before the robustness monitoring algorithm is described, we should note that it can also be used to decide formula satisfaction, since positive robustness implies formula satisfaction (and negative its invalidity). However, when $\rho({\varphi},s)=0$ no information about formula satisfaction can be derived. Additionally, robustness measure only underapproximates the robust neighbourhood, and so the robustness value may be zero even if clearly $s$ satisfies ${\varphi}$. Consequently, classical monitoring may produce more precise results. Algorithm \[alg:monitor\] computes robustness for a [STL\]{} formula and sufficiently long timed state sequence (which may constitute a sampled signal). It copies inductive definition of robustness with recursive calls of procedure [<span style="font-variant:small-caps;">Monitor</span>]{} (line \[alg:mon:monitor\]), which computes robustness only in the points of given state sequence. Therefore, instead of frozen time vector $t:\left({\mathbb{R}}^+_0\right)^{\mathcal{I}}$, frozen state vector $\iota:{\mathbb{N}}^{\mathcal{I}}$ is used. The computation starts at zero index and zero frozen state vector (line \[alg:mon:start\]), which ensures only robustness values needed for resulting robustness evaluation are computed. [STL\]{} formula ${\varphi}$ and timed state sequence $(\tau,\sigma)$ of length greater than $l({\varphi})$ (see Definition \[def:necessary\_length\]). The value of $\rho({\varphi},(\tau,\sigma))$. For any $i$ free in ${\varphi}$, ${\varphi}\gets{\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}i{\varphi}$. $P\gets\emptyset$ \[alg:mon:start\] \[alg:mon:monitor\] $+\infty$ $\rho(\mu,(\tau,\sigma),\iota,\iota)$ $-\mbox{\Call{Monitor}{${\varphi}_1,\iota,\iota$}}$ $\max\left(\mbox{\Call{Monitor}{${\varphi}_1,\iota,\iota$}},\mbox{\Call{Monitor}{${\varphi}_2,\iota,\iota$}}\right)$ \[alg:mon:until-beg\] $P({\varphi},\iota)(\iota)$ $\varrho\gets\mbox{\Call{PrecomputeUntil}{${\varphi}_1,{\varphi}_2,a,b,\iota$}}$ $P\gets P\cup(({\varphi},\iota),\varrho)$ $\varrho_\iota$ \[alg:mon:until-end\] \[alg:mon:until\] $i\gets 0$ $l\gets \max(l({\varphi}_1),l({\varphi}_2))$ $\varrho\gets\emptyset$ $j\gets 0$\[alg:mon:out-beg\] $r_1\gets \mbox{\Call{Monitor}{${\varphi}_1,i,\iota$}}$ $r_1\gets\min(r_1,\mbox{\Call{Monitor}{${\varphi}_1,i+j,\iota$}})$ $j\gets j+1$ $r\gets r_1$ \[alg:mon:in-beg\] $r_1\gets\min(r_1,\mbox{\Call{Monitor}{${\varphi}_1,i+j,\iota$}})$ $r_2\gets\mbox{\Call{Monitor}{${\varphi}_2,i+j,\iota$}}$ $r\gets\max(r,\min(r_1,r_2))$ $j\gets j+1$ \[alg:mon:in-end\] $\varrho\gets\varrho\cup\{(i,r)\}$ $i\gets i+1$ \[alg:mon:out-end\] Robustness values with respect to subformulae of input formula are not stored. Instead, they are computed every time procedure [<span style="font-variant:small-caps;">Monitor</span>]{} is called on a given subformula. The reasoning behind this practise is the following: For the majority of formulae, the value of robustness for given $\iota$ and $\iota$ is obtained by a simple – constant-time – operation on just a single value of robustness (or two in the case of $\lor$). Additionally, the robustness with respect to predicates can be computed in constant time. The only operator where robustness depends on robustness values over an interval is the until operator (and by extension all derived temporal operators). Consequently, robustness values associated with until operators are stored. Furthermore, when [<span style="font-variant:small-caps;">Monitor</span>]{}(${\varphi}_1{\@ifstar{{\ensuremath{\operatorname{\textbf{U}}_{}}}}{{{\ensuremath{\operatorname{\textbf{U}}_{[}}}]}}}I{\varphi}_2,\iota,\iota$) is called for the first time, robustness values with respect to ${\varphi}_1{\@ifstar{{\ensuremath{\operatorname{\textbf{U}}_{}}}}{{{\ensuremath{\operatorname{\textbf{U}}_{[}}}]}}}I{\varphi}_2$ for $\iota$ and all $\iota'$ are precomputed (see lines \[alg:mon:until-beg\]–\[alg:mon:until-end\]) by the procedure [<span style="font-variant:small-caps;">PrecomputeUntil</span>]{}, which constitutes an algorithmic version of robustness definition for until operator. These precomputed values are expected to be referred to later, since robustness computation is restricted to time interval $[0,l({\varphi})]$ which comprises all input values necessary to evaluate $\rho({\varphi},(\tau,\sigma))$. Complexity ---------- Apparently, the most time-consuming task of Algorithm \[alg:monitor\] is the [<span style="font-variant:small-caps;">PrecomputeUntil</span>]{} procedure, which is quadratic to the number of states in the input timed state sequence. In the worst case it is called for each $\iota$. Therefore, the complexity of Algorithm \[alg:monitor\] is in $\mathcal{O}\left(\|{\varphi}\|\cdot n^{2\|{\mathcal{I}}\|}\right)$ where $n$ is the size of input timed state sequence. For sampled signals, it may be expressed using necessary length, resulting in alternate complexity formulation: $\mathcal{O}\left(\|{\varphi}\|\cdot l({\varphi})^{2\|{\mathcal{I}}\|}\cdot f^{2\|{\mathcal{I}}\|}\right)$ where $f$ is the sampling rate of input signal, which correlates with the precision of robustness computation. Space complexity can be bounded by the same function. The parameter most adversely affecting the algorithm complexity is the size of frozen time index set $\|{\mathcal{I}}\|$. Naturally, ${\mathcal{I}}$ can be restricted to indices used in input formula. In most practical cases, their number will be small. This is supported by the following result: Any formula ${\varphi}$ can be rewritten into a semantically equivalent formula which uses only so many indices as is the maximum number of free indices in subformulae of ${\varphi}$. Note that the number of free indices may increase as we descend into subformulae. This statement derives from the fact that an index only serves to associate one freeze operator with a set of coefficients in one or more predicates and it is free on all paths between this freeze operator and all associated predicates. Therefore, indices which are never simultaneously free need not be different. The result of this theorem can be realized by an automatic procedure which renames frozen time indices in a formula while traversing its syntax tree (using DFS). This procedure stores pairs of indices $[k/l]$ corresponding to the renaming of source index $k$ in the original formula ${\varphi}$ to destination index $l$ in its optimized version ${\varphi}'$. When the procedure encounters freeze operator ${\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}i$, new pair $[i/m]$ is introduced where $m$ is the smallest unused destination index and the operator is changed to ${\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}m$. Whenever $k$ becomes free in ${\varphi}$, the pair $[k/l]$ is removed and $l$ can be reused. Upon reaching a predicate, all stored pairs are applied as a renaming. This procedure is described in greater detail in [@src] (page 44) where additional justification of its correctness can also be found. Together with freeze operator merging described in Section \[sec:merge\_indices\] (which does not increase number of free indices), this can considerably decrease the number of indices used in a formula and thus the time complexity of robustness monitoring. Although intelligent formula specification may result in already optimal formula, the existence of automatic optimization procedures reduces demands on writers of formulae. Implementation -------------- The algorithm has been implemented as an extension of the tool Parasim [@parasim]. Parasim is a highly modular Java-based open-source tool with graphical user interface for computing robustness of a model with respect to perturbations. Integrating the algorithm presented in this paper into an already existing tool has an additional advantage of facilitating the use of [STL\]{} robustness in practise. Given a model, [STL\]{} formula and perturbation set, Parasim samples the perturbation set into points and for each point simulates the model and computes robustness of the resulting signal with respect to [STL\]{} robustness measure. In the neighbourhood of signals with low robustness, additional points are sampled. Formula optimizing algorithms are implemented to maximize efficiency. Case Study {#sec:casestudy} ========== By employing the Parasim tool we have conducted several experiments on two simple population dynamics models. The experiments have also served us to briefly evaluate the algorithm performance (in the setting of the Parasim tool). SIR Model {#sec:6:SIR} --------- First, we demonstrate the robustness analysis on the model simulating an outbreak of an infectious disease in a population [@SIR]. The simulated population is divided into three categories: susceptible ($S$), infected ($I$) and recovered ($R$). A susceptible individual can become infected by contact with another infected individual and an infected individual may recover. The ODE model is the following: $$\begin{aligned} \frac{dS}{dt}&=-{\alpha}SI & \frac{dI}{dt}&={\alpha}SI-\beta I& \frac{dR}{dt}&=\beta I\end{aligned}$$ Where ${\alpha}$ is the contact rate which correlates to probability of disease transmission, while $\beta$, the recovery rate, takes into account the standard length of recovery. A typical simulation of this model (see \[fig:sir-course\]a) includes a rapid increase in infected individuals, which is then followed by their gradual recovery. In this case study, we compare robustness analysis based on a formula containing value-freezing with respect to a freezing-free formula analysis exploiting a similar behavioural pattern. In particular, we consider the following formulae: $$\begin{aligned} \text{STL}:{\varphi}_1&={\@ifstar{{\ensuremath{\operatorname{\textbf{F}}_{}}}}{{{\ensuremath{\operatorname{\textbf{F}}_{[}}}]}}}{1,5}(I\ge 50) & \text{STL}:{\varphi}_2&={\@ifstar{{\ensuremath{\operatorname{\textbf{F}}_{}}}}{{{\ensuremath{\operatorname{\textbf{F}}_{[}}}]}}}{1,5}\left(I\ge 50\land{\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}{\@ifstar{{\ensuremath{\operatorname{\textbf{G}}_{}}}}{{{\ensuremath{\operatorname{\textbf{G}}_{[}}}]}}}{0.25,5}(I\ge I)\right)\end{aligned}$$ Both formulae require the number of infected individuals to be greater than $50$ at some time in the interval $[1,5]$, while ${\varphi}_2$ also requires this number to be the local maximum (the number of infected individuals is required to decrease after reaching this maximum). The robustness with respect to both properties was analysed on perturbations of both contact rate and recovery rate. Results are presented in \[fig:sir-out\]. While the satisfaction sets of ${\varphi}_1$ and ${\varphi}_2$ (delineated by positive robustness) are essentially identical, the actual robustness values show a significant difference. Generally, when they are positive, the value of robustness with respect to ${\varphi}_1$ at given point is considerably greater than the corresponding value of robustness with respect to ${\varphi}_2$. In \[fig:sir-out\], this can be seen as lighter shade of green points in \[fig:sir-out:star\]. Also, lower robustness causes the apparent increase in the number of points. The reason for the rapid change in robustness comes from evaluation of the subformula ${\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}{\@ifstar{{\ensuremath{\operatorname{\textbf{G}}_{}}}}{{{\ensuremath{\operatorname{\textbf{G}}_{[}}}]}}}{0.25,5}(I\ge I)$ that describes the local extreme. When evaluated in time $t$, robustness is proportional to the difference $(I[t]-I[t+0.25])$ (by Definition \[def:predicate\_robustness\]). In practise, the difference is small provided that the descent of $I$ is not extremely steep. This causes such formulae to have typically low robustness values on common signals. Predator-Prey Model {#sec:6:lv} ------------------- In the second case study we analyse the predator-prey model [@lotka; @volterra], which attains oscillating behaviour for a wide variety of parameters. We use a variant of the Lotka-Volterra model represented by the following ordinary differential equations: $$\begin{aligned} \frac{dX}{dt}&=\nu X-{\alpha}XY & \frac{dY}{dt}&={\alpha}XY-\mu Y\end{aligned}$$ The model simulates a situation where a prey species $X$ is hunted by a predator species $Y$ with the simplifying assumption that predator birth rate and prey death rate are equal and proportional to the probability of prey and predator contact, and thus to the product of both species populations. We use the following coefficients: prey natality ($\nu$), predator mortality ($\mu$) and predation rate (${\alpha}$). Typical behaviour of this models constitutes periodic oscillations (see \[fig:sir-course\]b). We consider perturbation of two aforementioned coefficients, $\nu$ and ${\alpha}$, and compute robustness with respect to two properties specified by the following formulae: [$$\begin{aligned} \psi_1 &={\@ifstar{{\ensuremath{\operatorname{\textbf{G}}_{}}}}{{{\ensuremath{\operatorname{\textbf{G}}_{[}}}]}}}{0,300}{\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}{\@ifstar{{\ensuremath{\operatorname{\textbf{F}}_{}}}}{{{\ensuremath{\operatorname{\textbf{F}}_{[}}}]}}}{0,100}\left(X\ge Y\right)\\ \psi_2 &={{\@ifstar{{\ensuremath{\operatorname{\textbf{G}}_{}}}}{{{\ensuremath{\operatorname{\textbf{G}}_{[}}}]}}}{0,300}\left(X\ge 1\land Y\ge 1\land{\@ifstar{{\ensuremath{\operatorname{\textbf{F}}_{}}}}{{{\ensuremath{\operatorname{\textbf{F}}_{[}}}]}}}{0,50}{\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}\left({\@ifstar{{\ensuremath{\operatorname{\textbf{F}}_{}}}}{{{\ensuremath{\operatorname{\textbf{F}}_{[}}}]}}}{0,75}\left(X-X\ge 25\right)\land{\@ifstar{{\ensuremath{\operatorname{\textbf{F}}_{}}}}{{{\ensuremath{\operatorname{\textbf{F}}_{[}}}]}}}{0,75}\left(X-X\ge 25\right)\right)\right)}\end{aligned}$$]{} The property $\psi_1$ requires that for each time point $t\in[0,300]$, there is a subsequent time point $t'\in[t,t+100]$ such that population of prey in $t'$ is greater than population of predators in $t$. According to Definition \[def:predicate\_robustness\] its corresponding robustness can be expressed as follows: $$\rho({\varphi},s)=\min_{t\in[0,300]}\max_{t'\in[t,t+100]}\frac{X[t']-Y[t]}{2}$$ where $X[t']$ and $Y[t]$ denote values of $s$ associated with given species at given time. The robustness value is maximized with respect to $t'$ and minimized with respect to $t$, therefore, it uses maximal values of both $X$ and $Y$. Consequently, this property can be interpreted as maximum population of prey being greater then maximum population of predators (restricted to given intervals). Formula $\psi_2$ is based on the similar principle. While rejecting aberrant behaviour where population of one of the species drops below one individual, intuitively, it requires that there always is time in the future when population of prey can increase or decrease by 25 individuals, which is stated by the subformula ${\@ifstar{{\ensuremath{\operatorname{\textbf{F}}_{}}}}{{{\ensuremath{\operatorname{\textbf{F}}_{[}}}]}}}{0,50}{\@ifstar{\ensuremath{\operatorname{\ast}}}{\ensuremath{\operatorname{\ast}_{}}}}\left({\@ifstar{{\ensuremath{\operatorname{\textbf{F}}_{}}}}{{{\ensuremath{\operatorname{\textbf{F}}_{[}}}]}}}{0,75}\left(X-X\ge 25\right)\land{\@ifstar{{\ensuremath{\operatorname{\textbf{F}}_{}}}}{{{\ensuremath{\operatorname{\textbf{F}}_{[}}}]}}}{0,75}\left(X-X\ge 25\right)\right)$. Therefore, $\psi$ is satisfied when the difference between maximal and minimal prey population is greater than 50 and the associated robustness is proportional to this difference. Again, we have avoided use of the extreme property, which would adversely affect robustness value. ![Robustness of predator-prey model with respect to $\psi_1$ (left) and $\psi_2$ (right) for variable prey natality and predation rate. Robustness was positive in green points and negative in orange points. Darker colour represents greater absolute value of robustness.[]{data-label="fig:lv-out"}](maxrelation.pdf "fig:") ![Robustness of predator-prey model with respect to $\psi_1$ (left) and $\psi_2$ (right) for variable prey natality and predation rate. Robustness was positive in green points and negative in orange points. Darker colour represents greater absolute value of robustness.[]{data-label="fig:lv-out"}](lotkavG.pdf "fig:") Results of this analysis are presented in \[fig:lv-out\]. Here, we should point out that small prey natality produced behaviour where predator population approached zero and period of oscillations was greatly increased. For such behaviour, intervals used in $\psi_1$ and $\psi_2$ were shorter than one period. Apparently, satisfaction of $\psi_1$ is not affected by predation rate. More interestingly, when prey natality increases, predator population exceeds that of prey (see \[fig:lv-out\] (left)). \[fig:lv-out\] (right) shows that amplitude of prey population oscillation is affected by both prey natality and predation rate. The above results have been confirmed by simulation. Performance ----------- Performance of robustness analysis is summarized in Table \[tab:perf\]. All results have been obtained by executing the algorithm implementation on a $4$ core $2$ GHz CPU with $4$ GB RAM. Each computation has been arranged into $8$ threads. For each analysis we have set an optimal resolution of the trajectories (number of simulated points). The number of simulated trajectories has been bounded by the number of refinement iterations in the Parasim parameter space sampling procedure. It is worth noting that all analysed properties consist only of ${\@ifstar{{\ensuremath{\operatorname{\textbf{F}}_{}}}}{{{\ensuremath{\operatorname{\textbf{F}}_{[}}}]}}}{}$ and ${\@ifstar{{\ensuremath{\operatorname{\textbf{G}}_{}}}}{{{\ensuremath{\operatorname{\textbf{G}}_{[}}}]}}}{}$ operators for which the procedure is optimized by employing Lemire queues in the same way as proposed in [@DFM13]. This is based on an optimal streaming algorithm for computing maxima (resp. minima) of a numerical sequence and allows to reduce the quadratic complexity wrt formula size to linear. Property (model) formula size $\#$ trajectories $\#$ points per a trajectory time -------------------------- -------------- ------------------- ------------------------------ ----------- $\varphi_1$ (SIR) 2 250 500 $8.6$ s $\varphi_2$ (SIR) 6 1365 1000 $15.2$ s $\psi_1$ (Predator-Prey) 4 831 400 $85.4$ s $\psi_2$ (Predator-Prey) 12 1293 423 $309.4$ s : Performance of the robustness computation measured on the prototype implementation.[]{data-label="tab:perf"} The increase in computation time in the case of $\psi_1$ is caused by longer time intervals quantifying the temporal operators. Computation of the property $\psi_2$ has been slowed down due to insufficient memory. Conclusion ========== In this paper we have set up a robustness measure for a value-freezing extension of STL. The robustness of a signal with respect to a given [STL\]{} property is based on the distance of the signal from signals violating the property. We have introduced a measure that is proved to fulfil requirements imposed on robustness measures as defined in [@FP-robustness]. This guarantees that the robustness measure is defined correctly. We have derived the algorithm for [STL\]{} robustness computation from the discrete robustness and implemented it as an extension of the tool Parasim [@parasim]. Some of the properties from case studies required comparison of signal values at near frozen time points. Robustness of such properties is typically small. This is only natural as such properties represent stricter requirements on signals. However, this feature may also constitute a detriment for tools such as Parasim, which use robustness to direct perturbation set sampling. This is the exact case of analysed SIR model and property ${\varphi}_2$. It must be noted, though, that this problem is encompassed by the much broader issue of meaningful property design. In [@FP-robustness] the authors quantify error in robustness value caused by the approximate computation. We have not yet explored this possibility for [STL\]{} robustness measures and leave this for future work. However, results in [@FP-robustness] imply this error is inversely proportional to the rate of input signal sampling. [^1]: The work has been supported by the Grant Agency of Czech Republic grant GAP202/11/0312 and by the EC OP project No. CZ.1.07/2.3.00/20.0256.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The [Discrepancy]{} of a hypergraph is the minimum attainable value, over two-colorings of its vertices, of the maximum absolute imbalance of any hyperedge. The [Hereditary Discrepancy]{} of a hypergraph, defined as the maximum discrepancy of a restriction of the hypergraph to a subset of its vertices, is a measure of its complexity. Lovász, Spencer and Vesztergombi (1986) related the natural extension of this quantity to matrices to rounding algorithms for linear programs, and gave a determinant based lower bound on the hereditary discrepancy. Matoušek (2011) showed that this bound is tight up to a polylogarithmic factor, leaving open the question of actually computing this bound. Recent work by Nikolov, Talwar and Zhang (2013) showed a polynomial time $\tilde{O}(\log^3 n)$-approximation to hereditary discrepancy, as a by-product of their work in differential privacy. In this paper, we give a direct simple $O(\log^{3/2} n)$-approximation algorithm for this problem. We show that up to this approximation factor, the hereditary discrepancy of a matrix $A$ is characterized by the optimal value of simple geometric convex program that seeks to minimize the largest $\ell_{\infty}$ norm of any point in a ellipsoid containing the columns of $A$. This characterization promises to be a useful tool in discrepancy theory.' author: - \| Aleksandar Nikolov\ Rutgers University - \| Kunal Talwar\ Microsoft Research bibliography: - 'discrepancy.bib' title: Approximating Hereditary Discrepancy via Small Width Ellipsoids --- Introduction ============ Discrepancy theory, in the broadest sense, studies the fundamental limits to approximating a “complex measure” (i.e. continuous, or with large support) with a “simple measure” (i.e. counting measure, or a measure with small support) with respect to a class of “distinguishers”. A prototypical “continuous discrepancy” question is how uniform can a set $P$ of $n$ points in the unit square $[0, 1)^2$ be, where uniformity is measured with respect to a class of geometric shapes, e.g. axis-aligned rectangles [@schmidt]. A prototypical “discrete discrepancy” question asks whether we can color the $n$ vertices of a hypergraph of $O(n)$ edges with two colors, red and blue, so that each edge has approximately the same number of red vertices as blue vertices [@spencer-six]. The two kinds of questions are deeply related, and transference theorems between different discrepancy measures are known [@beck-sos]. Questions related to discrepancy theory are raised throughout mathematics, e.g. number theory, Diophantine approximation, numerical integration. Unsurprisingly, they also naturally appear in computer science – questions about the (im)possibility of approximating continuous, average objects with discrete ones are central to pseudorandomness, learning theory, communication complexity, approximation algorithms, among others. For a beautiful survey of applications of discrepancy theory to computer science, we refer the reader to Chazelle’s The Discrepancy Method [@Chazelle]. Despite discrepancy theory’s many applications in computer science, we have only recently began to understand the computational complexity of measures of discrepancy themselves. In this work, we address the problem of approximately computing hereditary discrepancy, one of the fundamental discrepancy measures. Hereditary discrepancy is a robust version of combinatorial discrepancy, which is the hypergraph coloring problem mentioned above. More precisely, the combinatorial discrepancy $\operatorname{disc}(\mathcal{H})$ of a hypergraph $\mathcal{H} = (H_1, \ldots, H_m)$ on the vertices $[n] = \{1, \ldots, n\}$ is the minimum over colorings $\chi:[n] \rightarrow \{-1, 1\}$ of the maximum “imbalance” over hyperedges $\max_{i = 1}^m{\|\sum_{j \in H_i}{\chi(i)}\|}$. While relatively simple, $\operatorname{disc}(\mathcal{H})$ is a brittle quantity, which can make it intractable to estimate.[[We may wish to say that $\operatorname{disc}(\mathcal{H})$ measures the complexity of $\mathcal{H}$, but it can be $0$ for intuitively complex $\mathcal{H}$ for trivial reasons. For example let $(V,E)$ be a complex hypergraph all of whose sets have equal size, such as ${[n] \choose n/2}$. Consider the hypergraph formed by taking two identical copies of $(V, E)$, say $(V_1, E_1)$ and $(V_2, E_2)$ and defining the new hypergraph as $(V_1 \cup V_2, E'\triangleq\{e_1 \cup e_2: e_1 \in E_1, e_2 \in E_2\})$. By coloring $V_1$ as $+1$ and $V_2$ as $-1$, we get discrepancy zero for each edge in $E'$, despite the intuitive complexity of $E$. ]{}]{}For this reason it is often more convenient to work with the more robust hereditary discrepancy. Hereditary discrepancy is the maximum discrepancy over restricted hypergraphs, i.e. $\operatorname{herdisc}(\mathcal{H}) = \max_{W\subseteq [n]} \operatorname{disc}(\mathcal{H}\|_W)$, where $\mathcal{H}\|_W = (H_1 \cap W, \ldots, H_m \cap W)$.[[Notice, for example, that the hereditary discrepancy of the above example is in fact $\Omega(n)$ – a more fitting measure of the complexity of the hypergraph. ]{}]{} Discrepancy and hereditary discrepancy have natural generalizations to matrices. The discrepancy of a matrix $A$ is equal to $\operatorname{disc}(A) = \min_{x\in \{-1, 1\}^n}{\\|Ax\\|_\infty}$, and hereditary discrepancy is equal to $\max_{S \subseteq [n]}{\operatorname{disc}(A\|_S)}$, where $A\|_S$ is the submatrix of $A$ consisting of columns indexed by elements of $S$. These quantities coincide with hypergraph discrepancy when evaluated on the incidence matrix of the hypergraph, and are also natural themselves. For example, a classical result of Lovász, Spencer, and Vesztergombi [@LSV] states that for any matrix $A$, any vector $c \in [-1, 1]^n$ can be rounded to $x \in \{-1,1\}^n$ so that $\\|Ax-Ac\\|_\infty \leq 2\operatorname{herdisc}(A)$. In the context of a linear program, this means that we can round fractional solutions to integral ones while still approximately satisfying the linear constraints defined by $A$. This fact was recently used by Rothvo[ß]{} to design an improved approximation algorithm for bin packing [@rothvoss-binpacking]. The robustness of hereditary discrepancy in comparison with discrepancy is evident in the hardness of approximating each of the two measures. By an important result of Spencer [@spencer-six], whenever the number of edges in $\mathcal{H}$ is $m = O(n)$, the discrepancy is at most $O(\sqrt{n})$. It turns our that it is $\mathsf{NP}$-hard to distinguish between this worst-case upper bound and discrepancy zero [@CNN]. By contrast, recently Nikolov, Talwar, and Zhang gave a polylogarithmic approximation to hereditary discrepancy [@NTZ]. At first glance, this is surprising, because hereditary discrepancy is a maximum over exponentially many $\mathsf{NP}$-hard minimization problems. Thus, hereditary discrepancy is not even obviously in $\mathsf{NP}$ (but is $\mathsf{NP}$-hard to approximate within a factor of 2, see [@AustrinGH13]). However, the structure and robustness of hereditary discrepancy explain its more tractable nature. As one classical illustration to this, we note that the hypergraphs with hereditary discrepancy $1$ are exactly the hypergraphs with totally unimodular incidence matrices [@gh-h-tum], and are recognizable by a polynomial time algorithm [@seymour-tum]. #### Our Results and Techniques. When approximating a function $f$, we need to provide (nearly matching) upper bounds and lower bounds on $f$. When we approximate $\mathsf{NP}$-optimization problems, usually proving either the upper (for maximization problems) or the lower bound (for minimization problems) is relatively straightforward: it is given by a combinatorial lower (or upper) bound or a convex relaxation. The challenge is to design a bound which is nearly tight. On the other hand, in a max-min problem like hereditary discrepancy, both upper and lower bounds are challenging to prove. Nevertheless, a convex relaxation of discrepancy still turns out to be very useful. The relaxation, vector discrepancy, is derived by relaxing the condition on the coloring $\chi:[n]\rightarrow \{-1, 1\}$ to the weaker $\chi:[n] \rightarrow {\mathbb{S}}^{n-1}$, where ${\mathbb{S}}^{n-1}$ is the unit sphere in $\R^n$. Then the vector discrepancy $\operatorname{vecdisc}(\mathcal{H})$ is the minimum over such $\chi$ of $\max_{i = 1}^m{\\|\sum_{j \in H_i}{\chi(i)}\\|_2}$. Similarly, the extension $\operatorname{vecdisc}(A)$ to matrices $A$ is the minimum over $\chi:[n] \rightarrow {\mathbb{S}}^{n-1}$ of $\max_{i = 1}^m{\\|\sum_{j=1}^n{A_{ij}\chi(i)}\\|_2}$. These quantities can be efficiently approximated to within any prescribed accuracy by solving a semidefinite program. The hereditary vector discrepancy $\operatorname{hvecdisc}(A)$ is defined analogously to hereditary discrepancy as the maximum vector discrepancy over submatrices. Clearly, $\operatorname{vecdisc}(A) \leq \operatorname{disc}(A)$ for any matrix $A$. There exist matrices $A$ with $m = O(n)$ for which $\operatorname{vecdisc}(A) = 0$ and $\operatorname{disc}(A) = \Omega(\sqrt{n})$[^1]. Nevertheless, in a recent breakthrough, Bansal showed that an upper bound on hereditary vector discrepancy is useful in efficient discrepancy minimization. \[thm:bansal\] Let $A$ be an $m$ by $n$ matrix with $\operatorname{hvecdisc}(A) \leq \lambda$. Then there exists a randomized polynomial time algorithm that computes $x \in \{-1, 1\}^n$ with discrepancy at most $\\|Ax\\|_\infty \leq O(\log m) \cdot\lambda$. Theorem \[thm:bansal\] implies that the gap between $\operatorname{hvecdisc}$ and $\operatorname{herdisc}$ is at most logarithmic. \[cor:bansal\] For any $m \times n$ matrix $A$ $$\operatorname{hvecdisc}(A) \leq \operatorname{herdisc}(A) \leq O(\log m) \operatorname{hvecdisc}(A).$$ While $\operatorname{vecdisc}(A)$ can be approximated to within any degree of accuracy in polynomial time, it is not clear if $\operatorname{hvecdisc}(A)$ can be computed efficiently: notice that hereditary vector discrepancy is the maximum of the objective functions of an exponential number of convex minimization problems. Nevertheless, by using vector discrepancy we remove one of the two quantifiers over exponentially large sets. In this paper we prove the following approximation result for $\operatorname{hvecdisc}$. \[thm:vecdisc-apx\] There exists a polynomial time algorithm that approximates $\operatorname{hvecdisc}(A)$ within a factor of $O(\log m)$ for any $m\times n$ matrix $A$. Moreover, the algorithm finds a submatrix $A\|_S$ of $A$, such that $\operatorname{hvecdisc}(A) = O(\log m) \operatorname{vecdisc}(A\|_S)$. Theorem \[thm:vecdisc-apx\] follows from a geometric characterization of hereditary vector discrepancy. We show that, up to a factor of $O(\log m)$, $\operatorname{hvecdisc}(A)$ is equal to the smallest value of $\\|E\\|_\infty$ over all ellipsoids that contain the columns of $A$. Here, $\\|E\\|_\infty$ is just the maximum $\ell_\infty^m$ norm of all points in $E$, or, equivalently, the maximum width of $E$ in the directions of the standard basis vectors $e_1, \ldots, e_m$. [A priori]{}, it is not clear how to relate this quantity in either direction to the $\operatorname{hvecdisc}(A)$, as it is not a fractional “relaxation” in the traditional sense. It is in fact non-trivial to prove either of the two inequalities relating the geometric quantity to $\operatorname{hvecdisc}(A)$. Proving that this quantity is an upper bound on hereditary discrepancy relies on a recent result of Nikolov that upper bounds the vector discrepancy of matrices with columns bounded in Euclidean norm by $1$ [@komlos-sdp]. We need a slight generalization of Nikolov’s result that shows that the vector discrepancy of such matrices can be bounded by $1$ in any direction. We then transform linearly the containing ellipsoid $E$ to a unit ball, so that Nikolov’s result applies; because of the transformation, we need to make sure that in the transformed space the vector discrepancy is low in a set of directions different from the standard basis. While, on the face of things, this argument only upper bounds the vector discrepancy of $A$, it in fact also upper bounds the vector discrepancy of any submatrix as well, because if $E$ contains all columns of $A$, it also contains all the columns of any submatrix of $A$. This simple observation is crucial to the success of our arguments. To show that the smallest value of $\\|E\\|_\infty$ over all containing ellipsoids also gives a lower bound on hereditary vector discrepancy, we analyze the convex dual of the problem of finding containing ellipsoids of small width and show that we can transform dual certificates for this problem to dual certificates for vector discrepancy of some submatrix of $A$. The dual of the problem of minimizing $\\|E\\|_\infty$ for a matrix $A$ is a problem of maximizing the nuclear norm (i.e. the sum of singular values) over re-weightings of the columns and rows of $A$. To get dual certificates for vector discrepancy for some submatrix, we need to be able to extract a submatrix with a large least singular value from a matrix of large nuclear norm. We accomplish this using the restricted invertibility principle of Bourgain and Tzafriri [@bour-tza]: a powerful theorem from functional analysis which states, roughly, that any submatrix with many approximately equal singular values contains a large well-conditioned submatrix. Using a constructive proof of the theorem by Spielman and Srivastava [@bt-constructive], we can also find the well-conditioned submatrix in deterministic polynomial time; this gives us a submatrix of $A$ on which hereditary vector discrepancy is approximately maximized. Theorem \[thm:vecdisc-apx\] immediately implies a $O(\log^2 m)$ approximation of $\operatorname{herdisc}$ via Bansal’s theorem. However, we can improve this bound to an $O(\log^{3/2} m)$ approximation. \[thm:herdisc-apx\] There exists a polynomial time algorithm that approximates $\operatorname{herdisc}(A)$ within a factor of $O(\log^{3/2} m)$ for any $m \times n$ matrix $A$. Moreover, the algorithm finds a submatrix $A\|_S$ of $A$, such that $\operatorname{herdisc}(A) \leq O(\log^{3/2} m) \operatorname{vecdisc}(A\|_S)$. To prove Theorem \[thm:herdisc-apx\], we lower bound hereditary vector discrepancy as before, in order to lower bound hereditary discrepancy. However, for the upper bound, rather than upper bounding vector discrepancy in terms of $\\|E\\|_\infty$ for a containing ellipsoid, and then upper bounding discrepancy in terms of vector discrepancy, we directly upper bound discrepancy in terms of $\\|E\\|_\infty$. For this purpose, we use another discrepancy bound – this time a theorem due to Banaszczyk [@bana] that shows that for any convex body $K$ of large Gaussian volume, and a matrix $A$ with columns of at most unit Euclidean norm, there exists a $x \in \{-1, 1\}^n$ such that $Ax \in CK$ for a constant $C$. We use this theorem analogously to the way we used Nikolov’s theorem: we linearly transform $E$ to the unit ball, and we specify a body $K$ such that if some $\pm 1$ combination of the columns of $A$ is in $K$ after the transformation, then in the preimage the combination is in an infinity ball scaled by $O(\sqrt{\log m})$. After a preliminary version of this paper was made available, Matoušek [@matousek-says] has shown that our analysis of both the upper and the lower bound on $\operatorname{herdisc}(A)$ in terms of the minimum of $\\|E\\|_\infty$ over containing ellipsoids $E$ is tight. He also used our characterization of $\operatorname{herdisc}(A)$ in terms of the minimum of $\\|E\\|_\infty$ and our analysis of the dual to give new proofs of classical and new results in discrepancy theory. #### Comparison with Related Works. Lovász, Spencer and Vesztergombi [@LSV] defined a determinant based lower bound on the hereditary discrepancy of a matrix. Matoušek [@Matousek11] showed that this lower bound is tight up to $O(\log^{3/2} m)$. These results did not immediately yield an approximation algorithm for hereditary discrepancy, as the determinant lower bound is a maximum over exponentially many quantities and not known to be efficiently computable. Nikolov, Talwar and Zhang [@NTZ] recently studied hereditary discrepancy as a tool for designing near optimal differentially private mechanisms for linear queries, and as a by-product, derived an $\tilde{O}(\log^3 n)$-approximation algorithm for hereditary discrepancy, where the $\tilde{O}$ notation hides sub-logarithmic factors. Small width containing ellipsoids were implicit in their work. The current paper is the first that explicitly considers this natural geometric object in the context of discrepancy. While the proof of the upper bound on discrepancy in [@NTZ] was via a connection between discrepancy and differential privacy due to Nikolov and Muthukrishnan [@halfspaces], here we give a tight and direct argument using results of Nikolov [@komlos-sdp] and Banaszczyk [@bana] on the Komlós problem. The arguments via differential privacy cannot give the tight relationship between the minimum width of a containing ellipsoid and hereditary discrepancy: they necessarily lose a logarithmic factor, because the relationship between discrepancy and privacy is itself not tight. Moreover, our arguments are simpler and more transparent. The proof of our lower bound is also more natural: we relate the duals of the two convex optimization problems under consideration, i.e. the problem of minimizing vector discrepancy, and the problem of minimizing the width of a containing ellipsoid. Via this approach we arrive at a new discrepancy lower bound, which is at least as strong as the determinant lower bound (up to a logarithmic factor), is tight with respect to hereditary discrepancy up to the same asymptotic factor of $O(\log^{3/2} m)$, and is efficiently computable. We believe our lower bound will have future applications in discrepancy theory. A separate but related line of recent works [@Bansal10; @lovettmeka; @Rothvoss14-giann] gives constructive versions of existence proofs in discrepancy theory. Preliminaries ============= We start by introducing some basic notation. For a $m \times n$ matrix $A$ and a set $S \subseteq [n]$, we denote by $A\|_S$ the submatrix of $A$ consisting of those columns of $A$ indexed by elements of $S$. $\mathcal{P}_k$ is the set of orthogonal projection matrices onto $k$-dimensional subspaces of $\R^m$. We use $\operatorname{range}(A)$ for the range, i.e. the span of the columns, of $A$. By $\sigma_{\min}(A)$ and $\sigma_{\max}(A)$ we denote, respectively, the smallest and largest singular value of $A$. I.e., $\sigma_{\min}(A) = \min_{x: \\|x\\|_2 = 1}{\\|Ax\\|_2}$ and $\sigma_{\max}(A) = \max_{x: \\|x\\|_2 = 1}{\\|Ax\\|_2}$. In general, we use $\sigma_i$ for the $i$-th largest singular value of $A$. By $X \succeq 0$ ($X\succ 0$) we denote that $X$ is a positive semidefinite (resp. positive definite) matrix, and by $X \preceq Y$ that $Y - X \succeq 0$. [[Recall that for a block matrix $$X = \left(\begin{array}{cc}A &B\\B^T &C\end{array}\right),$$ the Schur complement of an invertible block $C$ in $X$ is $A - B^TC^{-1}B$. When $C \succ 0$, $X \succeq 0$ if and only if $A - B^TC^{-1}B \succeq 0$. ]{}]{} For a positive semidefinite (PSD) matrix $X \succeq 0$, we denote by $X^{1/2}$ the principal square root of $X$, i.e. the matrix $Y \succeq 0$ such that $Y^2 = X$. Matrix Norms and Restricted Invertibility ----------------------------------------- The Schatten 1-norm of a matrix $A$, also known as the trace norm or the nuclear norm, is equal to $\\|A\\|_{S_1} = \sum_{i}{\sigma_i(A)} = \tr((AA^T)^{1/2})$. For a matrix $A$, we denote by $\\|A\\|_2=\sigma_{\max}(A)$ the spectral norm of $A$ and $\\|A\\|_{HS} = \sqrt{\sum_i \sigma_i^2(A)} = \sqrt{\sum_{i,j} a_{i,j}^2}$ the Hilbert-Schmidt (or Frobenius) norm of $A$. We use $\\|A\\|_{1 \rightarrow 2}$ for the maximum Euclidean length of the columns of the matrix $A = (a_i)_{i = 1}^n$, i.e. $\\|A\\|_{1 \rightarrow 2} = \max_{x: \\|x\\|_1 = 1}{\\|Ax\\|_2} = \max_{i \in [n]}{\\|A_i\\|_2}$. A matrix $A$ trivially contains an invertible submatrix of $k$ columns as long as $k \leq \rank(A)$. An important result of Bourgain and Tzafriri [@bour-tza] (later strengthened by Vershynin [@vershynin], and Spielman and Srivastava [@bt-constructive]) shows that when $k$ is strictly less than the robust rank $\\|A\\|_{HS}^2/\\|A\\|_2^2$ of $A$, we can find $k$ columns of $A$ that form a well-invertible submatrix. This result is usually called the restricted invertibility principle. Next we state a weighted version of it, which can be proved by slightly modifying the proof of Spielman and Srivastava [@bt-constructive]. Rather than describe the modification, we give a reduction of the weighted version to the standard statement in Appendix \[app:weighted-rip\]. \[thm:bt\] Let $\epsilon > 0$, let $A$ be an $m$ by $n$ real matrix, and let $Q$ be a diagonal matrix such that $Q \succeq 0$ and $\tr(Q) = 1$. For any integer $k$ such that $k \leq \epsilon^2 \frac{\\|AQ^{1/2}\\|_{HS}^2}{\\|AQ^{1/2}\\|_2^2}$ there exists a subset $S \subseteq [n]$ of size $S = k$ such that $\sigma_{\min}(A\|_S)^2 \geq (1-\epsilon)^2\\|AQ^{1/2}\\|_{HS}^2$. Moreover, $S$ can be computed in deterministic polynomial time. Geometry -------- Let $\operatorname{conv}\{a_1, \ldots a_n\}$ be the convex hull of the vectors $ a_1, \ldots, a_n$. A convex body is a convex compact subset of $\R^m$. For a convex body $K \subseteq \R^m$, the polar body $K^\circ$ is defined by $K^\circ = \{y: \langle y, x \rangle \leq 1~\forall x \in K\}$. A basic fact about polar bodies is that for any two convex bodies $K$ and $L$, $K \subseteq L \Leftrightarrow L^\circ \subseteq K^\circ$. Moreover, a symmetric convex body $K$ and its polar body are dual to each other, in the sense that $(K^\circ)^\circ = K$. A convex body $K$ is (centrally) symmetric if $-K = K$. The Minkowski norm $\\|x\\|_K$ induced by a symmetric convex body $K$ is defined as $\\|x\\|_K \triangleq \min\{r \in \R: x \in rK\}$. The Minkowski norm induced by the polar body $K^\circ$ of $K$ is the dual norm of $\\|x\\|_K$ and also has the form $\\|y\\|_{K^\circ} = \max_{x \in K}{\langle x, y\rangle}$. It follows that we can also write $\\|x\\|_K$ as $\\|x\\|_K = \max_{y \in K^\circ}{\langle x, y \rangle}$. For a vector $y$ of unit Euclidean length, $\\|y\\|_{K^\circ}$ is the width of $K$ in the direction of $y$, i.e. half the Euclidean distance between the two supporting hyperplanes of $K$ orthogonal to $y$. For symmetric body $K$, we denote by $\\|K\\| = \max_{x \in K}{\\|x\\|}$ the diameter of $K$ under the norm $\\|\cdot\\|$. Of special interest are the $\ell_p^m$ norms, defined for any $p \geq 1$ and any $x \in \R^m$ by $\\|x\\|_p = \left(\sum_{i = 1}^m{\|x\|^p}\right)^{1/p}$. The $\ell_\infty^m$ norm is defined for as $\\|x\\|_\infty = \max_{i = 1}^m{\|x_i\|}$. The norms $\ell_p^m$ and $\ell_q^m$ are dual if and only if $\frac{1}{p} + \frac{1}{q} = 1$, and $\ell_1^m$ is dual to $\ell_\infty^m$. We denote the unit ball of the $\ell_p^m$ norm by $B_p^m = \{x: \\|x\\|_p \leq 1\}$. As with the unit ball of any norm, $B_p^m$ is convex and centrally symmetric for $p \in [1, \infty]$. An ellipsoid in $\R^m$ is the image of the ball $B_2^m$ under an affine map. All ellipsoids we consider are symmetric, and therefore, are equal to an image $F B_2^m$ of the ball $B_2^m$ under a linear map $F$. A full dimensional ellipsoid $E = FB_2^d$ can be equivalently defined as $E = \{x: x^T(FF^T)^{-1}x \leq 1\}$. The polar body of a symmetric ellipsoid $E = F B_2^d$ is the ellipsoid $E^\circ = \{x: x^TFF^Tx \leq 1\}$. It follows that for $E = FB_2^m$ and for any $x$, $\\|x\\|_E = \sqrt{x^T(FF^T)^{-1}x}$ and for any $y$, $\\|y\\|_{E^\circ} = \sqrt{y^T(FF^T)y}$. Convex Duality -------------- Assume we are given the following optimization problem: $$\begin{aligned} &\text{Minimize } f_0(x)\label{eq:general-obj}\\ &\text{s.t.}\notag\\ &\forall 1\leq i \leq m: f_i(x) \leq 0.\label{eq:general-constr}\end{aligned}$$ The Lagrange dual function associated with – is defined as $g(y) = \inf_x f_0(x) + \sum_{i = 1}^m{y_if_i(x)}$, where the infimum is over the intersection of the domains of $f_1,\ldots,\ldots f_m$, and $y \in \R^m$, $y \geq 0$. Since $g(y)$ is the infimum of affine functions, it is a concave function. Moreover, $g$ is upper semi-continuous, and therefore continuous over the convex set $\{y: g(y) > -\infty\}$. For any $x$ which is feasible for –, and any $y \geq 0$, $g(y) \leq f_0(x)$. This fact is known as weak duality. The Lagrange dual problem is defined as $$\begin{aligned} &\text{Maximize } g(y) \text{ s.t. } y \geq 0.\label{eq:L-dual}\end{aligned}$$ Strong duality holds when the optimal value of equals the optimal value of –. Slater’s condition is a commonly used sufficient condition for strong duality. We state it next. \[thm:slater\] Assume $f_0, \ldots, f_m$ in the problem – are convex functions over their respective domains, and for some $k \geq 0$, $f_1, \ldots, f_k$ are affine functions. Let there be a point $x$ in the relative interior of the domains of $f_0, \ldots, f_m$, so that $f_i(x) \leq 0$ for $1 \leq i \leq k$ and $f_j(x) < 0$ for $k+1 \leq j \leq m$. Then the minimum of – equals the maximum of , and the maximum of is achieved if it is finite. For more information on convex programming and duality, we refer the reader to the book by Boyd and Vandenberghe. Hereditary Discrepancy and Relaxations -------------------------------------- Hereditary discrepancy and (hereditary) vector discrepancy are defined as in the Introduction. We note that $\operatorname{vecdisc}(A)$ can also be equivalently characterized as the minimum of $\max_{i = 1}^m{\sqrt{e_i^T(AXA^T)e_i}}$ over matrices $X \succeq 0$, where $e_i$ is the $i$-th standard basis vector. For a $m \times n$ matrix $A$, discrepancy and hereditary discrepancy are defined as $$\begin{aligned} &\operatorname{disc}(A) \triangleq \min_{x \in \{\pm 1\}^n}{\\|Ax\\|_\infty} &\operatorname{herdisc}(A) \triangleq \max_{S \subseteq [n]}{\operatorname{disc}(A\|_S)}.\end{aligned}$$ Vector discrepancy is a convex relaxation of discrepancy in which one can assign arbitrary unit vectors rather than $\pm 1$ to the columns of $A$. Formally, let ${\mathbb{S}}^{n-1}$ be the unit sphere in $\R^n$, and define $$\operatorname{vecdisc}(A) \triangleq \min_{u_1, \ldots, u_n \in {\mathbb{S}}^{n-1}} \max_{i = 1}^m \left\\| \sum_{j = 1}^n{A_{ij} u_j} \right\\|_2$$ From a computational complexity perspective, the important property of vector discrepancy $\operatorname{vecdisc}(A)$ is that it can be approximated to within an arbitrarily small additive constant $\epsilon$ in time polynomial in $\log \frac{1}{\epsilon}$ by (approximately) solving a semidefinite program. The following lower bound on vector discrepancy follows by weak convex duality for semidefinite programming (see [@Matousek11]). \[lm:speclb\] For any $m \times n$ matrix $A$, and any $m\times m$ diagonal matrix $P \geq 0$ with $\tr(P) = 1$, we have $\operatorname{vecdisc}(A) \geq \sqrt{n} \sigma_{\min}(P^{1/2}A)$. We define a spectral lower bound based on Lemma \[lm:speclb\]. $$\operatorname{specLB}(A) \triangleq \max_{k = 1}^n \max_{S \subseteq [n]: \|S\| = k} \max_P \sqrt{k} \sigma_{\min}(P^{1/2}A\|_S),$$ where $P$ ranges over positive (i.e. $P\succeq 0$) $m \times m$ diagonal matrices satisfying $\tr(P) = 1$. Lemma \[lm:speclb\] implies immediately that $\operatorname{hvecdisc}(A) \geq \operatorname{specLB}(A)$. Notice that it is not clear whether $\operatorname{specLB}(A)$ can be computed efficiently. One of our main contributions is to develop a lower bound on hereditary (vector) discrepancy which is tractable and can be related to $\operatorname{specLB}(A)$. Vector Balancing and Banaszczyk’s Theorem ----------------------------------------- A well-known conjecture by Komlós states that $\operatorname{disc}(A) \leq C \\|A\\|_{1\rightarrow 2}$, for an absolute constant $C$. The conjecture remains open, and the best known result towards resolving it is a discrepancy bound of $O(\sqrt{\log n}\ \\|A\\|_{1\rightarrow 2})$ due to Banaszczyk [@bana]. Banaszczyk’s result in fact concerns the more general vector balancing problem of determining sufficient conditions under which there exists an assignment of signs $x \in \{\pm 1\}^n$ such that $Ax \in K$ for given $m \times n$ matrix $A$ and convex body $K$. This general version of Banaszczyk’s result is crucial to our argument. \[thm:bana\] There exists a universal constant $C$ such that the following holds. Let $A$ be an $m$ by $n$ real matrix such that $\\|A\\|_{1 \rightarrow 2}$, and let $K$ be a convex body in $\R^m$ such that $\Pr[g \in K] \geq 1/2$ where $g \in \R^m$ is a standard $m$-dimensional Gaussian random vector, and the probability is taken over the choice of $g$. Then there exists $x \in \{-1, 1\}^n$ such that $Ax \in CK$. Another partial result towards the Komlós conjecture is a recent bound by Nikolov [@komlos-sdp] on the vector discrepancy of matrices $A$ satisfying the condition $\\|A\\|_{1 \rightarrow 2} \leq 1$. Here we state a version of the bound that is stronger than the one stated in [@komlos-sdp]. However, a minor variation of the same proof shows this stronger bound; we give the argument in Appendix \[app:vector-komlos\]. \[thm:komlos-sdp\] For any $m \times n$ matrix $A$ satisfying $\\|A\\|_{1 \rightarrow 2} \leq 1$, there exists a $n\times n$ matrix $X \succeq 0$ such that $\forall i \in [n]: X_{jj} = 1$ and $\\|AXA^T\\|_2 \leq 1$. Note that the above theorem implies $\operatorname{vecdisc}(A) \leq 1$ because $\\|AXA^T\\|_2 \leq 1$ implies $e_i^T(AXA^T)e_i \leq 1$ for all standard basis vectors $e_i$. However, the spectral norm bound is formally stronger then the vector discrepancy upper bound, and this will be essential in our proofs. Ellipsoid Upper Bounds on Discrepancy {#sect:ub} ===================================== In this section we show that small-width ellipsoids provide upper bounds on both hereditary vector discrepancy and hereditary discrepancy. Giving such an upper bound is in general challenging because it must hold for all submatrices simultaneously. The proofs use Theorems \[thm:bana\] and \[thm:komlos-sdp\]. We start with the two main technical lemmas. \[lm:ellips-bound-vector\] Let $A = (a_j)_{j=1}^n \in \R^{m \times n}$, and let $F\in \R^{m\times m}$ be a rank $m$ matrix such that $\forall j \in [n]: a_j \in E = FB_2^m$. Then there exists a matrix $X \succeq 0$ such that $\forall j \in [n]: X_{jj} = 1$ and $AXA^T \preceq FF^T$. Observe that, $a_j \in E \Leftrightarrow F^{-1}a_j \in B_2^m$. This implies $\\|F^{-1}A\\|_{1 \rightarrow 2} \leq 1$, and, by Theorem \[thm:komlos-sdp\], there exists an $X$ with $X_{jj}=1$ for all $j$ such that $(F^{-1}A)X(F^{-1}A)^T \preceq I$. Multiplying on the left by $F$ and on the right by $F^T$, we have $AXA^T \preceq FF^T$, and this completes the proof. Lemma \[lm:ellips-bound-vector\] is our main tool for approximating hereditary vector discrepancy. By the relationship between vector discrepancy and discrepancy established by Bansal (Corollary \[cor:bansal\]), this is sufficient for a poly-logarithmic approximation to hereditary discrepancy. However, to get tight upper bounds on discrepancy from small width ellipsoids (and improved approximation ratio), we give a direct argument using Banaszczyk’s theorem. \[lm:ellips-bound-bana\] Let $A = (a_j)_{j=1}^n \in \R^{m \times n}$, and let $F\in \R^{m\times m}$ be a rank $m$ matrix such that $\forall j \in [n]: a_j \in E = FB_2^m$. Then, for any set of vectors $v_1, \ldots, v_k \in \R^m$, there exists $x \in \{\pm 1\}^n$ such that $\forall i \in [k]: \|\langle Ax, v_i\rangle\| \leq C\sqrt{(v_i^TFF^Tv_i)\log k}$ for a universal constant $C$. Let $P = \{y: \|\langle y, v_i\rangle\| \leq \sqrt{v_i^TFF^Tv_i}\ \forall i \in [k]\}$. We need to prove that there exists an $x \in \{-1, 1\}^n$ such that $Ax \in (C\sqrt{\log k})P$ for a suitable constant $C$. Notice that the polar body of $P$ is $$P^\circ = \operatorname{conv}\{\pm (v_i^TFF^Tv_i)^{-1/2}v_i\}_{i = 1}^k.$$ Set $K = F^{-1}P$. To show that there exists an $x$ such that $Ax \in (C\sqrt{\log k})P$, we will show that there exists an $x \in \{-1, 1\}^n$ such that $F^{-1}Ax \in (C\sqrt{\log k})K$. For this, we will use Theorem \[thm:bana\]. As in the proof of Lemma \[lm:ellips-bound-vector\], $\\|F^{-1}A\\|_{1\rightarrow 2} \leq 1$. To use Theorem \[thm:bana\], we also need to argue that for a standard Gaussian $g$, $\Pr[g \in (C\sqrt{\log k})K] \geq \frac{1}{2}$. To this end, we compute the polar body of $K$ as $$\begin{aligned} K^\circ &= \{y: \langle y, F^{-1}x \rangle \leq 1\; \forall x \in P\} = \{y: \langle (F^T)^{-1}y, x \rangle \leq 1\; \forall x \in P\}\\ &= \{F^Tz: \langle z, x \rangle \leq 1\; \forall x \in P\} = F^TP^\circ \end{aligned}$$ By the definition of Minkowski norm, for any $t \in \R$, $y \in tK$, if and only if $$t \geq \\|y\\|_K = \sup_{z\in K^\circ}\langle y, z \rangle = \sup_{z\in P^\circ}\langle y, F^Tz \rangle = \max_{i=1}^k{\frac{1}{\sqrt{v_i^TFF^Tv_i}} \|\langle y, F^Tv_i \rangle\|},$$ where the first equality is by the duality of $\\|\cdot\\|_K$ and $\\|\cdot\\|_{K^\circ}$, and the final equality holds because the linear functional $\langle y, F^Tz \rangle$ is maximized at a vertex of $P^\circ$. We have then that $y \in tK$ if and only if $\forall i \in [k]: \|\langle y, F^Tv_i\rangle\|^2 \leq t^2(v_i^TFF^Tv_i)$. Let $g$ be a standard $m$-dimensional Gaussian vector. Then $\E_g \|\langle g, F^Tv_i\rangle\|^2 = v_iFF^Tv_i$; by standard concentration bounds, $\Pr[\|\langle g, F^Tv_i\rangle\|^2 > t^2 (v_iFF^Tv_i)] < \exp(-t^2/2)$. Setting $t = \sqrt{2\ln 2k}$ and taking a union bound over all $i \in [k]$ gives us that $\Pr[g \not \in \sqrt{2\ln 2k}\ K] < 1/2$. By Theorem \[thm:bana\], this implies that there exists an $x\in \{-1,1\}^n$ such that $F^{-1}Ax \in \sqrt{2\ln 2k}\ K$, and, by multiplying on both sides by $F$, it follows that $Ax \in \sqrt{2\ln 2k}\ P$. Notice that the quantity $\sqrt{v_i^TFF^Tv_i}$ is just the width of $E$ in the direction of $v_i$. The property that all columns of a matrix $A$ are contained in $E$ is hereditary: if it is satisfied for $A$, then it is satisfied for any submatrix of $A$. This elementary fact lends the power of Lemmas \[lm:ellips-bound-vector\] and \[lm:ellips-bound-bana\]: the bound given by ellipsoids is universal in the sense that the discrepancy bound for any direction $v_i$ holds for all submatrices $A\|_S$ of $A$ simultaneously. This fact makes it possible to upper bound hereditary discrepancy in arbitrary norms, and in the sequel we do this for $\ell_\infty^m$, which is the norm of interest for standard definitions of discrepancy. We consider ellipsoids $E$ that contain the columns of $A$ and minimize the quantity $\\|E\\|_\infty$: the largest $\ell_\infty$ norm of the points of $E$. Note that $\\|E\\|_\infty$, for an ellipsoid $E = FB_2^m$, can be written as $$\label{eq:ellips-infty} \\|E\\|_\infty = \max_{x \in E, y: \\|y\\|_1 = 1}{\langle x, y \rangle} = \max_{y:\\|y\\|_1 = 1}{\\|y\\|_{E^\circ}} = \max_{i \in [n]}{\sqrt{e_i^T FF^T e_i}},$$ where the first identity follows since $\ell_1$ is the dual norm to $\ell_\infty$, and the final identity follows from the formula for $\\|\cdot\\|_{E^\circ}$ and the fact that a convex function over the $\ell_1$ ball is always maximized at a vertex, i.e. a standard basis vector. The next theorem gives our main upper bound on hereditary (vector) discrepancy, which is in terms of $\\|E\\|_\infty$. \[thm:main-ub\] Let $A = (a_i)_{i=1}^n \in \R^{m \times n}$, and let $F$ be a rank $m$ matrix such that $\forall i \in [n]: a_i \in E = FB_2^m$. Then $\operatorname{hvecdisc}(A) \leq \\|E\\|_\infty$, and $\operatorname{herdisc}(A) = O(\sqrt{\log m}) \\|E\\|_\infty$. Let $A\|_S$ be an arbitrary submatrix of $A$ ($S \subseteq [n]$). Since all columns of $A$ are contained in $E$, this holds for all columns of $A\|_S$ as well, and by Lemma \[lm:ellips-bound-vector\], we have that there exists $X \succeq 0$ with $X_{jj} = 1$ for all $j \in S$, and $(A\|_S)X(A\|_S)^T \preceq FF^T$. Therefore, for all $i \in [m]$, $e_i^T(A\|_S)X(A\|_S)^Te_i \leq e_iFF^Te_i \leq \\|E\\|_\infty^2$, by . Since $S$ was arbitrary, this implies the bound on $\operatorname{hvecdisc}(A)$. For bounding $\operatorname{herdisc}(A)$, in Lemma \[lm:ellips-bound-bana\], set $k=m$ and $v_i = e_i$ for $i \in [m]$ and $e_i$ the $i$-th standard basis vector. Lower Bounds on Discrepancy =========================== In Section \[sect:ub\] we showed that the hereditary (vector) discrepancy of a matrix $A$ can be upper bounded in terms of the $\\|E \\|_\infty$ for any $E$ containing the columns of $A$. In this section we analyze the properties of the minimal such ellipsoid and show that it provides lower bounds for discrepancy as well. We use convex duality and the restricted invertibility theorem for this purpose. The lower bound we derive is new in discrepancy theory and of independent interest. The Ellipsoid Minimization Problem and Its Dual {#sect:min-ellips} ----------------------------------------------- To formulate the problem of minimizing $\\|E\\|_\infty = \max_{x \in E}{\\|x\\|_\infty}$ as a convex optimization problem we need the following well-known lemma, which shows that the matrix inverse is convex in the PSD sense. \[lm:inverse-convex\] For any two $m\times m$ matrices $X \succ 0$ and $Y \succ 0$, $(\frac{1}{2}X + \frac{1}{2}Y)^{-1} \preceq \frac{1}{2}X^{-1} + \frac{1}{2}Y^{-1}$. Define the matrices $$U = \left( \begin{array}{cc} X^{-1} &I\\ I &X \end{array} \right) \;\;\;\;\; V = \left( \begin{array}{cc} Y^{-1} &I\\ I &Y \end{array} \right).$$ The Schur complement of $X$ in $U$ is $0$, and therefore $U \succeq 0$, and analogously $V \succeq 0$. Therefore $U + V \succeq 0$, and the Schur complement of $X + Y$ in $U+V$ is also positive semidefinite, i.e. $X^{-1} + Y^{-1} - 4(X+Y)^{-1} \succeq 0$. This completes the proof, after re-arranging terms. Consider a matrix $A = (a_j)_{j=1}^n \in \R^{m\times n}$ of rank $m$. Let us formulate $\min \{\\|E\\|_\infty: \forall j \in [n]: a_j \in E\}$ as a convex minimization problem. The problem is defined as follows $$\begin{aligned} &\text{Minimize } t \label{eq:ellips-obj} \text{ s.t. }\\ &X\succ 0\\ &\forall i \in [m]: e_i^TX^{-1}e_i \leq t\label{eq:ellips-width}\\ &\forall j \in [n]: a_j^TXa_j \leq 1.\label{eq:ellips-enclose}\end{aligned}$$ \[lm:ellips-program\] For a rank $m$ matrix $A = (a_j)_{j = 1}^n \in \R^{m\times n}$, the optimal value of the optimization problem – is equal to $\min \{\\|E\\|_\infty^2: \forall j \in [n]: a_j \in E\}$. Moreover, the objective function and constraints – are convex over $t \in \R$ and $X \succ 0$. The equality of optimal values follows from the basic properties of ellipsoids and . The convexity of – is a consequence of the convexity of the matrix inverse, which is well-known, see e.g. [@bhatia-psdbook Chapter 1.5]. We give the full proof in Appendix \[app:dual\]. Let $\lambda$ be the optimal value of – and $\mu = \min\{\\|E\\|_\infty: \forall j \in [n]: a_j \in E\}$. Given a feasible $X$ for –, set $E = X^{-1/2}B_2^m$ (this is well-defined since $X \succ 0$). Then for any $j \in [n]$, $\\|a_j\\|_E = a_j^TXa_j \leq 1$ by , and, therefore, $a_j \in E$. Also, by , $\\|E\\|_\infty^2 = \max_{i = 1}^m e_i^TXe_i \leq t$. This shows that $\mu \leq \lambda$. In the reverse direction, let $E = FB_2^m$ be such that $\forall j\in [n]: a_j \in E$. Then, because $A$ is full rank, $F$ is also full rank and invertible, and we can define $X = (FF^T)^{-1}$ and $t = \\|E\\|_\infty^2$. Analogously to the calculations above, we can show that $X$ and $t$ are feasible, and therefore $\lambda \leq \mu$. The objective function and the constraints are affine, and therefore convex. To show are also convex, let $X_1, t_1$ and $X_2, t_2$ be two feasible solutions. Then, Lemma \[lm:inverse-convex\] implies that for any $i$, $e_i^T(\frac{1}{2}X_1 + \frac{1}{2}X_2)^{-1}e_i \leq \frac{1}{2}X_1^{-1} + \frac{1}{2}X_2^{-1} \leq \frac{1}{2}t_1 + \frac{1}{2}t_2$, and constraints are convex as well. \[thm:nuclear\] Let $A = (a_j)_{j = 1}^n \in \R^{m\times n}$ be a rank $m$ matrix, and let $\mu = \min \{\\|E\\|_\infty: \forall j \in [n]: a_j \in E\}$. Then, $$\begin{aligned} \mu^2 = &\max \\|P^{1/2}AQ^{1/2}\\|_{S_1}^2\label{eq:nuclear-obj2} \text{ s.t.}\\ &\tr(P) = \tr(Q) = 1\\ &P,Q \succeq 0; P,Q \text{ diagonal}. \label{eq:nuclear-pos2} \end{aligned}$$ We prove the theorem by showing that the convex optimization problem – satisfies Slater’s condition (so strong duality holds), and its Lagrange dual is equivalent to –. Strong duality then implies that the optimal value of – is equal to the optimal value of the Lagrange dual, which is equal to the optimal value of –. The Lagrange dual problem of – is to maximize $2\\|P^{1/2}AR^{1/2}\\|_{S_1} - \tr(R)$ over diagonal PSD matrices $P, R$, such that $\tr(P) = 1$. (Note that, as a Lagrange dual, this is a convex optimization problem.) The objective of the Lagrange dual is maximized when $\tr(R) = \\|P^{1/2}AR^{1/2}\\|_{S_1}^2$, and, therefore, the optimal value of the dual is equal to the optimal value of the optimization problem –. The full proof is given in Appendix \[app:dual\]. We shall prove the theorem by showing that the convex optimization problem – satisfies Slater’s condition, and its Lagrange dual is equivalent to –. Let us first verify Slater’s condition. We define the domain for constraints as the open cone $\{X: X \succ 0\}$, which makes the constraint $X \succ 0$ implicit. Let $r = \\|A\\|_{1\rightarrow 2}$, $X = \frac{1}{r}I$, and $t = r+\varepsilon$ for some $\varepsilon > 0$. Then the affine constraints are satisfied exactly, and the constraints are satisfied with slack since $\varepsilon > 0$. Moreover, by Lemma \[lm:ellips-program\], all the constraints and the objective function are convex. Therefore, – satisfies Slater’s condition, and consequently strong duality holds. The Lagrange dual function for – is $$g(p,q) = \inf_{t, X \succ 0}{t + \sum_{i = 1}^m{p_i(e_i^TX^{-1}e_i - t)} + \sum_{j = 1}^n{q_j(a_j^TXa_j - 1)} },$$ with dual variables $p \in \R^m$ and $q \in \R^n$, $p, q \geq 0$. Equivalently, writing $p$ as a diagonal matrix $P \in \R^{m\times m}$, $P \succeq 0$, $q$ as a diagonal matrix $R \in \R^{n\times n}$, $R \succeq 0$, we have $g(P,R) = \inf_{t,X\succ 0}{t + \tr(PX^{-1}) - \tr(tP) + \tr(ARA^TX) - \tr(R)}$. If $\tr(P) \neq 1$, then $g(P,R) = -\infty$, since we can take $t$ to $-\infty$ while keeping $X$ fixed. On the other hand, for $\tr(P) = 1$, the dual function simplifies to $$\label{eq:g-raw} g(P,R) = \inf_{X \succ 0}{\tr(PX^{-1}) + \tr(ARA^TX) - \tr(R)}.$$ Since $X\succ 0$ implies $X^{-1}\succ 0$, $g(P,R) \geq -\tr(R) > -\infty$ whenever $\tr(P) = 1$. Therefore, $g(P,R)$ is continuous over the set of diagonal positive semidefinite $P$, $R$ such that $\tr(P) = 1$. For the rest of the proof we assume that $P$ and $ARA^T$ are rank $m$. This is without loss of generality by the continuity of $g$ and because both assumptions can be satisfied by adding arbitrarily small perturbations to $P$ and $R$. (Here we use the fact that $A$ is rank $m$.) After differentiating the right hand side of with respect to $X$, we get the first-order optimality condition $$\label{eq:fo-optimality} X^{-1}PX^{-1} =ARA^T.$$ Multiplying by $P^{1/2}$ on the left and the right and taking square roots gives the equivalent condition $P^{1/2}X^{-1}P^{1/2} = (P^{1/2}ARA^TP^{1/2})^{1/2}$. This equation has a unique solution, since $P$ and $ARA^T$ were both assumed to be invertible. Since $\tr(PX^{-1}) = \tr(P^{1/2}X^{-1}P^{1/2})$ and also, by , $\tr(ARA^TX) = \tr(X^{-1}P) = \tr(PX^{-1})$, we simplify $g(P,R)$ to $$\label{eq:g-final} g(P,R) = 2\tr((P^{1/2}ARA^TP^{1/2})^{1/2}) - \tr(R) = 2\\|P^{1/2}AR^{1/2}\\|_{S_1} - \tr(R).$$ We showed that – satisfies Slater’s condition and therefore strong duality holds, so by Theorem \[thm:slater\] and Lemma \[lm:ellips-program\], $\mu^2 = \max\{g(P,R): \tr(P) = 1, P,R \succeq 0, \text{ diagonal}\}$. Let us define new variables $Q$ and $c$, where $c = \tr(R)$ and $Q = R/c$. Then we can re-write $g(P,R)$ as $$g(P,R) = g(P,Q,c) = 2\\|P^{1/2}A(cQ)^{1/2}\\|_{S_1} - \tr(cQ) = 2\sqrt{c}\\|P^{1/2}AQ^{1/2}\\|_{S_1} - c.$$ From the first-order optimality condition $\frac{dg}{dc} = 0$, we see that maximum of $g(P,Q,c)$ is achieved when $c = \\|P^{1/2}AQ^{1/2}\\|_{S_1}^2$ and is equal to $\\|P^{1/2}AQ^{1/2}\\|_{S_1}^2$. Therefore, maximizing $g(P,R)$ over diagonal positive semidefinite $P$ and $R$ such that $\tr(P) = 1$ is equivalent to the optimization problem –. This completes the proof. Spectral Lower Bounds via Restricted Invertibility -------------------------------------------------- In this subsection we relate the dual formulations of the min-ellipsoid problem from Section \[sect:min-ellips\] to the dual of vector discrepancy, and $\operatorname{specLB}$ in particular. The connection is via the restricted invertibility principle and gives our main lower bounds on hereditary (vector) discrepancy. \[lm:bt-lb\] Let $A$ be an $m$ by $n$ real matrix, and let $Q \succeq 0$ be a diagonal matrix such that $\tr(Q) = 1$. Then there exists a submatrix $A\|_S$ of $A$ such that $\|S\|\sigma_{\min}(A\|_S)^2 \geq \frac{c^2\\|AQ^{1/2}\\|_{S_1}^2}{(\log m)^2}$, for a universal constant $c > 0$. Moreover, given $A$ as input, $S$ can be computed in deterministic polynomial time. By homogeneity of the nuclear norm and the smallest singular value, it suffices to show that if $\\|AQ^{1/2}\\|_{S_1}^2 = 1$, then $\|S\|\sigma_{\min}(A\|_S)^2 \geq \frac{c^2}{(\log m)^2}$ for a set $S \subseteq [n]$. Let $T_k = \{i \in [m]: 2^{-k-1}\leq \sigma_i(AQ^{1/2}) \leq 2^{-k}\}$ for an integer $0 \leq k\leq \log_2 m$, and $R = \{i \in [m]: \sigma_i(AQ^{1/2}) \leq \frac{1}{2m}\}$. Then $$\sum_{k = 0}^{\log_2 m}\sum_{i \in T_k}{\sigma_i(AQ^{1/2})} = 1 - \sum_{i \in R}{\sigma_i(AQ^{1/2})}\geq 1/2,$$ since $\|R\| \leq m$. Therefore, by averaging, there exists a $k^$ such that $\sum_{i \in T_{k^}}{\sigma_i(AQ^{1/2})} \geq \frac{1}{2\log_2 m}$. Let $\Pi$ be the projection operator onto the span of the left singular vectors of $AQ^{1/2}$ corresponding to the singular values $\sigma_i(AQ^{1/2})$ for $i \in T_{k^}$. Setting $\tau = \frac{1}{2\log_2 m}$ and $r = \|T_{k^}\| = \rank(\Pi AQ^{1/2})$, we have $\\|\Pi AQ^{1/2}\\|_{S_1} \geq \tau$ by the choice of $k^$, and $\\|\Pi AQ^{1/2}\\|_2 \leq 2\tau/r$ because all values of $\Pi AQ^{1/2}$ are within a factor of $2$ from each other. Finally, applying Cauchy-Schwarz to the singular values of $\Pi AQ^{1/2}$, we have that $\\|\Pi AQ^{1/2}\\|_{HS} \geq \tau/r^{1/2}$. By Theorem \[thm:bt\] applied with $\epsilon = \frac{1}{2}$, there exists a set $S$ of size $\|S\| \geq r/16$ such that $\sigma_{\min}(\Pi A\|_S)^2 \geq \tau^2/4r$, implying that $$\|S\|\sigma_{\min}(A\|_S)^2 \geq \|S\|\sigma_{\min}(\Pi A\|_S)^2 \geq \frac{1}{64}\tau^2.$$ Moreover, $S$ can be computed in deterministic polynomial time. \[thm:main-lb\] Let $\mu= \min \{\\|E\\|_\infty: \forall j \in [n]: a_j \in E\}$ for a rank $m$ matrix $A = (a_j)_{j = 1}^n$. Then $$\mu = O(\log m)\ \operatorname{hvecdisc}(A).$$ Moreover, we can compute in deterministic polynomial time a set $S \subseteq [n]$ such that $\mu = O(\log m)\operatorname{vecdisc}(A\|_S)$. Let $P$ and $Q$ be optimal solutions for -. By Theorem \[thm:nuclear\], $\mu = \\|P^{1/2}AQ^{1/2}\\|_{S_1}$. Then, by Lemma \[lm:bt-lb\], applied to the matrices $P^{1/2}A$ and $Q$, there exists a set $S \subseteq [n]$, computable in deterministic polynomial time, such that $$\label{eq:main-lb} \operatorname{specLB}(A) \geq \sqrt{\|S\|}\sigma_{\min}(P^{1/2}A\|_S) \geq \frac{c \\|P^{1/2}AQ^{1/2}\\|_{S_1}}{\log m} = \frac{c\mu}{\log m}.$$ The determinant lower bound of Lovasz, Spencer, and Vesztergombi [@LSV] is equal to the maximum of $\|\det A_{S,T}\|$ over all submatrices $A_{S,T}$ of $A$. We note that up to the log factor, the lower bound is at least as strong. In particular, assume the determinant lower bound is maximized by a $k\times k$ submatrix $A_{S,T}$ induced by a subset $S$ of the rows and a subset $T$ of the columns. Then we can set $P = \frac{1}{k}\operatorname{diag}(\mathbf{1}_S)$ and $Q = \frac{1}{k}\operatorname{diag}(\mathbf{1}_T)$, and $\\|P^{1/2}AQ^{1/2}\\|_{S_1} = \frac{1}{k} \\|A_{S,T}\\|_{S_1}$ is at least as large as $\|\det A_{S,T}\|$ by the geometric mean-arithmetic mean inequality applied to the singular valus of $A_{S,T}$. The Approximation Algorithm =========================== We are now ready to give our approximation algorithm for hereditary vector discrepancy and hereditary discrepancy. In fact, the algorithm is a straightforward consequence of the upper and lower bounds we proved in the prior sections. There exists a polynomial time algorithm that, on input an $m \times n$ real matrix $A = (a_j)_{j=1}^n$, computes a value $\alpha$ such that the following inequalities hold $$\begin{aligned} \alpha &\leq \operatorname{hvecdisc}(A) \leq O(\log m) \cdot \alpha\\ \alpha &\leq \operatorname{herdisc}(A) \leq O(\log^{3/2} m) \cdot \alpha \end{aligned}$$ Moreover, the algorithm finds a submatrix $A\|_S$ of $A$, such that $\alpha \leq \operatorname{vecdisc}(A\|_S)$. We first ensure that the matrix $A$ is of rank $m$ by adding a tiny full rank perturbation to it, and adding extra columns if necessary[^2]. By making the perturbation small enough, we can ensure that it affects $\operatorname{herdisc}(A)$ and $\operatorname{hvecdisc}(A)$ negligibly. Let $\mu = \min\{\\|E\\|_\infty: \forall j \in [n]: a_j \in E\}$. The value $\alpha$ we output is $\alpha = \mu/(C\log m)$, where $C$ is a sufficiently large constant so that the asymptotic expression in Theorem \[thm:main-lb\] holds. The approximation guarantees follow from Theorems \[thm:main-ub\] and \[thm:main-lb\], and $S$ is computed as in Theorem \[thm:main-lb\]. To compute $\alpha$ in polynomial time, we solve –. By Lemma \[lm:ellips-program\], this is a convex minimization problem, and as such can be solved using the ellipsoid method up to an $\epsilon$-approximation in time polynomial in the input size and in $\log \epsilon^{-1}$. The optimal value is equal to $\mu$ by Lemma \[lm:ellips-program\], and, therefore, we can compute an arbitrarily good approximation to $\alpha$ in polynomial time. A Geometric Consequence ======================= In this section we derive a geometric consequence of Lemma \[lm:bt-lb\]. Specifically, we prove that any convex body $K$ is contained in an ellipsoid whose Gaussian width is bounded in terms of the Kolmogorov widths of $K$. While not necessary for our approximation algorithm, this result may be of independent interest. Let us first define the Kolmogorov widths for a convex body $K$. The Kolmogorov width $d_k(K)$ of a symmetric convex body $K\subseteq \R^n$ is equal to $d_k(K) \triangleq \min_{\Pi \in \mathcal{P}_{n-k+1}}{\\|\Pi K\\|_2}$, i.e. the minimum radius (in $\ell_2$) of any projection of $K$ of co-dimension $k-1$. We note that Kolmogorov width is more generally defined for linear operators between Banach spaces, and the definition above is the special case of the Kolmogorov width of the identity operator $I:X \rightarrow \ell_2$, where $X$ is a finite-dimensional Banach space with unit ball $K$. Lemma \[lm:bt-lb\] implies the following result. \[thm:geometric\] Let $K \subseteq \R^n$ be a symmetric convex body. There exists an ellipsoid $E=FB_2^n$ containing $K$ such that $$\\|F\\|_{HS} \leq (C\log n)\max_{k = 1}^n{\sqrt{k} d_k(K)} ,$$ for a universal constant $C$. The proof of the result relies on Lemma \[lm:bt-lb\] and an optimization problem over ellipsoids containing $K$ that is closely related to the problem of minimizing width in coordinate directions, discussed in prior sections. Let $v_1, \ldots, v_N$ be points in $\R^n$, and consider the problem of minizing $\\|F\\|_{HS}^2$ subject to $v_1, \ldots, v_N \in E = FB_2^n$. The problem can be formulated as $$\begin{aligned} &\text{Minimize } \tr(X^{-1}) \label{eq:trace-obj} \text{ s.t. }\\ &X\succ 0\\ &\forall i \in [N]: v_i^TXv_i \leq 1.\label{eq:enclose-tr}\end{aligned}$$ The constraints are affine, and the convexity of the objective follows from Lemma \[lm:inverse-convex\]. An argument analogous to the one in the proof of Theorem \[thm:nuclear\] shows that the Lagrange dual function for – is $$g(R) = \\|VR^{1/2}\\|_{S_1} - \tr(R),$$ where $V$ is the matrix whose columns are $v_1, \ldots, v_N$, and $R$ is a non-negative $N\times N$ diagonal matrix. Again analogously to Theorem \[thm:nuclear\], the Lagrange dual problem is to maximize $g(R)$ over all non-negative diagonal $R$, and by strong duality we have the equality $$\begin{aligned} \min\{\\|F\\|_{HS}^2: \forall i \in [N]\ v_i \in E = FB_2^n\} &= \max\{g(R): R \succeq 0, \text{diagonal}\} \label{eq:mintrace-dual}\\ &= \max\{ \\|VQ^{1/2}\\|^2_{S_1}: Q \succeq 0, \text{diagonal}, \tr(Q) = 1\}. \notag\end{aligned}$$ [Theorem \[thm:geometric\]]{} Let $v_1, \ldots, v_N$ be chosen to form a sufficiently dense net on the boundary of $K$ and let $E = FB_2^n$ be the ellipsoid containing $v_1, \ldots, v_N$ that minimizes $\\|F\\|_{HS}^2$; by taking $N$ sufficiently large, we can ensure that $K \subseteq (1+\epsilon)E$ for an arbitrary small $\epsilon$. Let $V$ be the matrix whose columns are the points $v_1, \ldots, v_N$. By and Lemma \[lm:bt-lb\] (with $V$ used in the role of $A$), there exists a set $S \subseteq [N]$ and an absolute constant $c$ such that $$\label{eq:matr-singval-lb} \|S\|\sigma_{\min}(V\|_S)^2 \geq \frac{c}{(\log n)^2} \\|F\\|_{HS}^2.$$ We claim that for $s = \|S\|$, $k = \lceil s/2 \rceil$, and any $\Pi \in \mathcal{P}_{n-k+1}$ there exists an $i \in S$ such that $\\|\Pi v_i\\|_2^2 \geq \sigma_{\min}(V\|_S)^2/2$. This suffices to prove the theorem, since, together with , it implies that $kd_{k}(K)^2 \geq \frac{c/2}{(\log n)^2} \\|F\\|_{HS}$. Define $M \triangleq (V\|_S)(V\|_S)^T$ and fix some $\Pi \in \mathcal{P}_{n-k+1}$. By averaging, it suffices to show that $$\frac{1}{s} \sum_{i \in S}{\\|\Pi v_i\\|_2^2} = \frac{1}{s}\tr((V\|_S)^T\Pi(V\|_S)) \geq \sigma_{\min}(V\|_S)^2/2.$$ Let $\Pi = UU^T$, where $U$ is a matrix with $n-k + 1$ mutually orthogonal unit columns. Then, by the Cauchy interlace theorem (see Lemma \[lm:interlace\] in Appendix \[app:vector-komlos\]), $$\lambda_{k}(U^TMU) \geq \lambda_{2k-1}(M) \geq \lambda_{s}(M) =\sigma_{\min}(V\|_S)^2.$$ Therefore, we have $$\begin{aligned} \frac{1}{s}\tr((V\|_S)^T\Pi (V\|_S)) &= \frac{1}{s}\tr((U^TV\|_S)^T(U^TV\|_S)) = \frac{1}{s}\tr((U^TV\|_S)(U^TV\|_S)^T)\\ &= \frac{1}{s}\tr(U^TMU) \geq \frac{k}{s}\lambda_k(U^TMU) \geq \frac{1}{2}\sigma_{\min}(V\|_S)^2. \end{aligned}$$ As remarked above, this completes the proof of the theorem together with . The Hilbert-Schmidt norm $\\|F\\|_{HS}$ has several natural geometric interpretations in terms of the ellipsoid $E = FB_2^n$. On one hand, $\\|F\\|_{HS}^2$ is equal to the sum of squared lengths (in $\ell_2$) of the major axes of $E$. By an easy calculation, $\\|F\\|_{HS}$ is also equivalent up to constants to the norm $\ell^(K) \triangleq \E \\|g\\|_{K^\circ} = \E \max_{x \in K}{\|\langle x, g\rangle \|}$, where $g$ is a standard $m$-dimensional Gaussian random variable (see, e.g. [@talagrand-chaining]). This quantity is also known as the Gaussian width of $K$. Phrased in these terms, Theorem \[thm:geometric\] shows that for any $n$-dimensional convex symmetric $K$ there exists an ellipsoid $E$ containing $K$ such that $$\label{eq:gauss-width} \ell^(E) \leq (C\log n) \max_{k = 1}^n{\sqrt{k} d_k(K)}.$$ A qualitatively weaker bound follows from a theorem of Carl and Dudley’s chaining bound. Carl [@carl-apxnum] showed that for an absolute constant $C_1$, $$\max_{k=1}^n{\sqrt{k} e_k(K)} \leq C_1 \max_{k = 1}^n{\sqrt{k}d_k},$$ where $e_k(K)$ is the $k$-th entropy number of $K$, i.e. the least $r$ such that $K$ can be covered by at most $2^{k-1}$ copies of $rB_2^n$. Dudley’s chaining argument [@dudley-chaining; @talagrand-chaining] implies that there exists a constant $C_2$ such that $\ell^(K) \leq (C_2\log n)\max_{k=1}^n{\sqrt{k} e_k(K)}$; combining the two bounds we have that $$\label{eq:carl-dudley} \ell^(K) \leq (C_3\log n) \max_{k = 1}^n{\sqrt{k} d_k(K)},$$ where $C_3 = C_1C_2$. This is readily implied by (up to the value of the constant), because $K \subseteq E$ implies $\ell^(K) \leq \ell^(E)$. However, there are examples where is near-tight while is loose. For example, for the $\ell_1^n$-ball $B_1^n$, $\max_{k=1}^n \sqrt{k} d_k(B_1^n) = \Omega(\sqrt{n})$ and $\ell^(B_1^n) = \Theta(\sqrt{\log n})$, while for any ellipsoid $E$ containing $B_1$ we have $\ell^(E) =\Omega(\sqrt{n})$, as can be seen from . Conclusion ========== We gave an $O(\log^{3/2} n)$-approximation algorithm for the hereditary discrepancy of a matrix $A$, by approximately characterizing the hereditary vector discrepancy of a matrix in terms of a simple convex program: that of minimizing $\\|E\\|_\infty$ over all $E$ that contain the columns of $A$. Our lower bound is “constructive”: we can construct in polynomial time a submatrix $A\|_S$ demonstrating the lower bound on $\operatorname{hvecdisc}(A)$ and hence on $\operatorname{herdisc}(A)$. Our upper bound for $\operatorname{hvecdisc}(A)$ is also “constructive” in that the ellipsoid $E^$ gives a recipe to construct a vector solution to the $\operatorname{vecdisc}(A\|_S)$ given any $S$. Our $O(\log^{3/2})$ upper bound for $\operatorname{herdisc}(A)$ is however non-constructive as it uses the result of Banaszczyk [@bana], whose proof does not yield an efficient algorithm for computing the sign vector $x$. We can however use the result of Bansal to algorithmically get a coloring for any given $S$, at the cost of losing a factor of $\sqrt{\log n}$ in the approximation. We leave open several questions of interest. One natural question is whether our approximation ratios can be improved. The best known hardness of approximating hereditary discrepancy is $2$, but we conjecture that the hardness is superconstant. Another interesting question is whether the guarantee for Bansal’s algorithm (Theorem \[thm:bansal\]) can be improved by a factor of $O(\sqrt{\log m})$, which would make it tight. This question was previously posed by Matoušek [@Matousek11]. Such an improvement would also imply a constructive proof of Banaszczyk’s theorem. A further question concerns the complexity of computing $\operatorname{herdisc}(A)$ exactly. Deciding whether $\operatorname{herdisc}(A) \leq t$ for any $t$ is naturally in $\Pi^{\mathsf{P}}_2$, but not know to be in $\mathsf{NP}$. Is this problem complete for $\Pi^{\mathsf{P}}_2$? Acknowledgements {#acknowledgements .unnumbered} ================ The first named author thanks Daniel Dadush for useful discussions, and Assaf Naor for bringing up the question about a geometric equivalent of Lemma \[lm:bt-lb\]. [^1]: This is the case, for example, for the matrix which contains three copies of each column of a Hadamard matrix. [^2]: There are other, more numerically stable ways to reduce to the full rank case, e.g. by projecting $A$ onto its range and modifying the norms we consider accordingly. We choose the perturbation approach for simplicity.	{ "pile_set_name": "ArXiv" }
--- abstract: 'This paper reports NMR measurements of the magnetic dipole moments of two high-K isomers, the 37/2$^-$, 51.4 m, 2740 keV state in $^{\rm 177}$Hf and the 8$^-$, 5.5 h, 1142 keV state in $^{\rm 180}$Hf by the method of on-line nuclear orientation. Also included are results on the angular distributions of gamma transitions in the decay of the $^{\rm 177}$Hf isotope. These yield high precision E2/M1 multipole mixing ratios for transitions in bands built on the 23/2$^+$, 1.1 s, isomer at 1315 keV and on the 9/2$^+$, 0.663 ns, isomer at 321 keV. The new results are discussed in the light of the recently reported finding of systematic dependence of the behavior of the g$_{\rm R}$ parameter upon the quasi-proton and quasi-neutron make up of high-K isomeric states in this region.' author: - 'S. Muto' - 'N. J. Stone' - 'C. R. Bingham' - 'J. R. Stone' - 'P. M. Walker' - 'G. Audi' - 'C. Gaulard' - 'U. Köster' - 'J. Nikolov' - 'K. Nishimura' - 'T. Ohtsubo' - 'Z. Podolyak' - 'L. Risegari' - 'G. S. Simpson' - 'M. Veskovic' - 'W. B. Walters' title: ' The magnetic properties of $^{\rm 177}$Hf and $^{\rm 180}$Hf in the strong coupling deformed model.' --- \[intro\]Introduction ===================== This paper reports measurement of the magnetic dipole moments of the 51.4 m, 2740 keV, 37/2$^-$ isomer in $^{\rm 177}$Hf and of the 5.5 h, 1142 keV, 8$^-$ isomer in $^{\rm 180}$Hf, using the method of NMR on oriented nuclei, at the on-line NICOLE facility, ISOLDE, CERN. In addition, measurements of the angular distribution of gamma transitions from the $^{\rm 177m2}$Hf isomer are analysed to give precise values of the E2/M1 mixing ratios in transitions in bands built upon the 23/2$^+$, 1.1 s, 1315 keV isomer and the 9/2$^+$, 321 keV state of this nucleus. Accurate measurements of the electromagnetic properties of ground states, isomers and rotational band states in deformed nuclei yield parameters of the deformed nuclear model used to describe these nuclei. In the region between Yb and W ground states and isomers have close-to-constant deformation and the parameters of the model can be explored thoroughly. In this work the main focus is on isomeric state, band-head, magnetic dipole moments to yield values of the single particle g-factor, g$_{\rm K}$, with good precision. In addition, analysis of the angular distribution of mixed M1/E2 transitions in the decay of the isomers studied yields values of the mixing ratio $\delta$ which is related to the g-factors g$_{\rm K}$, and g$_{\rm R}$, and intrinsic quadrupole moment Q$_{\rm 0}$, of the band-head (see Eq. \[delt\]). Interest in obtaining the magnetic parameters g$_{\rm K}$ and g$_{\rm R}$ in these states derives from their involvement with ideas of superconductivity in nuclei. In the strong coupling model of well-deformed nuclei the influence of the quasi-particle state upon the band properties arises through the dependence of the pairing strength, and hence the pairing gap $\Delta$, upon the available orbitals close to the Fermi surface. In the deformed potential each state is doubly degenerate and is available for pair scattering if empty, however the occupation of a state by a single quasi-particle renders it unavailable to pair scattering. This is the process of blocking [@dracoulis1998; @wu2011]. Since the pairing gap is determined by the number of pairs and the numbers of states between which they can scatter, increasing the number of quasi-particles increases blocking and reduces the gap $\Delta$. These considerations apply separately to protons and neutrons. Where the rotational energy levels of the bands are concerned, the relevant model parameter is the total moment of inertia $\mathcal{I}_{\rm tot}$ = $\mathcal{I}_{\rm p}$ +$\mathcal{I}_{\rm n}$. The stronger the pairing associated with the state the more $\mathcal{I}_{\rm tot}$ is reduced compared to the classical rigid body value. Blocking of either proton or neutron orbitals weakens the pairing and increases $\mathcal{I}_{\rm tot}$. When electromagnetic properties of the band are considered, transition matrix elements, intensity ratios, multipole mixing ratios and the collective g-factor g$_{\rm R}$ show different sensitivity to proton and neutron pairing, and the effects of blocking, through their differing effective charges, crudely 1 for protons and 0 for neutrons. In this limit g$_{\rm R}$ is given by the ratio $\mathcal{I}_{\rm p}$/$\mathcal{I}_{\rm tot}$, which yields the familiar simple result g$_{\rm R}$ $\sim$ Z/A for rigid rotation of the whole nucleus. It follows from these basic ideas that we can expect any increase in, for example, neutron blocking, by reducing the neutron pairing gap and increasing the neutron contribution to $\mathcal{I}_{\rm tot}$, to reduce g$_{\rm R}$ relative to Z/A, whilst increased proton blocking will have the opposite result. The new experimental results presented here, in combination with a much larger body of data from existing literature, have been used recently to explore these concepts [@stone2013]. It has been shown that, within the rather limited data presently available, the principle of additivity can be relied upon to give good estimates of the single particle g-factor, g$_{\rm K}$, in multi-quasi-particle isomers. The use of this principle to estimate g$_{\rm K}$ in many bands revealed a wide, systematic, dependence of the collective g-factor, g$_{\rm R}$, upon the neutron and proton quasi-particle make-up of the band-head states, presenting a new challenge to the theory of superconductivity, pairing and blocking in these deformed nuclei. This finding gives additional reason to extend, in particular, band-head magnetic moment measurements of good accuracy and to seek accurate mixing ratios to yield g$_{\rm R}$, as are reported here. The theoretical context of this work is given in Sec. \[theory\]. Experimental details and new data on the angular properties of gamma transitions in the decay of the 37/2$^-$ K-isomer of $^{\rm 177}$Hf are described in Secs. \[exp\] and \[res\] together with NMR results which yield the magnetic dipole moments of the 37/2$^-$ K-isomer and of the 8$^-$ isomer in $^{\rm 180}$Hf. Consistency of the new results with the findings of [@stone2013] is discussed in Sec. \[dis\]. \[theory\]Theoretical framework =============================== Relevant expressions from the strong coupling model concerning electromagnetic properties of the nuclear states and transitions are given here [@bm]. For the static magnetic dipole moment of a band state of spin I based on a band head of spin K $$\mu = g_{\rm R}I + (g_{\rm K} - g_{\rm R})\frac{K^{\rm 2}}{I + 1} \label{mu}$$ which, for the band head K = I, gives $$\mu = g_{\rm K} [I^{\rm 2}/(I + 1)] + g_{\rm R}[I/(I + 1)] \label{gk}$$ Since the values of spin I in this paper are large, this expression makes it clear that the moment has only weak dependence upon the ‘collectiveÕ g$_{\rm R}$ factor and is largely determined by the quasi-particle g$_{\rm K}$ factor. For a pure multi-quasi-particle state for which $K$ is the simple sum of the constituent quasi-particle K$_{\rm i}$’s, the quasi-particle g-factor g$_{\rm K}$ is given in terms of the constituent quasi-particle g$_{\rm Ki}$-factors by $$g_{\rm K} = \frac{1}{K}\sum_{i}K_{\rm i}g_{\rm K_{\rm i}}. \label{gks}$$ Two spectroscopic variables, the E2/M1 ratio $\delta$ in transitions from a level of spin I to a lower level of spin I - 1 and the branching ratio of transitions from a state of spin I to lower states of spin I - 1 and I - 2, depend upon the difference (g$_{\rm K}$ - g$_{\rm R}$) of the two (collective and quasi-particle) g-factors. These variables can both be written in terms of $\delta$ given by $$\delta = \frac{0.933E_{\rm \gamma}Q_{\rm 0}}{(g_{\rm K} - g_{\rm R}) \sqrt{(I^{\rm 2} - 1)}}, \label{delt}$$ where E$_{\rm \gamma}$ is the transition energy in MeV and Q$_{\rm 0}$ is the intrinsic quadrupole moment in $e$b. From these expressions it is seen that to separate the quasi-particle g-factor g$_{\rm K}$ from the rotation g-factor g$_{\rm R}$ it is necessary to have data on either transition branching ratios from states in the band to lower states of spin reduced by one and two units or M1/E2 mixing ratios for transitions between states of spin differing by one, combined with the static magnetic dipole moment of a state of the band. Of these, the static dipole moments are the rarest. In addition, knowledge of the intrinsic quadrupole moment Q$_{\rm 0}$ of the band states is required. \[exp\]Experimental details =========================== Nuclear orientation of the 37/2$^-$, 51.4 m, K-isomer in $^{\rm 177}$Hf ----------------------------------------------------------------------- At the ISOLDE isotope separator facility, CERN, new results have been obtained on the magnetic moment of the 37/2$^-$ high spin K-isomer and on gamma transitions in its decay to the ground state by the method of on-line nuclear orientation combined with NMR [@book]. The 1.4 GeV proton beam from CERN’s PSB synchrotrons was incident upon a mixed tantalum/tungsten foil target and separation of the Hf isotopes was achieved by introducing CF$_{\rm 4}$ gas into the plasma ion source and extracting a beam of HfF$_{\rm 3}^{+}$ ions [@koester2007]. The ions, accelerated through 60 keV, impinged upon a pure iron foil soldered to the cold finger of the NICOLE on-line nuclear orientation system [@eder1990] maintained at temperatures down to about 12 mK. The HfF$_{\rm 3}^{+}$ ions disintegrated at the foil surface and the Hf activity was implanted into the Fe lattice. By interrupting the implantation and allowing the dilution refrigerator to cool further, data were taken to about 6 mK as the activity decayed. The iron foil was magnetized by an applied magnetic field of 0.5 T. At mass A = 177 the Hf beam contains both the isomer $^{\rm 177}$Hf$^{\rm m2}$ (I$^{\rm \pi}$ = 37/2$^-$ T$_{\rm 1/2}$ = 51.4 m) and stable $^{\rm 177}$Hf ground state nuclei. The isomer activity made up about 1 % of the beam. The decay scheme of $^{\rm 177}$Hf$^{\rm m2}$ is shown in Fig \[decay\]. Angular distribution parameters for all transitions were obtained from measurements, as a function of the foil temperature, of the gamma photo-peak count rates in two pairs of high resolution Ge detectors placed at angles of 0, 90, 90, and 180 degrees to the axis of polarization. The counts were normalized to rates measured with the activity un-oriented at a temperature close to 1 K. The foil temperature was obtained from measurements of the angular distributions of the 1333 keV and 1173 keV transitions in the decay of $^{\rm 60}$Co activity in a Co single crystal soldered to the back of the cold finger. Since, for the $^{\rm 60}$Co$\underline{Co}$ system, all orientation parameters are known, these measurements act as an accurate thermometer in the region between about 50 mK and 5 mK. High degrees of polarization were achieved, produced by the large hyperfine field experienced by the Hf activity in the iron lattice. Examples of cold, oriented, gamma spectra compared with warm (1 K, un-oriented) spectra in 0$^{\rm 0}$ and 90$^{\rm 0}$ detectors are shown in Fig. \[spectra\], in which the strongly differing anisotropies in transitions between levels in different bands of the decay are clearly seen. In addition to gamma transition angular distribution measurements, NMR of the implanted $^{\rm 177}$Hf$^{\rm m2}$ parent state activity was sought by exposing the implanted nuclei to an RF field, transverse to the axis of polarization, produced by a simple two-turn coil. ![\[spectra\] (color online) Warm and cold spectra from 0$^{\rm 0}$ and 90$^{\rm 0}$ detectors for two series of E2/M1 transitions. Black/dark energy labels - in band built on 9/2$^+$ \[624\] state - note increases at 0$^{\rm 0}$. Blue/pale energy labels - in band built on 23/2$^+$ three-quasi-particle isomer - note decreases at 0$^{\rm 0}$.](newfig2.pdf){width="9cm" height="7cm"} Hyperfine interaction and nuclear spin-lattice relaxation time estimates. ------------------------------------------------------------------------- Implanted nuclei have a high probability of coming to rest in regular lattice sites and the well-defined hyperfine interaction allows observation of Nuclear Magnetic Resonance by the disturbance caused to the gamma decay angular distribution by resonant absorption of RF energy. However even though the foil lattice may be at millikelvin temperature the nuclei are implanted un-oriented and it takes a time, related to the nuclear spin-lattice relaxation time, T$_{\rm 1}$ for nuclear polarization to be established [@shaw1989]. Both the resonant NMR frequency, $\nu_{\rm res}$ and T$_{\rm 1}$ depend upon the strength of the hyperfine structure splitting of the state undergoing resonance and/or orientation, h$\nu_{\rm res}$ = $\mu$B$_{\rm hf}$/I. The nuclear moment of the 37/2$^-$ isomer can be estimated to be $\sim$7 $\mu_{\rm N}$ (see Sec. \[moment\] below) and the hyperfine field at Hf in Fe is B$_{\rm hf}$ = - 67.4(9) T ([@muto2004] improved using $^{\rm 175}$Hf moment 0.677(9) [@nieminen2002]). These values yield $\nu_{\rm res}$ in the region of 200 MHz. There have been no spin lattice relaxation time measurements for Hf isotopes in iron, however an empirical relationship, C$_{\rm K}$T$^{\rm 2}_{\rm int}$ = 1.4 x 10$^{\rm -4}$ sK$^{\rm 3}$, where T$_{\rm int}$ is the nuclear interaction strength, T$_{\rm int}$ = $\mu$B$_{\rm hf}$/kI, can be used to give estimates of the Korringa constant C$_{\rm K}$, usually found reliable to better than a factor of two [@shaw1989]. The resulting prediction is T$_{\rm 1}$ (37/2$^-$) $\sim$ 19 s at 10 mK which is much less than the isomer lifetime of 51.4 m, so full thermal equilibrium between the $^{\rm 177}$Hf$^{\rm m2}$ nuclei and the lattice will be established prior to decay. There remains the possibility that orientation may be perturbed by further interaction with the lattice during the lifetime of the 1.1 s 23/2$^+$ isomeric state at 1315 keV. Any such re-orientation would affect observed anisotropies in transitions below the isomer. However, an estimate of T$_{\rm 1}$ for the 23/2$^+$ isomer, based on an estimated 23/2$^+$ moment and the above T$_{\rm 1}$ (37/2$^-$) estimate gives the value T$_{\rm 1}$ (25/2$^+$) $\sim$ 8 s at 10 mK. This is substantially longer than the life time of the isomer, indicating that any re-orientation effect will be small. Nuclear orientation of the 8$^-$, 5.5 h, 1142 keV, state in $^{\rm 180}$Hf -------------------------------------------------------------------------- At mass 180 the Hf beam has a strong isomeric $^{\rm 180}$Hf component which decays by mixed E1/M2/E3 transitions to 8$^+$ and 6$^+$ states in the ground state band. These states decay by a sequence of pure E2 transitions which show strong anisotropy as has been reported previously [@stone2007]. Earlier measurements of the magnetic dipole moment of this isomer have not been of high accuracy. Körner et al. [@korner1971] reported 8.7(10) $\mu_{\rm N}$ by a Mössbauer method and Krane et al [@krane1976] 9.0(9) $\mu_{\rm N}$ by an integral nuclear orientation experiment on sources of $^{\rm 180m}$Hf in HfZrFe$_{\rm 2}$ alloys. Since NMR on the implanted activity would provide a more precise result for the moment, resonance was sought with the activity implanted into an Fe foil. Unfortunately, at the high frequency at which resonance is expected for this isomer (in the region of 520 MHz) the RF line into the Nicole dilution refrigerator shows strong standing wave resonances which result in non-nuclear-resonance heating of the sample and render true resonance unobservable. In a second experiment the activity was implanted into a pure Ni foil in which the magnetic hyperfine field at Hf nuclei is much reduced. Resonance was observed at a frequency of 92.2(2) MHz in an applied field of 0.10 T. The result is discussed further below. \[res\]Analysis of Results ========================== Gamma transition anisotropies ----------------------------- Photo-peak counts for each transition in each spectrum for each gamma detector were obtained using the fitting code DAMM. Counts taken with the sample at temperatures with appreciable orientation were normalized by taking ratios with counts from unoriented, warm (1 K), spectra. The gamma transition anisotropies, $A$, were obtained for each of the two pairs of detectors using the expressions $$W(\theta, T) = N(\theta, T)/N(\theta,{\rm warm})$$ and $$A = [W(0^{\rm 0},T)/W(90^{\rm 0},T) - 1]$$ The anisotropies were analysed further using the standard nuclear orientation formalism [@kranebook] $$W(\theta,T)=1+f\sum_{\lambda}{B_{\lambda}U_{\lambda}A_{\lambda}Q_{\lambda} P_{\lambda}(cos \theta)}.$$ Here $f$ is the fraction in good sites. The B$_{\lambda}$ factors describe orientation of the parent, isomeric, state and depend upon the strength of the hyperfine interaction, the spin and the temperature. The U$_{\lambda}$ factors are calculated for each state below the parent and require knowledge of the spins of any intervening states, the intensities of all transitions between them and their multipolarity, with admixtures treated without regard to phase. The A$_{\lambda}$ parameters describe the observed emission and depend upon the initial and final state spins and the multipolarity, treated with sensitivity to phase of any multipole mixing ratio $\delta$. Q$_{\lambda}$ are correction factors which account for the finite solid angle subtended by the detectors and P$_{\lambda}$ are the associated Legendre Polynomials. Since gamma emission conserves parity, the summation has only even terms (up to $\lambda$ = 2L$_{\rm max}$, where L$_{\rm max}$ is the highest multipolarity in the observed emission). In all on-line nuclear orientation experiments it has been found that the angular distributions can be well described as consisting of two components, the first from nuclei coming to rest in undisturbed (good) lattice sites and subject to the full hyperfine interaction and the second un-oriented at all temperatures. The un-oriented fraction describes nuclei which are stopped in the thin oxide layer which is always present on the implantation foil surface or are stopped in other very disturbed regions which are not magnetically ordered. Analysis of the pure E2, $\Delta$I = 2, transitions: determination of the fraction in good sites ------------------------------------------------------------------------------------------------ The first transition emitted by the 37/2$^{-}$ isomer is the 214 keV pure E3 transition. However this is an unresolved doublet with a mixed transition lower in the decay scheme so could not be used to evaluate the fraction $f$. The U$_\lambda$ values for this pure multipole transition are known, however, so $f$ can be obtained from the anisotropy of the pure E2, 638 keV, (31/2$^{+}$ - 27/2$^{+}$) transition. As the other parameters necessary to calculate U$_{\lambda}$ for lower states can all be extracted from the experimental data, values for $f$ can also be found from data on the other E2 transitions between pairs of lower states built on the 23/2$^{+}$ K-isomer. The results are given in Table \[fraction\]. As an example, in Fig. \[frac2\] are shown the data for 606.5 keV transition between the 29/2$^{+}$ and 25/2$^{+}$ states and calculation using the fit $f$ value, with its uncertainty. [llccc]{} & Energy & Fitted fraction $f$\ I$_{\rm 1}$ & I$_{\rm 2}$ & \[keV\] &\ \ 31/2 & 27/2 & 638.2 & 0.755(7)\ 29/2 & 25/2 & 606.5 & 0.786(12)\ 27/2 & 23/2 & 572.4 & 0.768(12)\ \ 23/2 & 19/2 & 228.5 & 0.773(6)\ 21/2 & 17/2 & 418.5 & 0.773(13) &\ 19/2 & 15/2 & 378.5 & 0.773(6) &\ 15/2 & 11/2 & 281.8 & 0.779(9) &\ 13/2 & 9/2 & 233.9 & 0.802(21) &\ ![\[frac2\](color online) Anisotropy of the 606.5 keV 29/2$^+$ - 25/2$^+$ pure E2 transition in band built on 23/2$^+$ isomer. Fitted curve with shaded band is for $f$ = 0.786(12). ](newfig3.pdf){width="9cm"} ![\[frac1\](color online) Anisotropy of the 228.5 keV 23/2$^+$ - 19/2$^+$ pure E2 transition from the 23/2$^+$ isomer to the band built on the 9/2$^+$ \[624\] band. Fitted curve with shaded band is for $f$ = 0.773(6). ](newfig4.pdf){width="9cm"} Re-orientation in the 23/2$^{+}$ isomeric state ----------------------------------------------- This state has lifetime 1.1 s and, although estimates given above suggest that the spin-lattice relaxation time T$_{\rm 1}$ of this state is considerably longer than the lifetime, this point requires further attention if anisotropies of gamma transitions between states below the isomer are to be quantitatively analysed. Data on anisotropies of the series of pure E2 transitions between states below the 23/2$^{+}$ isomer have all be treated to yield values of $f$ on the assumption that re-orientation may be neglected. The results are given in Table \[fraction\]. The close agreement between values of $f$ obtained from transitions above and below the isomer show clearly that the assumption is indeed valid. In Fig. \[frac1\] the data on the 228.5 keV transition between the 23/2$^{+}$ and 19/2$^{+}$ states are shown, with calculation using the fit $f$ value, with its uncertainty. ------ ------- -------------- ------------------ -------------------------- -------------------------- --------------------------------------- K Level E$_{\gamma}$ $\delta$(InBeam) $\delta$($^{\rm 177}$Lu) $\delta$($^{\rm 177}$Hf) (g$_{\rm K}$-g$_{\rm R}$)/Q$_{\rm 0}$ spin \[keV\] [@mullins1998] [@krane1974a] this work 37/2 41/2 420.9 0.74(23) 0.025(8) 43/2 440.0 0.74(16) 0.025(5) 23/2 25/2 277.3 0.302(4) 0.069(1) 27/2 295.1 0.31(2) 29/2 311.5 0.36(2) 0.285(5) 0.071(2) 31/2 326.5 0.30(1) 0.278(5) 0.071(2) 33/2 340.1 0.26(1) 35/2 351.9 0.21(1) 37/2 361.7 0.18(2) 39/2 369.4 0.20(2) 9/2 11/2 105.3 -0.36(4) -0.23(4) -0.083(15) 13/2 128.4 0.38(1) -0.37(6) -0.34(3) -0.055(5) 15/2 153.1 0.38(1) -0.33(5) -0.317(13) -0.061(3) 17/2 174.3 0.36(1) -0.32(4) -0.296(13) -0.065(3) 19/2 204.1 0.40(1) -0.33(5) -0.289(13) -0.070(4) 21/2 214.3 0.35(1) -0.29(2) 23/2 260.5 0.50(2) 25/2 241.8 0.40(2) 27/2 325.4 0.61(7) 29/2 249.4 0.27(2) ------ ------- -------------- ------------------ -------------------------- -------------------------- --------------------------------------- : \[delta\]E2/M1 mixing ratios in transitions from levels of $^{\rm 177}$Hf. The data from both pairs of detectors were in very close agreement. An average value $f$ = 0.775(5) was adopted and used throughout the analysis. ![\[anis1\](color online) Anisotropy of the 277.3 keV 25/2$^+$ - 23/2$^+$ mixed E2/M1 transition in band built on 23/2$^+$ isomer. Fitted curve with shaded band is for $\delta$ \[E2/M1\] = +0.302(4). ](newfig5.pdf){width="9.0cm"} ![\[anis2\](color online) Anisotropy of the 153.1 keV 15/2$^+$ - 13/2$^+$ mixed E2/M1 transition in band built on 9/2$^+$ \[624\] band. Fitted curve with shaded band is for $\delta$ \[E2/M1\] = -0.317(13). ](newfig6.pdf){width="9cm"} Analysis of the mixed multipole $\Delta$I = 1 transitions: determination of the E2/M1 mixing ratios. ---------------------------------------------------------------------------------------------------- Each state in the bands built upon the 23/2$^{+}$ multi-quasi-particle isomer and the 9/2$^{+}$ and 7/2 $^{-}$ single quasi-particle states decays both by a direct E2, $\Delta$I = 2, transition and a mixed E2/M1, $\Delta$I = 1, transition. The E2/M1 multipole mixing ratio, $\delta$, is a most useful quantity in establishing the parameters of the band, as described previously. For each of these transitions the anisotropies were analysed in a somewhat novel way to give the best fit values of $\delta$. Each data point in the temperature dependence (19 in all) was regarded as an individual experiment and fitted with $\delta$ as the only unknown. The resulting set of $\delta$Õs was treated statistically to give the best fit value and the error on this value. Results are shown in Table \[delta\], compared to results from other investigations [@mullins1998; @krane1974a]. The overall agreement is excellent and the new results have significantly smaller experimental uncertainties. Examples of the fits are shown in Figs. \[anis1\] and \[anis2\]. Fig. \[anis1\] shows data for the 277.3 keV transition between the 25/2$^{+}$ and 23/2$^{+}$ states in the 23/2$^{+}$ band. All mixed transitions in this band show similar negative anisotropies. Fig. \[anis2\] shows data for the 153.1 keV transition between the 15/2$^{+}$ and 13/2$^{+}$ states in the 9/2$^{+}$ band. All mixed transitions in this band show similar large positive anisotropies. The larger scatter and uncertainties for these transitions arise because they are measured on a large Compton-backscattered gamma intensity so that the small 90$^{\rm 0}$ counts at lower temperatures have large uncertainties (see Fig. \[spectra\]). ![\[fig:res177\](color online) Resonance observed in a combined signal from the stretched E2 transitions in the decay of the 37/2$^-$ isomer in $^{\rm 177}$Hf.](newfig7.pdf){width="9.0cm"} \[moment\]The magnetic moment of the 37/2$^{-}$ isomer in $^{\rm 177}$Hf ------------------------------------------------------------------------ The search for NMR of the 37/2$^{-}$ isomer was carried out with RF swept between 180 and 230 MHz in 1 MHz steps with RF modulation of 1 MHz. A positive signal was found, as shown in Fig. \[fig:res177\], with centre frequency $$\nu_{\rm res} (^{\rm 177}{\rm Hf}^{\rm m2} \underline{Fe}) = {\rm 203.4(2)\mbox{}MHz}.$$ This frequency, combined with the known hyperfine field -67.4(9) T, adjusted for the external field of 0.100 T applied to the Fe foil during the NMR measurements, yields the magnitude of the magnetic moment of the 37/2$^{-}$ isomer as $$\|\mu\|(^{\rm 177}{\rm Hf}^{\rm m2},\mbox{} 37/2^{-}, \mbox{} {\rm 51.4 m}) = {\rm 7.33(9)} \mbox{ }{\rm \mu_{\rm N}}$$ Note that any hyperfine anomaly between the isotope $^{\rm 175}$Hf, on which the hyperfine field measurement was made, and $^{\rm 177}$Hf$^{\rm m2}$ is not expected to exceed 0.1% and has been neglected. ![\[fig:res180\](color online) Resonance observed in percentage destruction of anisotropy of the 8$^-$ isomer in $^{\rm 180}$Hf.](newfig8.pdf){width="9cm"} The magnetic moment of the 8- isomer in $^{\rm 180}$Hf ------------------------------------------------------ Extraction of the magnetic moment from the observed resonance (Fig. \[fig:res180\]) requires knowledge of the hyperfine field acting on Hf nuclei in the Ni lattice. Several previous values for this field have been reported, none of them of the necessary accuracy. The best would appear to be the Mössbauer study by Aggarwal et al. [@aggrawal1983] who gave the value B$_{\rm hf}$(Hf$\underline{Ni}$) = 11.8(2.6) T. Using this value the observed $^{\rm 180m}$Hf resonance yields the moment of the isomer to be 8.3(1.8) $\mu_{\rm N}$. Other, lower, reported Hf$\underline{Ni}$ fields give unrealistically high values for the nuclear moment. Whilst the hyperfine fields at some neighbouring elements have better measurements in Nickel they do not allow an improved estimate for the field at the Hf nucleus. Clearly as it stands the present value is no improvement on the earlier results and the observed NMR result awaits an improved field value. This can be obtained, for example, by the observation of NMR in $^{\rm 177m2}$Hf activity in Ni where the frequency should be in the region of 35 MHz. Quasi-particle state Adopted g$_{\rm K}$ Basis of adopted value ---------------------- --------------------- ------------------------------------------------------------- -- Protons 7/2+\[404\] 0.765(25) moments of 7/2+ ground states in $^{\rm 175,177,179,181}$Ta 9/2-\[514\] 1.37(3) moment of 9/2- 6 keV state in $^{\rm 181}$Ta Neutrons 5/2-\[512\] -0.48(2) moment of 5/2- ground state of $^{\rm 175}$Hf 7/2-\[514\] 0.206(14) moment of 7/2- ground state of $^{\rm 177}$Hf 9/2+\[624\] -0.239(11) moment of 9/2+ ground state of $^{\rm 179}$Hf ![\[grplot\](color online) Plot of the collective g$_{rm R}$ as a function of $\Delta$ = N$_{\rm p}$ - N$_{\rm n}$. For ease of viewing the three data points at -1 have been slightly off-set along the x-axis. For discussion see text and Ref. [@stone2013].](newfig9.pdf){width="9.0cm"} ---------------- ------- --------- ------------- ------------- --------------------------- -------------------- ---------------- ------------- Isotope State Energy N$_{\rm p}$ N$_{\rm n}$ $\Delta$ Moment Spectroscopy g$_{\rm R}$ \[keV\] = N$_{\rm p}$-N$_{\rm n}$ Reference Reference $^{\rm 177}$Hf 37/2- 2740 2 3 -1 this work [@mullins1998] 0.21(4) $^{\rm 177}$Hf 7/2- 0 0 1 -1 [@buttgenbach1973] [@mullins1998] 0.24(2) $^{\rm 177}$Hf 9/2+ 321 0 1 -1 [@hubel1969] this work 0.21(2) $^{\rm 177}$Hf 23/2+ 1316 2 1 +1 additivity this work 0.30(3) ---------------- ------- --------- ------------- ------------- --------------------------- -------------------- ---------------- ------------- \[dis\]Discussion ================= The variation in g$_{\rm R}$ found in the analysis of this work (see below) prompted, in part, a comprehensive survey of nearly one hundred high-K bands in isotopes in this region [@stone2013]. The results have been recently published and draw two significant conclusions: that the assumption of additivity of quasi-particle contributions is a sound basis for estimation of the quasi-particle g-factor g$_{\rm K}$ in multi-quasi-particle isomers in the majority of cases and that, based on analysis of the mixing ratio and branching ratio data, there is a wide and systematic variation of the collective g-factor g$_{\rm R}$ which depends upon the quasi-particle make-up of the band head isomer. Underlying ideas concerning pairing and blocking in these isomers are discussed in the paper. In this section we review the extent to which the experimental results found in this work are consistent with the two conclusions described. Additivity of single quasi-particle g$_{\rm K}$ factors to form g$_{\rm K}$ for multi-quasi-particle states ----------------------------------------------------------------------------------------------------------- In Table \[states\] we give the g$_{\rm K}$ factors adopted in the broad survey for those single-quasiparticle states which are relevant to the two isomers $^{\rm 177m2}$Hf and $^{\rm 180m}$Hf for which the magnetic moments have been measured in this work. The g$_{\rm R}$ factor makes a small contribution to the total magnetic moment since it is multiplied by the factor I/(I+1) as compared to I$^{\rm 2}$/(I+1) for g$_{\rm K}$ (see Eq. \[gk\]) and the spins of the two isomers are high, 37/2 and 8, respectively. Taking g$_{\rm R}$ as 0.29(5) and assuming additivity, as described by Eq. \[gks\] yields predicted moments of these isomers of 7.28(13) $\mu_{\rm N}$ and 8.12(15) $\mu_{\rm N}$ to be compared to the measured values, 7.33(9) $\mu_{\rm N}$ and 8.3(18) $\mu_{\rm N}$ Both results show agreement with the predictions based on additivity, consistent with the findings in [@stone2013]. The variation of g$_{\rm R}$ ---------------------------- Taking average values of the parameter (g$_{\rm K}$ - g$_{\rm R}$/Q$_{\rm 0}$) with Q$_{\rm 0}$ = 7.2 $e$b and g$_{\rm K}$ assuming additivity in the multi-quasiparticle states, values of g$_{\rm R}$ were extracted from this work for bands built on the 9/2+, 23/2+ and 37/2- states in $^{\rm 177}$Hf. Results are given in Table \[tab:add\] which also includes g$_{\rm R}$ for the band built on the 7/2- ground state of $^{\rm 177}$Hf. We reproduce in Fig. \[grplot\] the evidence of variation of g$_{\rm R}$ as a function of the difference, N$_{\rm p}$ - N$_{\rm n}$, of the numbers of unpaired quasi-protons and quasi-neutrons in the isomer reported in Ref. [@stone2013], showing the broad band ranges which, it is suggested, are caused by variation in the contributions from breaking specific quasi-particle pairs and the individual results from Table \[tab:add\]. Conclusions =========== This paper reports new measurements of the magnetic properties of the nucleus $^{\rm 177}$Hf and $^{\rm 180}$Hf and their high K isomers.The magnetic moments of the 37/2- isomer in $^{\rm 177}$Hf and the 8$^{\rm -}$ isomer in $^{\rm 180}$Hf provide valuable additional evidence for the validity of additivity in estimating the quasi-particle g-factor g$_{\rm K}$ in these deformed nuclei. Values of the collective g-factor, g$_{\rm R}$, obtained from detailed gamma transition anisotropy measurements, are shown to be fully consistent with the recently revealed systematic dependence of this parameter upon the quasi-particle make-up of the bands involved. Dedication ========== We dedicate this paper to the memory of our colleague Suguru Muto, a most valuable member of the team who performed this work, who died in early 2013 after a long battle with cancer. Acknowledgements ================ We thank Ajay Deo for helpful assistance with the experiment. The research was supported by the US DOE Office of Science, the Ministry of Education and Science of Serbia (project No. 171002), the European Commission via the ENSAR project, SNRS and IN2P3 (France), JSPS KAKENHI (Grant Number 23540337) (Japan) and STFC (UK). [99]{} G. D. Dracoulis, F. G. Kondev and P. M. Walker, Phys. Lett. B419, 7 (1998) X. Wu, Z. H. Zhang, J. Y. Zeng and Y. A. Lei, Phys. Rev. C 83, 034323 (2011) N. J. Stone, J. R. Stone, P. M. Walker and C R Bingham, Phys. Lett. B726, 675 (2013) A. Bohr and B. Mottelson, Nuclear Structure, W. A. Benjamins, Inc, 1975 Low Temperature Nuclear Orientation, eds. N. J. Stone and H. Postma, North Holland, 1986 U. Köster et al., Eur.Phys.J. Special Topics, 150, 293 (2007) R. Eder and the NICOLE Collaboration, Hyp. Int. 59, 83 (1990) T Shaw and N. J. Stone ADNDT 42, 339 (1989) S. Muto, T. Ohtsubo, S. Ohya and K. Nishimura, Hyperfine Interactions 158, 195 (2004) A. Nieminen et al. Phys. Rev. Lett. 88, 094801 (2002) J. R. Stone et al., Phys. Rev. C76, 025502 (2007) H. J. Körner, F. E. Wagner and B. D. Dunlap, Phys.Rev.Lett. 27, 1593 (1971) K. S. Krane, S. S. Rosenblum and W. A. Steyert, Phys.Rev. C14, 650 (1976) K. S. Krane, in Low Temperature Nuclear Orientation, eds. N. J. Stone and H. Postma, North Holland, Chap. 2, 1986 S. M. Mullins, A. P. Byrne, G. D. Dracoulis, T. R. McGoram, and W. A. Seale , Phys. Rev. C58, 831 (1998) K. S. Krane, C. E. Olsen and W. A. Steyert, Phys. Rev. C10, 825 (1974) N. Aggarwal, B. Singh, A. K. Bhati, R. Bala, S. C. Bedi and H. S. Hans, Hyperfine Interactions 13, 271 (1983) N. J. Stone, Table of Nuclear Magnetic Dipole and Electric quadruple Moments, IAEA Nuclear Data Section Report INDC(NDS)-0594, April 2011 Y. Y. Chu, P. E. Haustein, and T. E. Ward, Phys. Rev. C6, 2259 (1972) S. Büttgenbach, M. Herschel, G. Meisel, E. Schrödl and W. Witte, Phys. Lett. B43, 479 (1973) H. Hübel, C. Günther, K. Krien, H. Toschinski, K.-H. Speidel, B. Klemme, G. Kumbartzki, L. Gidefeldt and E. Bodenstedt, Nucl. Phys. A127, 609, 609 (1969)	{ "pile_set_name": "ArXiv" }

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.